blob: 0f86f4714a17239ce0a3a294809cfcddc9ac9327 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700361 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
362 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
363 "src/qu8-dwconv/up1x9-minmax-scalar.c",
364 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
365 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
366 "src/qu8-gemm/2x2-minmax-scalar.c",
367 "src/qu8-igemm/2x2-minmax-scalar.c",
368 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700369 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700370 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700371 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700372 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700373 "src/x32-fill/scalar-float.c",
374 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700375 "src/x32-packx/x2-scalar.c",
376 "src/x32-packx/x3-scalar.c",
377 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700378 "src/x32-pad/scalar-int.c",
379 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700380 "src/x32-unpool/scalar.c",
381 "src/x32-zip/x2-scalar.c",
382 "src/x32-zip/x3-scalar.c",
383 "src/x32-zip/x4-scalar.c",
384 "src/x32-zip/xm-scalar.c",
385 "src/x8-lut/scalar.c",
386 "src/x8-zip/x2-scalar.c",
387 "src/x8-zip/x3-scalar.c",
388 "src/x8-zip/x4-scalar.c",
389 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700390 "src/requantization/precise-scalar.c",
391 "src/requantization/fp32-scalar.c",
392 "src/requantization/q31-scalar.c",
393 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700394]
395
Marat Dukhan436ebe62019-12-04 15:10:12 -0800396WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
398 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700399 "src/f32-clamp/gen/wasm-x1.c",
400 "src/f32-clamp/gen/wasm-x2.c",
401 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700402 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
403 "src/f32-dwconv/gen/up1x4-wasm.c",
404 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
405 "src/f32-dwconv/gen/up1x9-wasm.c",
406 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
407 "src/f32-dwconv/gen/up1x25-wasm.c",
408 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
409 "src/f32-dwconv/gen/up2x4-wasm.c",
410 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
411 "src/f32-dwconv/gen/up2x9-wasm.c",
412 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
413 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700414 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
416 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700418 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700420 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
422 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700424 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700426 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
427 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
429 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
430 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700431 "src/f32-gemm/gen/1x4-wasm.c",
432 "src/f32-gemm/gen/2x4-wasm.c",
433 "src/f32-gemm/gen/4x2-wasm.c",
434 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700435 "src/f32-gemm/gen/1x4-relu-wasm.c",
436 "src/f32-gemm/gen/2x4-relu-wasm.c",
437 "src/f32-gemm/gen/4x2-relu-wasm.c",
438 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700439 "src/f32-gemm/gen/1x4-minmax-wasm.c",
440 "src/f32-gemm/gen/2x4-minmax-wasm.c",
441 "src/f32-gemm/gen/4x2-minmax-wasm.c",
442 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700443 "src/f32-hswish/gen/hswish-wasm-x1.c",
444 "src/f32-hswish/gen/hswish-wasm-x2.c",
445 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700446 "src/f32-igemm/gen/1x4-wasm.c",
447 "src/f32-igemm/gen/2x4-wasm.c",
448 "src/f32-igemm/gen/4x2-wasm.c",
449 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700450 "src/f32-igemm/gen/1x4-relu-wasm.c",
451 "src/f32-igemm/gen/2x4-relu-wasm.c",
452 "src/f32-igemm/gen/4x2-relu-wasm.c",
453 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700454 "src/f32-igemm/gen/1x4-minmax-wasm.c",
455 "src/f32-igemm/gen/2x4-minmax-wasm.c",
456 "src/f32-igemm/gen/4x2-minmax-wasm.c",
457 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700458 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
459 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
460 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700461 "src/f32-prelu/gen/wasm-2x1.c",
462 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700463 "src/f32-relu/gen/wasm-x1.c",
464 "src/f32-relu/gen/wasm-x2.c",
465 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700466 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
467 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
468 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
469 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
470 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
471 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
472 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
473 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
474 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
475 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
476 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
477 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800478 "src/f32-vbinary/gen/vmax-wasm-x1.c",
479 "src/f32-vbinary/gen/vmax-wasm-x2.c",
480 "src/f32-vbinary/gen/vmax-wasm-x4.c",
481 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
482 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
483 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
484 "src/f32-vbinary/gen/vmin-wasm-x1.c",
485 "src/f32-vbinary/gen/vmin-wasm-x2.c",
486 "src/f32-vbinary/gen/vmin-wasm-x4.c",
487 "src/f32-vbinary/gen/vminc-wasm-x1.c",
488 "src/f32-vbinary/gen/vminc-wasm-x2.c",
489 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700490 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
491 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
492 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
493 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
494 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
495 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
496 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
497 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
498 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
499 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
500 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
501 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
502 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
503 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
504 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
505 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
506 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
507 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700508 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
509 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
510 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700511 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
512 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
513 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800514]
515
Marat Dukhan290055c2020-06-09 12:24:29 -0700516WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700517 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
518 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
519 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700520 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
521 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
522 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
523 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700524 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
525 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
526 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
527 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700528 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
529 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
530 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
531 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
532 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
533 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700534 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
535 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
536 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
537 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
538 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
539 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
540 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
541 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
542 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
547 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
549 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
551 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
553 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700558 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
559 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
560 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700561 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700562 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
563 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
564 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
565 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700566 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
567 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
568 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
569 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
570 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
571 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
572 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
573 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
574 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
575 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
576 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
577 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
578 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
579 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
580 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
581 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
582 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
583 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
584 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
585 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
586 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
587 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
588 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
589 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
590 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
591 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
592 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
593 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
594 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
595 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700596 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
597 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700598 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
599 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
600 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
601 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
602 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
603 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
604 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
605 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
606 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
607 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
608 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
609 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
610 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
611 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
612 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
613 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
614 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
615 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
616 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
617 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
618 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
619 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
620 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
621 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
622 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
623 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
624 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
625 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
626 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
627 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700628 "src/f32-ibilinear/gen/wasmsimd-c4.c",
629 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700630 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
631 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
632 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700633 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700634 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
635 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
636 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
637 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700638 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
639 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
640 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
641 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
642 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
643 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
644 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
645 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
646 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
647 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
648 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
649 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
650 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
651 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
652 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
653 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
654 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
655 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
656 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
657 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
658 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
659 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
660 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
661 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
662 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
663 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
664 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
665 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
666 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
667 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700668 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
669 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700670 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
671 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
672 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
673 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700674 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
675 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
676 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700677 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
678 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700679 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
680 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
681 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
682 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700683 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
684 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
685 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
686 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
691 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
692 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
693 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
694 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
695 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
696 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
697 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
698 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700699 "src/f32-relu/gen/wasmsimd-x4.c",
700 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700701 "src/f32-rmax/wasmsimd-arm.c",
702 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700703 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
704 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
705 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
706 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
707 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
708 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700709 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
710 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
711 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
712 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
713 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
714 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700715 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
716 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
717 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
718 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700719 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
720 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700721 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
722 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
723 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
724 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700725 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
726 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700727 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
728 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
729 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
730 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700731 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
732 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700733 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
734 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
735 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
736 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700737 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
738 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700739 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
740 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
741 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
742 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700743 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
744 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
745 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
746 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700747 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700751 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
752 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
753 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
754 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700755 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
756 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
757 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
758 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700759 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
760 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700761 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
762 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
763 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
764 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700765 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
766 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700767 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
768 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
769 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
770 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700771 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
772 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700773 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700777 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
778 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700779 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
780 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700781 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
782 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700783 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
784 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
785 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
786 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700787 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
788 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700789 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
790 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
791 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
792 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700793 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
794 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700795 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
796 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
797 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
798 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700799 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
800 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
801 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
802 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700803 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
804 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
805 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
806 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
807 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
808 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
809 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
810 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
811 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
812 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
813 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
814 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
815 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
816 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700817 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
818 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700819 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
820 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
821 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
822 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
823 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
824 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700825 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700826 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700827 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700828 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700829 "src/x32-zip/x2-wasmsimd.c",
830 "src/x32-zip/x3-wasmsimd.c",
831 "src/x32-zip/x4-wasmsimd.c",
832 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700833 "src/math/roundne-wasmsimd-addsub.c",
834 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700835 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700836 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700837 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700838 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700839 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700840 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
841 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700842 "src/requantization/fp32-wasmsimd.c",
843]
844
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800845PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800846 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800847 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800848 "src/f32-argmaxpool/9x-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700849 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700850 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700851 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700852 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700853 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700854 "src/f32-gavgpool-cw/psimd-x4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700855 "src/f32-spmm/gen/4x1-minmax-psimd.c",
856 "src/f32-spmm/gen/8x1-minmax-psimd.c",
857 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700858 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700859 "src/x32-packx/x4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700860]
861
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800862PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan249cd6a2020-07-17 14:05:30 -0700863 "src/requantization/precise-psimd.c",
864 "src/requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800865]
866
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867# ISA-specific micro-kernels
868NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700869 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
870 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700871 "src/f32-clamp/gen/neon-x4.c",
872 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700873 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
874 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700875 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
876 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700877 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
878 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700879 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
880 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700881 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
882 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
883 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
884 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700885 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
886 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
887 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
888 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700889 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
890 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
891 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
892 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700893 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700894 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
895 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700896 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
897 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
898 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
899 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
900 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
901 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
902 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
903 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
904 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
905 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
906 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
907 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
908 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
909 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
910 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
911 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
912 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
913 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
914 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
915 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
916 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
917 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
918 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
919 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
920 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
921 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
922 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
923 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
924 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
925 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
926 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700927 "src/f32-hswish/gen/hswish-neon-x4.c",
928 "src/f32-hswish/gen/hswish-neon-x8.c",
929 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700930 "src/f32-ibilinear/gen/neon-c4.c",
931 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700932 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
933 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
934 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
935 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
936 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
937 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
938 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
939 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
940 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
941 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
942 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
943 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
944 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
945 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
946 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
947 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700948 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
949 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
950 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700951 "src/f32-ppmm/gen/4x8-minmax-neon.c",
952 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800953 "src/f32-prelu/gen/neon-2x4.c",
954 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800955 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
956 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
957 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
958 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
959 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
960 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
961 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
962 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
963 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
964 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
965 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
966 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
967 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
968 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
969 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
970 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
971 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
972 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
973 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
974 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
975 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
976 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
977 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
978 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700979 "src/f32-relu/gen/neon-x4.c",
980 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700981 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800982 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800983 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
984 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
985 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
986 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
987 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
988 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
989 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
990 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
991 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
992 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
993 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
994 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
995 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
996 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
997 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
998 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
999 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1000 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001001 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1002 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1003 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1004 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001005 "src/f32-vbinary/gen/vmax-neon-x4.c",
1006 "src/f32-vbinary/gen/vmax-neon-x8.c",
1007 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1008 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1009 "src/f32-vbinary/gen/vmin-neon-x4.c",
1010 "src/f32-vbinary/gen/vmin-neon-x8.c",
1011 "src/f32-vbinary/gen/vminc-neon-x4.c",
1012 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001013 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1014 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1015 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1016 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1017 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1018 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001019 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1020 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1021 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1022 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001023 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1024 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1025 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1026 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001027 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1028 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001029 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1030 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001031 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1032 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1033 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1034 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1035 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1036 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1037 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1038 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001039 "src/f32-vunary/gen/vabs-neon-x4.c",
1040 "src/f32-vunary/gen/vabs-neon-x8.c",
1041 "src/f32-vunary/gen/vneg-neon-x4.c",
1042 "src/f32-vunary/gen/vneg-neon-x8.c",
1043 "src/f32-vunary/gen/vsqr-neon-x4.c",
1044 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001045 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1046 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1047 "src/qu8-dwconv/up8x9-minmax-neon.c",
1048 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1049 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1050 "src/qu8-gemm/4x8-minmax-neon.c",
1051 "src/qu8-gemm/8x8-minmax-neon.c",
1052 "src/qu8-igemm/4x8-minmax-neon.c",
1053 "src/qu8-igemm/8x8-minmax-neon.c",
1054 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001055 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001056 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001057 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001058 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001059 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001060 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001061 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001062 "src/x32-zip/x2-neon.c",
1063 "src/x32-zip/x3-neon.c",
1064 "src/x32-zip/x4-neon.c",
1065 "src/x32-zip/xm-neon.c",
1066 "src/x8-zip/x2-neon.c",
1067 "src/x8-zip/x3-neon.c",
1068 "src/x8-zip/x4-neon.c",
1069 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001070 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001071 "src/math/roundd-neon-addsub.c",
1072 "src/math/roundd-neon-cvt.c",
1073 "src/math/roundu-neon-addsub.c",
1074 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001075 "src/math/roundz-neon-addsub.c",
1076 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001077 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001078 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1079 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1080 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1081 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1082 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1083 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001084 "src/math/sqrt-neon-nr1rsqrts.c",
1085 "src/math/sqrt-neon-nr2rsqrts.c",
1086 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001087 "src/requantization/precise-neon.c",
1088 "src/requantization/fp32-neon.c",
1089 "src/requantization/q31-neon.c",
1090 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001091]
1092
1093NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001094 "src/f32-ibilinear/gen/neonfma-c4.c",
1095 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001096 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1097 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1098 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1099 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1100 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1101 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1102 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1103 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1104 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001105 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1106 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1107 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1108 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001109 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1110 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1111 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1112 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001113 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1114 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1115 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1116 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001117 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1118 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1119 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1120 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1121 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1122 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1123 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1124 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1125 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1126 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1127 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1128 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1129 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1130 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1131 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1132 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1133 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1134 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001135 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1136 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001137 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1138 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1139 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1140 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1141 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1142 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1143 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1144 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1145 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1146 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1147 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1148 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1149 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1150 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1151 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1152 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1153 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1154 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1155 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1156 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1157 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1158 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1159 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1160 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001161 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1162 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1163 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1164 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1165 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1166 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1167 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1168 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1169 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1170 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1171 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1172 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1173 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1174 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1175 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1176 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1177 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001215 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1216 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001217 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1218 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1219 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1220 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1221 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1222 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1223 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1224 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1225 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1226 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1227 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1228 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1229 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1230 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1231 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1232 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1233 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1234 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1235 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1236 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001237 "src/math/exp-neonfma-lut64-p2.c",
1238 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001239 "src/math/expminus-neonfma-lut2048-p1.c",
1240 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001241 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001242 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1243 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1244 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1245 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1246 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1247 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1248 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1249 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1250 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1251 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1252 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1253 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1254 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1255 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1256 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1257 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1258 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1259 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001260 "src/math/sqrt-neonfma-nr1fma.c",
1261 "src/math/sqrt-neonfma-nr2fma.c",
1262 "src/math/sqrt-neonfma-nr3fma.c",
1263 "src/math/sqrt-neonfma-nr2fma1adj.c",
1264 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001265]
1266
1267AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001268 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1269 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1270 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1271 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1272 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1273 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001274 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1275 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1276 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1277 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1278 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1279 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1280 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1281 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1282 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1283 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1284 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1285 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1286 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1287 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1288 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1289 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1290 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1291 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1292 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1293 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001294 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1295 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001296 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1297 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001298 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1299 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001300 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1301 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001302 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1303 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1304 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1305 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1306 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001307 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1308 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1309 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1310 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1311 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1312 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1313 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1314 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001325 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1326 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1327 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1328 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1329 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1330 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1331 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1332 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1333 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1334 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1335 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1336 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1337 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1338 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1339 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1340 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1341 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1342 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001343 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1344 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001345 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1346 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1347 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1348 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1349 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1350 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001351]
1352
Marat Dukhan8853b822020-05-07 12:19:01 -07001353NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001354 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1355 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1356 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1357 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1358 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1359 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1360 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1361 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001362 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001363 "src/math/roundd-neonv8.c",
1364 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001365 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001366]
1367
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001369 "src/f16-clamp/gen/neonfp16arith-x16.c",
1370 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001371 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1372 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1373 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1374 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1375 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1376 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1377 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1378 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1379 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1380 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1381 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1382 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001383 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1384 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001385 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1386 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001387 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1388 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001389 "src/f16-relu/gen/neonfp16arith-x16.c",
1390 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001391 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1392 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1393 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1394 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1395 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1396 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1397 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1398 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1399 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1400 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1401 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1402 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1403 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1404 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1405 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1406 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1407 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1408 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1409 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1410 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1411 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1412 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1413 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1414 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1415 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1416 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1417 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1418 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001419 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1420 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001421 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1422 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1423 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1424 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1425 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1426 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1427 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1428 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001429 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1430 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1431 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1432 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001433 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1434 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1435 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1436 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1437 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1438 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1439 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1440 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1441 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1442 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1443 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1444 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001445 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1446 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1447 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1448 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1449 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1450 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1451 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1452 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001453]
1454
1455SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001456 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1457 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001458 "src/f32-clamp/gen/sse-x4.c",
1459 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001460 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1461 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001462 "src/f32-dwconv-chw/3x3p1-sse.c",
1463 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001464 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1465 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1466 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1467 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1468 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1469 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1470 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1471 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1472 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1473 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1474 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1475 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001476 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001477 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1478 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001479 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1480 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1481 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1482 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1483 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1484 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1485 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1486 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1487 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1488 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1489 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1490 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1491 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001492 "src/f32-hswish/gen/hswish-sse-x4.c",
1493 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001494 "src/f32-ibilinear/gen/sse-c4.c",
1495 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001496 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1497 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1498 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1499 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1500 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1501 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1502 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001503 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1504 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1505 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001506 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001507 "src/f32-prelu/gen/sse-2x4.c",
1508 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001509 "src/f32-relu/gen/sse-x4.c",
1510 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001511 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001512 "src/f32-spmm/gen/4x1-minmax-sse.c",
1513 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001514 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001515 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1516 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1517 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1518 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1519 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1520 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1521 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1522 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001523 "src/f32-vbinary/gen/vmax-sse-x4.c",
1524 "src/f32-vbinary/gen/vmax-sse-x8.c",
1525 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1526 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1527 "src/f32-vbinary/gen/vmin-sse-x4.c",
1528 "src/f32-vbinary/gen/vmin-sse-x8.c",
1529 "src/f32-vbinary/gen/vminc-sse-x4.c",
1530 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001531 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1532 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1533 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1534 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1535 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1536 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1537 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1538 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001539 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1540 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1541 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1542 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001543 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1544 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1545 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1546 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001547 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1548 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001549 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1550 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001551 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1552 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001553 "src/f32-vunary/gen/vabs-sse-x4.c",
1554 "src/f32-vunary/gen/vabs-sse-x8.c",
1555 "src/f32-vunary/gen/vneg-sse-x4.c",
1556 "src/f32-vunary/gen/vneg-sse-x8.c",
1557 "src/f32-vunary/gen/vsqr-sse-x4.c",
1558 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001559 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001561 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001562 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001563 "src/math/roundd-sse-addsub.c",
1564 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001565 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001566 "src/math/sqrt-sse-hh1mac.c",
1567 "src/math/sqrt-sse-nr1mac.c",
1568 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569]
1570
1571SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001572 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1573 "src/f32-argmaxpool/4x-sse2-c4.c",
1574 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001575 "src/f32-prelu/gen/sse2-2x4.c",
1576 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001577 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1578 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1579 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1580 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1581 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1582 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1583 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1584 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1585 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1586 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1587 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1588 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001589 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001590 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001591 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001592 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001593 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1594 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001595 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1596 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001597 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1598 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1599 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1600 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1601 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1602 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1603 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1604 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001605 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1606 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1607 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1608 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1609 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1610 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1611 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1612 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1613 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001614 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001615 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001616 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001617 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618 "src/x32-zip/x2-sse2.c",
1619 "src/x32-zip/x3-sse2.c",
1620 "src/x32-zip/x4-sse2.c",
1621 "src/x32-zip/xm-sse2.c",
1622 "src/x8-zip/x2-sse2.c",
1623 "src/x8-zip/x3-sse2.c",
1624 "src/x8-zip/x4-sse2.c",
1625 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001626 "src/math/exp-sse2-p5.c",
1627 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001628 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001629 "src/math/roundd-sse2-cvt.c",
1630 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001631 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001632 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001633 "src/requantization/precise-sse2.c",
1634 "src/requantization/fp32-sse2.c",
1635 "src/requantization/q31-sse2.c",
1636 "src/requantization/gemmlowp-sse2.c",
1637]
1638
1639SSSE3_UKERNELS = [
1640 "src/requantization/precise-ssse3.c",
1641 "src/requantization/q31-ssse3.c",
1642 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001643]
1644
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001645SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001646 "src/f32-prelu/gen/sse41-2x4.c",
1647 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001648 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001649 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001650 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001651 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001652 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1653 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001654 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1655 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1657 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1658 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1659 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1660 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1661 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1662 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1663 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001664 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001665 "src/math/roundd-sse41.c",
1666 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001667 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001668 "src/requantization/precise-sse4.c",
1669 "src/requantization/q31-sse4.c",
1670 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001671]
1672
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001674 "src/f32-clamp/gen/avx-x8.c",
1675 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001676 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1677 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1678 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1679 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1680 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1681 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1682 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1683 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1684 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1685 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1686 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1687 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1688 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1689 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1690 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1691 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1692 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1693 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1694 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1695 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1696 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1697 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1698 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1699 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1700 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1701 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1702 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1703 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1704 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1705 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001706 "src/f32-hswish/gen/hswish-avx-x8.c",
1707 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001708 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1709 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1710 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1711 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1712 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1713 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1714 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1715 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1716 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001717 "src/f32-prelu/gen/avx-2x8.c",
1718 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001719 "src/f32-relu/gen/avx-x16.c",
1720 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001722 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1723 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1724 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1725 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1726 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1727 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1728 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1729 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001730 "src/f32-vbinary/gen/vmax-avx-x8.c",
1731 "src/f32-vbinary/gen/vmax-avx-x16.c",
1732 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1733 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1734 "src/f32-vbinary/gen/vmin-avx-x8.c",
1735 "src/f32-vbinary/gen/vmin-avx-x16.c",
1736 "src/f32-vbinary/gen/vminc-avx-x8.c",
1737 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001738 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1739 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1740 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1741 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1742 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1743 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1744 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1745 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001746 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1747 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1748 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1749 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001750 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1751 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1752 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1753 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001754 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1755 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001756 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001757 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1758 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1759 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1760 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1761 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1762 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1763 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1764 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001765 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1766 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001767 "src/f32-vunary/gen/vabs-avx-x8.c",
1768 "src/f32-vunary/gen/vabs-avx-x16.c",
1769 "src/f32-vunary/gen/vneg-avx-x8.c",
1770 "src/f32-vunary/gen/vneg-avx-x16.c",
1771 "src/f32-vunary/gen/vsqr-avx-x8.c",
1772 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773]
1774
Marat Dukhanfda12b82019-11-21 12:27:59 -08001775FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001776 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1777 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1778 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1779 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1780 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1781 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1782 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1783 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1784 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1785 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1786 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1787 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1788 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1789 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1790 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1791 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1792 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1793 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1794 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1795 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1796 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1797 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1798 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1799 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1800 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1801 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1802 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1803 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1804 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1805 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1806 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1807 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1808 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1809 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001816 "src/f32-hswish/gen/hswish-fma3-x8.c",
1817 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001818 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1819 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1820 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1821 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1822 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1823 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1824 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1825 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1826 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1827 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1828 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1829 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1830 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1831 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001832 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1833 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1834 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1835 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1836 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1837 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1838 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1839 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001840 "src/math/sqrt-fma3-nr1fma.c",
1841 "src/math/sqrt-fma3-nr2fma.c",
1842 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001843]
1844
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001845AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001846 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1847 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1848 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1849 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1850 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1851 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1852 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1853 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1854 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1855 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1856 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1857 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1858 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1859 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1860 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1861 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1862 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1863 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1864 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1865 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1866 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1867 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1868 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1869 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1870 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1871 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1872 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1873 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1874 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1875 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1876 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1877 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1878 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1879 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1880 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1881 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001882 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1883 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1884 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1885 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1886 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1887 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1888 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1889 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1890 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1891 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1892 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1893 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1894 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1895 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1896 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1897 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1898 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1899 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1900 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1901 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1902 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1903 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001912 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1913 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1914 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1915 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1916 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1917 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1918 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1919 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1920 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1921 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1924 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1925 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1926 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1927 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1928 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1929 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1930 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1931 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1932 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1933 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1934 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1935 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001936 "src/math/exp-avx2-p5.c",
1937 "src/math/exp-avx2-perm-p3.c",
1938 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001939 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001940 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001941 "src/math/sigmoid-avx2-rr2-p5-div.c",
1942 "src/math/sigmoid-avx2-rr1-p5-div.c",
1943 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1944 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1945 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1946 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001947]
1948
Marat Dukhan08c4a432019-10-03 09:29:21 -07001949AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001950 "src/f32-clamp/gen/avx512f-x16.c",
1951 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001952 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1953 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1954 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1955 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1956 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1957 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1958 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1959 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1960 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1961 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1962 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1963 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1964 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1965 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1966 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1967 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1968 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1969 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1970 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1971 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1972 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1973 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1974 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1975 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001976 "src/f32-hswish/gen/hswish-avx512f-x16.c",
1977 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001978 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1979 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1980 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1981 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1982 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1983 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001984 "src/f32-prelu/gen/avx512f-2x16.c",
1985 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001986 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1987 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1988 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1989 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1990 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1991 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1992 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1993 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1994 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1995 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1996 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1997 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1998 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1999 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2000 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2001 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2002 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2003 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2004 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2005 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2006 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2007 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2008 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2009 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2010 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2011 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2012 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2013 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2014 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2015 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2016 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2017 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2018 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2019 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002022 "src/f32-relu/gen/avx512f-x16.c",
2023 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002024 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002025 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2026 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2027 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2028 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2029 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2030 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2031 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2032 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002033 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2034 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2035 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2036 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2037 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2038 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2039 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2040 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002041 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2042 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2043 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2044 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2045 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2046 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2047 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2048 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002049 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2050 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2051 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2052 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002053 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2054 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2055 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2056 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002057 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2058 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002059 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002060 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2061 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2062 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2063 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2064 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2065 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2066 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2067 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2068 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2072 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2073 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2074 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2075 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2076 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2077 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2078 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2079 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2080 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002084 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2085 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2086 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2087 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2088 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2089 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2090 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2091 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002092 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2093 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2094 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2095 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2096 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2097 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2098 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2099 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002100 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2101 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2102 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2103 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2104 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2105 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002106 "src/math/exp-avx512f-p5-scalef.c",
2107 "src/math/exp-avx512f-p5.c",
2108 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002109 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002110 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002111 "src/math/sqrt-avx512f-nr1fma.c",
2112 "src/math/sqrt-avx512f-nr2fma.c",
2113 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002114]
2115
2116AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002117 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002118 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002119 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2120 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002121 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002122 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002123 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002124 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002125 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2126 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002127 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2128 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2129 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2130 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002131]
2132
2133AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002134 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002135 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2136 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002137 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002138 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2139 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002140 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2141 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2142 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002143 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002144 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2145 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2146 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002147 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002148 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2149 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002150 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002151 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2152 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2153 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2154 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2155 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2156 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2157 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2158 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2159 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2160 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2161 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2162 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2163 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2164 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2165 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2166 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2167 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2168 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002169 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2170 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002171 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002172 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2173 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2174 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2175 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2176 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2177 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2178 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2179 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2180 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2181 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2182 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2183 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2184 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2185 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2186 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2187 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2188 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2189 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002190 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2191 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2192 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2193 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2194 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2195 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2196 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2197 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2198 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2199 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2200 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2201 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2202 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2203 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2204 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2205 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2206 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2207 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208]
2209
Marat Dukhan1b354632020-03-23 12:50:22 -07002210INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002211 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002212 "src/xnnpack/argmaxpool.h",
2213 "src/xnnpack/avgpool.h",
2214 "src/xnnpack/clamp.h",
2215 "src/xnnpack/common.h",
2216 "src/xnnpack/conv.h",
2217 "src/xnnpack/dwconv.h",
2218 "src/xnnpack/gavgpool.h",
2219 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002220 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002221 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002222 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002223 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002224 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002225 "src/xnnpack/lut.h",
2226 "src/xnnpack/math.h",
2227 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002228 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229 "src/xnnpack/packx.h",
2230 "src/xnnpack/pad.h",
2231 "src/xnnpack/params.h",
2232 "src/xnnpack/pavgpool.h",
2233 "src/xnnpack/ppmm.h",
2234 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002235 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002236 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002237 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/xnnpack/rmax.h",
2239 "src/xnnpack/scalar-utils.h",
2240 "src/xnnpack/spmm.h",
2241 "src/xnnpack/unpool.h",
2242 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002243 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002245 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002246 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002247 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002248 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002250]
2251
2252INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002253 "include/xnnpack.h",
2254 "src/xnnpack/allocator.h",
2255 "src/xnnpack/compute.h",
2256 "src/xnnpack/im2col.h",
2257 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002258 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002259 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002260 "src/xnnpack/operator.h",
2261 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002262 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002263 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002264 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002265 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002266]
2267
Marat Dukhan1b354632020-03-23 12:50:22 -07002268ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002269 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002270]
2271
Marat Dukhan1b354632020-03-23 12:50:22 -07002272MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002273 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002274 "include/xnnpack.h",
2275]
2276
Marat Dukhan1b354632020-03-23 12:50:22 -07002277MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002279 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002280 "src/xnnpack/requantization.h",
2281 "include/xnnpack.h",
2282]
2283
2284OPERATOR_TEST_PARAMS_HDRS = [
2285 "src/xnnpack/params.h",
2286 "src/xnnpack/common.h",
2287]
2288
2289WEIGHTS_PACK_HDRS = [
2290 "src/xnnpack/pack.h",
2291 "src/xnnpack/operator.h",
2292 "src/xnnpack/compute.h",
2293]
2294
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002295LOGGING_COPTS = select({
2296 # No logging in optimized mode
2297 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2298 # Full logging in debug mode
2299 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2300 # Error-only logging in default (fastbuild) mode
2301 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2302})
2303
Marat Dukhan3b59de22020-06-03 20:15:19 -07002304LOGGING_SRCS = select({
2305 # No logging in optimized mode
2306 ":optimized_build": [],
2307 "//conditions:default": [
2308 "src/operator-strings.c",
2309 "src/subgraph-strings.c",
2310 ],
2311})
2312
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002313LOGGING_HDRS = [
2314 "src/xnnpack/log.h",
2315]
2316
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002318 name = "tables",
2319 srcs = TABLE_SRCS,
2320 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002321 gcc_copts = xnnpack_gcc_std_copts(),
2322 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002323)
2324
2325xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326 name = "scalar_ukernels",
2327 srcs = SCALAR_UKERNELS,
2328 hdrs = INTERNAL_HDRS,
2329 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002330 gcc_copts = xnnpack_gcc_std_copts(),
2331 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002333 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 "@FP16",
2335 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002336 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337 ],
2338)
2339
2340xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002341 name = "scalar_ukernels_test_mode",
2342 srcs = SCALAR_UKERNELS,
2343 hdrs = INTERNAL_HDRS,
2344 aarch32_copts = ["-marm"],
2345 copts = [
2346 "-UNDEBUG",
2347 "-DXNN_TEST_MODE=1",
2348 ],
2349 gcc_copts = xnnpack_gcc_std_copts(),
2350 msvc_copts = xnnpack_msvc_std_copts(),
2351 deps = [
2352 ":tables",
2353 "@FP16",
2354 "@FXdiv",
2355 "@pthreadpool",
2356 ],
2357)
2358
2359xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002360 name = "wasm_ukernels",
2361 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002362 gcc_copts = xnnpack_gcc_std_copts(),
2363 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002364 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002365 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002366 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002367 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002368 "@FP16",
2369 "@FXdiv",
2370 "@pthreadpool",
2371 ],
2372)
2373
2374xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002375 name = "wasm_ukernels_test_mode",
2376 hdrs = INTERNAL_HDRS,
2377 copts = [
2378 "-UNDEBUG",
2379 "-DXNN_TEST_MODE=1",
2380 ],
2381 gcc_copts = xnnpack_gcc_std_copts(),
2382 msvc_copts = xnnpack_msvc_std_copts(),
2383 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002384 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002385 deps = [
2386 ":tables",
2387 "@FP16",
2388 "@FXdiv",
2389 "@pthreadpool",
2390 ],
2391)
2392
2393xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002394 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002395 hdrs = INTERNAL_HDRS,
2396 aarch32_copts = [
2397 "-marm",
2398 "-mfpu=neon",
2399 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002400 gcc_copts = xnnpack_gcc_std_copts(),
2401 gcc_x86_copts = ["-msse2"],
2402 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002403 optimized_copts = [
2404 "-O3",
2405 "-ffast-math",
2406 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002407 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002408 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002409 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002410 "@FP16",
2411 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002412 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413 ],
2414)
2415
2416xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002417 name = "psimd_fastmath_ukernels_test_mode",
2418 hdrs = INTERNAL_HDRS,
2419 aarch32_copts = [
2420 "-marm",
2421 "-mfpu=neon",
2422 ],
2423 copts = [
2424 "-UNDEBUG",
2425 "-DXNN_TEST_MODE=1",
2426 ],
2427 gcc_copts = xnnpack_gcc_std_copts(),
2428 gcc_x86_copts = ["-msse2"],
2429 msvc_copts = xnnpack_msvc_std_copts(),
2430 optimized_copts = [
2431 "-O3",
2432 "-ffast-math",
2433 ],
2434 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2435 deps = [
2436 ":tables",
2437 "@FP16",
2438 "@psimd",
2439 "@pthreadpool",
2440 ],
2441)
2442
2443xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002444 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002445 hdrs = INTERNAL_HDRS,
2446 aarch32_copts = [
2447 "-marm",
2448 "-mfpu=neon",
2449 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002450 gcc_copts = xnnpack_gcc_std_copts(),
2451 gcc_x86_copts = ["-msse2"],
2452 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002453 optimized_copts = [
2454 "-O3",
2455 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002456 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002457 deps = [
2458 ":tables",
2459 "@FP16",
2460 "@psimd",
2461 "@pthreadpool",
2462 ],
2463)
2464
2465xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002466 name = "psimd_accmath_ukernels_test_mode",
2467 hdrs = INTERNAL_HDRS,
2468 aarch32_copts = [
2469 "-marm",
2470 "-mfpu=neon",
2471 ],
2472 copts = [
2473 "-UNDEBUG",
2474 "-DXNN_TEST_MODE=1",
2475 ],
2476 gcc_copts = xnnpack_gcc_std_copts(),
2477 gcc_x86_copts = ["-msse2"],
2478 msvc_copts = xnnpack_msvc_std_copts(),
2479 optimized_copts = [
2480 "-O3",
2481 ],
2482 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2483 deps = [
2484 ":tables",
2485 "@FP16",
2486 "@psimd",
2487 "@pthreadpool",
2488 ],
2489)
2490
2491xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002492 name = "neon_ukernels",
2493 hdrs = INTERNAL_HDRS,
2494 aarch32_copts = [
2495 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002496 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002497 "-mfpu=neon",
2498 ],
2499 aarch32_srcs = NEON_UKERNELS,
2500 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002501 gcc_copts = xnnpack_gcc_std_copts(),
2502 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002503 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002504 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002505 "@FP16",
2506 "@pthreadpool",
2507 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002508)
2509
2510xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002511 name = "neon_ukernels_test_mode",
2512 hdrs = INTERNAL_HDRS,
2513 aarch32_copts = [
2514 "-marm",
2515 "-march=armv7-a",
2516 "-mfpu=neon",
2517 ],
2518 aarch32_srcs = NEON_UKERNELS,
2519 aarch64_srcs = NEON_UKERNELS,
2520 copts = [
2521 "-UNDEBUG",
2522 "-DXNN_TEST_MODE=1",
2523 ],
2524 gcc_copts = xnnpack_gcc_std_copts(),
2525 msvc_copts = xnnpack_msvc_std_copts(),
2526 deps = [
2527 ":tables",
2528 "@FP16",
2529 "@pthreadpool",
2530 ],
2531)
2532
2533xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534 name = "neonfma_ukernels",
2535 hdrs = INTERNAL_HDRS,
2536 aarch32_copts = [
2537 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002538 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002539 "-mfpu=neon-vfpv4",
2540 ],
2541 aarch32_srcs = NEONFMA_UKERNELS,
2542 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002543 apple_aarch32_copts = [
2544 "-mcpu=swift",
2545 "-mtune=generic",
2546 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002547 gcc_copts = xnnpack_gcc_std_copts(),
2548 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002549 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002550 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002551 "@FP16",
2552 "@pthreadpool",
2553 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554)
2555
2556xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002557 name = "neonfma_ukernels_test_mode",
2558 hdrs = INTERNAL_HDRS,
2559 aarch32_copts = [
2560 "-marm",
2561 "-march=armv7-a",
2562 "-mfpu=neon-vfpv4",
2563 ],
2564 aarch32_srcs = NEONFMA_UKERNELS,
2565 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002566 apple_aarch32_copts = [
2567 "-mcpu=swift",
2568 "-mtune=generic",
2569 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002570 copts = [
2571 "-UNDEBUG",
2572 "-DXNN_TEST_MODE=1",
2573 ],
2574 gcc_copts = xnnpack_gcc_std_copts(),
2575 msvc_copts = xnnpack_msvc_std_copts(),
2576 deps = [
2577 ":tables",
2578 "@FP16",
2579 "@pthreadpool",
2580 ],
2581)
2582
2583xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002584 name = "neonv8_ukernels",
2585 hdrs = INTERNAL_HDRS,
2586 aarch32_copts = [
2587 "-marm",
2588 "-march=armv8-a",
2589 "-mfpu=neon-fp-armv8",
2590 ],
2591 aarch32_srcs = NEONV8_UKERNELS,
2592 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002593 apple_aarch32_copts = [
2594 "-mcpu=cyclone",
2595 "-mtune=generic",
2596 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002597 gcc_copts = xnnpack_gcc_std_copts(),
2598 msvc_copts = xnnpack_msvc_std_copts(),
2599 deps = [
2600 ":tables",
2601 "@FP16",
2602 "@pthreadpool",
2603 ],
2604)
2605
2606xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002607 name = "neonv8_ukernels_test_mode",
2608 hdrs = INTERNAL_HDRS,
2609 aarch32_copts = [
2610 "-marm",
2611 "-march=armv8-a",
2612 "-mfpu=neon-fp-armv8",
2613 ],
2614 aarch32_srcs = NEONV8_UKERNELS,
2615 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002616 apple_aarch32_copts = [
2617 "-mcpu=cyclone",
2618 "-mtune=generic",
2619 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002620 copts = [
2621 "-UNDEBUG",
2622 "-DXNN_TEST_MODE=1",
2623 ],
2624 gcc_copts = xnnpack_gcc_std_copts(),
2625 msvc_copts = xnnpack_msvc_std_copts(),
2626 deps = [
2627 ":tables",
2628 "@FP16",
2629 "@pthreadpool",
2630 ],
2631)
2632
2633xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002634 name = "neonfp16arith_ukernels",
2635 hdrs = INTERNAL_HDRS,
2636 aarch64_copts = ["-march=armv8.2-a+fp16"],
2637 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002638 gcc_copts = xnnpack_gcc_std_copts(),
2639 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002640 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002641 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002642 "@FP16",
2643 "@pthreadpool",
2644 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645)
2646
2647xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002648 name = "neonfp16arith_ukernels_test_mode",
2649 hdrs = INTERNAL_HDRS,
2650 aarch64_copts = ["-march=armv8.2-a+fp16"],
2651 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2652 copts = [
2653 "-UNDEBUG",
2654 "-DXNN_TEST_MODE=1",
2655 ],
2656 gcc_copts = xnnpack_gcc_std_copts(),
2657 msvc_copts = xnnpack_msvc_std_copts(),
2658 deps = [
2659 ":tables",
2660 "@FP16",
2661 "@pthreadpool",
2662 ],
2663)
2664
2665xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666 name = "sse2_ukernels",
2667 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002668 gcc_copts = xnnpack_gcc_std_copts(),
2669 gcc_x86_copts = ["-msse2"],
2670 msvc_copts = xnnpack_msvc_std_copts(),
2671 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002672 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002673 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002674 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002675 "@FP16",
2676 "@pthreadpool",
2677 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678)
2679
2680xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002681 name = "sse2_ukernels_test_mode",
2682 hdrs = INTERNAL_HDRS,
2683 copts = [
2684 "-UNDEBUG",
2685 "-DXNN_TEST_MODE=1",
2686 ],
2687 gcc_copts = xnnpack_gcc_std_copts(),
2688 gcc_x86_copts = ["-msse2"],
2689 msvc_copts = xnnpack_msvc_std_copts(),
2690 msvc_x86_32_copts = ["/arch:SSE2"],
2691 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2692 deps = [
2693 ":tables",
2694 "@FP16",
2695 "@pthreadpool",
2696 ],
2697)
2698
2699xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002700 name = "ssse3_ukernels",
2701 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002702 gcc_copts = xnnpack_gcc_std_copts(),
2703 gcc_x86_copts = ["-mssse3"],
2704 msvc_copts = xnnpack_msvc_std_copts(),
2705 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002706 x86_srcs = SSSE3_UKERNELS,
2707 deps = [
2708 ":tables",
2709 "@FP16",
2710 "@pthreadpool",
2711 ],
2712)
2713
2714xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002715 name = "ssse3_ukernels_test_mode",
2716 hdrs = INTERNAL_HDRS,
2717 copts = [
2718 "-UNDEBUG",
2719 "-DXNN_TEST_MODE=1",
2720 ],
2721 gcc_copts = xnnpack_gcc_std_copts(),
2722 gcc_x86_copts = ["-mssse3"],
2723 msvc_copts = xnnpack_msvc_std_copts(),
2724 msvc_x86_32_copts = ["/arch:SSE2"],
2725 x86_srcs = SSSE3_UKERNELS,
2726 deps = [
2727 ":tables",
2728 "@FP16",
2729 "@pthreadpool",
2730 ],
2731)
2732
2733xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002734 name = "sse41_ukernels",
2735 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002736 gcc_copts = xnnpack_gcc_std_copts(),
2737 gcc_x86_copts = ["-msse4.1"],
2738 msvc_copts = xnnpack_msvc_std_copts(),
2739 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002740 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002741 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002742 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002743 "@FP16",
2744 "@pthreadpool",
2745 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002746)
2747
2748xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002749 name = "sse41_ukernels_test_mode",
2750 hdrs = INTERNAL_HDRS,
2751 copts = [
2752 "-UNDEBUG",
2753 "-DXNN_TEST_MODE=1",
2754 ],
2755 gcc_copts = xnnpack_gcc_std_copts(),
2756 gcc_x86_copts = ["-msse4.1"],
2757 msvc_copts = xnnpack_msvc_std_copts(),
2758 msvc_x86_32_copts = ["/arch:SSE2"],
2759 x86_srcs = SSE41_UKERNELS,
2760 deps = [
2761 ":tables",
2762 "@FP16",
2763 "@pthreadpool",
2764 ],
2765)
2766
2767xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002768 name = "avx_ukernels",
2769 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002770 gcc_copts = xnnpack_gcc_std_copts(),
2771 gcc_x86_copts = ["-mavx"],
2772 msvc_copts = xnnpack_msvc_std_copts(),
2773 msvc_x86_32_copts = ["/arch:AVX"],
2774 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002775 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002776 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002777 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002778 "@FP16",
2779 "@pthreadpool",
2780 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002781)
2782
2783xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002784 name = "avx_ukernels_test_mode",
2785 hdrs = INTERNAL_HDRS,
2786 copts = [
2787 "-UNDEBUG",
2788 "-DXNN_TEST_MODE=1",
2789 ],
2790 gcc_copts = xnnpack_gcc_std_copts(),
2791 gcc_x86_copts = ["-mavx"],
2792 msvc_copts = xnnpack_msvc_std_copts(),
2793 msvc_x86_32_copts = ["/arch:AVX"],
2794 msvc_x86_64_copts = ["/arch:AVX"],
2795 x86_srcs = AVX_UKERNELS,
2796 deps = [
2797 ":tables",
2798 "@FP16",
2799 "@pthreadpool",
2800 ],
2801)
2802
2803xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002804 name = "fma3_ukernels",
2805 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002806 gcc_copts = xnnpack_gcc_std_copts(),
2807 gcc_x86_copts = ["-mfma"],
2808 msvc_copts = xnnpack_msvc_std_copts(),
2809 msvc_x86_32_copts = ["/arch:AVX"],
2810 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002811 x86_srcs = FMA3_UKERNELS,
2812 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002813 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002814 "@FP16",
2815 "@pthreadpool",
2816 ],
2817)
2818
2819xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002820 name = "fma3_ukernels_test_mode",
2821 hdrs = INTERNAL_HDRS,
2822 copts = [
2823 "-UNDEBUG",
2824 "-DXNN_TEST_MODE=1",
2825 ],
2826 gcc_copts = xnnpack_gcc_std_copts(),
2827 gcc_x86_copts = ["-mfma"],
2828 msvc_copts = xnnpack_msvc_std_copts(),
2829 msvc_x86_32_copts = ["/arch:AVX"],
2830 msvc_x86_64_copts = ["/arch:AVX"],
2831 x86_srcs = FMA3_UKERNELS,
2832 deps = [
2833 ":tables",
2834 "@FP16",
2835 "@pthreadpool",
2836 ],
2837)
2838
2839xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002840 name = "avx2_ukernels",
2841 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002842 gcc_copts = xnnpack_gcc_std_copts(),
2843 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002844 "-mfma",
2845 "-mavx2",
2846 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002847 msvc_copts = xnnpack_msvc_std_copts(),
2848 msvc_x86_32_copts = ["/arch:AVX2"],
2849 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002850 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002851 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002852 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002853 "@FP16",
2854 "@pthreadpool",
2855 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002856)
2857
2858xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002859 name = "avx2_ukernels_test_mode",
2860 hdrs = INTERNAL_HDRS,
2861 copts = [
2862 "-UNDEBUG",
2863 "-DXNN_TEST_MODE=1",
2864 ],
2865 gcc_copts = xnnpack_gcc_std_copts(),
2866 gcc_x86_copts = [
2867 "-mfma",
2868 "-mavx2",
2869 ],
2870 msvc_copts = xnnpack_msvc_std_copts(),
2871 msvc_x86_32_copts = ["/arch:AVX2"],
2872 msvc_x86_64_copts = ["/arch:AVX2"],
2873 x86_srcs = AVX2_UKERNELS,
2874 deps = [
2875 ":tables",
2876 "@FP16",
2877 "@pthreadpool",
2878 ],
2879)
2880
2881xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 name = "avx512f_ukernels",
2883 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002884 gcc_copts = xnnpack_gcc_std_copts(),
2885 gcc_x86_copts = ["-mavx512f"],
2886 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2887 msvc_copts = xnnpack_msvc_std_copts(),
2888 msvc_x86_32_copts = ["/arch:AVX512"],
2889 msvc_x86_64_copts = ["/arch:AVX512"],
2890 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002892 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002893 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002894 "@FP16",
2895 "@pthreadpool",
2896 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897)
2898
2899xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002900 name = "avx512f_ukernels_test_mode",
2901 hdrs = INTERNAL_HDRS,
2902 copts = [
2903 "-UNDEBUG",
2904 "-DXNN_TEST_MODE=1",
2905 ],
2906 gcc_copts = xnnpack_gcc_std_copts(),
2907 gcc_x86_copts = ["-mavx512f"],
2908 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2909 msvc_copts = xnnpack_msvc_std_copts(),
2910 msvc_x86_32_copts = ["/arch:AVX512"],
2911 msvc_x86_64_copts = ["/arch:AVX512"],
2912 msys_copts = ["-fno-asynchronous-unwind-tables"],
2913 x86_srcs = AVX512F_UKERNELS,
2914 deps = [
2915 ":tables",
2916 "@FP16",
2917 "@pthreadpool",
2918 ],
2919)
2920
2921xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 name = "asm_ukernels",
2923 hdrs = ["src/xnnpack/assembly.h"],
2924 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002925 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 aarch64_srcs = AARCH64_ASM_UKERNELS,
2927)
2928
Marat Dukhan3b59de22020-06-03 20:15:19 -07002929xnnpack_cc_library(
2930 name = "logging_utils",
2931 srcs = LOGGING_SRCS,
2932 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2933 copts = LOGGING_COPTS + [
2934 "-Isrc",
2935 "-Iinclude",
2936 ] + select({
2937 ":debug_build": [],
2938 "//conditions:default": xnnpack_min_size_copts(),
2939 }),
2940 gcc_copts = xnnpack_gcc_std_copts(),
2941 msvc_copts = xnnpack_msvc_std_copts(),
2942 visibility = xnnpack_visibility(),
2943 deps = [
2944 "@FP16",
2945 "@clog",
2946 "@pthreadpool",
2947 ],
2948)
2949
Marat Dukhan08c4a432019-10-03 09:29:21 -07002950xnnpack_aggregate_library(
2951 name = "ukernels",
2952 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002953 ":neon_ukernels",
2954 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002955 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 ":asm_ukernels",
2957 ],
2958 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002959 ":neon_ukernels",
2960 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002961 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 ":neonfp16arith_ukernels",
2963 ":asm_ukernels",
2964 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002965 generic_deps = [
2966 ":scalar_ukernels",
2967 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002968 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002969 ":psimd_fastmath_ukernels",
2970 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002972 wasm_deps = [
2973 ":wasm_ukernels",
2974 ],
2975 wasmsimd_deps = [
2976 ":wasm_ukernels",
2977 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002978 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002980 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002981 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002983 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002984 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 ":avx512f_ukernels",
2986 ],
2987)
2988
Marat Dukhan33fcf782020-05-24 14:27:15 -07002989xnnpack_aggregate_library(
2990 name = "ukernels_test_mode",
2991 aarch32_deps = [
2992 ":neon_ukernels_test_mode",
2993 ":neonfma_ukernels_test_mode",
2994 ":neonv8_ukernels_test_mode",
2995 ":asm_ukernels",
2996 ],
2997 aarch64_deps = [
2998 ":neon_ukernels_test_mode",
2999 ":neonfma_ukernels_test_mode",
3000 ":neonv8_ukernels_test_mode",
3001 ":neonfp16arith_ukernels_test_mode",
3002 ":asm_ukernels",
3003 ],
3004 generic_deps = [
3005 ":scalar_ukernels_test_mode",
3006 ],
3007 psimd_deps = [
3008 ":psimd_fastmath_ukernels_test_mode",
3009 ":psimd_accmath_ukernels_test_mode",
3010 ],
3011 wasm_deps = [
3012 ":wasm_ukernels_test_mode",
3013 ],
3014 wasmsimd_deps = [
3015 ":wasm_ukernels_test_mode",
3016 ],
3017 x86_deps = [
3018 ":sse2_ukernels_test_mode",
3019 ":ssse3_ukernels_test_mode",
3020 ":sse41_ukernels_test_mode",
3021 ":avx_ukernels_test_mode",
3022 ":fma3_ukernels_test_mode",
3023 ":avx2_ukernels_test_mode",
3024 ":avx512f_ukernels_test_mode",
3025 ],
3026)
3027
Marat Dukhan86df3842020-06-17 14:03:31 -07003028# This target must be compatible with Android NDK r17 and compile with Clang 6.
3029# It is similar to :ukernels target, but excludes microkernels which require
3030# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003031xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003032 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003033 aarch32_deps = [
3034 ":neon_ukernels",
3035 ":neonfma_ukernels",
3036 ":neonv8_ukernels",
3037 ":asm_ukernels",
3038 ],
3039 aarch64_deps = [
3040 ":neon_ukernels",
3041 ":neonfma_ukernels",
3042 ":neonv8_ukernels",
3043 ":asm_ukernels",
3044 ],
3045 generic_deps = [
3046 ":scalar_ukernels",
3047 ],
3048 psimd_deps = [
3049 ":psimd_fastmath_ukernels",
3050 ":psimd_accmath_ukernels",
3051 ],
3052 wasm_deps = [
3053 ":wasm_ukernels",
3054 ],
3055 wasmsimd_deps = [
3056 ":wasm_ukernels",
3057 ],
3058 x86_deps = [
3059 ":sse2_ukernels",
3060 ":ssse3_ukernels",
3061 ":sse41_ukernels",
3062 ":avx_ukernels",
3063 ":fma3_ukernels",
3064 ":avx2_ukernels",
3065 ":avx512f_ukernels",
3066 ],
3067)
3068
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069xnnpack_cc_library(
3070 name = "im2col",
3071 srcs = ["src/im2col.c"],
3072 hdrs = [
3073 "src/xnnpack/common.h",
3074 "src/xnnpack/im2col.h",
3075 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003076 gcc_copts = xnnpack_gcc_std_copts(),
3077 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078)
3079
3080xnnpack_cc_library(
3081 name = "indirection",
3082 srcs = ["src/indirection.c"],
3083 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003084 gcc_copts = xnnpack_gcc_std_copts(),
3085 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 deps = [
3087 "@FP16",
3088 "@FXdiv",
3089 "@pthreadpool",
3090 ],
3091)
3092
3093xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003094 name = "indirection_test_mode",
3095 srcs = ["src/indirection.c"],
3096 hdrs = INTERNAL_HDRS,
3097 copts = [
3098 "-UNDEBUG",
3099 "-DXNN_TEST_MODE=1",
3100 ],
3101 gcc_copts = xnnpack_gcc_std_copts(),
3102 msvc_copts = xnnpack_msvc_std_copts(),
3103 deps = [
3104 "@FP16",
3105 "@FXdiv",
3106 "@pthreadpool",
3107 ],
3108)
3109
3110xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003111 name = "packing",
3112 srcs = ["src/packing.c"],
3113 hdrs = INTERNAL_HDRS,
3114 gcc_copts = xnnpack_gcc_std_copts(),
3115 msvc_copts = xnnpack_msvc_std_copts(),
3116 deps = [
3117 "@FP16",
3118 "@FXdiv",
3119 "@pthreadpool",
3120 ],
3121)
3122
3123xnnpack_cc_library(
3124 name = "packing_test_mode",
3125 srcs = ["src/packing.c"],
3126 hdrs = INTERNAL_HDRS,
3127 copts = [
3128 "-UNDEBUG",
3129 "-DXNN_TEST_MODE=1",
3130 ],
3131 gcc_copts = xnnpack_gcc_std_copts(),
3132 msvc_copts = xnnpack_msvc_std_copts(),
3133 deps = [
3134 "@FP16",
3135 "@FXdiv",
3136 "@pthreadpool",
3137 ],
3138)
3139
3140xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003141 name = "operator_run",
3142 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003143 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003144 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003145 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3146 "//conditions:default": [],
3147 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003148 gcc_copts = xnnpack_gcc_std_copts(),
3149 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003151 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 "@FP16",
3153 "@FXdiv",
3154 "@clog",
3155 "@pthreadpool",
3156 ],
3157)
3158
Chao Mei6ddfc602020-05-13 22:29:36 -07003159xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003160 name = "operator_run_test_mode",
3161 srcs = ["src/operator-run.c"],
3162 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3163 copts = LOGGING_COPTS + [
3164 "-UNDEBUG",
3165 "-DXNN_TEST_MODE=1",
3166 ] + select({
3167 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3168 "//conditions:default": [],
3169 }),
3170 gcc_copts = xnnpack_gcc_std_copts(),
3171 msvc_copts = xnnpack_msvc_std_copts(),
3172 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003173 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003174 "@FP16",
3175 "@FXdiv",
3176 "@clog",
3177 "@pthreadpool",
3178 ],
3179)
3180
3181xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003182 name = "memory_planner",
3183 srcs = ["src/memory-planner.c"],
3184 hdrs = INTERNAL_HDRS,
3185 defines = select({
3186 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3187 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3188 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3189 }),
3190 gcc_copts = xnnpack_gcc_std_copts(),
3191 msvc_copts = xnnpack_msvc_std_copts(),
3192 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003193 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003194 "@pthreadpool",
3195 ],
3196)
3197
Marat Dukhan33fcf782020-05-24 14:27:15 -07003198xnnpack_cc_library(
3199 name = "memory_planner_test_mode",
3200 srcs = ["src/memory-planner.c"],
3201 hdrs = INTERNAL_HDRS,
3202 copts = [
3203 "-UNDEBUG",
3204 "-DXNN_TEST_MODE=1",
3205 ],
3206 defines = select({
3207 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3208 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3209 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3210 }),
3211 gcc_copts = xnnpack_gcc_std_copts(),
3212 msvc_copts = xnnpack_msvc_std_copts(),
3213 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003214 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003215 "@pthreadpool",
3216 ],
3217)
3218
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219cc_library(
3220 name = "enable_assembly",
3221 defines = select({
3222 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3223 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003224 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225 }),
3226)
3227
Marat Dukhan9de90e02020-06-18 16:04:12 -07003228cc_library(
3229 name = "enable_sparse",
3230 defines = select({
3231 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3232 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003233 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003234 }),
3235)
3236
Marat Dukhancf056b22019-10-07 10:26:29 -07003237xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238 name = "operators",
3239 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003240 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003241 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003242 ],
3243 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003244 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 "-Isrc",
3246 "-Iinclude",
3247 ] + select({
3248 ":debug_build": [],
3249 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003250 }) + select({
3251 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3252 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003254 gcc_copts = xnnpack_gcc_std_copts(),
3255 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003258 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003259 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 "@FP16",
3261 "@FXdiv",
3262 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003264 ],
3265)
3266
Marat Dukhan10a38082020-04-17 03:58:35 -07003267xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003268 name = "operators_test_mode",
3269 srcs = OPERATOR_SRCS + [
3270 "src/memory.c",
3271 "src/operator-delete.c",
3272 ],
3273 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3274 copts = LOGGING_COPTS + [
3275 "-Isrc",
3276 "-Iinclude",
3277 "-UNDEBUG",
3278 "-DXNN_TEST_MODE=1",
3279 ] + select({
3280 ":debug_build": [],
3281 "//conditions:default": xnnpack_min_size_copts(),
3282 }) + select({
3283 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3284 "//conditions:default": [],
3285 }),
3286 gcc_copts = xnnpack_gcc_std_copts(),
3287 msvc_copts = xnnpack_msvc_std_copts(),
3288 deps = [
3289 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003290 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003291 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003292 "@FP16",
3293 "@FXdiv",
3294 "@clog",
3295 "@pthreadpool",
3296 ],
3297)
3298
3299xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003300 name = "XNNPACK",
3301 srcs = [
3302 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003303 "src/runtime.c",
3304 "src/subgraph.c",
3305 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003306 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003307 hdrs = ["include/xnnpack.h"],
3308 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003309 "-Isrc",
3310 "-Iinclude",
3311 ] + select({
3312 ":debug_build": [],
3313 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003314 }) + select({
3315 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3316 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003317 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003318 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003319 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003320 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003321 visibility = xnnpack_visibility(),
3322 deps = [
3323 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003324 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003325 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003326 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003327 ":operator_run",
3328 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003329 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003330 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003331 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003332 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003333 ] + select({
3334 ":emscripten": [],
3335 "//conditions:default": ["@cpuinfo"],
3336 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337)
3338
Marat Dukhan10a38082020-04-17 03:58:35 -07003339xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003340 name = "XNNPACK_test_mode",
3341 srcs = [
3342 "src/init.c",
3343 "src/runtime.c",
3344 "src/subgraph.c",
3345 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003346 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003347 hdrs = ["include/xnnpack.h"],
3348 copts = LOGGING_COPTS + [
3349 "-Isrc",
3350 "-Iinclude",
3351 "-UNDEBUG",
3352 "-DXNN_TEST_MODE=1",
3353 ] + select({
3354 ":debug_build": [],
3355 "//conditions:default": xnnpack_min_size_copts(),
3356 }) + select({
3357 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3358 "//conditions:default": [],
3359 }),
3360 gcc_copts = xnnpack_gcc_std_copts(),
3361 includes = ["include"],
3362 msvc_copts = xnnpack_msvc_std_copts(),
3363 visibility = xnnpack_visibility(),
3364 deps = [
3365 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003366 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003367 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003368 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003369 ":operator_run_test_mode",
3370 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003371 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003372 "@clog",
3373 "@FP16",
3374 "@pthreadpool",
3375 ] + select({
3376 ":emscripten": [],
3377 "//conditions:default": ["@cpuinfo"],
3378 }),
3379)
3380
3381xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003382 name = "xnnpack_f32",
3383 srcs = [
3384 "src/init.c",
3385 "src/runtime.c",
3386 "src/subgraph.c",
3387 "src/tensor.c",
3388 ] + SUBGRAPH_SRCS,
3389 hdrs = ["include/xnnpack.h"],
3390 copts = LOGGING_COPTS + [
3391 "-Isrc",
3392 "-Iinclude",
3393 ] + select({
3394 ":debug_build": [],
3395 "//conditions:default": xnnpack_min_size_copts(),
3396 }) + select({
3397 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3398 "//conditions:default": [],
3399 }),
3400 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003401 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003402 "XNN_NO_U8_OPERATORS",
3403 "XNN_NO_X8_OPERATORS",
3404 "XNN_NO_F16_OPERATORS",
3405 "XNN_NO_X16_OPERATORS",
3406 ],
3407 gcc_copts = xnnpack_gcc_std_copts(),
3408 includes = ["include"],
3409 msvc_copts = xnnpack_msvc_std_copts(),
3410 visibility = xnnpack_visibility(),
3411 deps = [
3412 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003413 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003414 ":logging_utils",
3415 ":memory_planner",
3416 ":operator_run",
3417 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003418 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003419 "@clog",
3420 "@FP16",
3421 "@pthreadpool",
3422 ] + select({
3423 ":emscripten": [],
3424 "//conditions:default": ["@cpuinfo"],
3425 }),
3426)
3427
3428xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003429 name = "xnnpack_operators_nhwc_f32",
3430 srcs = [
3431 "src/init.c",
3432 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003433 hdrs = ["include/xnnpack.h"],
3434 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003435 "-Isrc",
3436 "-Iinclude",
3437 ] + select({
3438 ":debug_build": [],
3439 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003440 }) + select({
3441 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3442 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003443 }),
3444 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003445 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003446 "XNN_NO_U8_OPERATORS",
3447 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003448 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003449 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003450 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003451 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003452 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453 visibility = xnnpack_visibility(),
3454 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003455 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003456 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003457 ":operator_run",
3458 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003459 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003460 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003462 ] + select({
3463 ":emscripten": [],
3464 "//conditions:default": ["@cpuinfo"],
3465 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466)
3467
Marat Dukhancf056b22019-10-07 10:26:29 -07003468xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469 name = "bench_utils",
3470 srcs = ["bench/utils.cc"],
3471 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003472 deps = [
3473 "@com_google_benchmark//:benchmark",
3474 "@cpuinfo",
3475 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476)
3477
Frank Barchard7e955972019-10-11 10:34:25 -07003478######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479
3480xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003481 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003482 srcs = [
3483 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003484 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485 "src/xnnpack/AlignedAllocator.h",
3486 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003487 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003488 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489)
3490
3491xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003492 name = "f16_igemm_bench",
3493 srcs = [
3494 "bench/f16-igemm.cc",
3495 "bench/conv.h",
3496 "bench/google/conv.h",
3497 "src/xnnpack/AlignedAllocator.h",
3498 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003499 deps = MICROKERNEL_BENCHMARK_DEPS + [
3500 ":indirection",
3501 ":packing",
3502 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003503)
3504
3505xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506 name = "f16_gemm_bench",
3507 srcs = [
3508 "bench/f16-gemm.cc",
3509 "bench/gemm.h",
3510 "src/xnnpack/AlignedAllocator.h",
3511 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003512 deps = MICROKERNEL_BENCHMARK_DEPS + [
3513 ":packing",
3514 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515)
3516
3517xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003518 name = "f16_spmm_bench",
3519 srcs = [
3520 "bench/f16-spmm.cc",
3521 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003522 "src/xnnpack/AlignedAllocator.h",
3523 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003524 deps = MICROKERNEL_BENCHMARK_DEPS,
3525)
3526
3527xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528 name = "f32_igemm_bench",
3529 srcs = [
3530 "bench/f32-igemm.cc",
3531 "bench/conv.h",
3532 "src/xnnpack/AlignedAllocator.h",
3533 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003534 deps = MICROKERNEL_BENCHMARK_DEPS + [
3535 ":indirection",
3536 ":packing",
3537 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538)
3539
3540xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003541 name = "f16_relu_bench",
3542 srcs = [
3543 "bench/f16-relu.cc",
3544 "src/xnnpack/AlignedAllocator.h",
3545 ] + MICROKERNEL_BENCHMARK_HDRS,
3546 deps = MICROKERNEL_BENCHMARK_DEPS,
3547)
3548
3549xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003550 name = "f32_conv_hwc_bench",
3551 srcs = [
3552 "bench/f32-conv-hwc.cc",
3553 "bench/dconv.h",
3554 "src/xnnpack/AlignedAllocator.h",
3555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003556 deps = MICROKERNEL_BENCHMARK_DEPS + [
3557 ":packing",
3558 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559)
3560
3561xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003562 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003563 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003564 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003565 "bench/dconv.h",
3566 "src/xnnpack/AlignedAllocator.h",
3567 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003568 deps = MICROKERNEL_BENCHMARK_DEPS + [
3569 ":packing",
3570 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003571)
3572
3573xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003574 name = "f16_dwconv_bench",
3575 srcs = [
3576 "bench/f16-dwconv.cc",
3577 "bench/dwconv.h",
3578 "bench/google/dwconv.h",
3579 "src/xnnpack/AlignedAllocator.h",
3580 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003581 deps = MICROKERNEL_BENCHMARK_DEPS + [
3582 ":indirection",
3583 ":packing",
3584 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003585)
3586
3587xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 name = "f32_dwconv_bench",
3589 srcs = [
3590 "bench/f32-dwconv.cc",
3591 "bench/dwconv.h",
3592 "src/xnnpack/AlignedAllocator.h",
3593 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003594 deps = MICROKERNEL_BENCHMARK_DEPS + [
3595 ":indirection",
3596 ":packing",
3597 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598)
3599
3600xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003601 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003603 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604 "bench/dwconv.h",
3605 "src/xnnpack/AlignedAllocator.h",
3606 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003607 deps = MICROKERNEL_BENCHMARK_DEPS + [
3608 ":indirection",
3609 ":packing",
3610 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611)
3612
3613xnnpack_benchmark(
3614 name = "f32_gemm_bench",
3615 srcs = [
3616 "bench/f32-gemm.cc",
3617 "bench/gemm.h",
3618 "src/xnnpack/AlignedAllocator.h",
3619 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003620 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003621 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003622)
3623
3624xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003625 name = "f32_hswish_bench",
3626 srcs = [
3627 "bench/f32-hswish.cc",
3628 "src/xnnpack/AlignedAllocator.h",
3629 ] + MICROKERNEL_BENCHMARK_HDRS,
3630 deps = MICROKERNEL_BENCHMARK_DEPS,
3631)
3632
3633xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003634 name = "f32_raddexpminusmax_bench",
3635 srcs = [
3636 "bench/f32-raddexpminusmax.cc",
3637 "src/xnnpack/AlignedAllocator.h",
3638 ] + MICROKERNEL_BENCHMARK_HDRS,
3639 deps = MICROKERNEL_BENCHMARK_DEPS,
3640)
3641
3642xnnpack_benchmark(
3643 name = "f32_raddextexp_bench",
3644 srcs = [
3645 "bench/f32-raddextexp.cc",
3646 "src/xnnpack/AlignedAllocator.h",
3647 ] + MICROKERNEL_BENCHMARK_HDRS,
3648 deps = MICROKERNEL_BENCHMARK_DEPS,
3649)
3650
3651xnnpack_benchmark(
3652 name = "f32_raddstoreexpminusmax_bench",
3653 srcs = [
3654 "bench/f32-raddstoreexpminusmax.cc",
3655 "src/xnnpack/AlignedAllocator.h",
3656 ] + MICROKERNEL_BENCHMARK_HDRS,
3657 deps = MICROKERNEL_BENCHMARK_DEPS,
3658)
3659
3660xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003661 name = "f32_relu_bench",
3662 srcs = [
3663 "bench/f32-relu.cc",
3664 "src/xnnpack/AlignedAllocator.h",
3665 ] + MICROKERNEL_BENCHMARK_HDRS,
3666 deps = MICROKERNEL_BENCHMARK_DEPS,
3667)
3668
3669xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670 name = "f32_rmax_bench",
3671 srcs = [
3672 "bench/f32-rmax.cc",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + MICROKERNEL_BENCHMARK_HDRS,
3675 deps = MICROKERNEL_BENCHMARK_DEPS,
3676)
3677
3678xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003679 name = "f32_sigmoid_bench",
3680 srcs = [
3681 "bench/f32-sigmoid.cc",
3682 "src/xnnpack/AlignedAllocator.h",
3683 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003684 deps = MICROKERNEL_BENCHMARK_DEPS,
3685)
3686
3687xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003688 name = "f32_spmm_bench",
3689 srcs = [
3690 "bench/f32-spmm.cc",
3691 "bench/gemm.h",
3692 "src/xnnpack/AlignedAllocator.h",
3693 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003694 deps = MICROKERNEL_BENCHMARK_DEPS,
3695)
3696
3697xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003698 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003699 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003700 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003701 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003702 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003703 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003704)
3705
3706xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003707 name = "f32_vscaleexpminusmax_bench",
3708 srcs = [
3709 "bench/f32-vscaleexpminusmax.cc",
3710 "src/xnnpack/AlignedAllocator.h",
3711 ] + MICROKERNEL_BENCHMARK_HDRS,
3712 deps = MICROKERNEL_BENCHMARK_DEPS,
3713)
3714
3715xnnpack_benchmark(
3716 name = "f32_vscaleextexp_bench",
3717 srcs = [
3718 "bench/f32-vscaleextexp.cc",
3719 "src/xnnpack/AlignedAllocator.h",
3720 ] + MICROKERNEL_BENCHMARK_HDRS,
3721 deps = MICROKERNEL_BENCHMARK_DEPS,
3722)
3723
3724xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003725 name = "f32_vsqrt_bench",
3726 srcs = [
3727 "bench/f32-vsqrt.cc",
3728 "src/xnnpack/AlignedAllocator.h",
3729 ] + MICROKERNEL_BENCHMARK_HDRS,
3730 deps = MICROKERNEL_BENCHMARK_DEPS,
3731)
3732
3733xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003734 name = "f32_im2col_gemm_bench",
3735 srcs = [
3736 "bench/f32-im2col-gemm.cc",
3737 "bench/conv.h",
3738 "src/xnnpack/AlignedAllocator.h",
3739 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003740 deps = MICROKERNEL_BENCHMARK_DEPS + [
3741 ":im2col",
3742 ":packing",
3743 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003744)
3745
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003746xnnpack_benchmark(
3747 name = "requantization_bench",
3748 srcs = [
3749 "bench/requantization.cc",
3750 "src/xnnpack/requantization-stubs.h",
3751 "src/xnnpack/AlignedAllocator.h",
3752 ] + MICROKERNEL_BENCHMARK_HDRS,
3753 deps = MICROKERNEL_BENCHMARK_DEPS,
3754)
3755
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003756xnnpack_benchmark(
3757 name = "rounding_bench",
3758 srcs = [
3759 "bench/rounding.cc",
3760 "src/xnnpack/math-stubs.h",
3761 "src/xnnpack/AlignedAllocator.h",
3762 ] + MICROKERNEL_BENCHMARK_HDRS,
3763 deps = MICROKERNEL_BENCHMARK_DEPS,
3764)
3765
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766########################### Benchmarks for operators ###########################
3767
3768xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 name = "average_pooling_bench",
3770 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003771 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003772 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003773 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774)
3775
3776xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003777 name = "bankers_rounding_bench",
3778 srcs = ["bench/bankers-rounding.cc"],
3779 copts = xnnpack_optional_tflite_copts(),
3780 tags = ["nowin32"],
3781 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3782)
3783
3784xnnpack_benchmark(
3785 name = "ceiling_bench",
3786 srcs = ["bench/ceiling.cc"],
3787 copts = xnnpack_optional_tflite_copts(),
3788 tags = ["nowin32"],
3789 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3790)
3791
3792xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793 name = "channel_shuffle_bench",
3794 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003795 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003796)
3797
3798xnnpack_benchmark(
3799 name = "convolution_bench",
3800 srcs = ["bench/convolution.cc"],
3801 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003802 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003803 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804)
3805
3806xnnpack_benchmark(
3807 name = "deconvolution_bench",
3808 srcs = ["bench/deconvolution.cc"],
3809 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003810 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003811 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003812)
3813
3814xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003815 name = "floor_bench",
3816 srcs = ["bench/floor.cc"],
3817 copts = xnnpack_optional_tflite_copts(),
3818 tags = ["nowin32"],
3819 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3820)
3821
3822xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003823 name = "global_average_pooling_bench",
3824 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003825 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003826)
3827
3828xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003829 name = "hardswish_bench",
3830 srcs = ["bench/hardswish.cc"],
3831 copts = xnnpack_optional_tflite_copts(),
3832 tags = ["nowin32"],
3833 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3834)
3835
3836xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003837 name = "max_pooling_bench",
3838 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003839 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003840)
3841
3842xnnpack_benchmark(
3843 name = "sigmoid_bench",
3844 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003845 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003846 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003847 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003848)
3849
3850xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003851 name = "prelu_bench",
3852 srcs = ["bench/prelu.cc"],
3853 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003854 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003855 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003856)
3857
3858xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003859 name = "softmax_bench",
3860 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003861 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003862 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003863 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003864)
3865
Marat Dukhan87727142020-06-24 15:24:10 -07003866xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003867 name = "square_root_bench",
3868 srcs = ["bench/square-root.cc"],
3869 copts = xnnpack_optional_tflite_copts(),
3870 tags = ["nowin32"],
3871 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3872)
3873
3874xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003875 name = "truncation_bench",
3876 srcs = ["bench/truncation.cc"],
3877 deps = OPERATOR_BENCHMARK_DEPS,
3878)
3879
Marat Dukhanc068bb62019-10-04 13:24:39 -07003880############################# End-to-end benchmarks ############################
3881
3882cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003883 name = "fp32_mobilenet_v1",
3884 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003885 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003886 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003887 linkstatic = True,
3888 deps = [
3889 ":XNNPACK",
3890 "@pthreadpool",
3891 ],
3892)
3893
3894cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003895 name = "fp16_mobilenet_v1",
3896 srcs = ["models/fp16-mobilenet-v1.cc"],
3897 hdrs = ["models/models.h"],
3898 copts = xnnpack_std_cxxopts(),
3899 linkstatic = True,
3900 deps = [
3901 ":XNNPACK",
3902 "@FP16",
3903 "@pthreadpool",
3904 ],
3905)
3906
3907cc_library(
3908 name = "fp32_mobilenet_v2",
3909 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003910 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003911 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003912 linkstatic = True,
3913 deps = [
3914 ":XNNPACK",
3915 "@pthreadpool",
3916 ],
3917)
3918
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003919cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003920 name = "fp16_mobilenet_v2",
3921 srcs = ["models/fp16-mobilenet-v2.cc"],
3922 hdrs = ["models/models.h"],
3923 copts = xnnpack_std_cxxopts(),
3924 linkstatic = True,
3925 deps = [
3926 ":XNNPACK",
3927 "@FP16",
3928 "@pthreadpool",
3929 ],
3930)
3931
3932cc_library(
3933 name = "fp32_mobilenet_v3_large",
3934 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003935 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003936 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003937 linkstatic = True,
3938 deps = [
3939 ":XNNPACK",
3940 "@pthreadpool",
3941 ],
3942)
3943
3944cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003945 name = "fp32_mobilenet_v3_small",
3946 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003947 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003948 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003949 linkstatic = True,
3950 deps = [
3951 ":XNNPACK",
3952 "@pthreadpool",
3953 ],
3954)
3955
Marat Dukhanc068bb62019-10-04 13:24:39 -07003956xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003957 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003958 srcs = [
3959 "bench/f32-dwconv-e2e.cc",
3960 "bench/end2end.h",
3961 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003962 deps = MICROKERNEL_BENCHMARK_DEPS + [
3963 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003964 ":fp32_mobilenet_v1",
3965 ":fp32_mobilenet_v2",
3966 ":fp32_mobilenet_v3_large",
3967 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003968 ],
3969)
3970
3971xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003972 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003973 srcs = [
3974 "bench/f32-gemm-e2e.cc",
3975 "bench/end2end.h",
3976 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003977 deps = MICROKERNEL_BENCHMARK_DEPS + [
3978 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003979 ":fp32_mobilenet_v1",
3980 ":fp32_mobilenet_v2",
3981 ":fp32_mobilenet_v3_large",
3982 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003983 ],
3984)
3985
3986xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003987 name = "end2end_bench",
3988 srcs = ["bench/end2end.cc"],
3989 deps = [
3990 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003991 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003992 ":fp16_mobilenet_v1",
3993 ":fp16_mobilenet_v2",
3994 ":fp32_mobilenet_v1",
3995 ":fp32_mobilenet_v2",
3996 ":fp32_mobilenet_v3_large",
3997 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003998 "@pthreadpool",
3999 ],
4000)
4001
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004002#################### Accuracy evaluation for math functions ####################
4003
4004xnnpack_benchmark(
4005 name = "f32_exp_eval",
4006 srcs = [
4007 "eval/f32-exp.cc",
4008 "src/xnnpack/AlignedAllocator.h",
4009 ] + ACCURACY_EVAL_HDRS,
4010 deps = ACCURACY_EVAL_DEPS,
4011)
4012
Marat Dukhan515c9772019-10-17 18:07:57 -07004013xnnpack_benchmark(
4014 name = "f32_expminus_eval",
4015 srcs = [
4016 "eval/f32-expminus.cc",
4017 "src/xnnpack/AlignedAllocator.h",
4018 ] + ACCURACY_EVAL_HDRS,
4019 deps = ACCURACY_EVAL_DEPS,
4020)
4021
Marat Dukhan98ba4412019-10-23 02:14:28 -07004022xnnpack_benchmark(
4023 name = "f32_extexp_eval",
4024 srcs = [
4025 "eval/f32-extexp.cc",
4026 "src/xnnpack/AlignedAllocator.h",
4027 ] + ACCURACY_EVAL_HDRS,
4028 deps = ACCURACY_EVAL_DEPS,
4029)
4030
Marat Dukhan8853b822020-05-07 12:19:01 -07004031xnnpack_unit_test(
4032 name = "f32_roundne_eval",
4033 srcs = [
4034 "eval/f32-roundne.cc",
4035 "src/xnnpack/AlignedAllocator.h",
4036 "src/xnnpack/math-stubs.h",
4037 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004038 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004039 deps = MICROKERNEL_TEST_DEPS,
4040)
4041
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004042xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004043 name = "f32_roundd_eval",
4044 srcs = [
4045 "eval/f32-roundd.cc",
4046 "src/xnnpack/AlignedAllocator.h",
4047 "src/xnnpack/math-stubs.h",
4048 ] + MICROKERNEL_TEST_HDRS,
4049 automatic = False,
4050 deps = MICROKERNEL_TEST_DEPS,
4051)
4052
4053xnnpack_unit_test(
4054 name = "f32_roundu_eval",
4055 srcs = [
4056 "eval/f32-roundu.cc",
4057 "src/xnnpack/AlignedAllocator.h",
4058 "src/xnnpack/math-stubs.h",
4059 ] + MICROKERNEL_TEST_HDRS,
4060 automatic = False,
4061 deps = MICROKERNEL_TEST_DEPS,
4062)
4063
4064xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004065 name = "f32_roundz_eval",
4066 srcs = [
4067 "eval/f32-roundz.cc",
4068 "src/xnnpack/AlignedAllocator.h",
4069 "src/xnnpack/math-stubs.h",
4070 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004071 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004072 deps = MICROKERNEL_TEST_DEPS,
4073)
4074
Marat Dukhan346a9e52019-11-15 09:06:30 -08004075xnnpack_benchmark(
4076 name = "f32_sigmoid_eval",
4077 srcs = [
4078 "eval/f32-sigmoid.cc",
4079 "src/xnnpack/AlignedAllocator.h",
4080 ] + ACCURACY_EVAL_HDRS,
4081 deps = ACCURACY_EVAL_DEPS,
4082)
4083
Marat Dukhan84000762020-06-29 18:38:43 -07004084xnnpack_benchmark(
4085 name = "f32_sqrt_eval",
4086 srcs = [
4087 "eval/f32-sqrt.cc",
4088 "src/xnnpack/AlignedAllocator.h",
4089 ] + ACCURACY_EVAL_HDRS,
4090 deps = ACCURACY_EVAL_DEPS + [
4091 ":bench_utils",
4092 ],
4093)
4094
Marat Dukhan08c4a432019-10-03 09:29:21 -07004095######################### Unit tests for micro-kernels #########################
4096
4097xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004098 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004099 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004100 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004101 "test/gemm-microkernel-tester.h",
4102 "src/xnnpack/AlignedAllocator.h",
4103 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004104 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004105)
4106
4107xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004108 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004109 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004110 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004111 "test/spmm-microkernel-tester.h",
4112 "src/xnnpack/AlignedAllocator.h",
4113 ] + MICROKERNEL_TEST_HDRS,
4114 deps = MICROKERNEL_TEST_DEPS,
4115)
4116
4117xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004118 name = "f32_argmaxpool_test",
4119 srcs = [
4120 "test/f32-argmaxpool.cc",
4121 "test/argmaxpool-microkernel-tester.h",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + MICROKERNEL_TEST_HDRS,
4124 deps = MICROKERNEL_TEST_DEPS,
4125)
4126
4127xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004128 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004130 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 "test/avgpool-microkernel-tester.h",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + MICROKERNEL_TEST_HDRS,
4134 deps = MICROKERNEL_TEST_DEPS,
4135)
4136
4137xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004138 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004139 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004140 "test/f32-ibilinear.cc",
4141 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004142 "src/xnnpack/AlignedAllocator.h",
4143 ] + MICROKERNEL_TEST_HDRS,
4144 deps = MICROKERNEL_TEST_DEPS,
4145)
4146
4147xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004148 name = "f16_clamp_test",
4149 srcs = [
4150 "test/f16-clamp.cc",
4151 "test/clamp-microkernel-tester.h",
4152 ] + MICROKERNEL_TEST_HDRS,
4153 deps = MICROKERNEL_TEST_DEPS,
4154)
4155
4156xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 name = "f32_clamp_test",
4158 srcs = [
4159 "test/f32-clamp.cc",
4160 "test/clamp-microkernel-tester.h",
4161 ] + MICROKERNEL_TEST_HDRS,
4162 deps = MICROKERNEL_TEST_DEPS,
4163)
4164
4165xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004166 name = "f32_igemm_test",
4167 srcs = [
4168 "test/f32-igemm.cc",
4169 "test/gemm-microkernel-tester.h",
4170 "src/xnnpack/AlignedAllocator.h",
4171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004172 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004173)
4174
4175xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004176 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004178 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004179 "test/gemm-microkernel-tester.h",
4180 "src/xnnpack/AlignedAllocator.h",
4181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004182 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004183)
4184
4185xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004186 name = "f32_igemm_minmax_test",
4187 srcs = [
4188 "test/f32-igemm-minmax.cc",
4189 "test/gemm-microkernel-tester.h",
4190 "src/xnnpack/AlignedAllocator.h",
4191 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004192 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004193)
4194
4195xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004196 name = "f16_igemm_minmax_test",
4197 srcs = [
4198 "test/f16-igemm-minmax.cc",
4199 "test/gemm-microkernel-tester.h",
4200 "src/xnnpack/AlignedAllocator.h",
4201 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004202 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004203)
4204
4205xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004206 name = "f32_conv_hwc_test",
4207 srcs = [
4208 "test/f32-conv-hwc.cc",
4209 "test/conv-hwc-microkernel-tester.h",
4210 "src/xnnpack/AlignedAllocator.h",
4211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004212 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213)
4214
4215xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004216 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004218 "test/f32-conv-hwc2chw.cc",
4219 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004220 "src/xnnpack/AlignedAllocator.h",
4221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004222 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004223)
4224
4225xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004226 name = "f32_dwconv_test",
4227 srcs = [
4228 "test/f32-dwconv.cc",
4229 "test/dwconv-microkernel-tester.h",
4230 "src/xnnpack/AlignedAllocator.h",
4231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004232 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004233)
4234
4235xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004236 name = "f16_dwconv_minmax_test",
4237 srcs = [
4238 "test/f16-dwconv-minmax.cc",
4239 "test/dwconv-microkernel-tester.h",
4240 "src/xnnpack/AlignedAllocator.h",
4241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004242 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004243)
4244
4245xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004246 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004247 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004248 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004249 "test/dwconv-microkernel-tester.h",
4250 "src/xnnpack/AlignedAllocator.h",
4251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004252 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253)
4254
4255xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004256 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004257 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004258 "test/f32-dwconv-chw.cc",
4259 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 "src/xnnpack/AlignedAllocator.h",
4261 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004262 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263)
4264
4265xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004266 name = "f16_gavgpool_minmax_test",
4267 srcs = [
4268 "test/f16-gavgpool-minmax.cc",
4269 "test/gavgpool-microkernel-tester.h",
4270 "src/xnnpack/AlignedAllocator.h",
4271 ] + MICROKERNEL_TEST_HDRS,
4272 deps = MICROKERNEL_TEST_DEPS,
4273)
4274
4275xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004276 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004278 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279 "test/gavgpool-microkernel-tester.h",
4280 "src/xnnpack/AlignedAllocator.h",
4281 ] + MICROKERNEL_TEST_HDRS,
4282 deps = MICROKERNEL_TEST_DEPS,
4283)
4284
4285xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004286 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004288 "test/f32-gavgpool-cw.cc",
4289 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004290 "src/xnnpack/AlignedAllocator.h",
4291 ] + MICROKERNEL_TEST_HDRS,
4292 deps = MICROKERNEL_TEST_DEPS,
4293)
4294
4295xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004296 name = "f32_gemm_test",
4297 srcs = [
4298 "test/f32-gemm.cc",
4299 "test/gemm-microkernel-tester.h",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004302 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004303)
4304
4305xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004306 name = "f32_gemm_relu_test",
4307 srcs = [
4308 "test/f32-gemm-relu.cc",
4309 "test/gemm-microkernel-tester.h",
4310 "src/xnnpack/AlignedAllocator.h",
4311 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004312 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004313)
4314
4315xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004316 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004318 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 "test/gemm-microkernel-tester.h",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004322 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323)
4324
4325xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004326 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004328 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 "test/gemm-microkernel-tester.h",
4330 "src/xnnpack/AlignedAllocator.h",
4331 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004332 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333)
4334
4335xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004336 name = "f16_hswish_test",
4337 srcs = [
4338 "test/f16-hswish.cc",
4339 "test/hswish-microkernel-tester.h",
4340 ] + MICROKERNEL_TEST_HDRS,
4341 deps = MICROKERNEL_TEST_DEPS,
4342)
4343
4344xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 name = "f32_hswish_test",
4346 srcs = [
4347 "test/f32-hswish.cc",
4348 "test/hswish-microkernel-tester.h",
4349 ] + MICROKERNEL_TEST_HDRS,
4350 deps = MICROKERNEL_TEST_DEPS,
4351)
4352
4353xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004354 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004356 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 "test/maxpool-microkernel-tester.h",
4358 ] + MICROKERNEL_TEST_HDRS,
4359 deps = MICROKERNEL_TEST_DEPS,
4360)
4361
4362xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004363 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004365 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 "test/avgpool-microkernel-tester.h",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + MICROKERNEL_TEST_HDRS,
4369 deps = MICROKERNEL_TEST_DEPS,
4370)
4371
4372xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004373 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004374 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004375 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 "test/gemm-microkernel-tester.h",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004379 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380)
4381
4382xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004383 name = "f16_prelu_test",
4384 srcs = [
4385 "test/f16-prelu.cc",
4386 "test/prelu-microkernel-tester.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + MICROKERNEL_TEST_HDRS,
4389 deps = MICROKERNEL_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393 name = "f32_prelu_test",
4394 srcs = [
4395 "test/f32-prelu.cc",
4396 "test/prelu-microkernel-tester.h",
4397 "src/xnnpack/AlignedAllocator.h",
4398 ] + MICROKERNEL_TEST_HDRS,
4399 deps = MICROKERNEL_TEST_DEPS,
4400)
4401
4402xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004403 name = "f32_raddexpminusmax_test",
4404 srcs = [
4405 "test/f32-raddexpminusmax.cc",
4406 "test/raddexpminusmax-microkernel-tester.h",
4407 ] + MICROKERNEL_TEST_HDRS,
4408 deps = MICROKERNEL_TEST_DEPS,
4409)
4410
4411xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004412 name = "f32_raddextexp_test",
4413 srcs = [
4414 "test/f32-raddextexp.cc",
4415 "test/raddextexp-microkernel-tester.h",
4416 ] + MICROKERNEL_TEST_HDRS,
4417 deps = MICROKERNEL_TEST_DEPS,
4418)
4419
4420xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004421 name = "f32_raddstoreexpminusmax_test",
4422 srcs = [
4423 "test/f32-raddstoreexpminusmax.cc",
4424 "test/raddstoreexpminusmax-microkernel-tester.h",
4425 ] + MICROKERNEL_TEST_HDRS,
4426 deps = MICROKERNEL_TEST_DEPS,
4427)
4428
4429xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004430 name = "f32_relu_test",
4431 srcs = [
4432 "test/f32-relu.cc",
4433 "test/vunary-microkernel-tester.h",
4434 ] + MICROKERNEL_TEST_HDRS,
4435 deps = MICROKERNEL_TEST_DEPS,
4436)
4437
4438xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439 name = "f32_rmax_test",
4440 srcs = [
4441 "test/f32-rmax.cc",
4442 "test/rmax-microkernel-tester.h",
4443 ] + MICROKERNEL_TEST_HDRS,
4444 deps = MICROKERNEL_TEST_DEPS,
4445)
4446
4447xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004448 name = "f32_sigmoid_test",
4449 srcs = [
4450 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004451 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004452 ] + MICROKERNEL_TEST_HDRS,
4453 deps = MICROKERNEL_TEST_DEPS,
4454)
4455
4456xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004457 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004459 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 "test/spmm-microkernel-tester.h",
4461 "src/xnnpack/AlignedAllocator.h",
4462 ] + MICROKERNEL_TEST_HDRS,
4463 deps = MICROKERNEL_TEST_DEPS,
4464)
4465
4466xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004467 name = "f16_vadd_minmax_test",
4468 srcs = [
4469 "test/f16-vadd-minmax.cc",
4470 "test/vbinary-microkernel-tester.h",
4471 ] + MICROKERNEL_TEST_HDRS,
4472 deps = MICROKERNEL_TEST_DEPS,
4473)
4474
4475xnnpack_unit_test(
4476 name = "f16_vaddc_minmax_test",
4477 srcs = [
4478 "test/f16-vaddc-minmax.cc",
4479 "test/vbinaryc-microkernel-tester.h",
4480 ] + MICROKERNEL_TEST_HDRS,
4481 deps = MICROKERNEL_TEST_DEPS,
4482)
4483
4484xnnpack_unit_test(
4485 name = "f16_vdiv_minmax_test",
4486 srcs = [
4487 "test/f16-vdiv-minmax.cc",
4488 "test/vbinary-microkernel-tester.h",
4489 ] + MICROKERNEL_TEST_HDRS,
4490 deps = MICROKERNEL_TEST_DEPS,
4491)
4492
4493xnnpack_unit_test(
4494 name = "f16_vdivc_minmax_test",
4495 srcs = [
4496 "test/f16-vdivc-minmax.cc",
4497 "test/vbinaryc-microkernel-tester.h",
4498 ] + MICROKERNEL_TEST_HDRS,
4499 deps = MICROKERNEL_TEST_DEPS,
4500)
4501
4502xnnpack_unit_test(
4503 name = "f16_vrdivc_minmax_test",
4504 srcs = [
4505 "test/f16-vrdivc-minmax.cc",
4506 "test/vbinaryc-microkernel-tester.h",
4507 ] + MICROKERNEL_TEST_HDRS,
4508 deps = MICROKERNEL_TEST_DEPS,
4509)
4510
4511xnnpack_unit_test(
4512 name = "f16_vmax_test",
4513 srcs = [
4514 "test/f16-vmax.cc",
4515 "test/vbinary-microkernel-tester.h",
4516 ] + MICROKERNEL_TEST_HDRS,
4517 deps = MICROKERNEL_TEST_DEPS,
4518)
4519
4520xnnpack_unit_test(
4521 name = "f16_vmaxc_test",
4522 srcs = [
4523 "test/f16-vmaxc.cc",
4524 "test/vbinaryc-microkernel-tester.h",
4525 ] + MICROKERNEL_TEST_HDRS,
4526 deps = MICROKERNEL_TEST_DEPS,
4527)
4528
4529xnnpack_unit_test(
4530 name = "f16_vmin_test",
4531 srcs = [
4532 "test/f16-vmin.cc",
4533 "test/vbinary-microkernel-tester.h",
4534 ] + MICROKERNEL_TEST_HDRS,
4535 deps = MICROKERNEL_TEST_DEPS,
4536)
4537
4538xnnpack_unit_test(
4539 name = "f16_vminc_test",
4540 srcs = [
4541 "test/f16-vminc.cc",
4542 "test/vbinaryc-microkernel-tester.h",
4543 ] + MICROKERNEL_TEST_HDRS,
4544 deps = MICROKERNEL_TEST_DEPS,
4545)
4546
4547xnnpack_unit_test(
4548 name = "f16_vmul_minmax_test",
4549 srcs = [
4550 "test/f16-vmul-minmax.cc",
4551 "test/vbinary-microkernel-tester.h",
4552 ] + MICROKERNEL_TEST_HDRS,
4553 deps = MICROKERNEL_TEST_DEPS,
4554)
4555
4556xnnpack_unit_test(
4557 name = "f16_vmulc_minmax_test",
4558 srcs = [
4559 "test/f16-vmulc-minmax.cc",
4560 "test/vbinaryc-microkernel-tester.h",
4561 ] + MICROKERNEL_TEST_HDRS,
4562 deps = MICROKERNEL_TEST_DEPS,
4563)
4564
4565xnnpack_unit_test(
4566 name = "f16_vsub_minmax_test",
4567 srcs = [
4568 "test/f16-vsub-minmax.cc",
4569 "test/vbinary-microkernel-tester.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
4575 name = "f16_vsubc_minmax_test",
4576 srcs = [
4577 "test/f16-vsubc-minmax.cc",
4578 "test/vbinaryc-microkernel-tester.h",
4579 ] + MICROKERNEL_TEST_HDRS,
4580 deps = MICROKERNEL_TEST_DEPS,
4581)
4582
4583xnnpack_unit_test(
4584 name = "f16_vrsubc_minmax_test",
4585 srcs = [
4586 "test/f16-vrsubc-minmax.cc",
4587 "test/vbinaryc-microkernel-tester.h",
4588 ] + MICROKERNEL_TEST_HDRS,
4589 deps = MICROKERNEL_TEST_DEPS,
4590)
4591
4592xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004593 name = "f32_vabs_test",
4594 srcs = [
4595 "test/f32-vabs.cc",
4596 "test/vunary-microkernel-tester.h",
4597 ] + MICROKERNEL_TEST_HDRS,
4598 deps = MICROKERNEL_TEST_DEPS,
4599)
4600
4601xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004602 name = "f32_vadd_test",
4603 srcs = [
4604 "test/f32-vadd.cc",
4605 "test/vbinary-microkernel-tester.h",
4606 ] + MICROKERNEL_TEST_HDRS,
4607 deps = MICROKERNEL_TEST_DEPS,
4608)
4609
4610xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004611 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004612 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004613 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004614 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004615 ] + MICROKERNEL_TEST_HDRS,
4616 deps = MICROKERNEL_TEST_DEPS,
4617)
4618
4619xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004620 name = "f32_vaddc_test",
4621 srcs = [
4622 "test/f32-vaddc.cc",
4623 "test/vbinaryc-microkernel-tester.h",
4624 ] + MICROKERNEL_TEST_HDRS,
4625 deps = MICROKERNEL_TEST_DEPS,
4626)
4627
4628xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004629 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004630 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004631 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004632 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004633 ] + MICROKERNEL_TEST_HDRS,
4634 deps = MICROKERNEL_TEST_DEPS,
4635)
4636
4637xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004638 name = "f32_vdiv_test",
4639 srcs = [
4640 "test/f32-vdiv.cc",
4641 "test/vbinary-microkernel-tester.h",
4642 ] + MICROKERNEL_TEST_HDRS,
4643 deps = MICROKERNEL_TEST_DEPS,
4644)
4645
4646xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004647 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004648 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004649 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004650 "test/vbinary-microkernel-tester.h",
4651 ] + MICROKERNEL_TEST_HDRS,
4652 deps = MICROKERNEL_TEST_DEPS,
4653)
4654
4655xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004656 name = "f32_vdivc_test",
4657 srcs = [
4658 "test/f32-vdivc.cc",
4659 "test/vbinaryc-microkernel-tester.h",
4660 ] + MICROKERNEL_TEST_HDRS,
4661 deps = MICROKERNEL_TEST_DEPS,
4662)
4663
4664xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004665 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004666 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004667 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004668 "test/vbinaryc-microkernel-tester.h",
4669 ] + MICROKERNEL_TEST_HDRS,
4670 deps = MICROKERNEL_TEST_DEPS,
4671)
4672
4673xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004674 name = "f32_vrdivc_test",
4675 srcs = [
4676 "test/f32-vrdivc.cc",
4677 "test/vbinaryc-microkernel-tester.h",
4678 ] + MICROKERNEL_TEST_HDRS,
4679 deps = MICROKERNEL_TEST_DEPS,
4680)
4681
4682xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004683 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004684 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004685 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004686 "test/vbinaryc-microkernel-tester.h",
4687 ] + MICROKERNEL_TEST_HDRS,
4688 deps = MICROKERNEL_TEST_DEPS,
4689)
4690
4691xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004692 name = "f32_vmax_test",
4693 srcs = [
4694 "test/f32-vmax.cc",
4695 "test/vbinary-microkernel-tester.h",
4696 ] + MICROKERNEL_TEST_HDRS,
4697 deps = MICROKERNEL_TEST_DEPS,
4698)
4699
4700xnnpack_unit_test(
4701 name = "f32_vmaxc_test",
4702 srcs = [
4703 "test/f32-vmaxc.cc",
4704 "test/vbinaryc-microkernel-tester.h",
4705 ] + MICROKERNEL_TEST_HDRS,
4706 deps = MICROKERNEL_TEST_DEPS,
4707)
4708
4709xnnpack_unit_test(
4710 name = "f32_vmin_test",
4711 srcs = [
4712 "test/f32-vmin.cc",
4713 "test/vbinary-microkernel-tester.h",
4714 ] + MICROKERNEL_TEST_HDRS,
4715 deps = MICROKERNEL_TEST_DEPS,
4716)
4717
4718xnnpack_unit_test(
4719 name = "f32_vminc_test",
4720 srcs = [
4721 "test/f32-vminc.cc",
4722 "test/vbinaryc-microkernel-tester.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004728 name = "f32_vmul_test",
4729 srcs = [
4730 "test/f32-vmul.cc",
4731 "test/vbinary-microkernel-tester.h",
4732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004737 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004739 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004740 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004741 ] + MICROKERNEL_TEST_HDRS,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004746 name = "f32_vmulc_test",
4747 srcs = [
4748 "test/f32-vmulc.cc",
4749 "test/vbinaryc-microkernel-tester.h",
4750 ] + MICROKERNEL_TEST_HDRS,
4751 deps = MICROKERNEL_TEST_DEPS,
4752)
4753
4754xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004755 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004756 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004757 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004758 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 ] + MICROKERNEL_TEST_HDRS,
4760 deps = MICROKERNEL_TEST_DEPS,
4761)
4762
4763xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004764 name = "f16_vmulcaddc_minmax_test",
4765 srcs = [
4766 "test/f16-vmulcaddc-minmax.cc",
4767 "test/vmulcaddc-microkernel-tester.h",
4768 "src/xnnpack/AlignedAllocator.h",
4769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004770 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004771)
4772
4773xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004774 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004775 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004776 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777 "test/vmulcaddc-microkernel-tester.h",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004780 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781)
4782
4783xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004784 name = "f32_vlrelu_test",
4785 srcs = [
4786 "test/f32-vlrelu.cc",
4787 "test/vunary-microkernel-tester.h",
4788 ] + MICROKERNEL_TEST_HDRS,
4789 deps = MICROKERNEL_TEST_DEPS,
4790)
4791
4792xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004793 name = "f32_vneg_test",
4794 srcs = [
4795 "test/f32-vneg.cc",
4796 "test/vunary-microkernel-tester.h",
4797 ] + MICROKERNEL_TEST_HDRS,
4798 deps = MICROKERNEL_TEST_DEPS,
4799)
4800
4801xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004802 name = "f32_vrndne_test",
4803 srcs = [
4804 "test/f32-vrndne.cc",
4805 "test/vunary-microkernel-tester.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
4811 name = "f32_vrndz_test",
4812 srcs = [
4813 "test/f32-vrndz.cc",
4814 "test/vunary-microkernel-tester.h",
4815 ] + MICROKERNEL_TEST_HDRS,
4816 deps = MICROKERNEL_TEST_DEPS,
4817)
4818
4819xnnpack_unit_test(
4820 name = "f32_vrndu_test",
4821 srcs = [
4822 "test/f32-vrndu.cc",
4823 "test/vunary-microkernel-tester.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
4829 name = "f32_vrndd_test",
4830 srcs = [
4831 "test/f32-vrndd.cc",
4832 "test/vunary-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004838 name = "f32_vscale_test",
4839 srcs = [
4840 "test/f32-vscale.cc",
4841 "test/vscale-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004847 name = "f32_vscaleexpminusmax_test",
4848 srcs = [
4849 "test/f32-vscaleexpminusmax.cc",
4850 "test/vscaleexpminusmax-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
4855xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004856 name = "f32_vscaleextexp_test",
4857 srcs = [
4858 "test/f32-vscaleextexp.cc",
4859 "test/vscaleextexp-microkernel-tester.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004865 name = "f32_vsqr_test",
4866 srcs = [
4867 "test/f32-vsqr.cc",
4868 "test/vunary-microkernel-tester.h",
4869 ] + MICROKERNEL_TEST_HDRS,
4870 deps = MICROKERNEL_TEST_DEPS,
4871)
4872
4873xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004874 name = "f32_vsqrdiff_test",
4875 srcs = [
4876 "test/f32-vsqrdiff.cc",
4877 "test/vbinary-microkernel-tester.h",
4878 ] + MICROKERNEL_TEST_HDRS,
4879 deps = MICROKERNEL_TEST_DEPS,
4880)
4881
4882xnnpack_unit_test(
4883 name = "f32_vsqrdiffc_test",
4884 srcs = [
4885 "test/f32-vsqrdiffc.cc",
4886 "test/vbinaryc-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004892 name = "f32_vsqrt_test",
4893 srcs = [
4894 "test/f32-vsqrt.cc",
4895 "test/vunary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004901 name = "f32_vsub_test",
4902 srcs = [
4903 "test/f32-vsub.cc",
4904 "test/vbinary-microkernel-tester.h",
4905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004910 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004911 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004912 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004913 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004919 name = "f32_vsubc_test",
4920 srcs = [
4921 "test/f32-vsubc.cc",
4922 "test/vbinaryc-microkernel-tester.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004928 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004929 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004930 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004931 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004937 name = "f32_vrsubc_test",
4938 srcs = [
4939 "test/f32-vrsubc.cc",
4940 "test/vbinaryc-microkernel-tester.h",
4941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004946 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004947 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004948 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004949 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004955 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004956 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004957 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958 "test/avgpool-microkernel-tester.h",
4959 "src/xnnpack/AlignedAllocator.h",
4960 ] + MICROKERNEL_TEST_HDRS,
4961 deps = MICROKERNEL_TEST_DEPS,
4962)
4963
4964xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004965 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004966 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004967 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 "test/gemm-microkernel-tester.h",
4969 "src/xnnpack/AlignedAllocator.h",
4970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004971 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972)
4973
4974xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004975 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004977 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978 "test/dwconv-microkernel-tester.h",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004981 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004982)
4983
4984xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004985 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004987 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004988 "test/gavgpool-microkernel-tester.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + MICROKERNEL_TEST_HDRS,
4991 deps = MICROKERNEL_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004995 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004997 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998 "test/gemm-microkernel-tester.h",
4999 "src/xnnpack/AlignedAllocator.h",
5000 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005001 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002)
5003
5004xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005005 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005006 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005007 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008 "test/vadd-microkernel-tester.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
5014 name = "u8_clamp_test",
5015 srcs = [
5016 "test/u8-clamp.cc",
5017 "test/clamp-microkernel-tester.h",
5018 ] + MICROKERNEL_TEST_HDRS,
5019 deps = MICROKERNEL_TEST_DEPS,
5020)
5021
5022xnnpack_unit_test(
5023 name = "u8_lut32norm_test",
5024 srcs = [
5025 "test/u8-lut32norm.cc",
5026 "test/lut-norm-microkernel-tester.h",
5027 ] + MICROKERNEL_TEST_HDRS,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005032 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005034 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 "test/maxpool-microkernel-tester.h",
5036 ] + MICROKERNEL_TEST_HDRS,
5037 deps = MICROKERNEL_TEST_DEPS,
5038)
5039
5040xnnpack_unit_test(
5041 name = "u8_rmax_test",
5042 srcs = [
5043 "test/u8-rmax.cc",
5044 "test/rmax-microkernel-tester.h",
5045 ] + MICROKERNEL_TEST_HDRS,
5046 deps = MICROKERNEL_TEST_DEPS,
5047)
5048
5049xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005050 name = "x32_fill_test",
5051 srcs = [
5052 "test/x32-fill.cc",
5053 "test/fill-microkernel-tester.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059 name = "x32_packx_test",
5060 srcs = [
5061 "test/x32-packx.cc",
5062 "test/pack-microkernel-tester.h",
5063 "src/xnnpack/AlignedAllocator.h",
5064 ] + MICROKERNEL_TEST_HDRS,
5065 deps = MICROKERNEL_TEST_DEPS,
5066)
5067
5068xnnpack_unit_test(
5069 name = "x32_pad_test",
5070 srcs = [
5071 "test/x32-pad.cc",
5072 "test/pad-microkernel-tester.h",
5073 ] + MICROKERNEL_TEST_HDRS,
5074 deps = MICROKERNEL_TEST_DEPS,
5075)
5076
5077xnnpack_unit_test(
5078 name = "x32_unpool_test",
5079 srcs = [
5080 "test/x32-unpool.cc",
5081 "test/unpool-microkernel-tester.h",
5082 ] + MICROKERNEL_TEST_HDRS,
5083 deps = MICROKERNEL_TEST_DEPS,
5084)
5085
5086xnnpack_unit_test(
5087 name = "x32_zip_test",
5088 srcs = [
5089 "test/x32-zip.cc",
5090 "test/zip-microkernel-tester.h",
5091 ] + MICROKERNEL_TEST_HDRS,
5092 deps = MICROKERNEL_TEST_DEPS,
5093)
5094
5095xnnpack_unit_test(
5096 name = "x8_lut_test",
5097 srcs = [
5098 "test/x8-lut.cc",
5099 "test/lut-microkernel-tester.h",
5100 ] + MICROKERNEL_TEST_HDRS,
5101 deps = MICROKERNEL_TEST_DEPS,
5102)
5103
5104xnnpack_unit_test(
5105 name = "x8_zip_test",
5106 srcs = [
5107 "test/x8-zip.cc",
5108 "test/zip-microkernel-tester.h",
5109 ] + MICROKERNEL_TEST_HDRS,
5110 deps = MICROKERNEL_TEST_DEPS,
5111)
5112
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005113xnnpack_unit_test(
5114 name = "requantization_test",
5115 srcs = [
5116 "src/xnnpack/requantization-stubs.h",
5117 "test/requantization.cc",
5118 "test/requantization-tester.h",
5119 ] + MICROKERNEL_TEST_HDRS,
5120 deps = MICROKERNEL_TEST_DEPS,
5121)
5122
Marat Dukhan20c3b922020-03-10 03:45:06 -07005123########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005124
5125xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005126 name = "operator_size_test",
5127 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005128 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129)
5130
Marat Dukhan20c3b922020-03-10 03:45:06 -07005131xnnpack_binary(
5132 name = "subgraph_size_test",
5133 srcs = ["test/subgraph-size.c"],
5134 deps = [":XNNPACK"],
5135)
5136
5137########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005138
5139xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005140 name = "abs_nc_test",
5141 srcs = [
5142 "test/abs-nc.cc",
5143 "test/abs-operator-tester.h",
5144 ],
5145 deps = OPERATOR_TEST_DEPS,
5146)
5147
5148xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005149 name = "add_nd_test",
5150 srcs = [
5151 "test/add-nd.cc",
5152 "test/binary-elementwise-operator-tester.h",
5153 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005154 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005155)
5156
5157xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005158 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005160 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161 "test/argmax-pooling-operator-tester.h",
5162 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005163 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164)
5165
5166xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005167 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005169 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005170 "test/average-pooling-operator-tester.h",
5171 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005172 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173)
5174
5175xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005176 name = "bankers_rounding_nc_test",
5177 srcs = [
5178 "test/bankers-rounding-nc.cc",
5179 "test/bankers-rounding-operator-tester.h",
5180 ],
5181 deps = OPERATOR_TEST_DEPS,
5182)
5183
5184xnnpack_unit_test(
5185 name = "ceiling_nc_test",
5186 srcs = [
5187 "test/ceiling-nc.cc",
5188 "test/ceiling-operator-tester.h",
5189 ],
5190 deps = OPERATOR_TEST_DEPS,
5191)
5192
5193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005194 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005196 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 "test/channel-shuffle-operator-tester.h",
5198 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200)
5201
5202xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005203 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005205 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 "test/clamp-operator-tester.h",
5207 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005208 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209)
5210
5211xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005212 name = "constant_pad_nd_test",
5213 srcs = [
5214 "test/constant-pad-nd.cc",
5215 "test/constant-pad-operator-tester.h",
5216 ],
5217 deps = OPERATOR_TEST_DEPS,
5218)
5219
5220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005221 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005223 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 "test/convolution-operator-tester.h",
5225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227)
5228
5229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005230 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005232 "test/convolution-nchw.cc",
5233 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005236)
5237
5238xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005239 name = "copy_nc_test",
5240 srcs = [
5241 "test/copy-nc.cc",
5242 "test/copy-operator-tester.h",
5243 ],
5244 deps = OPERATOR_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005248 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005250 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251 "test/deconvolution-operator-tester.h",
5252 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254)
5255
5256xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005257 name = "divide_nd_test",
5258 srcs = [
5259 "test/binary-elementwise-operator-tester.h",
5260 "test/divide-nd.cc",
5261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005263)
5264
5265xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005266 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005268 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269 "test/fully-connected-operator-tester.h",
5270 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005271 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272)
5273
5274xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005275 name = "floor_nc_test",
5276 srcs = [
5277 "test/floor-nc.cc",
5278 "test/floor-operator-tester.h",
5279 ],
5280 deps = OPERATOR_TEST_DEPS,
5281)
5282
5283xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005284 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005286 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005288 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290)
5291
5292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005293 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005295 "test/global-average-pooling-ncw.cc",
5296 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005302 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005304 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 "test/hardswish-operator-tester.h",
5306 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308)
5309
5310xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005311 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005313 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314 "test/leaky-relu-operator-tester.h",
5315 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005316 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317)
5318
5319xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005320 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005322 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 "test/max-pooling-operator-tester.h",
5324 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005326)
5327
5328xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005329 name = "maximum_nd_test",
5330 srcs = [
5331 "test/binary-elementwise-operator-tester.h",
5332 "test/maximum-nd.cc",
5333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005335)
5336
5337xnnpack_unit_test(
5338 name = "minimum_nd_test",
5339 srcs = [
5340 "test/binary-elementwise-operator-tester.h",
5341 "test/minimum-nd.cc",
5342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005344)
5345
5346xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005347 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005348 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005349 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005350 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005351 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005352 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005353)
5354
5355xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005356 name = "negate_nc_test",
5357 srcs = [
5358 "test/negate-nc.cc",
5359 "test/negate-operator-tester.h",
5360 ],
5361 deps = OPERATOR_TEST_DEPS,
5362)
5363
5364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005365 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005367 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005368 "test/prelu-operator-tester.h",
5369 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371)
5372
5373xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005374 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005375 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005376 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005377 "test/resize-bilinear-operator-tester.h",
5378 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005380)
5381
5382xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005383 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005385 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386 "test/sigmoid-operator-tester.h",
5387 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005388 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389)
5390
5391xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005392 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005394 "test/softmax-nc.cc",
5395 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005397 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005398)
5399
5400xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005401 name = "square_nc_test",
5402 srcs = [
5403 "test/square-nc.cc",
5404 "test/square-operator-tester.h",
5405 ],
5406 deps = OPERATOR_TEST_DEPS,
5407)
5408
5409xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005410 name = "square_root_nc_test",
5411 srcs = [
5412 "test/square-root-nc.cc",
5413 "test/square-root-operator-tester.h",
5414 ],
5415 deps = OPERATOR_TEST_DEPS,
5416)
5417
5418xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005419 name = "squared_difference_nd_test",
5420 srcs = [
5421 "test/binary-elementwise-operator-tester.h",
5422 "test/squared-difference-nd.cc",
5423 ],
5424 deps = OPERATOR_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005428 name = "subtract_nd_test",
5429 srcs = [
5430 "test/binary-elementwise-operator-tester.h",
5431 "test/subtract-nd.cc",
5432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005434)
5435
5436xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005437 name = "truncation_nc_test",
5438 srcs = [
5439 "test/truncation-nc.cc",
5440 "test/truncation-operator-tester.h",
5441 ],
5442 deps = OPERATOR_TEST_DEPS,
5443)
5444
5445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005446 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005448 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005449 "test/unpooling-operator-tester.h",
5450 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005451 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005452)
5453
Chao Mei6ddfc602020-05-13 22:29:36 -07005454############################### Misc unit tests ###############################
5455
5456xnnpack_unit_test(
5457 name = "memory_planner_test",
5458 srcs = [
5459 "test/memory-planner-test.cc",
5460 ],
5461 deps = [
5462 ":XNNPACK",
5463 ":memory_planner",
5464 ],
5465)
5466
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467############################# Build configurations #############################
5468
Marat Dukhanb8642352019-10-30 15:43:02 -07005469# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005470config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005471 name = "xnn_enable_assembly_explicit_true",
5472 define_values = {"xnn_enable_assembly": "true"},
5473)
5474
5475# Disables usage of assembly kernels.
5476config_setting(
5477 name = "xnn_enable_assembly_explicit_false",
5478 define_values = {"xnn_enable_assembly": "false"},
5479)
5480
Marat Dukhan9de90e02020-06-18 16:04:12 -07005481# Enables usage of sparse inference.
5482config_setting(
5483 name = "xnn_enable_sparse_explicit_true",
5484 define_values = {"xnn_enable_sparse": "true"},
5485)
5486
5487# Disables usage of sparse inference.
5488config_setting(
5489 name = "xnn_enable_sparse_explicit_false",
5490 define_values = {"xnn_enable_sparse": "false"},
5491)
5492
Marat Dukhan05702cf2020-03-26 15:41:33 -07005493# Disables usage of HMP-aware optimizations.
5494config_setting(
5495 name = "xnn_enable_hmp_explicit_false",
5496 define_values = {"xnn_enable_hmp": "false"},
5497)
5498
Chao Mei6ddfc602020-05-13 22:29:36 -07005499# Enable usage of optimized memory allocation
5500config_setting(
5501 name = "xnn_enable_memopt_explicit_true",
5502 define_values = {"xnnpack_enable_memopt": "true"},
5503)
5504
5505# Disable usage of optimized memory allocation
5506config_setting(
5507 name = "xnn_enable_memopt_explicit_false",
5508 define_values = {"xnnpack_enable_memopt": "false"},
5509)
5510
Marat Dukhanb8642352019-10-30 15:43:02 -07005511# Builds with -c dbg
5512config_setting(
5513 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005514 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005515 "compilation_mode": "dbg",
5516 },
5517)
5518
5519# Builds with -c opt
5520config_setting(
5521 name = "optimized_build",
5522 values = {
5523 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005524 },
5525)
5526
5527config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005528 name = "linux_k8",
5529 values = {"cpu": "k8"},
5530)
5531
5532config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005533 name = "linux_arm",
5534 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005535)
5536
5537config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005538 name = "linux_armeabi",
5539 values = {"cpu": "armeabi"},
5540)
5541
5542config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005543 name = "linux_armhf",
5544 values = {"cpu": "armhf"},
5545)
5546
5547config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005548 name = "linux_armv7a",
5549 values = {"cpu": "armv7a"},
5550)
5551
5552config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005553 name = "linux_aarch64",
5554 values = {"cpu": "aarch64"},
5555)
5556
5557config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558 name = "android",
5559 values = {"crosstool_top": "//external:android/crosstool"},
5560)
5561
5562config_setting(
5563 name = "android_armv7",
5564 values = {
5565 "crosstool_top": "//external:android/crosstool",
5566 "cpu": "armeabi-v7a",
5567 },
5568)
5569
5570config_setting(
5571 name = "android_arm64",
5572 values = {
5573 "crosstool_top": "//external:android/crosstool",
5574 "cpu": "arm64-v8a",
5575 },
5576)
5577
5578config_setting(
5579 name = "android_x86",
5580 values = {
5581 "crosstool_top": "//external:android/crosstool",
5582 "cpu": "x86",
5583 },
5584)
5585
5586config_setting(
5587 name = "android_x86_64",
5588 values = {
5589 "crosstool_top": "//external:android/crosstool",
5590 "cpu": "x86_64",
5591 },
5592)
5593
5594config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005595 name = "windows_x86_64",
5596 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005597)
5598
5599config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005600 name = "windows_x86_64_clang",
5601 values = {
5602 "compiler": "clang-cl",
5603 "cpu": "x64_windows",
5604 },
5605)
5606
5607config_setting(
5608 name = "windows_x86_64_mingw",
5609 values = {
5610 "compiler": "mingw-gcc",
5611 "cpu": "x64_windows",
5612 },
5613)
5614
5615config_setting(
5616 name = "windows_x86_64_msys",
5617 values = {
5618 "compiler": "msys-gcc",
5619 "cpu": "x64_windows",
5620 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005621)
5622
5623config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005624 name = "macos_x86_64",
5625 values = {
5626 "apple_platform_type": "macos",
5627 "cpu": "darwin",
5628 },
5629)
5630
5631config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005632 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005633 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634)
5635
5636config_setting(
5637 name = "emscripten_wasm",
5638 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005639 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640 "cpu": "wasm",
5641 },
5642)
5643
5644config_setting(
5645 name = "emscripten_wasmsimd",
5646 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005647 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005648 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005649 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005650 },
5651)
5652
5653config_setting(
5654 name = "emscripten_asmjs",
5655 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005656 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005657 "cpu": "asmjs",
5658 },
5659)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005660
5661config_setting(
5662 name = "ios_armv7",
5663 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005664 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005665 "cpu": "ios_armv7",
5666 },
5667)
5668
5669config_setting(
5670 name = "ios_arm64",
5671 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005672 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005673 "cpu": "ios_arm64",
5674 },
5675)
5676
5677config_setting(
5678 name = "ios_arm64e",
5679 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005680 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005681 "cpu": "ios_arm64e",
5682 },
5683)
5684
5685config_setting(
5686 name = "ios_x86",
5687 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005688 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005689 "cpu": "ios_i386",
5690 },
5691)
5692
5693config_setting(
5694 name = "ios_x86_64",
5695 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005696 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005697 "cpu": "ios_x86_64",
5698 },
5699)
5700
5701config_setting(
5702 name = "watchos_armv7k",
5703 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005704 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005705 "cpu": "watchos_armv7k",
5706 },
5707)
5708
5709config_setting(
5710 name = "watchos_arm64_32",
5711 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005712 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005713 "cpu": "watchos_arm64_32",
5714 },
5715)
5716
5717config_setting(
5718 name = "watchos_x86",
5719 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005720 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005721 "cpu": "watchos_i386",
5722 },
5723)
5724
5725config_setting(
5726 name = "watchos_x86_64",
5727 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005728 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005729 "cpu": "watchos_x86_64",
5730 },
5731)
5732
5733config_setting(
5734 name = "tvos_arm64",
5735 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005736 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005737 "cpu": "tvos_arm64",
5738 },
5739)
5740
5741config_setting(
5742 name = "tvos_x86_64",
5743 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005744 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005745 "cpu": "tvos_x86_64",
5746 },
5747)