blob: f1ac5744f8b0b8216d53c8f353ed8babbb61faea [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700851 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
852 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
853 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
854 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
855 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
856 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
857 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
858 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
859 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
860 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
861 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
862 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
863 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
864 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
865 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700866 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700867 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700868 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700869 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700870 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700871 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700872 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700873 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700874 "src/x32-zip/x2-wasmsimd.c",
875 "src/x32-zip/x3-wasmsimd.c",
876 "src/x32-zip/x4-wasmsimd.c",
877 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700878 "src/math/roundne-wasmsimd-addsub.c",
879 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700880 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700881 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700882 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700883 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700884 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700885 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
886 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700887]
888
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800889PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700890 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700891 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700892 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700893 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700894 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700895 "src/f32-spmm/gen/4x1-minmax-psimd.c",
896 "src/f32-spmm/gen/8x1-minmax-psimd.c",
897 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700898]
899
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800900PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700901 "src/qs8-requantization/precise-psimd.c",
902 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700903 "src/qu8-requantization/precise-psimd.c",
904 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800905]
906
Marat Dukhan08c4a432019-10-03 09:29:21 -0700907# ISA-specific micro-kernels
908NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700909 "src/f32-argmaxpool/4x-neon-c4.c",
910 "src/f32-argmaxpool/9p8x-neon-c4.c",
911 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700912 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
913 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700914 "src/f32-clamp/gen/neon-x4.c",
915 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700916 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
917 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700918 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
919 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700920 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
921 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700922 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
923 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700924 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
925 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
926 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
927 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700928 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
929 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
930 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
931 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700932 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
933 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
934 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
935 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700936 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700937 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
938 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700939 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
940 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
942 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
945 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
946 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
948 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
949 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
950 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
951 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
952 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
953 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
954 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
955 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
956 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
957 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
959 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
960 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
961 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
962 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
963 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
966 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
967 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
969 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700970 "src/f32-hswish/gen/hswish-neon-x4.c",
971 "src/f32-hswish/gen/hswish-neon-x8.c",
972 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700973 "src/f32-ibilinear/gen/neon-c4.c",
974 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700975 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
976 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
977 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
978 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
979 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
980 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
981 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
982 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
983 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
984 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
985 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
986 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
987 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
988 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
989 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
990 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700991 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
992 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
993 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700994 "src/f32-ppmm/gen/4x8-minmax-neon.c",
995 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700996 "src/f32-prelu/gen/neon-1x4.c",
997 "src/f32-prelu/gen/neon-1x8.c",
998 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800999 "src/f32-prelu/gen/neon-2x4.c",
1000 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001001 "src/f32-prelu/gen/neon-2x16.c",
1002 "src/f32-prelu/gen/neon-4x4.c",
1003 "src/f32-prelu/gen/neon-4x8.c",
1004 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001005 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1014 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1015 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1016 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1017 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1018 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001029 "src/f32-relu/gen/neon-x4.c",
1030 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001031 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001032 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001033 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1034 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1035 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1036 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1037 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1038 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1039 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1040 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1041 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1042 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1043 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1044 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1045 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1046 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1047 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1048 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1049 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1050 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001051 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1052 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1053 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1054 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001055 "src/f32-vbinary/gen/vmax-neon-x4.c",
1056 "src/f32-vbinary/gen/vmax-neon-x8.c",
1057 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1058 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1059 "src/f32-vbinary/gen/vmin-neon-x4.c",
1060 "src/f32-vbinary/gen/vmin-neon-x8.c",
1061 "src/f32-vbinary/gen/vminc-neon-x4.c",
1062 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001063 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1064 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1065 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1066 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1067 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1068 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001069 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1070 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1071 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1072 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001073 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1074 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1075 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1076 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001077 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1078 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001079 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1080 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001081 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1082 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1083 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1084 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1085 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1086 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1087 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1088 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001089 "src/f32-vunary/gen/vabs-neon-x4.c",
1090 "src/f32-vunary/gen/vabs-neon-x8.c",
1091 "src/f32-vunary/gen/vneg-neon-x4.c",
1092 "src/f32-vunary/gen/vneg-neon-x8.c",
1093 "src/f32-vunary/gen/vsqr-neon-x4.c",
1094 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001095 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1096 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1097 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1098 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1099 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1100 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1101 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1102 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001103 "src/qs8-requantization/precise-neon.c",
1104 "src/qs8-requantization/fp32-neon.c",
1105 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001106 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1107 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1108 "src/qu8-dwconv/up8x9-minmax-neon.c",
1109 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1110 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1111 "src/qu8-gemm/4x8-minmax-neon.c",
1112 "src/qu8-gemm/8x8-minmax-neon.c",
1113 "src/qu8-igemm/4x8-minmax-neon.c",
1114 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001115 "src/qu8-requantization/precise-neon.c",
1116 "src/qu8-requantization/fp32-neon.c",
1117 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001118 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001119 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001120 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001122 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001123 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001124 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001125 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126 "src/x32-zip/x2-neon.c",
1127 "src/x32-zip/x3-neon.c",
1128 "src/x32-zip/x4-neon.c",
1129 "src/x32-zip/xm-neon.c",
1130 "src/x8-zip/x2-neon.c",
1131 "src/x8-zip/x3-neon.c",
1132 "src/x8-zip/x4-neon.c",
1133 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001134 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001135 "src/math/roundd-neon-addsub.c",
1136 "src/math/roundd-neon-cvt.c",
1137 "src/math/roundu-neon-addsub.c",
1138 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001139 "src/math/roundz-neon-addsub.c",
1140 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001141 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001142 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1143 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1144 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1145 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1146 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1147 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001148 "src/math/sqrt-neon-nr1rsqrts.c",
1149 "src/math/sqrt-neon-nr2rsqrts.c",
1150 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001151]
1152
1153NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001154 "src/f32-ibilinear/gen/neonfma-c4.c",
1155 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001156 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1157 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1158 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1159 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1160 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1161 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1162 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1163 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1164 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001165 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1166 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1167 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1168 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001169 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1170 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1171 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1172 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001173 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1174 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1175 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1176 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001177 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1178 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1179 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1180 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1181 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1182 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1183 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1184 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1185 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1186 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1187 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1188 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1189 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1190 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1191 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1192 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1193 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1194 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001195 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1196 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001197 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1198 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1199 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1200 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001221 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001275 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1276 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001277 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1278 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1279 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1280 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1281 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1282 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1283 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1284 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1285 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1286 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1287 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1288 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1289 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1290 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1291 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1292 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1293 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1294 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1295 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1296 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001297 "src/math/exp-neonfma-lut64-p2.c",
1298 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001299 "src/math/expminus-neonfma-lut2048-p1.c",
1300 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001301 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001302 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1303 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1304 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1305 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1306 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1307 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1308 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1309 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1310 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1311 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1312 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1313 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1314 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1315 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1316 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1317 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1318 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1319 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001320 "src/math/sqrt-neonfma-nr1fma.c",
1321 "src/math/sqrt-neonfma-nr2fma.c",
1322 "src/math/sqrt-neonfma-nr3fma.c",
1323 "src/math/sqrt-neonfma-nr2fma1adj.c",
1324 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001325]
1326
1327AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001328 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1329 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1330 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1331 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1332 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1333 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001334 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1335 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1336 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1337 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1338 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1339 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1340 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1341 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1342 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1343 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1344 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1345 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1346 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1347 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1348 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1349 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1350 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1351 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1352 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1353 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001354 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1355 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001356 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1357 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001358 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1359 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001360 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1361 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001362 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1363 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1364 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1365 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1366 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001367 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001385 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1386 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1387 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1388 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1389 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1390 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1391 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1392 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1393 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1394 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1395 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1396 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1397 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1398 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1399 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1400 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1401 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1402 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001403 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1404 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001405 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1406 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1407 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1408 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1409 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1410 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001411]
1412
Marat Dukhan8853b822020-05-07 12:19:01 -07001413NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001414 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1415 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1416 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1417 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1418 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1419 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1420 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1421 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001422 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001423 "src/math/roundd-neonv8.c",
1424 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001425 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001426]
1427
Marat Dukhan08c4a432019-10-03 09:29:21 -07001428AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001429 "src/f16-clamp/gen/neonfp16arith-x16.c",
1430 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001431 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1432 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1433 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1434 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1435 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1436 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1437 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1438 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1439 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1440 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1441 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1442 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001443 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1444 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001445 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1446 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001447 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001448 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001449 "src/f16-relu/gen/neonfp16arith-x16.c",
1450 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001451 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1455 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1456 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1457 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1458 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1459 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1460 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1461 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1462 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1463 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1464 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1465 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1466 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1467 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1468 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1469 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1470 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1471 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1472 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1473 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1474 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1475 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1476 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1477 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1478 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001479 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1480 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001481 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1482 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1483 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1484 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1487 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001489 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1490 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1491 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1492 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001493 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1494 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1495 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1496 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1497 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1498 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1499 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1501 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1502 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1503 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1504 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001505 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1506 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1507 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1508 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1509 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1510 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1511 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1512 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001513]
1514
1515SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001516 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1517 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001518 "src/f32-clamp/gen/sse-x4.c",
1519 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001520 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1521 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001522 "src/f32-dwconv-chw/3x3p1-sse.c",
1523 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1525 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1526 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1527 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1528 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1529 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1530 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1531 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1532 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1533 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1534 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1535 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001536 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001537 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1538 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001539 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1540 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1541 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1542 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1543 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1544 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1545 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1546 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1547 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1548 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1549 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1550 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001552 "src/f32-hswish/gen/hswish-sse-x4.c",
1553 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001554 "src/f32-ibilinear/gen/sse-c4.c",
1555 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001556 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1557 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1558 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1559 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1560 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1561 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1562 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001563 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1564 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1565 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001566 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001567 "src/f32-prelu/gen/sse-2x4.c",
1568 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001569 "src/f32-relu/gen/sse-x4.c",
1570 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001571 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001572 "src/f32-spmm/gen/4x1-minmax-sse.c",
1573 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001574 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001575 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1576 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1577 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1578 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1579 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1581 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001583 "src/f32-vbinary/gen/vmax-sse-x4.c",
1584 "src/f32-vbinary/gen/vmax-sse-x8.c",
1585 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1586 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1587 "src/f32-vbinary/gen/vmin-sse-x4.c",
1588 "src/f32-vbinary/gen/vmin-sse-x8.c",
1589 "src/f32-vbinary/gen/vminc-sse-x4.c",
1590 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001591 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1592 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1593 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1594 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1595 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1596 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1597 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1598 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001599 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1600 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1601 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1602 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001603 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1604 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1605 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1606 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001607 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1608 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001609 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1610 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001611 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1612 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001613 "src/f32-vunary/gen/vabs-sse-x4.c",
1614 "src/f32-vunary/gen/vabs-sse-x8.c",
1615 "src/f32-vunary/gen/vneg-sse-x4.c",
1616 "src/f32-vunary/gen/vneg-sse-x8.c",
1617 "src/f32-vunary/gen/vsqr-sse-x4.c",
1618 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001619 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001620 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001621 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001622 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001623 "src/math/roundd-sse-addsub.c",
1624 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001625 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001626 "src/math/sqrt-sse-hh1mac.c",
1627 "src/math/sqrt-sse-nr1mac.c",
1628 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001629]
1630
1631SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001632 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1633 "src/f32-argmaxpool/4x-sse2-c4.c",
1634 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001635 "src/f32-prelu/gen/sse2-2x4.c",
1636 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001637 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1638 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1639 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1640 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1641 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1642 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1643 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1644 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1645 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1646 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1647 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1648 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001649 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001650 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001651 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001652 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001653 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1654 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001655 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1656 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001657 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1658 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1659 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1660 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1661 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1662 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1663 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1664 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001665 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1666 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1667 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001668 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1669 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001670 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1671 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001672 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1673 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001674 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1675 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001676 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001677 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1678 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001679 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001680 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1681 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1682 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001683 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1684 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001685 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1686 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001687 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1688 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001689 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001690 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1691 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001692 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001693 "src/qs8-requantization/precise-sse2.c",
1694 "src/qs8-requantization/fp32-sse2.c",
1695 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001696 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1697 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1698 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1699 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1700 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1701 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1702 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1703 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001704 "src/qu8-requantization/precise-sse2.c",
1705 "src/qu8-requantization/fp32-sse2.c",
1706 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001707 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001708 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001709 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001710 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001711 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 "src/x32-zip/x2-sse2.c",
1713 "src/x32-zip/x3-sse2.c",
1714 "src/x32-zip/x4-sse2.c",
1715 "src/x32-zip/xm-sse2.c",
1716 "src/x8-zip/x2-sse2.c",
1717 "src/x8-zip/x3-sse2.c",
1718 "src/x8-zip/x4-sse2.c",
1719 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001720 "src/math/exp-sse2-p5.c",
1721 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001722 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001723 "src/math/roundd-sse2-cvt.c",
1724 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001725 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001726 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001727]
1728
1729SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001730 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1731 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1732 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001733 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1734 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001735 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1736 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001737 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1738 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001739 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1740 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001741 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001742 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1743 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001744 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001745 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1746 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1747 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001748 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1749 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001750 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1751 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001752 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1753 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001754 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001755 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1756 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001757 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001758 "src/qs8-requantization/precise-ssse3.c",
1759 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001760 "src/qu8-requantization/precise-ssse3.c",
1761 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762]
1763
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001764SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001765 "src/f32-prelu/gen/sse41-2x4.c",
1766 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001767 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001768 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001769 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001770 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001771 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1772 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001773 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1774 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001775 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1776 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1777 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1778 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1779 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1780 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1781 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1782 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001783 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1784 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1785 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001786 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1787 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001788 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1789 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001790 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1791 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001792 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1793 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001794 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001795 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1796 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001797 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001798 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1799 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1800 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001801 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1802 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001803 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1804 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001805 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1806 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001807 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001808 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1809 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001810 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/fp32-sse4.c",
1812 "src/qs8-requantization/precise-sse4.c",
1813 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001814 "src/qu8-requantization/precise-sse4.c",
1815 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001816 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001817 "src/math/roundd-sse41.c",
1818 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001819 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001820]
1821
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001823 "src/f32-clamp/gen/avx-x8.c",
1824 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001825 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1826 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1827 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1828 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1829 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1830 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1831 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1832 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1833 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1834 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1835 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1836 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1837 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1838 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1839 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1840 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1841 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1842 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1843 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1844 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1845 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1846 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1847 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1848 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1849 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1850 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1851 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1852 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1853 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1854 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001855 "src/f32-hswish/gen/hswish-avx-x8.c",
1856 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001857 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1858 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1859 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1860 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1861 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1862 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1863 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1864 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1865 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001866 "src/f32-prelu/gen/avx-2x8.c",
1867 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001868 "src/f32-relu/gen/avx-x16.c",
1869 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001870 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001871 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1872 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1873 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1874 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1875 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1876 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1877 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1878 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001879 "src/f32-vbinary/gen/vmax-avx-x8.c",
1880 "src/f32-vbinary/gen/vmax-avx-x16.c",
1881 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1882 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1883 "src/f32-vbinary/gen/vmin-avx-x8.c",
1884 "src/f32-vbinary/gen/vmin-avx-x16.c",
1885 "src/f32-vbinary/gen/vminc-avx-x8.c",
1886 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001887 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1888 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1889 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1890 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1891 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1892 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1893 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1894 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001895 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1896 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1897 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1898 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001899 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1900 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1901 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1902 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001903 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1904 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001905 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001906 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1907 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1908 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1909 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1910 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1911 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1912 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1913 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001914 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1915 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001916 "src/f32-vunary/gen/vabs-avx-x8.c",
1917 "src/f32-vunary/gen/vabs-avx-x16.c",
1918 "src/f32-vunary/gen/vneg-avx-x8.c",
1919 "src/f32-vunary/gen/vneg-avx-x16.c",
1920 "src/f32-vunary/gen/vsqr-avx-x8.c",
1921 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001922]
1923
Marat Dukhan1566fee2020-08-02 21:55:41 -07001924XOP_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001925 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul16.c",
1926 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul16.c",
1927 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul16.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001928 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1929 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1930 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1931 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001932 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1933 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001934 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1935 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001936 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001937 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1938 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001939 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001940 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1941 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1942 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001943 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1944 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1945 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1946 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1947 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1948 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001949 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001950 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1951 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001952 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001953]
1954
Marat Dukhanfda12b82019-11-21 12:27:59 -08001955FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001956 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1957 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1958 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1959 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1960 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1961 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1962 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1963 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1964 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1965 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1966 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1967 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1968 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1969 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1970 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1971 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1972 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1973 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1974 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1975 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1976 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1977 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1978 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1979 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1980 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1981 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1982 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1983 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1984 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1985 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1986 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1987 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1988 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1989 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1990 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1991 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1992 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1993 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1994 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1995 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001996 "src/f32-hswish/gen/hswish-fma3-x8.c",
1997 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001998 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1999 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2000 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2001 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2002 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2003 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2004 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2005 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2006 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2007 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2008 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2009 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2010 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2011 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002012 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2014 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2015 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2016 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2017 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2018 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2019 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002020 "src/math/sqrt-fma3-nr1fma.c",
2021 "src/math/sqrt-fma3-nr2fma.c",
2022 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002023]
2024
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002025AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002026 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2027 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2028 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2029 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2030 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2031 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2032 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2033 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2034 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2035 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2036 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2037 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2038 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2039 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2040 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2041 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2042 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2043 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2044 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2045 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2046 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2047 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2048 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2049 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2050 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2051 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2052 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2053 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2054 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2055 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2056 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2057 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2059 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2060 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2061 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002062 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2063 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2064 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2065 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2066 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2067 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2068 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2069 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2070 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2071 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2083 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2084 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2085 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2086 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2087 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2088 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2089 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2090 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2091 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002092 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2093 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2094 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2095 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2096 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2097 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2098 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2099 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2100 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2101 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2102 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2103 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2104 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2105 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2106 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2107 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2108 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2109 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2110 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2111 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2112 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2113 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2114 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2115 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002116 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2117 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2118 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2119 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002120 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2121 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2122 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002123 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2124 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2125 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002126 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2127 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2128 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002129 "src/math/exp-avx2-p5.c",
2130 "src/math/exp-avx2-perm-p3.c",
2131 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002132 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002133 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002134 "src/math/sigmoid-avx2-rr2-p5-div.c",
2135 "src/math/sigmoid-avx2-rr1-p5-div.c",
2136 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2137 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2138 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2139 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002140]
2141
Marat Dukhan08c4a432019-10-03 09:29:21 -07002142AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002143 "src/f32-clamp/gen/avx512f-x16.c",
2144 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002145 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2146 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2147 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2148 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2149 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2150 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2151 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2152 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2153 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2154 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2155 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2156 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2157 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2158 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2159 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2160 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2161 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2162 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2163 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2164 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2165 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2166 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2167 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2168 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002169 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2170 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002171 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2172 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2173 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2174 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2175 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2176 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002177 "src/f32-prelu/gen/avx512f-2x16.c",
2178 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002179 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2180 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2181 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2182 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2183 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2184 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2185 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2186 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2187 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2188 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2189 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2190 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2191 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2192 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2193 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2194 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2195 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2196 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2197 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2198 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2199 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2200 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2201 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2202 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2203 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2204 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2205 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2206 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2207 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2208 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2209 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2210 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2211 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2212 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2213 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2214 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002215 "src/f32-relu/gen/avx512f-x16.c",
2216 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002218 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2219 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2220 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2221 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2222 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2223 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2224 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2225 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002226 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2227 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2228 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2229 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2230 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2231 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2232 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2233 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002234 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2235 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2236 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2237 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2238 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2239 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2240 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2241 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002242 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2243 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2244 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2245 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002246 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2247 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2248 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2249 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002250 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2251 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002252 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002253 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2254 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2255 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2256 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2257 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2258 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2259 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2260 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2261 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2262 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2263 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2264 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2265 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2266 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2267 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2268 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2269 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2270 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2271 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2272 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2273 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2274 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2275 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2276 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002277 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2278 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2279 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2280 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2281 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2282 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2283 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2284 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002285 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2286 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2287 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2288 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2289 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2290 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2291 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2292 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002293 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2294 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2295 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2296 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2297 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2298 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002299 "src/math/exp-avx512f-p5-scalef.c",
2300 "src/math/exp-avx512f-p5.c",
2301 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002302 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002303 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002304 "src/math/sqrt-avx512f-nr1fma.c",
2305 "src/math/sqrt-avx512f-nr2fma.c",
2306 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307]
2308
2309AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002310 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002311 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002312 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2313 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002314 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002315 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002316 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002317 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002318 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2319 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002320 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2321 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2322 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2323 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324]
2325
2326AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002327 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002328 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2329 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002330 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002331 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2332 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002333 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2334 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2335 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002336 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002337 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2338 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2339 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002340 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002341 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2342 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002343 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002344 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2345 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2346 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2347 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2348 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2349 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2350 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2351 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2352 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2353 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2354 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2355 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2356 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2357 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2358 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2359 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2360 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2361 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002362 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2363 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002364 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2366 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2367 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2368 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2369 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2370 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2371 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2372 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2373 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2374 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2375 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2376 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2377 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2378 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2379 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2380 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2381 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2382 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2384 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2385 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2386 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2387 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2388 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2389 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2390 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2391 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2392 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2393 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2394 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2395 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2396 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2397 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2398 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2399 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2400 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002401]
2402
Marat Dukhan1b354632020-03-23 12:50:22 -07002403INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 "src/xnnpack/argmaxpool.h",
2405 "src/xnnpack/avgpool.h",
2406 "src/xnnpack/clamp.h",
2407 "src/xnnpack/common.h",
2408 "src/xnnpack/conv.h",
2409 "src/xnnpack/dwconv.h",
2410 "src/xnnpack/gavgpool.h",
2411 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002412 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002414 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002415 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002416 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 "src/xnnpack/lut.h",
2418 "src/xnnpack/math.h",
2419 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002420 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421 "src/xnnpack/packx.h",
2422 "src/xnnpack/pad.h",
2423 "src/xnnpack/params.h",
2424 "src/xnnpack/pavgpool.h",
2425 "src/xnnpack/ppmm.h",
2426 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002427 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002428 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002429 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430 "src/xnnpack/rmax.h",
2431 "src/xnnpack/scalar-utils.h",
2432 "src/xnnpack/spmm.h",
2433 "src/xnnpack/unpool.h",
2434 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002435 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002437 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002438 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002439 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002440 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002441 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002442]
2443
2444INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002445 "include/xnnpack.h",
2446 "src/xnnpack/allocator.h",
2447 "src/xnnpack/compute.h",
2448 "src/xnnpack/im2col.h",
2449 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002450 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002451 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002452 "src/xnnpack/operator.h",
2453 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002454 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002456 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002457 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002458]
2459
Marat Dukhan1b354632020-03-23 12:50:22 -07002460ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002461 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002462]
2463
Marat Dukhan1b354632020-03-23 12:50:22 -07002464MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002465 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002466 "include/xnnpack.h",
2467]
2468
Marat Dukhan1b354632020-03-23 12:50:22 -07002469MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002470 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002471 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002472 "src/xnnpack/requantization.h",
2473 "include/xnnpack.h",
2474]
2475
2476OPERATOR_TEST_PARAMS_HDRS = [
2477 "src/xnnpack/params.h",
2478 "src/xnnpack/common.h",
2479]
2480
2481WEIGHTS_PACK_HDRS = [
2482 "src/xnnpack/pack.h",
2483 "src/xnnpack/operator.h",
2484 "src/xnnpack/compute.h",
2485]
2486
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002487LOGGING_COPTS = select({
2488 # No logging in optimized mode
2489 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2490 # Full logging in debug mode
2491 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2492 # Error-only logging in default (fastbuild) mode
2493 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2494})
2495
Marat Dukhan3b59de22020-06-03 20:15:19 -07002496LOGGING_SRCS = select({
2497 # No logging in optimized mode
2498 ":optimized_build": [],
2499 "//conditions:default": [
2500 "src/operator-strings.c",
2501 "src/subgraph-strings.c",
2502 ],
2503})
2504
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002505LOGGING_HDRS = [
2506 "src/xnnpack/log.h",
2507]
2508
Marat Dukhan08c4a432019-10-03 09:29:21 -07002509xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002510 name = "tables",
2511 srcs = TABLE_SRCS,
2512 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002513 gcc_copts = xnnpack_gcc_std_copts(),
2514 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002515)
2516
2517xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002518 name = "scalar_ukernels",
2519 srcs = SCALAR_UKERNELS,
2520 hdrs = INTERNAL_HDRS,
2521 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002522 gcc_copts = xnnpack_gcc_std_copts(),
2523 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002524 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002525 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002526 "@FP16",
2527 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002528 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002529 ],
2530)
2531
2532xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002533 name = "scalar_ukernels_test_mode",
2534 srcs = SCALAR_UKERNELS,
2535 hdrs = INTERNAL_HDRS,
2536 aarch32_copts = ["-marm"],
2537 copts = [
2538 "-UNDEBUG",
2539 "-DXNN_TEST_MODE=1",
2540 ],
2541 gcc_copts = xnnpack_gcc_std_copts(),
2542 msvc_copts = xnnpack_msvc_std_copts(),
2543 deps = [
2544 ":tables",
2545 "@FP16",
2546 "@FXdiv",
2547 "@pthreadpool",
2548 ],
2549)
2550
2551xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002552 name = "wasm_ukernels",
2553 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002554 gcc_copts = xnnpack_gcc_std_copts(),
2555 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002556 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002557 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002558 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002559 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002560 "@FP16",
2561 "@FXdiv",
2562 "@pthreadpool",
2563 ],
2564)
2565
2566xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002567 name = "wasm_ukernels_test_mode",
2568 hdrs = INTERNAL_HDRS,
2569 copts = [
2570 "-UNDEBUG",
2571 "-DXNN_TEST_MODE=1",
2572 ],
2573 gcc_copts = xnnpack_gcc_std_copts(),
2574 msvc_copts = xnnpack_msvc_std_copts(),
2575 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002576 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002577 deps = [
2578 ":tables",
2579 "@FP16",
2580 "@FXdiv",
2581 "@pthreadpool",
2582 ],
2583)
2584
2585xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002586 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002587 hdrs = INTERNAL_HDRS,
2588 aarch32_copts = [
2589 "-marm",
2590 "-mfpu=neon",
2591 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002592 gcc_copts = xnnpack_gcc_std_copts(),
2593 gcc_x86_copts = ["-msse2"],
2594 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595 optimized_copts = [
2596 "-O3",
2597 "-ffast-math",
2598 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002599 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002601 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002602 "@FP16",
2603 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002604 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605 ],
2606)
2607
2608xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002609 name = "psimd_fastmath_ukernels_test_mode",
2610 hdrs = INTERNAL_HDRS,
2611 aarch32_copts = [
2612 "-marm",
2613 "-mfpu=neon",
2614 ],
2615 copts = [
2616 "-UNDEBUG",
2617 "-DXNN_TEST_MODE=1",
2618 ],
2619 gcc_copts = xnnpack_gcc_std_copts(),
2620 gcc_x86_copts = ["-msse2"],
2621 msvc_copts = xnnpack_msvc_std_copts(),
2622 optimized_copts = [
2623 "-O3",
2624 "-ffast-math",
2625 ],
2626 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2627 deps = [
2628 ":tables",
2629 "@FP16",
2630 "@psimd",
2631 "@pthreadpool",
2632 ],
2633)
2634
2635xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002636 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002637 hdrs = INTERNAL_HDRS,
2638 aarch32_copts = [
2639 "-marm",
2640 "-mfpu=neon",
2641 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002642 gcc_copts = xnnpack_gcc_std_copts(),
2643 gcc_x86_copts = ["-msse2"],
2644 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002645 optimized_copts = [
2646 "-O3",
2647 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002648 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002649 deps = [
2650 ":tables",
2651 "@FP16",
2652 "@psimd",
2653 "@pthreadpool",
2654 ],
2655)
2656
2657xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002658 name = "psimd_accmath_ukernels_test_mode",
2659 hdrs = INTERNAL_HDRS,
2660 aarch32_copts = [
2661 "-marm",
2662 "-mfpu=neon",
2663 ],
2664 copts = [
2665 "-UNDEBUG",
2666 "-DXNN_TEST_MODE=1",
2667 ],
2668 gcc_copts = xnnpack_gcc_std_copts(),
2669 gcc_x86_copts = ["-msse2"],
2670 msvc_copts = xnnpack_msvc_std_copts(),
2671 optimized_copts = [
2672 "-O3",
2673 ],
2674 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2675 deps = [
2676 ":tables",
2677 "@FP16",
2678 "@psimd",
2679 "@pthreadpool",
2680 ],
2681)
2682
2683xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684 name = "neon_ukernels",
2685 hdrs = INTERNAL_HDRS,
2686 aarch32_copts = [
2687 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002688 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002689 "-mfpu=neon",
2690 ],
2691 aarch32_srcs = NEON_UKERNELS,
2692 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002693 gcc_copts = xnnpack_gcc_std_copts(),
2694 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002695 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002696 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002697 "@FP16",
2698 "@pthreadpool",
2699 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002700)
2701
2702xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002703 name = "neon_ukernels_test_mode",
2704 hdrs = INTERNAL_HDRS,
2705 aarch32_copts = [
2706 "-marm",
2707 "-march=armv7-a",
2708 "-mfpu=neon",
2709 ],
2710 aarch32_srcs = NEON_UKERNELS,
2711 aarch64_srcs = NEON_UKERNELS,
2712 copts = [
2713 "-UNDEBUG",
2714 "-DXNN_TEST_MODE=1",
2715 ],
2716 gcc_copts = xnnpack_gcc_std_copts(),
2717 msvc_copts = xnnpack_msvc_std_copts(),
2718 deps = [
2719 ":tables",
2720 "@FP16",
2721 "@pthreadpool",
2722 ],
2723)
2724
2725xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 name = "neonfma_ukernels",
2727 hdrs = INTERNAL_HDRS,
2728 aarch32_copts = [
2729 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002730 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002731 "-mfpu=neon-vfpv4",
2732 ],
2733 aarch32_srcs = NEONFMA_UKERNELS,
2734 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002735 apple_aarch32_copts = [
2736 "-mcpu=swift",
2737 "-mtune=generic",
2738 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002739 gcc_copts = xnnpack_gcc_std_copts(),
2740 msvc_copts = xnnpack_msvc_std_copts(),
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 Dukhan08c4a432019-10-03 09:29:21 -07002746)
2747
2748xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002749 name = "neonfma_ukernels_test_mode",
2750 hdrs = INTERNAL_HDRS,
2751 aarch32_copts = [
2752 "-marm",
2753 "-march=armv7-a",
2754 "-mfpu=neon-vfpv4",
2755 ],
2756 aarch32_srcs = NEONFMA_UKERNELS,
2757 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002758 apple_aarch32_copts = [
2759 "-mcpu=swift",
2760 "-mtune=generic",
2761 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002762 copts = [
2763 "-UNDEBUG",
2764 "-DXNN_TEST_MODE=1",
2765 ],
2766 gcc_copts = xnnpack_gcc_std_copts(),
2767 msvc_copts = xnnpack_msvc_std_copts(),
2768 deps = [
2769 ":tables",
2770 "@FP16",
2771 "@pthreadpool",
2772 ],
2773)
2774
2775xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002776 name = "neonv8_ukernels",
2777 hdrs = INTERNAL_HDRS,
2778 aarch32_copts = [
2779 "-marm",
2780 "-march=armv8-a",
2781 "-mfpu=neon-fp-armv8",
2782 ],
2783 aarch32_srcs = NEONV8_UKERNELS,
2784 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002785 apple_aarch32_copts = [
2786 "-mcpu=cyclone",
2787 "-mtune=generic",
2788 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002789 gcc_copts = xnnpack_gcc_std_copts(),
2790 msvc_copts = xnnpack_msvc_std_copts(),
2791 deps = [
2792 ":tables",
2793 "@FP16",
2794 "@pthreadpool",
2795 ],
2796)
2797
2798xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002799 name = "neonv8_ukernels_test_mode",
2800 hdrs = INTERNAL_HDRS,
2801 aarch32_copts = [
2802 "-marm",
2803 "-march=armv8-a",
2804 "-mfpu=neon-fp-armv8",
2805 ],
2806 aarch32_srcs = NEONV8_UKERNELS,
2807 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002808 apple_aarch32_copts = [
2809 "-mcpu=cyclone",
2810 "-mtune=generic",
2811 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002812 copts = [
2813 "-UNDEBUG",
2814 "-DXNN_TEST_MODE=1",
2815 ],
2816 gcc_copts = xnnpack_gcc_std_copts(),
2817 msvc_copts = xnnpack_msvc_std_copts(),
2818 deps = [
2819 ":tables",
2820 "@FP16",
2821 "@pthreadpool",
2822 ],
2823)
2824
2825xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 name = "neonfp16arith_ukernels",
2827 hdrs = INTERNAL_HDRS,
2828 aarch64_copts = ["-march=armv8.2-a+fp16"],
2829 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002830 gcc_copts = xnnpack_gcc_std_copts(),
2831 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002832 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002833 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002834 "@FP16",
2835 "@pthreadpool",
2836 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837)
2838
2839xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002840 name = "neonfp16arith_ukernels_test_mode",
2841 hdrs = INTERNAL_HDRS,
2842 aarch64_copts = ["-march=armv8.2-a+fp16"],
2843 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2844 copts = [
2845 "-UNDEBUG",
2846 "-DXNN_TEST_MODE=1",
2847 ],
2848 gcc_copts = xnnpack_gcc_std_copts(),
2849 msvc_copts = xnnpack_msvc_std_copts(),
2850 deps = [
2851 ":tables",
2852 "@FP16",
2853 "@pthreadpool",
2854 ],
2855)
2856
2857xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002858 name = "sse2_ukernels",
2859 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002860 gcc_copts = xnnpack_gcc_std_copts(),
2861 gcc_x86_copts = ["-msse2"],
2862 msvc_copts = xnnpack_msvc_std_copts(),
2863 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002864 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002865 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002866 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002867 "@FP16",
2868 "@pthreadpool",
2869 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002870)
2871
2872xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002873 name = "sse2_ukernels_test_mode",
2874 hdrs = INTERNAL_HDRS,
2875 copts = [
2876 "-UNDEBUG",
2877 "-DXNN_TEST_MODE=1",
2878 ],
2879 gcc_copts = xnnpack_gcc_std_copts(),
2880 gcc_x86_copts = ["-msse2"],
2881 msvc_copts = xnnpack_msvc_std_copts(),
2882 msvc_x86_32_copts = ["/arch:SSE2"],
2883 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2884 deps = [
2885 ":tables",
2886 "@FP16",
2887 "@pthreadpool",
2888 ],
2889)
2890
2891xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002892 name = "ssse3_ukernels",
2893 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002894 gcc_copts = xnnpack_gcc_std_copts(),
2895 gcc_x86_copts = ["-mssse3"],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002898 x86_srcs = SSSE3_UKERNELS,
2899 deps = [
2900 ":tables",
2901 "@FP16",
2902 "@pthreadpool",
2903 ],
2904)
2905
2906xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002907 name = "ssse3_ukernels_test_mode",
2908 hdrs = INTERNAL_HDRS,
2909 copts = [
2910 "-UNDEBUG",
2911 "-DXNN_TEST_MODE=1",
2912 ],
2913 gcc_copts = xnnpack_gcc_std_copts(),
2914 gcc_x86_copts = ["-mssse3"],
2915 msvc_copts = xnnpack_msvc_std_copts(),
2916 msvc_x86_32_copts = ["/arch:SSE2"],
2917 x86_srcs = SSSE3_UKERNELS,
2918 deps = [
2919 ":tables",
2920 "@FP16",
2921 "@pthreadpool",
2922 ],
2923)
2924
2925xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002926 name = "sse41_ukernels",
2927 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002928 gcc_copts = xnnpack_gcc_std_copts(),
2929 gcc_x86_copts = ["-msse4.1"],
2930 msvc_copts = xnnpack_msvc_std_copts(),
2931 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002932 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002933 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002934 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002935 "@FP16",
2936 "@pthreadpool",
2937 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002938)
2939
2940xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002941 name = "sse41_ukernels_test_mode",
2942 hdrs = INTERNAL_HDRS,
2943 copts = [
2944 "-UNDEBUG",
2945 "-DXNN_TEST_MODE=1",
2946 ],
2947 gcc_copts = xnnpack_gcc_std_copts(),
2948 gcc_x86_copts = ["-msse4.1"],
2949 msvc_copts = xnnpack_msvc_std_copts(),
2950 msvc_x86_32_copts = ["/arch:SSE2"],
2951 x86_srcs = SSE41_UKERNELS,
2952 deps = [
2953 ":tables",
2954 "@FP16",
2955 "@pthreadpool",
2956 ],
2957)
2958
2959xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 name = "avx_ukernels",
2961 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002962 gcc_copts = xnnpack_gcc_std_copts(),
2963 gcc_x86_copts = ["-mavx"],
2964 msvc_copts = xnnpack_msvc_std_copts(),
2965 msvc_x86_32_copts = ["/arch:AVX"],
2966 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002968 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002969 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002970 "@FP16",
2971 "@pthreadpool",
2972 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973)
2974
2975xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002976 name = "avx_ukernels_test_mode",
2977 hdrs = INTERNAL_HDRS,
2978 copts = [
2979 "-UNDEBUG",
2980 "-DXNN_TEST_MODE=1",
2981 ],
2982 gcc_copts = xnnpack_gcc_std_copts(),
2983 gcc_x86_copts = ["-mavx"],
2984 msvc_copts = xnnpack_msvc_std_copts(),
2985 msvc_x86_32_copts = ["/arch:AVX"],
2986 msvc_x86_64_copts = ["/arch:AVX"],
2987 x86_srcs = AVX_UKERNELS,
2988 deps = [
2989 ":tables",
2990 "@FP16",
2991 "@pthreadpool",
2992 ],
2993)
2994
2995xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002996 name = "xop_ukernels",
2997 hdrs = INTERNAL_HDRS,
2998 gcc_copts = xnnpack_gcc_std_copts(),
2999 gcc_x86_copts = ["-mxop"],
3000 msvc_copts = xnnpack_msvc_std_copts(),
3001 msvc_x86_32_copts = ["/arch:AVX"],
3002 msvc_x86_64_copts = ["/arch:AVX"],
3003 x86_srcs = XOP_UKERNELS,
3004 deps = [
3005 ":tables",
3006 "@FP16",
3007 "@pthreadpool",
3008 ],
3009)
3010
3011xnnpack_cc_library(
3012 name = "xop_ukernels_test_mode",
3013 hdrs = INTERNAL_HDRS,
3014 copts = [
3015 "-UNDEBUG",
3016 "-DXNN_TEST_MODE=1",
3017 ],
3018 gcc_copts = xnnpack_gcc_std_copts(),
3019 gcc_x86_copts = ["-mxop"],
3020 msvc_copts = xnnpack_msvc_std_copts(),
3021 msvc_x86_32_copts = ["/arch:AVX"],
3022 msvc_x86_64_copts = ["/arch:AVX"],
3023 x86_srcs = XOP_UKERNELS,
3024 deps = [
3025 ":tables",
3026 "@FP16",
3027 "@pthreadpool",
3028 ],
3029)
3030
3031xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003032 name = "fma3_ukernels",
3033 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003034 gcc_copts = xnnpack_gcc_std_copts(),
3035 gcc_x86_copts = ["-mfma"],
3036 msvc_copts = xnnpack_msvc_std_copts(),
3037 msvc_x86_32_copts = ["/arch:AVX"],
3038 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003039 x86_srcs = FMA3_UKERNELS,
3040 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003041 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003042 "@FP16",
3043 "@pthreadpool",
3044 ],
3045)
3046
3047xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003048 name = "fma3_ukernels_test_mode",
3049 hdrs = INTERNAL_HDRS,
3050 copts = [
3051 "-UNDEBUG",
3052 "-DXNN_TEST_MODE=1",
3053 ],
3054 gcc_copts = xnnpack_gcc_std_copts(),
3055 gcc_x86_copts = ["-mfma"],
3056 msvc_copts = xnnpack_msvc_std_copts(),
3057 msvc_x86_32_copts = ["/arch:AVX"],
3058 msvc_x86_64_copts = ["/arch:AVX"],
3059 x86_srcs = FMA3_UKERNELS,
3060 deps = [
3061 ":tables",
3062 "@FP16",
3063 "@pthreadpool",
3064 ],
3065)
3066
3067xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003068 name = "avx2_ukernels",
3069 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003070 gcc_copts = xnnpack_gcc_std_copts(),
3071 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003072 "-mfma",
3073 "-mavx2",
3074 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003075 msvc_copts = xnnpack_msvc_std_copts(),
3076 msvc_x86_32_copts = ["/arch:AVX2"],
3077 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003078 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003079 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003080 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003081 "@FP16",
3082 "@pthreadpool",
3083 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003084)
3085
3086xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003087 name = "avx2_ukernels_test_mode",
3088 hdrs = INTERNAL_HDRS,
3089 copts = [
3090 "-UNDEBUG",
3091 "-DXNN_TEST_MODE=1",
3092 ],
3093 gcc_copts = xnnpack_gcc_std_copts(),
3094 gcc_x86_copts = [
3095 "-mfma",
3096 "-mavx2",
3097 ],
3098 msvc_copts = xnnpack_msvc_std_copts(),
3099 msvc_x86_32_copts = ["/arch:AVX2"],
3100 msvc_x86_64_copts = ["/arch:AVX2"],
3101 x86_srcs = AVX2_UKERNELS,
3102 deps = [
3103 ":tables",
3104 "@FP16",
3105 "@pthreadpool",
3106 ],
3107)
3108
3109xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 name = "avx512f_ukernels",
3111 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003112 gcc_copts = xnnpack_gcc_std_copts(),
3113 gcc_x86_copts = ["-mavx512f"],
3114 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3115 msvc_copts = xnnpack_msvc_std_copts(),
3116 msvc_x86_32_copts = ["/arch:AVX512"],
3117 msvc_x86_64_copts = ["/arch:AVX512"],
3118 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003120 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003121 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003122 "@FP16",
3123 "@pthreadpool",
3124 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125)
3126
3127xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003128 name = "avx512f_ukernels_test_mode",
3129 hdrs = INTERNAL_HDRS,
3130 copts = [
3131 "-UNDEBUG",
3132 "-DXNN_TEST_MODE=1",
3133 ],
3134 gcc_copts = xnnpack_gcc_std_copts(),
3135 gcc_x86_copts = ["-mavx512f"],
3136 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3137 msvc_copts = xnnpack_msvc_std_copts(),
3138 msvc_x86_32_copts = ["/arch:AVX512"],
3139 msvc_x86_64_copts = ["/arch:AVX512"],
3140 msys_copts = ["-fno-asynchronous-unwind-tables"],
3141 x86_srcs = AVX512F_UKERNELS,
3142 deps = [
3143 ":tables",
3144 "@FP16",
3145 "@pthreadpool",
3146 ],
3147)
3148
3149xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 name = "asm_ukernels",
3151 hdrs = ["src/xnnpack/assembly.h"],
3152 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003153 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154 aarch64_srcs = AARCH64_ASM_UKERNELS,
3155)
3156
Marat Dukhan3b59de22020-06-03 20:15:19 -07003157xnnpack_cc_library(
3158 name = "logging_utils",
3159 srcs = LOGGING_SRCS,
3160 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3161 copts = LOGGING_COPTS + [
3162 "-Isrc",
3163 "-Iinclude",
3164 ] + select({
3165 ":debug_build": [],
3166 "//conditions:default": xnnpack_min_size_copts(),
3167 }),
3168 gcc_copts = xnnpack_gcc_std_copts(),
3169 msvc_copts = xnnpack_msvc_std_copts(),
3170 visibility = xnnpack_visibility(),
3171 deps = [
3172 "@FP16",
3173 "@clog",
3174 "@pthreadpool",
3175 ],
3176)
3177
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178xnnpack_aggregate_library(
3179 name = "ukernels",
3180 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 ":neon_ukernels",
3182 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003183 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 ":asm_ukernels",
3185 ],
3186 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 ":neon_ukernels",
3188 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003189 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 ":neonfp16arith_ukernels",
3191 ":asm_ukernels",
3192 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003193 generic_deps = [
3194 ":scalar_ukernels",
3195 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003196 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003197 ":psimd_fastmath_ukernels",
3198 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003200 wasm_deps = [
3201 ":wasm_ukernels",
3202 ],
3203 wasmsimd_deps = [
3204 ":wasm_ukernels",
3205 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003208 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003209 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003211 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003212 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003213 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 ":avx512f_ukernels",
3215 ],
3216)
3217
Marat Dukhan33fcf782020-05-24 14:27:15 -07003218xnnpack_aggregate_library(
3219 name = "ukernels_test_mode",
3220 aarch32_deps = [
3221 ":neon_ukernels_test_mode",
3222 ":neonfma_ukernels_test_mode",
3223 ":neonv8_ukernels_test_mode",
3224 ":asm_ukernels",
3225 ],
3226 aarch64_deps = [
3227 ":neon_ukernels_test_mode",
3228 ":neonfma_ukernels_test_mode",
3229 ":neonv8_ukernels_test_mode",
3230 ":neonfp16arith_ukernels_test_mode",
3231 ":asm_ukernels",
3232 ],
3233 generic_deps = [
3234 ":scalar_ukernels_test_mode",
3235 ],
3236 psimd_deps = [
3237 ":psimd_fastmath_ukernels_test_mode",
3238 ":psimd_accmath_ukernels_test_mode",
3239 ],
3240 wasm_deps = [
3241 ":wasm_ukernels_test_mode",
3242 ],
3243 wasmsimd_deps = [
3244 ":wasm_ukernels_test_mode",
3245 ],
3246 x86_deps = [
3247 ":sse2_ukernels_test_mode",
3248 ":ssse3_ukernels_test_mode",
3249 ":sse41_ukernels_test_mode",
3250 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003251 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003252 ":fma3_ukernels_test_mode",
3253 ":avx2_ukernels_test_mode",
3254 ":avx512f_ukernels_test_mode",
3255 ],
3256)
3257
Marat Dukhan86df3842020-06-17 14:03:31 -07003258# This target must be compatible with Android NDK r17 and compile with Clang 6.
3259# It is similar to :ukernels target, but excludes microkernels which require
3260# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003261xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003262 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003263 aarch32_deps = [
3264 ":neon_ukernels",
3265 ":neonfma_ukernels",
3266 ":neonv8_ukernels",
3267 ":asm_ukernels",
3268 ],
3269 aarch64_deps = [
3270 ":neon_ukernels",
3271 ":neonfma_ukernels",
3272 ":neonv8_ukernels",
3273 ":asm_ukernels",
3274 ],
3275 generic_deps = [
3276 ":scalar_ukernels",
3277 ],
3278 psimd_deps = [
3279 ":psimd_fastmath_ukernels",
3280 ":psimd_accmath_ukernels",
3281 ],
3282 wasm_deps = [
3283 ":wasm_ukernels",
3284 ],
3285 wasmsimd_deps = [
3286 ":wasm_ukernels",
3287 ],
3288 x86_deps = [
3289 ":sse2_ukernels",
3290 ":ssse3_ukernels",
3291 ":sse41_ukernels",
3292 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003293 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003294 ":fma3_ukernels",
3295 ":avx2_ukernels",
3296 ":avx512f_ukernels",
3297 ],
3298)
3299
Marat Dukhan08c4a432019-10-03 09:29:21 -07003300xnnpack_cc_library(
3301 name = "im2col",
3302 srcs = ["src/im2col.c"],
3303 hdrs = [
3304 "src/xnnpack/common.h",
3305 "src/xnnpack/im2col.h",
3306 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003307 gcc_copts = xnnpack_gcc_std_copts(),
3308 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309)
3310
3311xnnpack_cc_library(
3312 name = "indirection",
3313 srcs = ["src/indirection.c"],
3314 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003315 gcc_copts = xnnpack_gcc_std_copts(),
3316 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 deps = [
3318 "@FP16",
3319 "@FXdiv",
3320 "@pthreadpool",
3321 ],
3322)
3323
3324xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003325 name = "indirection_test_mode",
3326 srcs = ["src/indirection.c"],
3327 hdrs = INTERNAL_HDRS,
3328 copts = [
3329 "-UNDEBUG",
3330 "-DXNN_TEST_MODE=1",
3331 ],
3332 gcc_copts = xnnpack_gcc_std_copts(),
3333 msvc_copts = xnnpack_msvc_std_copts(),
3334 deps = [
3335 "@FP16",
3336 "@FXdiv",
3337 "@pthreadpool",
3338 ],
3339)
3340
3341xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003342 name = "packing",
3343 srcs = ["src/packing.c"],
3344 hdrs = INTERNAL_HDRS,
3345 gcc_copts = xnnpack_gcc_std_copts(),
3346 msvc_copts = xnnpack_msvc_std_copts(),
3347 deps = [
3348 "@FP16",
3349 "@FXdiv",
3350 "@pthreadpool",
3351 ],
3352)
3353
3354xnnpack_cc_library(
3355 name = "packing_test_mode",
3356 srcs = ["src/packing.c"],
3357 hdrs = INTERNAL_HDRS,
3358 copts = [
3359 "-UNDEBUG",
3360 "-DXNN_TEST_MODE=1",
3361 ],
3362 gcc_copts = xnnpack_gcc_std_copts(),
3363 msvc_copts = xnnpack_msvc_std_copts(),
3364 deps = [
3365 "@FP16",
3366 "@FXdiv",
3367 "@pthreadpool",
3368 ],
3369)
3370
3371xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 name = "operator_run",
3373 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003374 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003375 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003376 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3377 "//conditions:default": [],
3378 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003379 gcc_copts = xnnpack_gcc_std_copts(),
3380 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003382 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 "@FP16",
3384 "@FXdiv",
3385 "@clog",
3386 "@pthreadpool",
3387 ],
3388)
3389
Chao Mei6ddfc602020-05-13 22:29:36 -07003390xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003391 name = "operator_run_test_mode",
3392 srcs = ["src/operator-run.c"],
3393 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3394 copts = LOGGING_COPTS + [
3395 "-UNDEBUG",
3396 "-DXNN_TEST_MODE=1",
3397 ] + select({
3398 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3399 "//conditions:default": [],
3400 }),
3401 gcc_copts = xnnpack_gcc_std_copts(),
3402 msvc_copts = xnnpack_msvc_std_copts(),
3403 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003404 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003405 "@FP16",
3406 "@FXdiv",
3407 "@clog",
3408 "@pthreadpool",
3409 ],
3410)
3411
3412xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003413 name = "memory_planner",
3414 srcs = ["src/memory-planner.c"],
3415 hdrs = INTERNAL_HDRS,
3416 defines = select({
3417 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3418 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3419 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3420 }),
3421 gcc_copts = xnnpack_gcc_std_copts(),
3422 msvc_copts = xnnpack_msvc_std_copts(),
3423 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003424 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003425 "@pthreadpool",
3426 ],
3427)
3428
Marat Dukhan33fcf782020-05-24 14:27:15 -07003429xnnpack_cc_library(
3430 name = "memory_planner_test_mode",
3431 srcs = ["src/memory-planner.c"],
3432 hdrs = INTERNAL_HDRS,
3433 copts = [
3434 "-UNDEBUG",
3435 "-DXNN_TEST_MODE=1",
3436 ],
3437 defines = select({
3438 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3439 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3440 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3441 }),
3442 gcc_copts = xnnpack_gcc_std_copts(),
3443 msvc_copts = xnnpack_msvc_std_copts(),
3444 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003445 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003446 "@pthreadpool",
3447 ],
3448)
3449
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450cc_library(
3451 name = "enable_assembly",
3452 defines = select({
3453 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3454 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003455 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003456 }),
3457)
3458
Marat Dukhan9de90e02020-06-18 16:04:12 -07003459cc_library(
3460 name = "enable_sparse",
3461 defines = select({
3462 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3463 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003464 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003465 }),
3466)
3467
Marat Dukhancf056b22019-10-07 10:26:29 -07003468xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003469 name = "operators",
3470 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003471 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003473 ],
3474 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003475 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476 "-Isrc",
3477 "-Iinclude",
3478 ] + select({
3479 ":debug_build": [],
3480 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003481 }) + select({
3482 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3483 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003485 gcc_copts = xnnpack_gcc_std_copts(),
3486 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003489 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003490 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491 "@FP16",
3492 "@FXdiv",
3493 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003494 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 ],
3496)
3497
Marat Dukhan10a38082020-04-17 03:58:35 -07003498xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003499 name = "operators_test_mode",
3500 srcs = OPERATOR_SRCS + [
3501 "src/memory.c",
3502 "src/operator-delete.c",
3503 ],
3504 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3505 copts = LOGGING_COPTS + [
3506 "-Isrc",
3507 "-Iinclude",
3508 "-UNDEBUG",
3509 "-DXNN_TEST_MODE=1",
3510 ] + select({
3511 ":debug_build": [],
3512 "//conditions:default": xnnpack_min_size_copts(),
3513 }) + select({
3514 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3515 "//conditions:default": [],
3516 }),
3517 gcc_copts = xnnpack_gcc_std_copts(),
3518 msvc_copts = xnnpack_msvc_std_copts(),
3519 deps = [
3520 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003521 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003522 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003523 "@FP16",
3524 "@FXdiv",
3525 "@clog",
3526 "@pthreadpool",
3527 ],
3528)
3529
3530xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003531 name = "XNNPACK",
3532 srcs = [
3533 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003534 "src/runtime.c",
3535 "src/subgraph.c",
3536 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003537 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003538 hdrs = ["include/xnnpack.h"],
3539 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003540 "-Isrc",
3541 "-Iinclude",
3542 ] + select({
3543 ":debug_build": [],
3544 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003545 }) + select({
3546 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3547 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003548 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003549 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003550 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003551 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003552 visibility = xnnpack_visibility(),
3553 deps = [
3554 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003555 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003556 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003557 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003558 ":operator_run",
3559 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003560 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003561 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003562 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003563 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003564 ] + select({
3565 ":emscripten": [],
3566 "//conditions:default": ["@cpuinfo"],
3567 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568)
3569
Marat Dukhan10a38082020-04-17 03:58:35 -07003570xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003571 name = "XNNPACK_test_mode",
3572 srcs = [
3573 "src/init.c",
3574 "src/runtime.c",
3575 "src/subgraph.c",
3576 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003577 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003578 hdrs = ["include/xnnpack.h"],
3579 copts = LOGGING_COPTS + [
3580 "-Isrc",
3581 "-Iinclude",
3582 "-UNDEBUG",
3583 "-DXNN_TEST_MODE=1",
3584 ] + select({
3585 ":debug_build": [],
3586 "//conditions:default": xnnpack_min_size_copts(),
3587 }) + select({
3588 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3589 "//conditions:default": [],
3590 }),
3591 gcc_copts = xnnpack_gcc_std_copts(),
3592 includes = ["include"],
3593 msvc_copts = xnnpack_msvc_std_copts(),
3594 visibility = xnnpack_visibility(),
3595 deps = [
3596 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003597 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003598 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003599 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003600 ":operator_run_test_mode",
3601 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003602 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003603 "@clog",
3604 "@FP16",
3605 "@pthreadpool",
3606 ] + select({
3607 ":emscripten": [],
3608 "//conditions:default": ["@cpuinfo"],
3609 }),
3610)
3611
3612xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003613 name = "xnnpack_f32",
3614 srcs = [
3615 "src/init.c",
3616 "src/runtime.c",
3617 "src/subgraph.c",
3618 "src/tensor.c",
3619 ] + SUBGRAPH_SRCS,
3620 hdrs = ["include/xnnpack.h"],
3621 copts = LOGGING_COPTS + [
3622 "-Isrc",
3623 "-Iinclude",
3624 ] + select({
3625 ":debug_build": [],
3626 "//conditions:default": xnnpack_min_size_copts(),
3627 }) + select({
3628 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3629 "//conditions:default": [],
3630 }),
3631 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003632 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003633 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003634 "XNN_NO_U8_OPERATORS",
3635 "XNN_NO_X8_OPERATORS",
3636 "XNN_NO_F16_OPERATORS",
3637 "XNN_NO_X16_OPERATORS",
3638 ],
3639 gcc_copts = xnnpack_gcc_std_copts(),
3640 includes = ["include"],
3641 msvc_copts = xnnpack_msvc_std_copts(),
3642 visibility = xnnpack_visibility(),
3643 deps = [
3644 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003645 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003646 ":logging_utils",
3647 ":memory_planner",
3648 ":operator_run",
3649 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003650 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003651 "@clog",
3652 "@FP16",
3653 "@pthreadpool",
3654 ] + select({
3655 ":emscripten": [],
3656 "//conditions:default": ["@cpuinfo"],
3657 }),
3658)
3659
3660xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003661 name = "xnnpack_operators_nhwc_f32",
3662 srcs = [
3663 "src/init.c",
3664 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003665 hdrs = ["include/xnnpack.h"],
3666 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003667 "-Isrc",
3668 "-Iinclude",
3669 ] + select({
3670 ":debug_build": [],
3671 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003672 }) + select({
3673 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3674 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003675 }),
3676 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003677 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003678 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003679 "XNN_NO_U8_OPERATORS",
3680 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003681 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003682 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003683 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003685 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686 visibility = xnnpack_visibility(),
3687 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003688 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003689 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690 ":operator_run",
3691 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003692 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003693 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003694 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003695 ] + select({
3696 ":emscripten": [],
3697 "//conditions:default": ["@cpuinfo"],
3698 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699)
3700
Marat Dukhancf056b22019-10-07 10:26:29 -07003701xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702 name = "bench_utils",
3703 srcs = ["bench/utils.cc"],
3704 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003705 deps = [
3706 "@com_google_benchmark//:benchmark",
3707 "@cpuinfo",
3708 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709)
3710
Frank Barchard7e955972019-10-11 10:34:25 -07003711######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712
3713xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003714 name = "qs8_gemm_bench",
3715 srcs = [
3716 "bench/gemm.h",
3717 "bench/qs8-gemm.cc",
3718 "src/xnnpack/AlignedAllocator.h",
3719 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3720 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3721)
3722
3723xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003724 name = "qs8_requantization_bench",
3725 srcs = [
3726 "bench/qs8-requantization.cc",
3727 "src/xnnpack/requantization-stubs.h",
3728 "src/xnnpack/AlignedAllocator.h",
3729 ] + MICROKERNEL_BENCHMARK_HDRS,
3730 deps = MICROKERNEL_BENCHMARK_DEPS,
3731)
3732
3733xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003734 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 srcs = [
3736 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003737 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738 "src/xnnpack/AlignedAllocator.h",
3739 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003740 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003741 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742)
3743
3744xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003745 name = "qu8_requantization_bench",
3746 srcs = [
3747 "bench/qu8-requantization.cc",
3748 "src/xnnpack/requantization-stubs.h",
3749 "src/xnnpack/AlignedAllocator.h",
3750 ] + MICROKERNEL_BENCHMARK_HDRS,
3751 deps = MICROKERNEL_BENCHMARK_DEPS,
3752)
3753
3754xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003755 name = "f16_igemm_bench",
3756 srcs = [
3757 "bench/f16-igemm.cc",
3758 "bench/conv.h",
3759 "bench/google/conv.h",
3760 "src/xnnpack/AlignedAllocator.h",
3761 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003762 deps = MICROKERNEL_BENCHMARK_DEPS + [
3763 ":indirection",
3764 ":packing",
3765 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003766)
3767
3768xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 name = "f16_gemm_bench",
3770 srcs = [
3771 "bench/f16-gemm.cc",
3772 "bench/gemm.h",
3773 "src/xnnpack/AlignedAllocator.h",
3774 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003775 deps = MICROKERNEL_BENCHMARK_DEPS + [
3776 ":packing",
3777 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778)
3779
3780xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003781 name = "f16_spmm_bench",
3782 srcs = [
3783 "bench/f16-spmm.cc",
3784 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003785 "src/xnnpack/AlignedAllocator.h",
3786 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003787 deps = MICROKERNEL_BENCHMARK_DEPS,
3788)
3789
3790xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791 name = "f32_igemm_bench",
3792 srcs = [
3793 "bench/f32-igemm.cc",
3794 "bench/conv.h",
3795 "src/xnnpack/AlignedAllocator.h",
3796 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003797 deps = MICROKERNEL_BENCHMARK_DEPS + [
3798 ":indirection",
3799 ":packing",
3800 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003801)
3802
3803xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003804 name = "f16_relu_bench",
3805 srcs = [
3806 "bench/f16-relu.cc",
3807 "src/xnnpack/AlignedAllocator.h",
3808 ] + MICROKERNEL_BENCHMARK_HDRS,
3809 deps = MICROKERNEL_BENCHMARK_DEPS,
3810)
3811
3812xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813 name = "f32_conv_hwc_bench",
3814 srcs = [
3815 "bench/f32-conv-hwc.cc",
3816 "bench/dconv.h",
3817 "src/xnnpack/AlignedAllocator.h",
3818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003819 deps = MICROKERNEL_BENCHMARK_DEPS + [
3820 ":packing",
3821 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822)
3823
3824xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003825 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003826 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003827 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003828 "bench/dconv.h",
3829 "src/xnnpack/AlignedAllocator.h",
3830 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003831 deps = MICROKERNEL_BENCHMARK_DEPS + [
3832 ":packing",
3833 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003834)
3835
3836xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003837 name = "f16_dwconv_bench",
3838 srcs = [
3839 "bench/f16-dwconv.cc",
3840 "bench/dwconv.h",
3841 "bench/google/dwconv.h",
3842 "src/xnnpack/AlignedAllocator.h",
3843 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003844 deps = MICROKERNEL_BENCHMARK_DEPS + [
3845 ":indirection",
3846 ":packing",
3847 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003848)
3849
3850xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003851 name = "f32_dwconv_bench",
3852 srcs = [
3853 "bench/f32-dwconv.cc",
3854 "bench/dwconv.h",
3855 "src/xnnpack/AlignedAllocator.h",
3856 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003857 deps = MICROKERNEL_BENCHMARK_DEPS + [
3858 ":indirection",
3859 ":packing",
3860 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003861)
3862
3863xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003864 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003866 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003867 "bench/dwconv.h",
3868 "src/xnnpack/AlignedAllocator.h",
3869 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003870 deps = MICROKERNEL_BENCHMARK_DEPS + [
3871 ":indirection",
3872 ":packing",
3873 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874)
3875
3876xnnpack_benchmark(
3877 name = "f32_gemm_bench",
3878 srcs = [
3879 "bench/f32-gemm.cc",
3880 "bench/gemm.h",
3881 "src/xnnpack/AlignedAllocator.h",
3882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003883 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003884 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003885)
3886
3887xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003888 name = "f32_hswish_bench",
3889 srcs = [
3890 "bench/f32-hswish.cc",
3891 "src/xnnpack/AlignedAllocator.h",
3892 ] + MICROKERNEL_BENCHMARK_HDRS,
3893 deps = MICROKERNEL_BENCHMARK_DEPS,
3894)
3895
3896xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003897 name = "f32_raddexpminusmax_bench",
3898 srcs = [
3899 "bench/f32-raddexpminusmax.cc",
3900 "src/xnnpack/AlignedAllocator.h",
3901 ] + MICROKERNEL_BENCHMARK_HDRS,
3902 deps = MICROKERNEL_BENCHMARK_DEPS,
3903)
3904
3905xnnpack_benchmark(
3906 name = "f32_raddextexp_bench",
3907 srcs = [
3908 "bench/f32-raddextexp.cc",
3909 "src/xnnpack/AlignedAllocator.h",
3910 ] + MICROKERNEL_BENCHMARK_HDRS,
3911 deps = MICROKERNEL_BENCHMARK_DEPS,
3912)
3913
3914xnnpack_benchmark(
3915 name = "f32_raddstoreexpminusmax_bench",
3916 srcs = [
3917 "bench/f32-raddstoreexpminusmax.cc",
3918 "src/xnnpack/AlignedAllocator.h",
3919 ] + MICROKERNEL_BENCHMARK_HDRS,
3920 deps = MICROKERNEL_BENCHMARK_DEPS,
3921)
3922
3923xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003924 name = "f32_relu_bench",
3925 srcs = [
3926 "bench/f32-relu.cc",
3927 "src/xnnpack/AlignedAllocator.h",
3928 ] + MICROKERNEL_BENCHMARK_HDRS,
3929 deps = MICROKERNEL_BENCHMARK_DEPS,
3930)
3931
3932xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003933 name = "f32_rmax_bench",
3934 srcs = [
3935 "bench/f32-rmax.cc",
3936 "src/xnnpack/AlignedAllocator.h",
3937 ] + MICROKERNEL_BENCHMARK_HDRS,
3938 deps = MICROKERNEL_BENCHMARK_DEPS,
3939)
3940
3941xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003942 name = "f32_sigmoid_bench",
3943 srcs = [
3944 "bench/f32-sigmoid.cc",
3945 "src/xnnpack/AlignedAllocator.h",
3946 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003947 deps = MICROKERNEL_BENCHMARK_DEPS,
3948)
3949
3950xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 name = "f32_spmm_bench",
3952 srcs = [
3953 "bench/f32-spmm.cc",
3954 "bench/gemm.h",
3955 "src/xnnpack/AlignedAllocator.h",
3956 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003957 deps = MICROKERNEL_BENCHMARK_DEPS,
3958)
3959
3960xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003961 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003962 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003963 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003964 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003965 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003966 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003967)
3968
3969xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003970 name = "f32_vscaleexpminusmax_bench",
3971 srcs = [
3972 "bench/f32-vscaleexpminusmax.cc",
3973 "src/xnnpack/AlignedAllocator.h",
3974 ] + MICROKERNEL_BENCHMARK_HDRS,
3975 deps = MICROKERNEL_BENCHMARK_DEPS,
3976)
3977
3978xnnpack_benchmark(
3979 name = "f32_vscaleextexp_bench",
3980 srcs = [
3981 "bench/f32-vscaleextexp.cc",
3982 "src/xnnpack/AlignedAllocator.h",
3983 ] + MICROKERNEL_BENCHMARK_HDRS,
3984 deps = MICROKERNEL_BENCHMARK_DEPS,
3985)
3986
3987xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003988 name = "f32_vsqrt_bench",
3989 srcs = [
3990 "bench/f32-vsqrt.cc",
3991 "src/xnnpack/AlignedAllocator.h",
3992 ] + MICROKERNEL_BENCHMARK_HDRS,
3993 deps = MICROKERNEL_BENCHMARK_DEPS,
3994)
3995
3996xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997 name = "f32_im2col_gemm_bench",
3998 srcs = [
3999 "bench/f32-im2col-gemm.cc",
4000 "bench/conv.h",
4001 "src/xnnpack/AlignedAllocator.h",
4002 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004003 deps = MICROKERNEL_BENCHMARK_DEPS + [
4004 ":im2col",
4005 ":packing",
4006 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004007)
4008
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004009xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004010 name = "rounding_bench",
4011 srcs = [
4012 "bench/rounding.cc",
4013 "src/xnnpack/math-stubs.h",
4014 "src/xnnpack/AlignedAllocator.h",
4015 ] + MICROKERNEL_BENCHMARK_HDRS,
4016 deps = MICROKERNEL_BENCHMARK_DEPS,
4017)
4018
Marat Dukhan08c4a432019-10-03 09:29:21 -07004019########################### Benchmarks for operators ###########################
4020
4021xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022 name = "average_pooling_bench",
4023 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004024 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004025 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004026 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004027)
4028
4029xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004030 name = "bankers_rounding_bench",
4031 srcs = ["bench/bankers-rounding.cc"],
4032 copts = xnnpack_optional_tflite_copts(),
4033 tags = ["nowin32"],
4034 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4035)
4036
4037xnnpack_benchmark(
4038 name = "ceiling_bench",
4039 srcs = ["bench/ceiling.cc"],
4040 copts = xnnpack_optional_tflite_copts(),
4041 tags = ["nowin32"],
4042 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4043)
4044
4045xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004046 name = "channel_shuffle_bench",
4047 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004048 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004049)
4050
4051xnnpack_benchmark(
4052 name = "convolution_bench",
4053 srcs = ["bench/convolution.cc"],
4054 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004055 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004056 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057)
4058
4059xnnpack_benchmark(
4060 name = "deconvolution_bench",
4061 srcs = ["bench/deconvolution.cc"],
4062 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004063 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004064 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004065)
4066
4067xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004068 name = "floor_bench",
4069 srcs = ["bench/floor.cc"],
4070 copts = xnnpack_optional_tflite_copts(),
4071 tags = ["nowin32"],
4072 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4073)
4074
4075xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004076 name = "global_average_pooling_bench",
4077 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004078 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004079)
4080
4081xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004082 name = "hardswish_bench",
4083 srcs = ["bench/hardswish.cc"],
4084 copts = xnnpack_optional_tflite_copts(),
4085 tags = ["nowin32"],
4086 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4087)
4088
4089xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 name = "max_pooling_bench",
4091 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004092 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093)
4094
4095xnnpack_benchmark(
4096 name = "sigmoid_bench",
4097 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004098 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004099 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004100 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004101)
4102
4103xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004104 name = "prelu_bench",
4105 srcs = ["bench/prelu.cc"],
4106 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004107 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004108 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004109)
4110
4111xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004112 name = "softmax_bench",
4113 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004114 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004115 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004116 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004117)
4118
Marat Dukhan87727142020-06-24 15:24:10 -07004119xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004120 name = "square_root_bench",
4121 srcs = ["bench/square-root.cc"],
4122 copts = xnnpack_optional_tflite_copts(),
4123 tags = ["nowin32"],
4124 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4125)
4126
4127xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004128 name = "truncation_bench",
4129 srcs = ["bench/truncation.cc"],
4130 deps = OPERATOR_BENCHMARK_DEPS,
4131)
4132
Marat Dukhanc068bb62019-10-04 13:24:39 -07004133############################# End-to-end benchmarks ############################
4134
4135cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004136 name = "fp32_mobilenet_v1",
4137 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004138 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004139 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004140 linkstatic = True,
4141 deps = [
4142 ":XNNPACK",
4143 "@pthreadpool",
4144 ],
4145)
4146
4147cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004148 name = "fp16_mobilenet_v1",
4149 srcs = ["models/fp16-mobilenet-v1.cc"],
4150 hdrs = ["models/models.h"],
4151 copts = xnnpack_std_cxxopts(),
4152 linkstatic = True,
4153 deps = [
4154 ":XNNPACK",
4155 "@FP16",
4156 "@pthreadpool",
4157 ],
4158)
4159
4160cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004161 name = "qs8_mobilenet_v1",
4162 srcs = ["models/qs8-mobilenet-v1.cc"],
4163 hdrs = ["models/models.h"],
4164 copts = xnnpack_std_cxxopts(),
4165 linkstatic = True,
4166 deps = [
4167 ":XNNPACK",
4168 "@pthreadpool",
4169 ],
4170)
4171
4172cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004173 name = "fp32_mobilenet_v2",
4174 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004175 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004176 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004177 linkstatic = True,
4178 deps = [
4179 ":XNNPACK",
4180 "@pthreadpool",
4181 ],
4182)
4183
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004184cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004185 name = "fp16_mobilenet_v2",
4186 srcs = ["models/fp16-mobilenet-v2.cc"],
4187 hdrs = ["models/models.h"],
4188 copts = xnnpack_std_cxxopts(),
4189 linkstatic = True,
4190 deps = [
4191 ":XNNPACK",
4192 "@FP16",
4193 "@pthreadpool",
4194 ],
4195)
4196
4197cc_library(
4198 name = "fp32_mobilenet_v3_large",
4199 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004200 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004201 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004202 linkstatic = True,
4203 deps = [
4204 ":XNNPACK",
4205 "@pthreadpool",
4206 ],
4207)
4208
4209cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004210 name = "fp32_mobilenet_v3_small",
4211 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004212 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004213 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004214 linkstatic = True,
4215 deps = [
4216 ":XNNPACK",
4217 "@pthreadpool",
4218 ],
4219)
4220
Marat Dukhanc068bb62019-10-04 13:24:39 -07004221xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004222 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004223 srcs = [
4224 "bench/f32-dwconv-e2e.cc",
4225 "bench/end2end.h",
4226 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004227 deps = MICROKERNEL_BENCHMARK_DEPS + [
4228 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004229 ":fp32_mobilenet_v1",
4230 ":fp32_mobilenet_v2",
4231 ":fp32_mobilenet_v3_large",
4232 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004233 ],
4234)
4235
4236xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004237 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004238 srcs = [
4239 "bench/f32-gemm-e2e.cc",
4240 "bench/end2end.h",
4241 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004242 deps = MICROKERNEL_BENCHMARK_DEPS + [
4243 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004244 ":fp32_mobilenet_v1",
4245 ":fp32_mobilenet_v2",
4246 ":fp32_mobilenet_v3_large",
4247 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004248 ],
4249)
4250
4251xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004252 name = "end2end_bench",
4253 srcs = ["bench/end2end.cc"],
4254 deps = [
4255 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004256 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004257 ":fp16_mobilenet_v1",
4258 ":fp16_mobilenet_v2",
4259 ":fp32_mobilenet_v1",
4260 ":fp32_mobilenet_v2",
4261 ":fp32_mobilenet_v3_large",
4262 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004263 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004264 "@pthreadpool",
4265 ],
4266)
4267
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004268#################### Accuracy evaluation for math functions ####################
4269
4270xnnpack_benchmark(
4271 name = "f32_exp_eval",
4272 srcs = [
4273 "eval/f32-exp.cc",
4274 "src/xnnpack/AlignedAllocator.h",
4275 ] + ACCURACY_EVAL_HDRS,
4276 deps = ACCURACY_EVAL_DEPS,
4277)
4278
Marat Dukhan515c9772019-10-17 18:07:57 -07004279xnnpack_benchmark(
4280 name = "f32_expminus_eval",
4281 srcs = [
4282 "eval/f32-expminus.cc",
4283 "src/xnnpack/AlignedAllocator.h",
4284 ] + ACCURACY_EVAL_HDRS,
4285 deps = ACCURACY_EVAL_DEPS,
4286)
4287
Marat Dukhan98ba4412019-10-23 02:14:28 -07004288xnnpack_benchmark(
4289 name = "f32_extexp_eval",
4290 srcs = [
4291 "eval/f32-extexp.cc",
4292 "src/xnnpack/AlignedAllocator.h",
4293 ] + ACCURACY_EVAL_HDRS,
4294 deps = ACCURACY_EVAL_DEPS,
4295)
4296
Marat Dukhan8853b822020-05-07 12:19:01 -07004297xnnpack_unit_test(
4298 name = "f32_roundne_eval",
4299 srcs = [
4300 "eval/f32-roundne.cc",
4301 "src/xnnpack/AlignedAllocator.h",
4302 "src/xnnpack/math-stubs.h",
4303 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004304 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004305 deps = MICROKERNEL_TEST_DEPS,
4306)
4307
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004308xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004309 name = "f32_roundd_eval",
4310 srcs = [
4311 "eval/f32-roundd.cc",
4312 "src/xnnpack/AlignedAllocator.h",
4313 "src/xnnpack/math-stubs.h",
4314 ] + MICROKERNEL_TEST_HDRS,
4315 automatic = False,
4316 deps = MICROKERNEL_TEST_DEPS,
4317)
4318
4319xnnpack_unit_test(
4320 name = "f32_roundu_eval",
4321 srcs = [
4322 "eval/f32-roundu.cc",
4323 "src/xnnpack/AlignedAllocator.h",
4324 "src/xnnpack/math-stubs.h",
4325 ] + MICROKERNEL_TEST_HDRS,
4326 automatic = False,
4327 deps = MICROKERNEL_TEST_DEPS,
4328)
4329
4330xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004331 name = "f32_roundz_eval",
4332 srcs = [
4333 "eval/f32-roundz.cc",
4334 "src/xnnpack/AlignedAllocator.h",
4335 "src/xnnpack/math-stubs.h",
4336 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004337 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004338 deps = MICROKERNEL_TEST_DEPS,
4339)
4340
Marat Dukhan346a9e52019-11-15 09:06:30 -08004341xnnpack_benchmark(
4342 name = "f32_sigmoid_eval",
4343 srcs = [
4344 "eval/f32-sigmoid.cc",
4345 "src/xnnpack/AlignedAllocator.h",
4346 ] + ACCURACY_EVAL_HDRS,
4347 deps = ACCURACY_EVAL_DEPS,
4348)
4349
Marat Dukhan84000762020-06-29 18:38:43 -07004350xnnpack_benchmark(
4351 name = "f32_sqrt_eval",
4352 srcs = [
4353 "eval/f32-sqrt.cc",
4354 "src/xnnpack/AlignedAllocator.h",
4355 ] + ACCURACY_EVAL_HDRS,
4356 deps = ACCURACY_EVAL_DEPS + [
4357 ":bench_utils",
4358 ],
4359)
4360
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361######################### Unit tests for micro-kernels #########################
4362
4363xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004364 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004366 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 "test/gemm-microkernel-tester.h",
4368 "src/xnnpack/AlignedAllocator.h",
4369 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004370 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371)
4372
4373xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004374 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004375 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004376 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004377 "test/spmm-microkernel-tester.h",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + MICROKERNEL_TEST_HDRS,
4380 deps = MICROKERNEL_TEST_DEPS,
4381)
4382
4383xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_argmaxpool_test",
4385 srcs = [
4386 "test/f32-argmaxpool.cc",
4387 "test/argmaxpool-microkernel-tester.h",
4388 "src/xnnpack/AlignedAllocator.h",
4389 ] + MICROKERNEL_TEST_HDRS,
4390 deps = MICROKERNEL_TEST_DEPS,
4391)
4392
4393xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004394 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004396 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004397 "test/avgpool-microkernel-tester.h",
4398 "src/xnnpack/AlignedAllocator.h",
4399 ] + MICROKERNEL_TEST_HDRS,
4400 deps = MICROKERNEL_TEST_DEPS,
4401)
4402
4403xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004404 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004405 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004406 "test/f32-ibilinear.cc",
4407 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004408 "src/xnnpack/AlignedAllocator.h",
4409 ] + MICROKERNEL_TEST_HDRS,
4410 deps = MICROKERNEL_TEST_DEPS,
4411)
4412
4413xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004414 name = "f16_clamp_test",
4415 srcs = [
4416 "test/f16-clamp.cc",
4417 "test/clamp-microkernel-tester.h",
4418 ] + MICROKERNEL_TEST_HDRS,
4419 deps = MICROKERNEL_TEST_DEPS,
4420)
4421
4422xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423 name = "f32_clamp_test",
4424 srcs = [
4425 "test/f32-clamp.cc",
4426 "test/clamp-microkernel-tester.h",
4427 ] + MICROKERNEL_TEST_HDRS,
4428 deps = MICROKERNEL_TEST_DEPS,
4429)
4430
4431xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004432 name = "f32_igemm_test",
4433 srcs = [
4434 "test/f32-igemm.cc",
4435 "test/gemm-microkernel-tester.h",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004438 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004439)
4440
4441xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004442 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004444 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445 "test/gemm-microkernel-tester.h",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004448 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004449)
4450
4451xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004452 name = "f32_igemm_minmax_test",
4453 srcs = [
4454 "test/f32-igemm-minmax.cc",
4455 "test/gemm-microkernel-tester.h",
4456 "src/xnnpack/AlignedAllocator.h",
4457 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004458 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004459)
4460
4461xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004462 name = "f16_igemm_minmax_test",
4463 srcs = [
4464 "test/f16-igemm-minmax.cc",
4465 "test/gemm-microkernel-tester.h",
4466 "src/xnnpack/AlignedAllocator.h",
4467 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004468 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004469)
4470
4471xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472 name = "f32_conv_hwc_test",
4473 srcs = [
4474 "test/f32-conv-hwc.cc",
4475 "test/conv-hwc-microkernel-tester.h",
4476 "src/xnnpack/AlignedAllocator.h",
4477 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004478 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479)
4480
4481xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004482 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004484 "test/f32-conv-hwc2chw.cc",
4485 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 "src/xnnpack/AlignedAllocator.h",
4487 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004488 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004489)
4490
4491xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004492 name = "f32_dwconv_test",
4493 srcs = [
4494 "test/f32-dwconv.cc",
4495 "test/dwconv-microkernel-tester.h",
4496 "src/xnnpack/AlignedAllocator.h",
4497 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004498 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004499)
4500
4501xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004502 name = "f16_dwconv_minmax_test",
4503 srcs = [
4504 "test/f16-dwconv-minmax.cc",
4505 "test/dwconv-microkernel-tester.h",
4506 "src/xnnpack/AlignedAllocator.h",
4507 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004508 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004509)
4510
4511xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004512 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004513 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004514 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004515 "test/dwconv-microkernel-tester.h",
4516 "src/xnnpack/AlignedAllocator.h",
4517 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004518 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004519)
4520
4521xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004522 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004523 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004524 "test/f32-dwconv-chw.cc",
4525 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 "src/xnnpack/AlignedAllocator.h",
4527 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004528 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529)
4530
4531xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004532 name = "f16_gavgpool_minmax_test",
4533 srcs = [
4534 "test/f16-gavgpool-minmax.cc",
4535 "test/gavgpool-microkernel-tester.h",
4536 "src/xnnpack/AlignedAllocator.h",
4537 ] + MICROKERNEL_TEST_HDRS,
4538 deps = MICROKERNEL_TEST_DEPS,
4539)
4540
4541xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004542 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004544 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545 "test/gavgpool-microkernel-tester.h",
4546 "src/xnnpack/AlignedAllocator.h",
4547 ] + MICROKERNEL_TEST_HDRS,
4548 deps = MICROKERNEL_TEST_DEPS,
4549)
4550
4551xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004552 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004554 "test/f32-gavgpool-cw.cc",
4555 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 "src/xnnpack/AlignedAllocator.h",
4557 ] + MICROKERNEL_TEST_HDRS,
4558 deps = MICROKERNEL_TEST_DEPS,
4559)
4560
4561xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004562 name = "f32_gemm_test",
4563 srcs = [
4564 "test/f32-gemm.cc",
4565 "test/gemm-microkernel-tester.h",
4566 "src/xnnpack/AlignedAllocator.h",
4567 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004568 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004569)
4570
4571xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004572 name = "f32_gemm_relu_test",
4573 srcs = [
4574 "test/f32-gemm-relu.cc",
4575 "test/gemm-microkernel-tester.h",
4576 "src/xnnpack/AlignedAllocator.h",
4577 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004578 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004579)
4580
4581xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004582 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004583 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004584 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585 "test/gemm-microkernel-tester.h",
4586 "src/xnnpack/AlignedAllocator.h",
4587 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004588 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004589)
4590
4591xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004592 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004594 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004595 "test/gemm-microkernel-tester.h",
4596 "src/xnnpack/AlignedAllocator.h",
4597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004598 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004599)
4600
4601xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004602 name = "f16_hswish_test",
4603 srcs = [
4604 "test/f16-hswish.cc",
4605 "test/hswish-microkernel-tester.h",
4606 ] + MICROKERNEL_TEST_HDRS,
4607 deps = MICROKERNEL_TEST_DEPS,
4608)
4609
4610xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004611 name = "f32_hswish_test",
4612 srcs = [
4613 "test/f32-hswish.cc",
4614 "test/hswish-microkernel-tester.h",
4615 ] + MICROKERNEL_TEST_HDRS,
4616 deps = MICROKERNEL_TEST_DEPS,
4617)
4618
4619xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004620 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004622 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623 "test/maxpool-microkernel-tester.h",
4624 ] + MICROKERNEL_TEST_HDRS,
4625 deps = MICROKERNEL_TEST_DEPS,
4626)
4627
4628xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004629 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004630 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004631 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 "test/avgpool-microkernel-tester.h",
4633 "src/xnnpack/AlignedAllocator.h",
4634 ] + MICROKERNEL_TEST_HDRS,
4635 deps = MICROKERNEL_TEST_DEPS,
4636)
4637
4638xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004639 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004640 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004641 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004642 "test/gemm-microkernel-tester.h",
4643 "src/xnnpack/AlignedAllocator.h",
4644 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004645 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646)
4647
4648xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004649 name = "f16_prelu_test",
4650 srcs = [
4651 "test/f16-prelu.cc",
4652 "test/prelu-microkernel-tester.h",
4653 "src/xnnpack/AlignedAllocator.h",
4654 ] + MICROKERNEL_TEST_HDRS,
4655 deps = MICROKERNEL_TEST_DEPS,
4656)
4657
4658xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659 name = "f32_prelu_test",
4660 srcs = [
4661 "test/f32-prelu.cc",
4662 "test/prelu-microkernel-tester.h",
4663 "src/xnnpack/AlignedAllocator.h",
4664 ] + MICROKERNEL_TEST_HDRS,
4665 deps = MICROKERNEL_TEST_DEPS,
4666)
4667
4668xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004669 name = "f32_raddexpminusmax_test",
4670 srcs = [
4671 "test/f32-raddexpminusmax.cc",
4672 "test/raddexpminusmax-microkernel-tester.h",
4673 ] + MICROKERNEL_TEST_HDRS,
4674 deps = MICROKERNEL_TEST_DEPS,
4675)
4676
4677xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004678 name = "f32_raddextexp_test",
4679 srcs = [
4680 "test/f32-raddextexp.cc",
4681 "test/raddextexp-microkernel-tester.h",
4682 ] + MICROKERNEL_TEST_HDRS,
4683 deps = MICROKERNEL_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004687 name = "f32_raddstoreexpminusmax_test",
4688 srcs = [
4689 "test/f32-raddstoreexpminusmax.cc",
4690 "test/raddstoreexpminusmax-microkernel-tester.h",
4691 ] + MICROKERNEL_TEST_HDRS,
4692 deps = MICROKERNEL_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004696 name = "f32_relu_test",
4697 srcs = [
4698 "test/f32-relu.cc",
4699 "test/vunary-microkernel-tester.h",
4700 ] + MICROKERNEL_TEST_HDRS,
4701 deps = MICROKERNEL_TEST_DEPS,
4702)
4703
4704xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004705 name = "f32_rmax_test",
4706 srcs = [
4707 "test/f32-rmax.cc",
4708 "test/rmax-microkernel-tester.h",
4709 ] + MICROKERNEL_TEST_HDRS,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004714 name = "f32_sigmoid_test",
4715 srcs = [
4716 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004717 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004718 ] + MICROKERNEL_TEST_HDRS,
4719 deps = MICROKERNEL_TEST_DEPS,
4720)
4721
4722xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004723 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004724 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004725 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004726 "test/spmm-microkernel-tester.h",
4727 "src/xnnpack/AlignedAllocator.h",
4728 ] + MICROKERNEL_TEST_HDRS,
4729 deps = MICROKERNEL_TEST_DEPS,
4730)
4731
4732xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004733 name = "f16_vadd_minmax_test",
4734 srcs = [
4735 "test/f16-vadd-minmax.cc",
4736 "test/vbinary-microkernel-tester.h",
4737 ] + MICROKERNEL_TEST_HDRS,
4738 deps = MICROKERNEL_TEST_DEPS,
4739)
4740
4741xnnpack_unit_test(
4742 name = "f16_vaddc_minmax_test",
4743 srcs = [
4744 "test/f16-vaddc-minmax.cc",
4745 "test/vbinaryc-microkernel-tester.h",
4746 ] + MICROKERNEL_TEST_HDRS,
4747 deps = MICROKERNEL_TEST_DEPS,
4748)
4749
4750xnnpack_unit_test(
4751 name = "f16_vdiv_minmax_test",
4752 srcs = [
4753 "test/f16-vdiv-minmax.cc",
4754 "test/vbinary-microkernel-tester.h",
4755 ] + MICROKERNEL_TEST_HDRS,
4756 deps = MICROKERNEL_TEST_DEPS,
4757)
4758
4759xnnpack_unit_test(
4760 name = "f16_vdivc_minmax_test",
4761 srcs = [
4762 "test/f16-vdivc-minmax.cc",
4763 "test/vbinaryc-microkernel-tester.h",
4764 ] + MICROKERNEL_TEST_HDRS,
4765 deps = MICROKERNEL_TEST_DEPS,
4766)
4767
4768xnnpack_unit_test(
4769 name = "f16_vrdivc_minmax_test",
4770 srcs = [
4771 "test/f16-vrdivc-minmax.cc",
4772 "test/vbinaryc-microkernel-tester.h",
4773 ] + MICROKERNEL_TEST_HDRS,
4774 deps = MICROKERNEL_TEST_DEPS,
4775)
4776
4777xnnpack_unit_test(
4778 name = "f16_vmax_test",
4779 srcs = [
4780 "test/f16-vmax.cc",
4781 "test/vbinary-microkernel-tester.h",
4782 ] + MICROKERNEL_TEST_HDRS,
4783 deps = MICROKERNEL_TEST_DEPS,
4784)
4785
4786xnnpack_unit_test(
4787 name = "f16_vmaxc_test",
4788 srcs = [
4789 "test/f16-vmaxc.cc",
4790 "test/vbinaryc-microkernel-tester.h",
4791 ] + MICROKERNEL_TEST_HDRS,
4792 deps = MICROKERNEL_TEST_DEPS,
4793)
4794
4795xnnpack_unit_test(
4796 name = "f16_vmin_test",
4797 srcs = [
4798 "test/f16-vmin.cc",
4799 "test/vbinary-microkernel-tester.h",
4800 ] + MICROKERNEL_TEST_HDRS,
4801 deps = MICROKERNEL_TEST_DEPS,
4802)
4803
4804xnnpack_unit_test(
4805 name = "f16_vminc_test",
4806 srcs = [
4807 "test/f16-vminc.cc",
4808 "test/vbinaryc-microkernel-tester.h",
4809 ] + MICROKERNEL_TEST_HDRS,
4810 deps = MICROKERNEL_TEST_DEPS,
4811)
4812
4813xnnpack_unit_test(
4814 name = "f16_vmul_minmax_test",
4815 srcs = [
4816 "test/f16-vmul-minmax.cc",
4817 "test/vbinary-microkernel-tester.h",
4818 ] + MICROKERNEL_TEST_HDRS,
4819 deps = MICROKERNEL_TEST_DEPS,
4820)
4821
4822xnnpack_unit_test(
4823 name = "f16_vmulc_minmax_test",
4824 srcs = [
4825 "test/f16-vmulc-minmax.cc",
4826 "test/vbinaryc-microkernel-tester.h",
4827 ] + MICROKERNEL_TEST_HDRS,
4828 deps = MICROKERNEL_TEST_DEPS,
4829)
4830
4831xnnpack_unit_test(
4832 name = "f16_vsub_minmax_test",
4833 srcs = [
4834 "test/f16-vsub-minmax.cc",
4835 "test/vbinary-microkernel-tester.h",
4836 ] + MICROKERNEL_TEST_HDRS,
4837 deps = MICROKERNEL_TEST_DEPS,
4838)
4839
4840xnnpack_unit_test(
4841 name = "f16_vsubc_minmax_test",
4842 srcs = [
4843 "test/f16-vsubc-minmax.cc",
4844 "test/vbinaryc-microkernel-tester.h",
4845 ] + MICROKERNEL_TEST_HDRS,
4846 deps = MICROKERNEL_TEST_DEPS,
4847)
4848
4849xnnpack_unit_test(
4850 name = "f16_vrsubc_minmax_test",
4851 srcs = [
4852 "test/f16-vrsubc-minmax.cc",
4853 "test/vbinaryc-microkernel-tester.h",
4854 ] + MICROKERNEL_TEST_HDRS,
4855 deps = MICROKERNEL_TEST_DEPS,
4856)
4857
4858xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004859 name = "f32_vabs_test",
4860 srcs = [
4861 "test/f32-vabs.cc",
4862 "test/vunary-microkernel-tester.h",
4863 ] + MICROKERNEL_TEST_HDRS,
4864 deps = MICROKERNEL_TEST_DEPS,
4865)
4866
4867xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004868 name = "f32_vadd_test",
4869 srcs = [
4870 "test/f32-vadd.cc",
4871 "test/vbinary-microkernel-tester.h",
4872 ] + MICROKERNEL_TEST_HDRS,
4873 deps = MICROKERNEL_TEST_DEPS,
4874)
4875
4876xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004877 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004879 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004880 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004881 ] + MICROKERNEL_TEST_HDRS,
4882 deps = MICROKERNEL_TEST_DEPS,
4883)
4884
4885xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004886 name = "f32_vaddc_test",
4887 srcs = [
4888 "test/f32-vaddc.cc",
4889 "test/vbinaryc-microkernel-tester.h",
4890 ] + MICROKERNEL_TEST_HDRS,
4891 deps = MICROKERNEL_TEST_DEPS,
4892)
4893
4894xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004895 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004896 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004897 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004898 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 ] + MICROKERNEL_TEST_HDRS,
4900 deps = MICROKERNEL_TEST_DEPS,
4901)
4902
4903xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004904 name = "f32_vdiv_test",
4905 srcs = [
4906 "test/f32-vdiv.cc",
4907 "test/vbinary-microkernel-tester.h",
4908 ] + MICROKERNEL_TEST_HDRS,
4909 deps = MICROKERNEL_TEST_DEPS,
4910)
4911
4912xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004913 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004914 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004915 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004916 "test/vbinary-microkernel-tester.h",
4917 ] + MICROKERNEL_TEST_HDRS,
4918 deps = MICROKERNEL_TEST_DEPS,
4919)
4920
4921xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004922 name = "f32_vdivc_test",
4923 srcs = [
4924 "test/f32-vdivc.cc",
4925 "test/vbinaryc-microkernel-tester.h",
4926 ] + MICROKERNEL_TEST_HDRS,
4927 deps = MICROKERNEL_TEST_DEPS,
4928)
4929
4930xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004931 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004932 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004933 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004934 "test/vbinaryc-microkernel-tester.h",
4935 ] + MICROKERNEL_TEST_HDRS,
4936 deps = MICROKERNEL_TEST_DEPS,
4937)
4938
4939xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004940 name = "f32_vrdivc_test",
4941 srcs = [
4942 "test/f32-vrdivc.cc",
4943 "test/vbinaryc-microkernel-tester.h",
4944 ] + MICROKERNEL_TEST_HDRS,
4945 deps = MICROKERNEL_TEST_DEPS,
4946)
4947
4948xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004949 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004950 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004951 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004952 "test/vbinaryc-microkernel-tester.h",
4953 ] + MICROKERNEL_TEST_HDRS,
4954 deps = MICROKERNEL_TEST_DEPS,
4955)
4956
4957xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004958 name = "f32_vmax_test",
4959 srcs = [
4960 "test/f32-vmax.cc",
4961 "test/vbinary-microkernel-tester.h",
4962 ] + MICROKERNEL_TEST_HDRS,
4963 deps = MICROKERNEL_TEST_DEPS,
4964)
4965
4966xnnpack_unit_test(
4967 name = "f32_vmaxc_test",
4968 srcs = [
4969 "test/f32-vmaxc.cc",
4970 "test/vbinaryc-microkernel-tester.h",
4971 ] + MICROKERNEL_TEST_HDRS,
4972 deps = MICROKERNEL_TEST_DEPS,
4973)
4974
4975xnnpack_unit_test(
4976 name = "f32_vmin_test",
4977 srcs = [
4978 "test/f32-vmin.cc",
4979 "test/vbinary-microkernel-tester.h",
4980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
4985 name = "f32_vminc_test",
4986 srcs = [
4987 "test/f32-vminc.cc",
4988 "test/vbinaryc-microkernel-tester.h",
4989 ] + MICROKERNEL_TEST_HDRS,
4990 deps = MICROKERNEL_TEST_DEPS,
4991)
4992
4993xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004994 name = "f32_vmul_test",
4995 srcs = [
4996 "test/f32-vmul.cc",
4997 "test/vbinary-microkernel-tester.h",
4998 ] + MICROKERNEL_TEST_HDRS,
4999 deps = MICROKERNEL_TEST_DEPS,
5000)
5001
5002xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005003 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005004 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005005 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005006 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005007 ] + MICROKERNEL_TEST_HDRS,
5008 deps = MICROKERNEL_TEST_DEPS,
5009)
5010
5011xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005012 name = "f32_vmulc_test",
5013 srcs = [
5014 "test/f32-vmulc.cc",
5015 "test/vbinaryc-microkernel-tester.h",
5016 ] + MICROKERNEL_TEST_HDRS,
5017 deps = MICROKERNEL_TEST_DEPS,
5018)
5019
5020xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005021 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005022 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005023 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005024 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 ] + MICROKERNEL_TEST_HDRS,
5026 deps = MICROKERNEL_TEST_DEPS,
5027)
5028
5029xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005030 name = "f16_vmulcaddc_minmax_test",
5031 srcs = [
5032 "test/f16-vmulcaddc-minmax.cc",
5033 "test/vmulcaddc-microkernel-tester.h",
5034 "src/xnnpack/AlignedAllocator.h",
5035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005036 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005037)
5038
5039xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005040 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005042 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005043 "test/vmulcaddc-microkernel-tester.h",
5044 "src/xnnpack/AlignedAllocator.h",
5045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005046 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047)
5048
5049xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005050 name = "f32_vlrelu_test",
5051 srcs = [
5052 "test/f32-vlrelu.cc",
5053 "test/vunary-microkernel-tester.h",
5054 ] + MICROKERNEL_TEST_HDRS,
5055 deps = MICROKERNEL_TEST_DEPS,
5056)
5057
5058xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005059 name = "f32_vneg_test",
5060 srcs = [
5061 "test/f32-vneg.cc",
5062 "test/vunary-microkernel-tester.h",
5063 ] + MICROKERNEL_TEST_HDRS,
5064 deps = MICROKERNEL_TEST_DEPS,
5065)
5066
5067xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005068 name = "f32_vrndne_test",
5069 srcs = [
5070 "test/f32-vrndne.cc",
5071 "test/vunary-microkernel-tester.h",
5072 ] + MICROKERNEL_TEST_HDRS,
5073 deps = MICROKERNEL_TEST_DEPS,
5074)
5075
5076xnnpack_unit_test(
5077 name = "f32_vrndz_test",
5078 srcs = [
5079 "test/f32-vrndz.cc",
5080 "test/vunary-microkernel-tester.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
5086 name = "f32_vrndu_test",
5087 srcs = [
5088 "test/f32-vrndu.cc",
5089 "test/vunary-microkernel-tester.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
5095 name = "f32_vrndd_test",
5096 srcs = [
5097 "test/f32-vrndd.cc",
5098 "test/vunary-microkernel-tester.h",
5099 ] + MICROKERNEL_TEST_HDRS,
5100 deps = MICROKERNEL_TEST_DEPS,
5101)
5102
5103xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005104 name = "f32_vscale_test",
5105 srcs = [
5106 "test/f32-vscale.cc",
5107 "test/vscale-microkernel-tester.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005113 name = "f32_vscaleexpminusmax_test",
5114 srcs = [
5115 "test/f32-vscaleexpminusmax.cc",
5116 "test/vscaleexpminusmax-microkernel-tester.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005122 name = "f32_vscaleextexp_test",
5123 srcs = [
5124 "test/f32-vscaleextexp.cc",
5125 "test/vscaleextexp-microkernel-tester.h",
5126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005131 name = "f32_vsqr_test",
5132 srcs = [
5133 "test/f32-vsqr.cc",
5134 "test/vunary-microkernel-tester.h",
5135 ] + MICROKERNEL_TEST_HDRS,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005140 name = "f32_vsqrdiff_test",
5141 srcs = [
5142 "test/f32-vsqrdiff.cc",
5143 "test/vbinary-microkernel-tester.h",
5144 ] + MICROKERNEL_TEST_HDRS,
5145 deps = MICROKERNEL_TEST_DEPS,
5146)
5147
5148xnnpack_unit_test(
5149 name = "f32_vsqrdiffc_test",
5150 srcs = [
5151 "test/f32-vsqrdiffc.cc",
5152 "test/vbinaryc-microkernel-tester.h",
5153 ] + MICROKERNEL_TEST_HDRS,
5154 deps = MICROKERNEL_TEST_DEPS,
5155)
5156
5157xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005158 name = "f32_vsqrt_test",
5159 srcs = [
5160 "test/f32-vsqrt.cc",
5161 "test/vunary-microkernel-tester.h",
5162 ] + MICROKERNEL_TEST_HDRS,
5163 deps = MICROKERNEL_TEST_DEPS,
5164)
5165
5166xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005167 name = "f32_vsub_test",
5168 srcs = [
5169 "test/f32-vsub.cc",
5170 "test/vbinary-microkernel-tester.h",
5171 ] + MICROKERNEL_TEST_HDRS,
5172 deps = MICROKERNEL_TEST_DEPS,
5173)
5174
5175xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005176 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005177 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005178 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005179 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005180 ] + MICROKERNEL_TEST_HDRS,
5181 deps = MICROKERNEL_TEST_DEPS,
5182)
5183
5184xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005185 name = "f32_vsubc_test",
5186 srcs = [
5187 "test/f32-vsubc.cc",
5188 "test/vbinaryc-microkernel-tester.h",
5189 ] + MICROKERNEL_TEST_HDRS,
5190 deps = MICROKERNEL_TEST_DEPS,
5191)
5192
5193xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005194 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005195 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005196 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005197 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005198 ] + MICROKERNEL_TEST_HDRS,
5199 deps = MICROKERNEL_TEST_DEPS,
5200)
5201
5202xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005203 name = "f32_vrsubc_test",
5204 srcs = [
5205 "test/f32-vrsubc.cc",
5206 "test/vbinaryc-microkernel-tester.h",
5207 ] + MICROKERNEL_TEST_HDRS,
5208 deps = MICROKERNEL_TEST_DEPS,
5209)
5210
5211xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005212 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005213 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005214 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005215 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005216 ] + MICROKERNEL_TEST_HDRS,
5217 deps = MICROKERNEL_TEST_DEPS,
5218)
5219
5220xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005221 name = "qs8_dwconv_minmax_test",
5222 srcs = [
5223 "test/qs8-dwconv-minmax.cc",
5224 "test/dwconv-microkernel-tester.h",
5225 "src/xnnpack/AlignedAllocator.h",
5226 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5227 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5228)
5229
5230xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005231 name = "qs8_gemm_minmax_test",
5232 srcs = [
5233 "test/qs8-gemm-minmax.cc",
5234 "test/gemm-microkernel-tester.h",
5235 "src/xnnpack/AlignedAllocator.h",
5236 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5237 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5238)
5239
5240xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005241 name = "qs8_igemm_minmax_test",
5242 srcs = [
5243 "test/qs8-igemm-minmax.cc",
5244 "test/gemm-microkernel-tester.h",
5245 "src/xnnpack/AlignedAllocator.h",
5246 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5248)
5249
5250xnnpack_unit_test(
5251 name = "qs8_requantization_test",
5252 srcs = [
5253 "src/xnnpack/requantization-stubs.h",
5254 "test/qs8-requantization.cc",
5255 "test/requantization-tester.h",
5256 ] + MICROKERNEL_TEST_HDRS,
5257 deps = MICROKERNEL_TEST_DEPS,
5258)
5259
5260xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005261 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005263 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 "test/avgpool-microkernel-tester.h",
5265 "src/xnnpack/AlignedAllocator.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005271 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005273 "test/qu8-dwconv-minmax.cc",
5274 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 "src/xnnpack/AlignedAllocator.h",
5276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005277 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278)
5279
5280xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005281 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005283 "test/qu8-igemm-minmax.cc",
5284 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285 "src/xnnpack/AlignedAllocator.h",
5286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005287 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288)
5289
5290xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005291 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005293 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 "test/gavgpool-microkernel-tester.h",
5295 "src/xnnpack/AlignedAllocator.h",
5296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005301 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005303 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 "test/gemm-microkernel-tester.h",
5305 "src/xnnpack/AlignedAllocator.h",
5306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005307 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308)
5309
5310xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005311 name = "qu8_requantization_test",
5312 srcs = [
5313 "src/xnnpack/requantization-stubs.h",
5314 "test/qu8-requantization.cc",
5315 "test/requantization-tester.h",
5316 ] + MICROKERNEL_TEST_HDRS,
5317 deps = MICROKERNEL_TEST_DEPS,
5318)
5319
5320xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005321 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005323 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005324 "test/vadd-microkernel-tester.h",
5325 ] + MICROKERNEL_TEST_HDRS,
5326 deps = MICROKERNEL_TEST_DEPS,
5327)
5328
5329xnnpack_unit_test(
5330 name = "u8_clamp_test",
5331 srcs = [
5332 "test/u8-clamp.cc",
5333 "test/clamp-microkernel-tester.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
5339 name = "u8_lut32norm_test",
5340 srcs = [
5341 "test/u8-lut32norm.cc",
5342 "test/lut-norm-microkernel-tester.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005348 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005350 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005351 "test/maxpool-microkernel-tester.h",
5352 ] + MICROKERNEL_TEST_HDRS,
5353 deps = MICROKERNEL_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
5357 name = "u8_rmax_test",
5358 srcs = [
5359 "test/u8-rmax.cc",
5360 "test/rmax-microkernel-tester.h",
5361 ] + MICROKERNEL_TEST_HDRS,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
5365xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005366 name = "x32_fill_test",
5367 srcs = [
5368 "test/x32-fill.cc",
5369 "test/fill-microkernel-tester.h",
5370 ] + MICROKERNEL_TEST_HDRS,
5371 deps = MICROKERNEL_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 name = "x32_packx_test",
5376 srcs = [
5377 "test/x32-packx.cc",
5378 "test/pack-microkernel-tester.h",
5379 "src/xnnpack/AlignedAllocator.h",
5380 ] + MICROKERNEL_TEST_HDRS,
5381 deps = MICROKERNEL_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
5385 name = "x32_pad_test",
5386 srcs = [
5387 "test/x32-pad.cc",
5388 "test/pad-microkernel-tester.h",
5389 ] + MICROKERNEL_TEST_HDRS,
5390 deps = MICROKERNEL_TEST_DEPS,
5391)
5392
5393xnnpack_unit_test(
5394 name = "x32_unpool_test",
5395 srcs = [
5396 "test/x32-unpool.cc",
5397 "test/unpool-microkernel-tester.h",
5398 ] + MICROKERNEL_TEST_HDRS,
5399 deps = MICROKERNEL_TEST_DEPS,
5400)
5401
5402xnnpack_unit_test(
5403 name = "x32_zip_test",
5404 srcs = [
5405 "test/x32-zip.cc",
5406 "test/zip-microkernel-tester.h",
5407 ] + MICROKERNEL_TEST_HDRS,
5408 deps = MICROKERNEL_TEST_DEPS,
5409)
5410
5411xnnpack_unit_test(
5412 name = "x8_lut_test",
5413 srcs = [
5414 "test/x8-lut.cc",
5415 "test/lut-microkernel-tester.h",
5416 ] + MICROKERNEL_TEST_HDRS,
5417 deps = MICROKERNEL_TEST_DEPS,
5418)
5419
5420xnnpack_unit_test(
5421 name = "x8_zip_test",
5422 srcs = [
5423 "test/x8-zip.cc",
5424 "test/zip-microkernel-tester.h",
5425 ] + MICROKERNEL_TEST_HDRS,
5426 deps = MICROKERNEL_TEST_DEPS,
5427)
5428
Marat Dukhan20c3b922020-03-10 03:45:06 -07005429########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430
5431xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005432 name = "operator_size_test",
5433 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005434 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435)
5436
Marat Dukhan20c3b922020-03-10 03:45:06 -07005437xnnpack_binary(
5438 name = "subgraph_size_test",
5439 srcs = ["test/subgraph-size.c"],
5440 deps = [":XNNPACK"],
5441)
5442
5443########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005444
5445xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005446 name = "abs_nc_test",
5447 srcs = [
5448 "test/abs-nc.cc",
5449 "test/abs-operator-tester.h",
5450 ],
5451 deps = OPERATOR_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005455 name = "add_nd_test",
5456 srcs = [
5457 "test/add-nd.cc",
5458 "test/binary-elementwise-operator-tester.h",
5459 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005460 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005461)
5462
5463xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005464 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005466 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 "test/argmax-pooling-operator-tester.h",
5468 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005470)
5471
5472xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005475 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 "test/average-pooling-operator-tester.h",
5477 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479)
5480
5481xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005482 name = "bankers_rounding_nc_test",
5483 srcs = [
5484 "test/bankers-rounding-nc.cc",
5485 "test/bankers-rounding-operator-tester.h",
5486 ],
5487 deps = OPERATOR_TEST_DEPS,
5488)
5489
5490xnnpack_unit_test(
5491 name = "ceiling_nc_test",
5492 srcs = [
5493 "test/ceiling-nc.cc",
5494 "test/ceiling-operator-tester.h",
5495 ],
5496 deps = OPERATOR_TEST_DEPS,
5497)
5498
5499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005500 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005502 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005503 "test/channel-shuffle-operator-tester.h",
5504 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005506)
5507
5508xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005509 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005511 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005512 "test/clamp-operator-tester.h",
5513 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005514 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005515)
5516
5517xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005518 name = "constant_pad_nd_test",
5519 srcs = [
5520 "test/constant-pad-nd.cc",
5521 "test/constant-pad-operator-tester.h",
5522 ],
5523 deps = OPERATOR_TEST_DEPS,
5524)
5525
5526xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005527 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005528 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005529 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005530 "test/convolution-operator-tester.h",
5531 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533)
5534
5535xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005536 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005538 "test/convolution-nchw.cc",
5539 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005541 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005542)
5543
5544xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005545 name = "copy_nc_test",
5546 srcs = [
5547 "test/copy-nc.cc",
5548 "test/copy-operator-tester.h",
5549 ],
5550 deps = OPERATOR_TEST_DEPS,
5551)
5552
5553xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005554 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005556 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005557 "test/deconvolution-operator-tester.h",
5558 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005559 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005560)
5561
5562xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005563 name = "divide_nd_test",
5564 srcs = [
5565 "test/binary-elementwise-operator-tester.h",
5566 "test/divide-nd.cc",
5567 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005568 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005569)
5570
5571xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005572 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005574 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005575 "test/fully-connected-operator-tester.h",
5576 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005577 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578)
5579
5580xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005581 name = "floor_nc_test",
5582 srcs = [
5583 "test/floor-nc.cc",
5584 "test/floor-operator-tester.h",
5585 ],
5586 deps = OPERATOR_TEST_DEPS,
5587)
5588
5589xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005590 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005592 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005593 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005594 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005595 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005596)
5597
5598xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005599 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005601 "test/global-average-pooling-ncw.cc",
5602 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005603 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005604 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005605)
5606
5607xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005608 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005610 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005611 "test/hardswish-operator-tester.h",
5612 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005613 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005614)
5615
5616xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005617 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005618 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005619 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005620 "test/leaky-relu-operator-tester.h",
5621 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005622 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005623)
5624
5625xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005626 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005628 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005629 "test/max-pooling-operator-tester.h",
5630 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005631 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005632)
5633
5634xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005635 name = "maximum_nd_test",
5636 srcs = [
5637 "test/binary-elementwise-operator-tester.h",
5638 "test/maximum-nd.cc",
5639 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005640 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005641)
5642
5643xnnpack_unit_test(
5644 name = "minimum_nd_test",
5645 srcs = [
5646 "test/binary-elementwise-operator-tester.h",
5647 "test/minimum-nd.cc",
5648 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005649 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005650)
5651
5652xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005653 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005654 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005655 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005656 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005657 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005658 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005659)
5660
5661xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005662 name = "negate_nc_test",
5663 srcs = [
5664 "test/negate-nc.cc",
5665 "test/negate-operator-tester.h",
5666 ],
5667 deps = OPERATOR_TEST_DEPS,
5668)
5669
5670xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005671 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005673 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674 "test/prelu-operator-tester.h",
5675 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005676 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677)
5678
5679xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005680 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005681 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005682 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005683 "test/resize-bilinear-operator-tester.h",
5684 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005685 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005686)
5687
5688xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005689 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005690 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005691 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005692 "test/sigmoid-operator-tester.h",
5693 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005694 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005695)
5696
5697xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005698 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005699 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005700 "test/softmax-nc.cc",
5701 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005703 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005704)
5705
5706xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005707 name = "square_nc_test",
5708 srcs = [
5709 "test/square-nc.cc",
5710 "test/square-operator-tester.h",
5711 ],
5712 deps = OPERATOR_TEST_DEPS,
5713)
5714
5715xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005716 name = "square_root_nc_test",
5717 srcs = [
5718 "test/square-root-nc.cc",
5719 "test/square-root-operator-tester.h",
5720 ],
5721 deps = OPERATOR_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005725 name = "squared_difference_nd_test",
5726 srcs = [
5727 "test/binary-elementwise-operator-tester.h",
5728 "test/squared-difference-nd.cc",
5729 ],
5730 deps = OPERATOR_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005734 name = "subtract_nd_test",
5735 srcs = [
5736 "test/binary-elementwise-operator-tester.h",
5737 "test/subtract-nd.cc",
5738 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005739 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005740)
5741
5742xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005743 name = "truncation_nc_test",
5744 srcs = [
5745 "test/truncation-nc.cc",
5746 "test/truncation-operator-tester.h",
5747 ],
5748 deps = OPERATOR_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005752 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005753 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005754 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005755 "test/unpooling-operator-tester.h",
5756 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005757 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758)
5759
Chao Mei6ddfc602020-05-13 22:29:36 -07005760############################### Misc unit tests ###############################
5761
5762xnnpack_unit_test(
5763 name = "memory_planner_test",
5764 srcs = [
5765 "test/memory-planner-test.cc",
5766 ],
5767 deps = [
5768 ":XNNPACK",
5769 ":memory_planner",
5770 ],
5771)
5772
Marat Dukhan08c4a432019-10-03 09:29:21 -07005773############################# Build configurations #############################
5774
Marat Dukhanb8642352019-10-30 15:43:02 -07005775# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005776config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005777 name = "xnn_enable_assembly_explicit_true",
5778 define_values = {"xnn_enable_assembly": "true"},
5779)
5780
5781# Disables usage of assembly kernels.
5782config_setting(
5783 name = "xnn_enable_assembly_explicit_false",
5784 define_values = {"xnn_enable_assembly": "false"},
5785)
5786
Marat Dukhan9de90e02020-06-18 16:04:12 -07005787# Enables usage of sparse inference.
5788config_setting(
5789 name = "xnn_enable_sparse_explicit_true",
5790 define_values = {"xnn_enable_sparse": "true"},
5791)
5792
5793# Disables usage of sparse inference.
5794config_setting(
5795 name = "xnn_enable_sparse_explicit_false",
5796 define_values = {"xnn_enable_sparse": "false"},
5797)
5798
Marat Dukhan05702cf2020-03-26 15:41:33 -07005799# Disables usage of HMP-aware optimizations.
5800config_setting(
5801 name = "xnn_enable_hmp_explicit_false",
5802 define_values = {"xnn_enable_hmp": "false"},
5803)
5804
Chao Mei6ddfc602020-05-13 22:29:36 -07005805# Enable usage of optimized memory allocation
5806config_setting(
5807 name = "xnn_enable_memopt_explicit_true",
5808 define_values = {"xnnpack_enable_memopt": "true"},
5809)
5810
5811# Disable usage of optimized memory allocation
5812config_setting(
5813 name = "xnn_enable_memopt_explicit_false",
5814 define_values = {"xnnpack_enable_memopt": "false"},
5815)
5816
Marat Dukhanb8642352019-10-30 15:43:02 -07005817# Builds with -c dbg
5818config_setting(
5819 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005820 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005821 "compilation_mode": "dbg",
5822 },
5823)
5824
5825# Builds with -c opt
5826config_setting(
5827 name = "optimized_build",
5828 values = {
5829 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005830 },
5831)
5832
5833config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005834 name = "linux_k8",
5835 values = {"cpu": "k8"},
5836)
5837
5838config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005839 name = "linux_arm",
5840 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005841)
5842
5843config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005844 name = "linux_armeabi",
5845 values = {"cpu": "armeabi"},
5846)
5847
5848config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005849 name = "linux_armhf",
5850 values = {"cpu": "armhf"},
5851)
5852
5853config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005854 name = "linux_armv7a",
5855 values = {"cpu": "armv7a"},
5856)
5857
5858config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005859 name = "linux_aarch64",
5860 values = {"cpu": "aarch64"},
5861)
5862
5863config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005864 name = "android",
5865 values = {"crosstool_top": "//external:android/crosstool"},
5866)
5867
5868config_setting(
5869 name = "android_armv7",
5870 values = {
5871 "crosstool_top": "//external:android/crosstool",
5872 "cpu": "armeabi-v7a",
5873 },
5874)
5875
5876config_setting(
5877 name = "android_arm64",
5878 values = {
5879 "crosstool_top": "//external:android/crosstool",
5880 "cpu": "arm64-v8a",
5881 },
5882)
5883
5884config_setting(
5885 name = "android_x86",
5886 values = {
5887 "crosstool_top": "//external:android/crosstool",
5888 "cpu": "x86",
5889 },
5890)
5891
5892config_setting(
5893 name = "android_x86_64",
5894 values = {
5895 "crosstool_top": "//external:android/crosstool",
5896 "cpu": "x86_64",
5897 },
5898)
5899
5900config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005901 name = "windows_x86_64",
5902 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005903)
5904
5905config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005906 name = "windows_x86_64_clang",
5907 values = {
5908 "compiler": "clang-cl",
5909 "cpu": "x64_windows",
5910 },
5911)
5912
5913config_setting(
5914 name = "windows_x86_64_mingw",
5915 values = {
5916 "compiler": "mingw-gcc",
5917 "cpu": "x64_windows",
5918 },
5919)
5920
5921config_setting(
5922 name = "windows_x86_64_msys",
5923 values = {
5924 "compiler": "msys-gcc",
5925 "cpu": "x64_windows",
5926 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005927)
5928
5929config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005930 name = "macos_x86_64",
5931 values = {
5932 "apple_platform_type": "macos",
5933 "cpu": "darwin",
5934 },
5935)
5936
5937config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005938 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005939 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005940)
5941
5942config_setting(
5943 name = "emscripten_wasm",
5944 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005945 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005946 "cpu": "wasm",
5947 },
5948)
5949
5950config_setting(
5951 name = "emscripten_wasmsimd",
5952 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005953 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005954 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005955 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005956 },
5957)
5958
5959config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005960 name = "ios_armv7",
5961 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005962 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005963 "cpu": "ios_armv7",
5964 },
5965)
5966
5967config_setting(
5968 name = "ios_arm64",
5969 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005970 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005971 "cpu": "ios_arm64",
5972 },
5973)
5974
5975config_setting(
5976 name = "ios_arm64e",
5977 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005978 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005979 "cpu": "ios_arm64e",
5980 },
5981)
5982
5983config_setting(
5984 name = "ios_x86",
5985 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005986 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005987 "cpu": "ios_i386",
5988 },
5989)
5990
5991config_setting(
5992 name = "ios_x86_64",
5993 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005994 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005995 "cpu": "ios_x86_64",
5996 },
5997)
5998
5999config_setting(
6000 name = "watchos_armv7k",
6001 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006002 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006003 "cpu": "watchos_armv7k",
6004 },
6005)
6006
6007config_setting(
6008 name = "watchos_arm64_32",
6009 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006010 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006011 "cpu": "watchos_arm64_32",
6012 },
6013)
6014
6015config_setting(
6016 name = "watchos_x86",
6017 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006018 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006019 "cpu": "watchos_i386",
6020 },
6021)
6022
6023config_setting(
6024 name = "watchos_x86_64",
6025 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006026 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006027 "cpu": "watchos_x86_64",
6028 },
6029)
6030
6031config_setting(
6032 name = "tvos_arm64",
6033 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006034 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006035 "cpu": "tvos_arm64",
6036 },
6037)
6038
6039config_setting(
6040 name = "tvos_x86_64",
6041 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006042 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006043 "cpu": "tvos_x86_64",
6044 },
6045)