blob: d250c83161f803087f5c3f59cce47ebece6f3f4d [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700851 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700852 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700853 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700854 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700855 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700856 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700857 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700858 "src/x32-zip/x2-wasmsimd.c",
859 "src/x32-zip/x3-wasmsimd.c",
860 "src/x32-zip/x4-wasmsimd.c",
861 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700862 "src/math/roundne-wasmsimd-addsub.c",
863 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700864 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700865 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700866 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700867 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700868 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700869 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
870 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700871]
872
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800873PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700874 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700875 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700876 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700877 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700878 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700879 "src/f32-spmm/gen/4x1-minmax-psimd.c",
880 "src/f32-spmm/gen/8x1-minmax-psimd.c",
881 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700882]
883
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800884PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700885 "src/qs8-requantization/precise-psimd.c",
886 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700887 "src/qu8-requantization/precise-psimd.c",
888 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800889]
890
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891# ISA-specific micro-kernels
892NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700893 "src/f32-argmaxpool/4x-neon-c4.c",
894 "src/f32-argmaxpool/9p8x-neon-c4.c",
895 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700896 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
897 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700898 "src/f32-clamp/gen/neon-x4.c",
899 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700904 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
905 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700906 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
907 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700908 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
909 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
910 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
911 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700912 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
913 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
914 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
915 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700916 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
917 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
918 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
919 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700920 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700921 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
922 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700923 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
926 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
927 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
929 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
930 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
931 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
932 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
933 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
934 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
935 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
936 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
938 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
939 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
940 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
941 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
942 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
945 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
947 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
948 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
949 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
950 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
951 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
953 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700954 "src/f32-hswish/gen/hswish-neon-x4.c",
955 "src/f32-hswish/gen/hswish-neon-x8.c",
956 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700957 "src/f32-ibilinear/gen/neon-c4.c",
958 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700959 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
960 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
962 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
963 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
964 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
965 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
966 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
967 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
968 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
969 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
970 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
971 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
972 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
974 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700975 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
976 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
977 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700978 "src/f32-ppmm/gen/4x8-minmax-neon.c",
979 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700980 "src/f32-prelu/gen/neon-1x4.c",
981 "src/f32-prelu/gen/neon-1x8.c",
982 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800983 "src/f32-prelu/gen/neon-2x4.c",
984 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700985 "src/f32-prelu/gen/neon-2x16.c",
986 "src/f32-prelu/gen/neon-4x4.c",
987 "src/f32-prelu/gen/neon-4x8.c",
988 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001013 "src/f32-relu/gen/neon-x4.c",
1014 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001015 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001016 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001017 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1022 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1034 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001035 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1036 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1037 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1038 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001039 "src/f32-vbinary/gen/vmax-neon-x4.c",
1040 "src/f32-vbinary/gen/vmax-neon-x8.c",
1041 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1042 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1043 "src/f32-vbinary/gen/vmin-neon-x4.c",
1044 "src/f32-vbinary/gen/vmin-neon-x8.c",
1045 "src/f32-vbinary/gen/vminc-neon-x4.c",
1046 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001047 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1048 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1050 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1051 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1052 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001053 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1054 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1055 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1056 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001057 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1058 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1059 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1060 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001061 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1062 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001063 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1064 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001065 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1066 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1067 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1068 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1069 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1070 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1071 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1072 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001073 "src/f32-vunary/gen/vabs-neon-x4.c",
1074 "src/f32-vunary/gen/vabs-neon-x8.c",
1075 "src/f32-vunary/gen/vneg-neon-x4.c",
1076 "src/f32-vunary/gen/vneg-neon-x8.c",
1077 "src/f32-vunary/gen/vsqr-neon-x4.c",
1078 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001079 "src/qs8-requantization/precise-neon.c",
1080 "src/qs8-requantization/fp32-neon.c",
1081 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001082 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1083 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1084 "src/qu8-dwconv/up8x9-minmax-neon.c",
1085 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1086 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1087 "src/qu8-gemm/4x8-minmax-neon.c",
1088 "src/qu8-gemm/8x8-minmax-neon.c",
1089 "src/qu8-igemm/4x8-minmax-neon.c",
1090 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001091 "src/qu8-requantization/precise-neon.c",
1092 "src/qu8-requantization/fp32-neon.c",
1093 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001094 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001095 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001096 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001097 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001098 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001099 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001100 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001101 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001102 "src/x32-zip/x2-neon.c",
1103 "src/x32-zip/x3-neon.c",
1104 "src/x32-zip/x4-neon.c",
1105 "src/x32-zip/xm-neon.c",
1106 "src/x8-zip/x2-neon.c",
1107 "src/x8-zip/x3-neon.c",
1108 "src/x8-zip/x4-neon.c",
1109 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001110 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001111 "src/math/roundd-neon-addsub.c",
1112 "src/math/roundd-neon-cvt.c",
1113 "src/math/roundu-neon-addsub.c",
1114 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001115 "src/math/roundz-neon-addsub.c",
1116 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001117 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001118 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1119 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1120 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1121 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1123 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001124 "src/math/sqrt-neon-nr1rsqrts.c",
1125 "src/math/sqrt-neon-nr2rsqrts.c",
1126 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001127]
1128
1129NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001130 "src/f32-ibilinear/gen/neonfma-c4.c",
1131 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001132 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1133 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1134 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1136 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1137 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1138 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1140 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001141 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1142 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1144 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001145 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1146 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1148 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001149 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1150 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1152 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1154 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1155 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1157 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1158 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1159 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1162 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1164 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1167 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1168 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1170 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001171 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1172 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001251 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1252 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001253 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1262 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1263 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1272 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001273 "src/math/exp-neonfma-lut64-p2.c",
1274 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001275 "src/math/expminus-neonfma-lut2048-p1.c",
1276 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001277 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001278 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1280 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1281 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1283 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1284 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1286 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1287 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1289 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1290 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1292 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1293 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1295 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001296 "src/math/sqrt-neonfma-nr1fma.c",
1297 "src/math/sqrt-neonfma-nr2fma.c",
1298 "src/math/sqrt-neonfma-nr3fma.c",
1299 "src/math/sqrt-neonfma-nr2fma1adj.c",
1300 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001301]
1302
1303AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001304 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1305 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1306 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1307 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1308 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1309 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001310 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1312 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1313 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1316 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1317 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1318 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1319 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1322 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1323 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1324 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1326 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1327 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1329 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001330 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1333 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001334 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1335 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001336 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1337 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001338 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1339 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1340 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1341 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1342 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001343 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001361 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1362 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1363 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1364 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1366 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1367 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1368 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1369 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1371 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1372 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1373 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1374 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1376 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1377 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1378 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001379 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1380 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001381 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1382 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1383 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1384 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1385 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1386 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001387]
1388
Marat Dukhan8853b822020-05-07 12:19:01 -07001389NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001390 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1391 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1392 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1393 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1394 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1395 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1396 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1397 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001398 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001399 "src/math/roundd-neonv8.c",
1400 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001401 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001402]
1403
Marat Dukhan08c4a432019-10-03 09:29:21 -07001404AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001405 "src/f16-clamp/gen/neonfp16arith-x16.c",
1406 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001407 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1408 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1409 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1410 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1411 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1412 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1413 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1414 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1415 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1416 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1417 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1418 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001419 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1420 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001421 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1422 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001423 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001424 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001425 "src/f16-relu/gen/neonfp16arith-x16.c",
1426 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001427 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1428 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1429 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1430 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1431 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1432 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1433 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1434 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1435 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1436 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1437 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1438 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1439 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1440 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1441 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1442 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1443 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1444 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1445 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1446 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1447 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1448 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1449 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1450 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1451 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001455 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1456 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001457 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1464 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001465 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1468 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001469 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1480 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001481 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1482 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1483 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1484 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1485 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1486 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1487 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1488 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489]
1490
1491SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001492 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1493 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001494 "src/f32-clamp/gen/sse-x4.c",
1495 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001496 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1497 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001498 "src/f32-dwconv-chw/3x3p1-sse.c",
1499 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001500 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1501 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1502 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1503 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1504 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1505 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1506 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1507 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1508 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1509 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1510 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1511 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001512 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001513 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1514 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001515 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1516 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1517 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1518 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1519 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1520 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1521 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1522 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1523 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1524 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1526 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1527 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001528 "src/f32-hswish/gen/hswish-sse-x4.c",
1529 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001530 "src/f32-ibilinear/gen/sse-c4.c",
1531 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001532 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1533 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1534 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1535 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1536 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1537 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1538 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001539 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1540 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1541 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001542 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001543 "src/f32-prelu/gen/sse-2x4.c",
1544 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001545 "src/f32-relu/gen/sse-x4.c",
1546 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001547 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001548 "src/f32-spmm/gen/4x1-minmax-sse.c",
1549 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001550 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001551 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1552 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1553 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1554 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1556 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1558 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001559 "src/f32-vbinary/gen/vmax-sse-x4.c",
1560 "src/f32-vbinary/gen/vmax-sse-x8.c",
1561 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1562 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1563 "src/f32-vbinary/gen/vmin-sse-x4.c",
1564 "src/f32-vbinary/gen/vmin-sse-x8.c",
1565 "src/f32-vbinary/gen/vminc-sse-x4.c",
1566 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001567 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1568 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1569 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1570 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1571 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1572 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1573 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1574 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001575 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1576 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1577 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1578 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001579 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1581 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001583 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1584 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001585 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1586 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001587 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1588 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001589 "src/f32-vunary/gen/vabs-sse-x4.c",
1590 "src/f32-vunary/gen/vabs-sse-x8.c",
1591 "src/f32-vunary/gen/vneg-sse-x4.c",
1592 "src/f32-vunary/gen/vneg-sse-x8.c",
1593 "src/f32-vunary/gen/vsqr-sse-x4.c",
1594 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001595 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001597 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001598 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001599 "src/math/roundd-sse-addsub.c",
1600 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001601 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001602 "src/math/sqrt-sse-hh1mac.c",
1603 "src/math/sqrt-sse-nr1mac.c",
1604 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001605]
1606
1607SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001608 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1609 "src/f32-argmaxpool/4x-sse2-c4.c",
1610 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001611 "src/f32-prelu/gen/sse2-2x4.c",
1612 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1624 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001625 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001629 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1630 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001631 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1632 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001633 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1634 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1636 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1637 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1638 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1639 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1640 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001641 "src/qs8-requantization/precise-sse2.c",
1642 "src/qs8-requantization/fp32-sse2.c",
1643 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001644 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1645 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001646 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1647 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001648 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1649 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001650 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001651 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1652 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001653 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001654 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1655 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001656 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1657 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001658 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1659 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001660 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001661 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1662 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001663 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001664 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1665 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1666 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1667 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1668 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1669 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1670 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1671 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001672 "src/qu8-requantization/precise-sse2.c",
1673 "src/qu8-requantization/fp32-sse2.c",
1674 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001675 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001676 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001677 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001678 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001679 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 "src/x32-zip/x2-sse2.c",
1681 "src/x32-zip/x3-sse2.c",
1682 "src/x32-zip/x4-sse2.c",
1683 "src/x32-zip/xm-sse2.c",
1684 "src/x8-zip/x2-sse2.c",
1685 "src/x8-zip/x3-sse2.c",
1686 "src/x8-zip/x4-sse2.c",
1687 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001688 "src/math/exp-sse2-p5.c",
1689 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001690 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001691 "src/math/roundd-sse2-cvt.c",
1692 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001693 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001694 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001695]
1696
1697SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001698 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1699 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001700 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1701 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001702 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1703 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001704 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001705 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1706 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001707 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001708 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1709 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001710 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1711 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001712 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1713 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001714 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001715 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1716 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001717 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001718 "src/qs8-requantization/precise-ssse3.c",
1719 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001720 "src/qu8-requantization/precise-ssse3.c",
1721 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722]
1723
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001724SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001725 "src/f32-prelu/gen/sse41-2x4.c",
1726 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001727 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001728 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001729 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001730 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001731 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1732 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001733 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1734 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001735 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1736 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1737 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1738 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1739 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1740 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1741 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1742 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001743 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1744 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001745 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1746 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001747 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1748 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001749 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001750 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1751 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001752 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001753 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1754 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001755 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1756 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001757 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1758 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001759 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001760 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1761 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001762 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001763 "src/qs8-requantization/fp32-sse4.c",
1764 "src/qs8-requantization/precise-sse4.c",
1765 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001766 "src/qu8-requantization/precise-sse4.c",
1767 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001768 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001769 "src/math/roundd-sse41.c",
1770 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001771 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001772]
1773
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001775 "src/f32-clamp/gen/avx-x8.c",
1776 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001777 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1778 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1779 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1780 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1781 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1782 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1783 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1784 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1785 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1786 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1787 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1788 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1789 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1790 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1791 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1794 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1795 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1796 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1797 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1798 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1799 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1800 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1802 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1803 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1804 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1806 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001807 "src/f32-hswish/gen/hswish-avx-x8.c",
1808 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001809 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1810 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1811 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1812 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1813 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1814 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1815 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1816 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1817 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001818 "src/f32-prelu/gen/avx-2x8.c",
1819 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001820 "src/f32-relu/gen/avx-x16.c",
1821 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001823 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1824 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1825 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1826 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1827 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1828 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1829 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1830 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001831 "src/f32-vbinary/gen/vmax-avx-x8.c",
1832 "src/f32-vbinary/gen/vmax-avx-x16.c",
1833 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1834 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1835 "src/f32-vbinary/gen/vmin-avx-x8.c",
1836 "src/f32-vbinary/gen/vmin-avx-x16.c",
1837 "src/f32-vbinary/gen/vminc-avx-x8.c",
1838 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001839 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1840 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1841 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1842 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1843 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1844 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1845 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1846 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001847 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1848 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1849 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1850 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001851 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1852 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1853 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1854 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001855 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1856 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001857 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001858 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1859 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1860 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1861 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1862 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1863 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1864 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1865 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001866 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1867 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001868 "src/f32-vunary/gen/vabs-avx-x8.c",
1869 "src/f32-vunary/gen/vabs-avx-x16.c",
1870 "src/f32-vunary/gen/vneg-avx-x8.c",
1871 "src/f32-vunary/gen/vneg-avx-x16.c",
1872 "src/f32-vunary/gen/vsqr-avx-x8.c",
1873 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874]
1875
Marat Dukhan1566fee2020-08-02 21:55:41 -07001876XOP_UKERNELS = [
1877 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1878 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1879 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1880 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
1881 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1882 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001883 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001884 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1885 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001886 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001887 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1888 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1889 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1890 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1891 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1892 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001893 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001894 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1895 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001896 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001897]
1898
Marat Dukhanfda12b82019-11-21 12:27:59 -08001899FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001900 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1901 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1902 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1903 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1904 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1905 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1906 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1907 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1908 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1909 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1910 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1911 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1912 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1923 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1924 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1925 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1926 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1927 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1928 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1929 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1930 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1931 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1932 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1933 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1934 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1935 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1936 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1937 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1938 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1939 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001940 "src/f32-hswish/gen/hswish-fma3-x8.c",
1941 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001942 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1943 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1944 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1945 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1946 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1947 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1948 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1949 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1950 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1951 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1952 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1953 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1954 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1955 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001956 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1957 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1958 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1959 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1960 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1961 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1962 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1963 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001964 "src/math/sqrt-fma3-nr1fma.c",
1965 "src/math/sqrt-fma3-nr2fma.c",
1966 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001967]
1968
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001969AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001970 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1971 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1972 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1973 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1974 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1975 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1976 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1977 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1978 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1979 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1980 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1981 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1982 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1983 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1984 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1985 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1986 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1987 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1988 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1989 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1990 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1991 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1992 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1993 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1994 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1995 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1996 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1997 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1998 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1999 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2000 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2001 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2002 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2003 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2004 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2005 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002006 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2019 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2020 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2021 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2022 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2023 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2024 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2025 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2026 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2027 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2029 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2030 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2031 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2032 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2033 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2034 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2035 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002036 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2037 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2038 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2039 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2040 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2041 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2042 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2043 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2044 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2045 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2046 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2047 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2048 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2049 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2050 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2051 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2052 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2053 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2054 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2055 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2056 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2057 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2058 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2059 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002060 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2061 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2062 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
2063 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2064 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2065 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002066 "src/math/exp-avx2-p5.c",
2067 "src/math/exp-avx2-perm-p3.c",
2068 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002069 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002070 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002071 "src/math/sigmoid-avx2-rr2-p5-div.c",
2072 "src/math/sigmoid-avx2-rr1-p5-div.c",
2073 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2074 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2075 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2076 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002077]
2078
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002080 "src/f32-clamp/gen/avx512f-x16.c",
2081 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002082 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2083 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2084 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2085 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2086 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2087 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2088 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2089 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2090 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2091 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2092 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2093 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2094 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2095 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2096 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2097 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2098 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2099 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2100 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2101 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2102 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2103 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2104 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2105 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002106 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2107 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002108 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2109 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2110 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2111 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2112 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2113 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002114 "src/f32-prelu/gen/avx512f-2x16.c",
2115 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002116 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2117 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2118 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2119 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2120 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2121 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2122 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2123 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2124 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2125 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2126 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2127 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2128 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2129 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2130 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2131 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2132 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2133 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2134 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2135 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2136 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2137 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2138 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2139 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2140 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2141 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2142 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2143 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2144 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2145 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2146 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2147 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2148 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2149 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2150 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2151 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002152 "src/f32-relu/gen/avx512f-x16.c",
2153 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002155 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2156 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2157 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2158 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2159 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2160 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2161 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2162 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002163 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2164 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2165 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2166 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2167 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2168 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2169 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2170 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002171 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2172 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2173 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2174 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2175 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2176 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2177 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2178 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002179 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2180 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2181 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2182 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002183 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2184 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2185 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2186 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002187 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2188 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002189 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002190 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2191 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2192 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2193 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2194 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2195 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2196 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2197 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2198 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2199 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2200 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2201 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2202 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2203 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2204 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2205 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2206 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2207 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2208 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2209 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2210 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2211 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2212 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2213 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002214 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2215 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2216 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2217 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2218 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2219 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2220 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2221 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002222 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2223 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2224 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2225 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2226 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2227 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2228 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2229 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002230 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2231 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2232 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2233 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2234 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2235 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002236 "src/math/exp-avx512f-p5-scalef.c",
2237 "src/math/exp-avx512f-p5.c",
2238 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002239 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002240 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002241 "src/math/sqrt-avx512f-nr1fma.c",
2242 "src/math/sqrt-avx512f-nr2fma.c",
2243 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244]
2245
2246AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002247 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002248 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002249 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2250 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002251 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002252 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002253 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002254 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002255 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2256 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002257 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2258 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2259 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2260 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261]
2262
2263AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002264 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002265 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2266 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002267 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002268 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2269 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002270 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2271 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2272 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002273 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002274 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2275 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2276 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002277 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002278 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2279 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002280 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002281 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2282 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2283 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2284 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2285 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2286 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2287 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2288 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2289 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2290 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2291 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2292 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2293 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2294 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2295 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2296 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2297 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2298 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002299 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2300 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002301 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2303 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2304 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2305 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2306 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2307 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2308 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2309 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2310 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2311 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2312 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2313 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2314 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2315 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2316 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2317 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2318 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2319 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002320 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2321 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2322 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2323 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2324 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2325 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2326 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2327 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2328 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2329 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2330 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2331 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2332 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2333 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2334 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2335 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2336 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2337 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002338]
2339
Marat Dukhan1b354632020-03-23 12:50:22 -07002340INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 "src/xnnpack/argmaxpool.h",
2342 "src/xnnpack/avgpool.h",
2343 "src/xnnpack/clamp.h",
2344 "src/xnnpack/common.h",
2345 "src/xnnpack/conv.h",
2346 "src/xnnpack/dwconv.h",
2347 "src/xnnpack/gavgpool.h",
2348 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002349 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002350 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002351 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002353 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354 "src/xnnpack/lut.h",
2355 "src/xnnpack/math.h",
2356 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002357 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002358 "src/xnnpack/packx.h",
2359 "src/xnnpack/pad.h",
2360 "src/xnnpack/params.h",
2361 "src/xnnpack/pavgpool.h",
2362 "src/xnnpack/ppmm.h",
2363 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002364 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002365 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002366 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002367 "src/xnnpack/rmax.h",
2368 "src/xnnpack/scalar-utils.h",
2369 "src/xnnpack/spmm.h",
2370 "src/xnnpack/unpool.h",
2371 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002372 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002374 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002375 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002376 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002377 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002379]
2380
2381INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382 "include/xnnpack.h",
2383 "src/xnnpack/allocator.h",
2384 "src/xnnpack/compute.h",
2385 "src/xnnpack/im2col.h",
2386 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002387 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002388 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002389 "src/xnnpack/operator.h",
2390 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002391 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002392 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002393 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002394 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002395]
2396
Marat Dukhan1b354632020-03-23 12:50:22 -07002397ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002398 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399]
2400
Marat Dukhan1b354632020-03-23 12:50:22 -07002401MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002402 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002403 "include/xnnpack.h",
2404]
2405
Marat Dukhan1b354632020-03-23 12:50:22 -07002406MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002407 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002408 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002409 "src/xnnpack/requantization.h",
2410 "include/xnnpack.h",
2411]
2412
2413OPERATOR_TEST_PARAMS_HDRS = [
2414 "src/xnnpack/params.h",
2415 "src/xnnpack/common.h",
2416]
2417
2418WEIGHTS_PACK_HDRS = [
2419 "src/xnnpack/pack.h",
2420 "src/xnnpack/operator.h",
2421 "src/xnnpack/compute.h",
2422]
2423
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002424LOGGING_COPTS = select({
2425 # No logging in optimized mode
2426 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2427 # Full logging in debug mode
2428 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2429 # Error-only logging in default (fastbuild) mode
2430 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2431})
2432
Marat Dukhan3b59de22020-06-03 20:15:19 -07002433LOGGING_SRCS = select({
2434 # No logging in optimized mode
2435 ":optimized_build": [],
2436 "//conditions:default": [
2437 "src/operator-strings.c",
2438 "src/subgraph-strings.c",
2439 ],
2440})
2441
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002442LOGGING_HDRS = [
2443 "src/xnnpack/log.h",
2444]
2445
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002447 name = "tables",
2448 srcs = TABLE_SRCS,
2449 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002450 gcc_copts = xnnpack_gcc_std_copts(),
2451 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002452)
2453
2454xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455 name = "scalar_ukernels",
2456 srcs = SCALAR_UKERNELS,
2457 hdrs = INTERNAL_HDRS,
2458 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002459 gcc_copts = xnnpack_gcc_std_copts(),
2460 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002462 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463 "@FP16",
2464 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002465 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002466 ],
2467)
2468
2469xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002470 name = "scalar_ukernels_test_mode",
2471 srcs = SCALAR_UKERNELS,
2472 hdrs = INTERNAL_HDRS,
2473 aarch32_copts = ["-marm"],
2474 copts = [
2475 "-UNDEBUG",
2476 "-DXNN_TEST_MODE=1",
2477 ],
2478 gcc_copts = xnnpack_gcc_std_copts(),
2479 msvc_copts = xnnpack_msvc_std_copts(),
2480 deps = [
2481 ":tables",
2482 "@FP16",
2483 "@FXdiv",
2484 "@pthreadpool",
2485 ],
2486)
2487
2488xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002489 name = "wasm_ukernels",
2490 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002491 gcc_copts = xnnpack_gcc_std_copts(),
2492 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002493 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002494 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002495 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002496 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002497 "@FP16",
2498 "@FXdiv",
2499 "@pthreadpool",
2500 ],
2501)
2502
2503xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002504 name = "wasm_ukernels_test_mode",
2505 hdrs = INTERNAL_HDRS,
2506 copts = [
2507 "-UNDEBUG",
2508 "-DXNN_TEST_MODE=1",
2509 ],
2510 gcc_copts = xnnpack_gcc_std_copts(),
2511 msvc_copts = xnnpack_msvc_std_copts(),
2512 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002513 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002514 deps = [
2515 ":tables",
2516 "@FP16",
2517 "@FXdiv",
2518 "@pthreadpool",
2519 ],
2520)
2521
2522xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002523 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002524 hdrs = INTERNAL_HDRS,
2525 aarch32_copts = [
2526 "-marm",
2527 "-mfpu=neon",
2528 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002529 gcc_copts = xnnpack_gcc_std_copts(),
2530 gcc_x86_copts = ["-msse2"],
2531 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002532 optimized_copts = [
2533 "-O3",
2534 "-ffast-math",
2535 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002536 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002537 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002538 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002539 "@FP16",
2540 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002541 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 ],
2543)
2544
2545xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002546 name = "psimd_fastmath_ukernels_test_mode",
2547 hdrs = INTERNAL_HDRS,
2548 aarch32_copts = [
2549 "-marm",
2550 "-mfpu=neon",
2551 ],
2552 copts = [
2553 "-UNDEBUG",
2554 "-DXNN_TEST_MODE=1",
2555 ],
2556 gcc_copts = xnnpack_gcc_std_copts(),
2557 gcc_x86_copts = ["-msse2"],
2558 msvc_copts = xnnpack_msvc_std_copts(),
2559 optimized_copts = [
2560 "-O3",
2561 "-ffast-math",
2562 ],
2563 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2564 deps = [
2565 ":tables",
2566 "@FP16",
2567 "@psimd",
2568 "@pthreadpool",
2569 ],
2570)
2571
2572xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002573 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002574 hdrs = INTERNAL_HDRS,
2575 aarch32_copts = [
2576 "-marm",
2577 "-mfpu=neon",
2578 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002579 gcc_copts = xnnpack_gcc_std_copts(),
2580 gcc_x86_copts = ["-msse2"],
2581 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002582 optimized_copts = [
2583 "-O3",
2584 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002585 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002586 deps = [
2587 ":tables",
2588 "@FP16",
2589 "@psimd",
2590 "@pthreadpool",
2591 ],
2592)
2593
2594xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002595 name = "psimd_accmath_ukernels_test_mode",
2596 hdrs = INTERNAL_HDRS,
2597 aarch32_copts = [
2598 "-marm",
2599 "-mfpu=neon",
2600 ],
2601 copts = [
2602 "-UNDEBUG",
2603 "-DXNN_TEST_MODE=1",
2604 ],
2605 gcc_copts = xnnpack_gcc_std_copts(),
2606 gcc_x86_copts = ["-msse2"],
2607 msvc_copts = xnnpack_msvc_std_copts(),
2608 optimized_copts = [
2609 "-O3",
2610 ],
2611 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2612 deps = [
2613 ":tables",
2614 "@FP16",
2615 "@psimd",
2616 "@pthreadpool",
2617 ],
2618)
2619
2620xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 name = "neon_ukernels",
2622 hdrs = INTERNAL_HDRS,
2623 aarch32_copts = [
2624 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002625 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 "-mfpu=neon",
2627 ],
2628 aarch32_srcs = NEON_UKERNELS,
2629 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002630 gcc_copts = xnnpack_gcc_std_copts(),
2631 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002632 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002633 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002634 "@FP16",
2635 "@pthreadpool",
2636 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002637)
2638
2639xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002640 name = "neon_ukernels_test_mode",
2641 hdrs = INTERNAL_HDRS,
2642 aarch32_copts = [
2643 "-marm",
2644 "-march=armv7-a",
2645 "-mfpu=neon",
2646 ],
2647 aarch32_srcs = NEON_UKERNELS,
2648 aarch64_srcs = NEON_UKERNELS,
2649 copts = [
2650 "-UNDEBUG",
2651 "-DXNN_TEST_MODE=1",
2652 ],
2653 gcc_copts = xnnpack_gcc_std_copts(),
2654 msvc_copts = xnnpack_msvc_std_copts(),
2655 deps = [
2656 ":tables",
2657 "@FP16",
2658 "@pthreadpool",
2659 ],
2660)
2661
2662xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002663 name = "neonfma_ukernels",
2664 hdrs = INTERNAL_HDRS,
2665 aarch32_copts = [
2666 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002667 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002668 "-mfpu=neon-vfpv4",
2669 ],
2670 aarch32_srcs = NEONFMA_UKERNELS,
2671 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002672 apple_aarch32_copts = [
2673 "-mcpu=swift",
2674 "-mtune=generic",
2675 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002676 gcc_copts = xnnpack_gcc_std_copts(),
2677 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002678 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002679 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002680 "@FP16",
2681 "@pthreadpool",
2682 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002683)
2684
2685xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002686 name = "neonfma_ukernels_test_mode",
2687 hdrs = INTERNAL_HDRS,
2688 aarch32_copts = [
2689 "-marm",
2690 "-march=armv7-a",
2691 "-mfpu=neon-vfpv4",
2692 ],
2693 aarch32_srcs = NEONFMA_UKERNELS,
2694 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002695 apple_aarch32_copts = [
2696 "-mcpu=swift",
2697 "-mtune=generic",
2698 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002699 copts = [
2700 "-UNDEBUG",
2701 "-DXNN_TEST_MODE=1",
2702 ],
2703 gcc_copts = xnnpack_gcc_std_copts(),
2704 msvc_copts = xnnpack_msvc_std_copts(),
2705 deps = [
2706 ":tables",
2707 "@FP16",
2708 "@pthreadpool",
2709 ],
2710)
2711
2712xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002713 name = "neonv8_ukernels",
2714 hdrs = INTERNAL_HDRS,
2715 aarch32_copts = [
2716 "-marm",
2717 "-march=armv8-a",
2718 "-mfpu=neon-fp-armv8",
2719 ],
2720 aarch32_srcs = NEONV8_UKERNELS,
2721 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002722 apple_aarch32_copts = [
2723 "-mcpu=cyclone",
2724 "-mtune=generic",
2725 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002726 gcc_copts = xnnpack_gcc_std_copts(),
2727 msvc_copts = xnnpack_msvc_std_copts(),
2728 deps = [
2729 ":tables",
2730 "@FP16",
2731 "@pthreadpool",
2732 ],
2733)
2734
2735xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002736 name = "neonv8_ukernels_test_mode",
2737 hdrs = INTERNAL_HDRS,
2738 aarch32_copts = [
2739 "-marm",
2740 "-march=armv8-a",
2741 "-mfpu=neon-fp-armv8",
2742 ],
2743 aarch32_srcs = NEONV8_UKERNELS,
2744 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002745 apple_aarch32_copts = [
2746 "-mcpu=cyclone",
2747 "-mtune=generic",
2748 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002749 copts = [
2750 "-UNDEBUG",
2751 "-DXNN_TEST_MODE=1",
2752 ],
2753 gcc_copts = xnnpack_gcc_std_copts(),
2754 msvc_copts = xnnpack_msvc_std_copts(),
2755 deps = [
2756 ":tables",
2757 "@FP16",
2758 "@pthreadpool",
2759 ],
2760)
2761
2762xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002763 name = "neonfp16arith_ukernels",
2764 hdrs = INTERNAL_HDRS,
2765 aarch64_copts = ["-march=armv8.2-a+fp16"],
2766 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002767 gcc_copts = xnnpack_gcc_std_copts(),
2768 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002769 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002770 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002771 "@FP16",
2772 "@pthreadpool",
2773 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002774)
2775
2776xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002777 name = "neonfp16arith_ukernels_test_mode",
2778 hdrs = INTERNAL_HDRS,
2779 aarch64_copts = ["-march=armv8.2-a+fp16"],
2780 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2781 copts = [
2782 "-UNDEBUG",
2783 "-DXNN_TEST_MODE=1",
2784 ],
2785 gcc_copts = xnnpack_gcc_std_copts(),
2786 msvc_copts = xnnpack_msvc_std_copts(),
2787 deps = [
2788 ":tables",
2789 "@FP16",
2790 "@pthreadpool",
2791 ],
2792)
2793
2794xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 name = "sse2_ukernels",
2796 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002797 gcc_copts = xnnpack_gcc_std_copts(),
2798 gcc_x86_copts = ["-msse2"],
2799 msvc_copts = xnnpack_msvc_std_copts(),
2800 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002801 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002802 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002803 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002804 "@FP16",
2805 "@pthreadpool",
2806 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002807)
2808
2809xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002810 name = "sse2_ukernels_test_mode",
2811 hdrs = INTERNAL_HDRS,
2812 copts = [
2813 "-UNDEBUG",
2814 "-DXNN_TEST_MODE=1",
2815 ],
2816 gcc_copts = xnnpack_gcc_std_copts(),
2817 gcc_x86_copts = ["-msse2"],
2818 msvc_copts = xnnpack_msvc_std_copts(),
2819 msvc_x86_32_copts = ["/arch:SSE2"],
2820 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2821 deps = [
2822 ":tables",
2823 "@FP16",
2824 "@pthreadpool",
2825 ],
2826)
2827
2828xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002829 name = "ssse3_ukernels",
2830 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002831 gcc_copts = xnnpack_gcc_std_copts(),
2832 gcc_x86_copts = ["-mssse3"],
2833 msvc_copts = xnnpack_msvc_std_copts(),
2834 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002835 x86_srcs = SSSE3_UKERNELS,
2836 deps = [
2837 ":tables",
2838 "@FP16",
2839 "@pthreadpool",
2840 ],
2841)
2842
2843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002844 name = "ssse3_ukernels_test_mode",
2845 hdrs = INTERNAL_HDRS,
2846 copts = [
2847 "-UNDEBUG",
2848 "-DXNN_TEST_MODE=1",
2849 ],
2850 gcc_copts = xnnpack_gcc_std_copts(),
2851 gcc_x86_copts = ["-mssse3"],
2852 msvc_copts = xnnpack_msvc_std_copts(),
2853 msvc_x86_32_copts = ["/arch:SSE2"],
2854 x86_srcs = SSSE3_UKERNELS,
2855 deps = [
2856 ":tables",
2857 "@FP16",
2858 "@pthreadpool",
2859 ],
2860)
2861
2862xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002863 name = "sse41_ukernels",
2864 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002865 gcc_copts = xnnpack_gcc_std_copts(),
2866 gcc_x86_copts = ["-msse4.1"],
2867 msvc_copts = xnnpack_msvc_std_copts(),
2868 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002869 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002870 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002871 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002872 "@FP16",
2873 "@pthreadpool",
2874 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002875)
2876
2877xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002878 name = "sse41_ukernels_test_mode",
2879 hdrs = INTERNAL_HDRS,
2880 copts = [
2881 "-UNDEBUG",
2882 "-DXNN_TEST_MODE=1",
2883 ],
2884 gcc_copts = xnnpack_gcc_std_copts(),
2885 gcc_x86_copts = ["-msse4.1"],
2886 msvc_copts = xnnpack_msvc_std_copts(),
2887 msvc_x86_32_copts = ["/arch:SSE2"],
2888 x86_srcs = SSE41_UKERNELS,
2889 deps = [
2890 ":tables",
2891 "@FP16",
2892 "@pthreadpool",
2893 ],
2894)
2895
2896xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 name = "avx_ukernels",
2898 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002899 gcc_copts = xnnpack_gcc_std_copts(),
2900 gcc_x86_copts = ["-mavx"],
2901 msvc_copts = xnnpack_msvc_std_copts(),
2902 msvc_x86_32_copts = ["/arch:AVX"],
2903 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002904 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002905 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002906 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002907 "@FP16",
2908 "@pthreadpool",
2909 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002910)
2911
2912xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002913 name = "avx_ukernels_test_mode",
2914 hdrs = INTERNAL_HDRS,
2915 copts = [
2916 "-UNDEBUG",
2917 "-DXNN_TEST_MODE=1",
2918 ],
2919 gcc_copts = xnnpack_gcc_std_copts(),
2920 gcc_x86_copts = ["-mavx"],
2921 msvc_copts = xnnpack_msvc_std_copts(),
2922 msvc_x86_32_copts = ["/arch:AVX"],
2923 msvc_x86_64_copts = ["/arch:AVX"],
2924 x86_srcs = AVX_UKERNELS,
2925 deps = [
2926 ":tables",
2927 "@FP16",
2928 "@pthreadpool",
2929 ],
2930)
2931
2932xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002933 name = "xop_ukernels",
2934 hdrs = INTERNAL_HDRS,
2935 gcc_copts = xnnpack_gcc_std_copts(),
2936 gcc_x86_copts = ["-mxop"],
2937 msvc_copts = xnnpack_msvc_std_copts(),
2938 msvc_x86_32_copts = ["/arch:AVX"],
2939 msvc_x86_64_copts = ["/arch:AVX"],
2940 x86_srcs = XOP_UKERNELS,
2941 deps = [
2942 ":tables",
2943 "@FP16",
2944 "@pthreadpool",
2945 ],
2946)
2947
2948xnnpack_cc_library(
2949 name = "xop_ukernels_test_mode",
2950 hdrs = INTERNAL_HDRS,
2951 copts = [
2952 "-UNDEBUG",
2953 "-DXNN_TEST_MODE=1",
2954 ],
2955 gcc_copts = xnnpack_gcc_std_copts(),
2956 gcc_x86_copts = ["-mxop"],
2957 msvc_copts = xnnpack_msvc_std_copts(),
2958 msvc_x86_32_copts = ["/arch:AVX"],
2959 msvc_x86_64_copts = ["/arch:AVX"],
2960 x86_srcs = XOP_UKERNELS,
2961 deps = [
2962 ":tables",
2963 "@FP16",
2964 "@pthreadpool",
2965 ],
2966)
2967
2968xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002969 name = "fma3_ukernels",
2970 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002971 gcc_copts = xnnpack_gcc_std_copts(),
2972 gcc_x86_copts = ["-mfma"],
2973 msvc_copts = xnnpack_msvc_std_copts(),
2974 msvc_x86_32_copts = ["/arch:AVX"],
2975 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002976 x86_srcs = FMA3_UKERNELS,
2977 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002978 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002979 "@FP16",
2980 "@pthreadpool",
2981 ],
2982)
2983
2984xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002985 name = "fma3_ukernels_test_mode",
2986 hdrs = INTERNAL_HDRS,
2987 copts = [
2988 "-UNDEBUG",
2989 "-DXNN_TEST_MODE=1",
2990 ],
2991 gcc_copts = xnnpack_gcc_std_copts(),
2992 gcc_x86_copts = ["-mfma"],
2993 msvc_copts = xnnpack_msvc_std_copts(),
2994 msvc_x86_32_copts = ["/arch:AVX"],
2995 msvc_x86_64_copts = ["/arch:AVX"],
2996 x86_srcs = FMA3_UKERNELS,
2997 deps = [
2998 ":tables",
2999 "@FP16",
3000 "@pthreadpool",
3001 ],
3002)
3003
3004xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003005 name = "avx2_ukernels",
3006 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003007 gcc_copts = xnnpack_gcc_std_copts(),
3008 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003009 "-mfma",
3010 "-mavx2",
3011 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003012 msvc_copts = xnnpack_msvc_std_copts(),
3013 msvc_x86_32_copts = ["/arch:AVX2"],
3014 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003015 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003016 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003017 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003018 "@FP16",
3019 "@pthreadpool",
3020 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003021)
3022
3023xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003024 name = "avx2_ukernels_test_mode",
3025 hdrs = INTERNAL_HDRS,
3026 copts = [
3027 "-UNDEBUG",
3028 "-DXNN_TEST_MODE=1",
3029 ],
3030 gcc_copts = xnnpack_gcc_std_copts(),
3031 gcc_x86_copts = [
3032 "-mfma",
3033 "-mavx2",
3034 ],
3035 msvc_copts = xnnpack_msvc_std_copts(),
3036 msvc_x86_32_copts = ["/arch:AVX2"],
3037 msvc_x86_64_copts = ["/arch:AVX2"],
3038 x86_srcs = AVX2_UKERNELS,
3039 deps = [
3040 ":tables",
3041 "@FP16",
3042 "@pthreadpool",
3043 ],
3044)
3045
3046xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003047 name = "avx512f_ukernels",
3048 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003049 gcc_copts = xnnpack_gcc_std_copts(),
3050 gcc_x86_copts = ["-mavx512f"],
3051 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3052 msvc_copts = xnnpack_msvc_std_copts(),
3053 msvc_x86_32_copts = ["/arch:AVX512"],
3054 msvc_x86_64_copts = ["/arch:AVX512"],
3055 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003056 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003057 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003058 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003059 "@FP16",
3060 "@pthreadpool",
3061 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062)
3063
3064xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003065 name = "avx512f_ukernels_test_mode",
3066 hdrs = INTERNAL_HDRS,
3067 copts = [
3068 "-UNDEBUG",
3069 "-DXNN_TEST_MODE=1",
3070 ],
3071 gcc_copts = xnnpack_gcc_std_copts(),
3072 gcc_x86_copts = ["-mavx512f"],
3073 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3074 msvc_copts = xnnpack_msvc_std_copts(),
3075 msvc_x86_32_copts = ["/arch:AVX512"],
3076 msvc_x86_64_copts = ["/arch:AVX512"],
3077 msys_copts = ["-fno-asynchronous-unwind-tables"],
3078 x86_srcs = AVX512F_UKERNELS,
3079 deps = [
3080 ":tables",
3081 "@FP16",
3082 "@pthreadpool",
3083 ],
3084)
3085
3086xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087 name = "asm_ukernels",
3088 hdrs = ["src/xnnpack/assembly.h"],
3089 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003090 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 aarch64_srcs = AARCH64_ASM_UKERNELS,
3092)
3093
Marat Dukhan3b59de22020-06-03 20:15:19 -07003094xnnpack_cc_library(
3095 name = "logging_utils",
3096 srcs = LOGGING_SRCS,
3097 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3098 copts = LOGGING_COPTS + [
3099 "-Isrc",
3100 "-Iinclude",
3101 ] + select({
3102 ":debug_build": [],
3103 "//conditions:default": xnnpack_min_size_copts(),
3104 }),
3105 gcc_copts = xnnpack_gcc_std_copts(),
3106 msvc_copts = xnnpack_msvc_std_copts(),
3107 visibility = xnnpack_visibility(),
3108 deps = [
3109 "@FP16",
3110 "@clog",
3111 "@pthreadpool",
3112 ],
3113)
3114
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115xnnpack_aggregate_library(
3116 name = "ukernels",
3117 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118 ":neon_ukernels",
3119 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003120 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 ":asm_ukernels",
3122 ],
3123 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124 ":neon_ukernels",
3125 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003126 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 ":neonfp16arith_ukernels",
3128 ":asm_ukernels",
3129 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003130 generic_deps = [
3131 ":scalar_ukernels",
3132 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003133 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003134 ":psimd_fastmath_ukernels",
3135 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003137 wasm_deps = [
3138 ":wasm_ukernels",
3139 ],
3140 wasmsimd_deps = [
3141 ":wasm_ukernels",
3142 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003143 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003145 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003146 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003148 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003149 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003150 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 ":avx512f_ukernels",
3152 ],
3153)
3154
Marat Dukhan33fcf782020-05-24 14:27:15 -07003155xnnpack_aggregate_library(
3156 name = "ukernels_test_mode",
3157 aarch32_deps = [
3158 ":neon_ukernels_test_mode",
3159 ":neonfma_ukernels_test_mode",
3160 ":neonv8_ukernels_test_mode",
3161 ":asm_ukernels",
3162 ],
3163 aarch64_deps = [
3164 ":neon_ukernels_test_mode",
3165 ":neonfma_ukernels_test_mode",
3166 ":neonv8_ukernels_test_mode",
3167 ":neonfp16arith_ukernels_test_mode",
3168 ":asm_ukernels",
3169 ],
3170 generic_deps = [
3171 ":scalar_ukernels_test_mode",
3172 ],
3173 psimd_deps = [
3174 ":psimd_fastmath_ukernels_test_mode",
3175 ":psimd_accmath_ukernels_test_mode",
3176 ],
3177 wasm_deps = [
3178 ":wasm_ukernels_test_mode",
3179 ],
3180 wasmsimd_deps = [
3181 ":wasm_ukernels_test_mode",
3182 ],
3183 x86_deps = [
3184 ":sse2_ukernels_test_mode",
3185 ":ssse3_ukernels_test_mode",
3186 ":sse41_ukernels_test_mode",
3187 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003188 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003189 ":fma3_ukernels_test_mode",
3190 ":avx2_ukernels_test_mode",
3191 ":avx512f_ukernels_test_mode",
3192 ],
3193)
3194
Marat Dukhan86df3842020-06-17 14:03:31 -07003195# This target must be compatible with Android NDK r17 and compile with Clang 6.
3196# It is similar to :ukernels target, but excludes microkernels which require
3197# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003198xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003199 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003200 aarch32_deps = [
3201 ":neon_ukernels",
3202 ":neonfma_ukernels",
3203 ":neonv8_ukernels",
3204 ":asm_ukernels",
3205 ],
3206 aarch64_deps = [
3207 ":neon_ukernels",
3208 ":neonfma_ukernels",
3209 ":neonv8_ukernels",
3210 ":asm_ukernels",
3211 ],
3212 generic_deps = [
3213 ":scalar_ukernels",
3214 ],
3215 psimd_deps = [
3216 ":psimd_fastmath_ukernels",
3217 ":psimd_accmath_ukernels",
3218 ],
3219 wasm_deps = [
3220 ":wasm_ukernels",
3221 ],
3222 wasmsimd_deps = [
3223 ":wasm_ukernels",
3224 ],
3225 x86_deps = [
3226 ":sse2_ukernels",
3227 ":ssse3_ukernels",
3228 ":sse41_ukernels",
3229 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003230 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003231 ":fma3_ukernels",
3232 ":avx2_ukernels",
3233 ":avx512f_ukernels",
3234 ],
3235)
3236
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237xnnpack_cc_library(
3238 name = "im2col",
3239 srcs = ["src/im2col.c"],
3240 hdrs = [
3241 "src/xnnpack/common.h",
3242 "src/xnnpack/im2col.h",
3243 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003244 gcc_copts = xnnpack_gcc_std_copts(),
3245 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246)
3247
3248xnnpack_cc_library(
3249 name = "indirection",
3250 srcs = ["src/indirection.c"],
3251 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003252 gcc_copts = xnnpack_gcc_std_copts(),
3253 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 deps = [
3255 "@FP16",
3256 "@FXdiv",
3257 "@pthreadpool",
3258 ],
3259)
3260
3261xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003262 name = "indirection_test_mode",
3263 srcs = ["src/indirection.c"],
3264 hdrs = INTERNAL_HDRS,
3265 copts = [
3266 "-UNDEBUG",
3267 "-DXNN_TEST_MODE=1",
3268 ],
3269 gcc_copts = xnnpack_gcc_std_copts(),
3270 msvc_copts = xnnpack_msvc_std_copts(),
3271 deps = [
3272 "@FP16",
3273 "@FXdiv",
3274 "@pthreadpool",
3275 ],
3276)
3277
3278xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003279 name = "packing",
3280 srcs = ["src/packing.c"],
3281 hdrs = INTERNAL_HDRS,
3282 gcc_copts = xnnpack_gcc_std_copts(),
3283 msvc_copts = xnnpack_msvc_std_copts(),
3284 deps = [
3285 "@FP16",
3286 "@FXdiv",
3287 "@pthreadpool",
3288 ],
3289)
3290
3291xnnpack_cc_library(
3292 name = "packing_test_mode",
3293 srcs = ["src/packing.c"],
3294 hdrs = INTERNAL_HDRS,
3295 copts = [
3296 "-UNDEBUG",
3297 "-DXNN_TEST_MODE=1",
3298 ],
3299 gcc_copts = xnnpack_gcc_std_copts(),
3300 msvc_copts = xnnpack_msvc_std_copts(),
3301 deps = [
3302 "@FP16",
3303 "@FXdiv",
3304 "@pthreadpool",
3305 ],
3306)
3307
3308xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309 name = "operator_run",
3310 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003311 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003312 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003313 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3314 "//conditions:default": [],
3315 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003316 gcc_copts = xnnpack_gcc_std_copts(),
3317 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003318 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003319 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320 "@FP16",
3321 "@FXdiv",
3322 "@clog",
3323 "@pthreadpool",
3324 ],
3325)
3326
Chao Mei6ddfc602020-05-13 22:29:36 -07003327xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003328 name = "operator_run_test_mode",
3329 srcs = ["src/operator-run.c"],
3330 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3331 copts = LOGGING_COPTS + [
3332 "-UNDEBUG",
3333 "-DXNN_TEST_MODE=1",
3334 ] + select({
3335 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3336 "//conditions:default": [],
3337 }),
3338 gcc_copts = xnnpack_gcc_std_copts(),
3339 msvc_copts = xnnpack_msvc_std_copts(),
3340 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003341 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003342 "@FP16",
3343 "@FXdiv",
3344 "@clog",
3345 "@pthreadpool",
3346 ],
3347)
3348
3349xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003350 name = "memory_planner",
3351 srcs = ["src/memory-planner.c"],
3352 hdrs = INTERNAL_HDRS,
3353 defines = select({
3354 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3355 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3356 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3357 }),
3358 gcc_copts = xnnpack_gcc_std_copts(),
3359 msvc_copts = xnnpack_msvc_std_copts(),
3360 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003361 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003362 "@pthreadpool",
3363 ],
3364)
3365
Marat Dukhan33fcf782020-05-24 14:27:15 -07003366xnnpack_cc_library(
3367 name = "memory_planner_test_mode",
3368 srcs = ["src/memory-planner.c"],
3369 hdrs = INTERNAL_HDRS,
3370 copts = [
3371 "-UNDEBUG",
3372 "-DXNN_TEST_MODE=1",
3373 ],
3374 defines = select({
3375 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3376 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3377 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3378 }),
3379 gcc_copts = xnnpack_gcc_std_copts(),
3380 msvc_copts = xnnpack_msvc_std_copts(),
3381 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003382 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003383 "@pthreadpool",
3384 ],
3385)
3386
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387cc_library(
3388 name = "enable_assembly",
3389 defines = select({
3390 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3391 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003392 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 }),
3394)
3395
Marat Dukhan9de90e02020-06-18 16:04:12 -07003396cc_library(
3397 name = "enable_sparse",
3398 defines = select({
3399 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3400 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003401 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003402 }),
3403)
3404
Marat Dukhancf056b22019-10-07 10:26:29 -07003405xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 name = "operators",
3407 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003408 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003410 ],
3411 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003412 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003413 "-Isrc",
3414 "-Iinclude",
3415 ] + select({
3416 ":debug_build": [],
3417 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003418 }) + select({
3419 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3420 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003422 gcc_copts = xnnpack_gcc_std_copts(),
3423 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003424 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003426 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003427 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003428 "@FP16",
3429 "@FXdiv",
3430 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003432 ],
3433)
3434
Marat Dukhan10a38082020-04-17 03:58:35 -07003435xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003436 name = "operators_test_mode",
3437 srcs = OPERATOR_SRCS + [
3438 "src/memory.c",
3439 "src/operator-delete.c",
3440 ],
3441 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3442 copts = LOGGING_COPTS + [
3443 "-Isrc",
3444 "-Iinclude",
3445 "-UNDEBUG",
3446 "-DXNN_TEST_MODE=1",
3447 ] + select({
3448 ":debug_build": [],
3449 "//conditions:default": xnnpack_min_size_copts(),
3450 }) + select({
3451 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3452 "//conditions:default": [],
3453 }),
3454 gcc_copts = xnnpack_gcc_std_copts(),
3455 msvc_copts = xnnpack_msvc_std_copts(),
3456 deps = [
3457 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003458 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003459 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003460 "@FP16",
3461 "@FXdiv",
3462 "@clog",
3463 "@pthreadpool",
3464 ],
3465)
3466
3467xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003468 name = "XNNPACK",
3469 srcs = [
3470 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003471 "src/runtime.c",
3472 "src/subgraph.c",
3473 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003474 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003475 hdrs = ["include/xnnpack.h"],
3476 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003477 "-Isrc",
3478 "-Iinclude",
3479 ] + select({
3480 ":debug_build": [],
3481 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003482 }) + select({
3483 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3484 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003485 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003486 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003487 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003488 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003489 visibility = xnnpack_visibility(),
3490 deps = [
3491 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003492 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003493 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003494 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 ":operator_run",
3496 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003497 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003498 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003499 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003500 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003501 ] + select({
3502 ":emscripten": [],
3503 "//conditions:default": ["@cpuinfo"],
3504 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505)
3506
Marat Dukhan10a38082020-04-17 03:58:35 -07003507xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003508 name = "XNNPACK_test_mode",
3509 srcs = [
3510 "src/init.c",
3511 "src/runtime.c",
3512 "src/subgraph.c",
3513 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003514 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003515 hdrs = ["include/xnnpack.h"],
3516 copts = LOGGING_COPTS + [
3517 "-Isrc",
3518 "-Iinclude",
3519 "-UNDEBUG",
3520 "-DXNN_TEST_MODE=1",
3521 ] + select({
3522 ":debug_build": [],
3523 "//conditions:default": xnnpack_min_size_copts(),
3524 }) + select({
3525 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3526 "//conditions:default": [],
3527 }),
3528 gcc_copts = xnnpack_gcc_std_copts(),
3529 includes = ["include"],
3530 msvc_copts = xnnpack_msvc_std_copts(),
3531 visibility = xnnpack_visibility(),
3532 deps = [
3533 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003534 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003535 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003536 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003537 ":operator_run_test_mode",
3538 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003539 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003540 "@clog",
3541 "@FP16",
3542 "@pthreadpool",
3543 ] + select({
3544 ":emscripten": [],
3545 "//conditions:default": ["@cpuinfo"],
3546 }),
3547)
3548
3549xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003550 name = "xnnpack_f32",
3551 srcs = [
3552 "src/init.c",
3553 "src/runtime.c",
3554 "src/subgraph.c",
3555 "src/tensor.c",
3556 ] + SUBGRAPH_SRCS,
3557 hdrs = ["include/xnnpack.h"],
3558 copts = LOGGING_COPTS + [
3559 "-Isrc",
3560 "-Iinclude",
3561 ] + select({
3562 ":debug_build": [],
3563 "//conditions:default": xnnpack_min_size_copts(),
3564 }) + select({
3565 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3566 "//conditions:default": [],
3567 }),
3568 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003569 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003570 "XNN_NO_U8_OPERATORS",
3571 "XNN_NO_X8_OPERATORS",
3572 "XNN_NO_F16_OPERATORS",
3573 "XNN_NO_X16_OPERATORS",
3574 ],
3575 gcc_copts = xnnpack_gcc_std_copts(),
3576 includes = ["include"],
3577 msvc_copts = xnnpack_msvc_std_copts(),
3578 visibility = xnnpack_visibility(),
3579 deps = [
3580 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003581 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003582 ":logging_utils",
3583 ":memory_planner",
3584 ":operator_run",
3585 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003586 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003587 "@clog",
3588 "@FP16",
3589 "@pthreadpool",
3590 ] + select({
3591 ":emscripten": [],
3592 "//conditions:default": ["@cpuinfo"],
3593 }),
3594)
3595
3596xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003597 name = "xnnpack_operators_nhwc_f32",
3598 srcs = [
3599 "src/init.c",
3600 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003601 hdrs = ["include/xnnpack.h"],
3602 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003603 "-Isrc",
3604 "-Iinclude",
3605 ] + select({
3606 ":debug_build": [],
3607 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003608 }) + select({
3609 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3610 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003611 }),
3612 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003613 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003614 "XNN_NO_U8_OPERATORS",
3615 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003616 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003617 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003618 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003620 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621 visibility = xnnpack_visibility(),
3622 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003623 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003624 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003625 ":operator_run",
3626 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003627 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003628 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003630 ] + select({
3631 ":emscripten": [],
3632 "//conditions:default": ["@cpuinfo"],
3633 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634)
3635
Marat Dukhancf056b22019-10-07 10:26:29 -07003636xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637 name = "bench_utils",
3638 srcs = ["bench/utils.cc"],
3639 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003640 deps = [
3641 "@com_google_benchmark//:benchmark",
3642 "@cpuinfo",
3643 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644)
3645
Frank Barchard7e955972019-10-11 10:34:25 -07003646######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647
3648xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003649 name = "qs8_gemm_bench",
3650 srcs = [
3651 "bench/gemm.h",
3652 "bench/qs8-gemm.cc",
3653 "src/xnnpack/AlignedAllocator.h",
3654 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3655 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3656)
3657
3658xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003659 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 srcs = [
3661 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003662 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 "src/xnnpack/AlignedAllocator.h",
3664 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003665 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003666 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667)
3668
3669xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003670 name = "qu8_requantization_bench",
3671 srcs = [
3672 "bench/qu8-requantization.cc",
3673 "src/xnnpack/requantization-stubs.h",
3674 "src/xnnpack/AlignedAllocator.h",
3675 ] + MICROKERNEL_BENCHMARK_HDRS,
3676 deps = MICROKERNEL_BENCHMARK_DEPS,
3677)
3678
3679xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003680 name = "f16_igemm_bench",
3681 srcs = [
3682 "bench/f16-igemm.cc",
3683 "bench/conv.h",
3684 "bench/google/conv.h",
3685 "src/xnnpack/AlignedAllocator.h",
3686 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003687 deps = MICROKERNEL_BENCHMARK_DEPS + [
3688 ":indirection",
3689 ":packing",
3690 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003691)
3692
3693xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003694 name = "f16_gemm_bench",
3695 srcs = [
3696 "bench/f16-gemm.cc",
3697 "bench/gemm.h",
3698 "src/xnnpack/AlignedAllocator.h",
3699 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003700 deps = MICROKERNEL_BENCHMARK_DEPS + [
3701 ":packing",
3702 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703)
3704
3705xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003706 name = "f16_spmm_bench",
3707 srcs = [
3708 "bench/f16-spmm.cc",
3709 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003710 "src/xnnpack/AlignedAllocator.h",
3711 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003712 deps = MICROKERNEL_BENCHMARK_DEPS,
3713)
3714
3715xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 name = "f32_igemm_bench",
3717 srcs = [
3718 "bench/f32-igemm.cc",
3719 "bench/conv.h",
3720 "src/xnnpack/AlignedAllocator.h",
3721 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003722 deps = MICROKERNEL_BENCHMARK_DEPS + [
3723 ":indirection",
3724 ":packing",
3725 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726)
3727
3728xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003729 name = "f16_relu_bench",
3730 srcs = [
3731 "bench/f16-relu.cc",
3732 "src/xnnpack/AlignedAllocator.h",
3733 ] + MICROKERNEL_BENCHMARK_HDRS,
3734 deps = MICROKERNEL_BENCHMARK_DEPS,
3735)
3736
3737xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738 name = "f32_conv_hwc_bench",
3739 srcs = [
3740 "bench/f32-conv-hwc.cc",
3741 "bench/dconv.h",
3742 "src/xnnpack/AlignedAllocator.h",
3743 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003744 deps = MICROKERNEL_BENCHMARK_DEPS + [
3745 ":packing",
3746 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747)
3748
3749xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003750 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003751 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003752 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003753 "bench/dconv.h",
3754 "src/xnnpack/AlignedAllocator.h",
3755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003756 deps = MICROKERNEL_BENCHMARK_DEPS + [
3757 ":packing",
3758 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003759)
3760
3761xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003762 name = "f16_dwconv_bench",
3763 srcs = [
3764 "bench/f16-dwconv.cc",
3765 "bench/dwconv.h",
3766 "bench/google/dwconv.h",
3767 "src/xnnpack/AlignedAllocator.h",
3768 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003769 deps = MICROKERNEL_BENCHMARK_DEPS + [
3770 ":indirection",
3771 ":packing",
3772 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003773)
3774
3775xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 name = "f32_dwconv_bench",
3777 srcs = [
3778 "bench/f32-dwconv.cc",
3779 "bench/dwconv.h",
3780 "src/xnnpack/AlignedAllocator.h",
3781 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003782 deps = MICROKERNEL_BENCHMARK_DEPS + [
3783 ":indirection",
3784 ":packing",
3785 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786)
3787
3788xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003789 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003791 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003792 "bench/dwconv.h",
3793 "src/xnnpack/AlignedAllocator.h",
3794 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003795 deps = MICROKERNEL_BENCHMARK_DEPS + [
3796 ":indirection",
3797 ":packing",
3798 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799)
3800
3801xnnpack_benchmark(
3802 name = "f32_gemm_bench",
3803 srcs = [
3804 "bench/f32-gemm.cc",
3805 "bench/gemm.h",
3806 "src/xnnpack/AlignedAllocator.h",
3807 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003809 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810)
3811
3812xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003813 name = "f32_hswish_bench",
3814 srcs = [
3815 "bench/f32-hswish.cc",
3816 "src/xnnpack/AlignedAllocator.h",
3817 ] + MICROKERNEL_BENCHMARK_HDRS,
3818 deps = MICROKERNEL_BENCHMARK_DEPS,
3819)
3820
3821xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003822 name = "f32_raddexpminusmax_bench",
3823 srcs = [
3824 "bench/f32-raddexpminusmax.cc",
3825 "src/xnnpack/AlignedAllocator.h",
3826 ] + MICROKERNEL_BENCHMARK_HDRS,
3827 deps = MICROKERNEL_BENCHMARK_DEPS,
3828)
3829
3830xnnpack_benchmark(
3831 name = "f32_raddextexp_bench",
3832 srcs = [
3833 "bench/f32-raddextexp.cc",
3834 "src/xnnpack/AlignedAllocator.h",
3835 ] + MICROKERNEL_BENCHMARK_HDRS,
3836 deps = MICROKERNEL_BENCHMARK_DEPS,
3837)
3838
3839xnnpack_benchmark(
3840 name = "f32_raddstoreexpminusmax_bench",
3841 srcs = [
3842 "bench/f32-raddstoreexpminusmax.cc",
3843 "src/xnnpack/AlignedAllocator.h",
3844 ] + MICROKERNEL_BENCHMARK_HDRS,
3845 deps = MICROKERNEL_BENCHMARK_DEPS,
3846)
3847
3848xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003849 name = "f32_relu_bench",
3850 srcs = [
3851 "bench/f32-relu.cc",
3852 "src/xnnpack/AlignedAllocator.h",
3853 ] + MICROKERNEL_BENCHMARK_HDRS,
3854 deps = MICROKERNEL_BENCHMARK_DEPS,
3855)
3856
3857xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 name = "f32_rmax_bench",
3859 srcs = [
3860 "bench/f32-rmax.cc",
3861 "src/xnnpack/AlignedAllocator.h",
3862 ] + MICROKERNEL_BENCHMARK_HDRS,
3863 deps = MICROKERNEL_BENCHMARK_DEPS,
3864)
3865
3866xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003867 name = "f32_sigmoid_bench",
3868 srcs = [
3869 "bench/f32-sigmoid.cc",
3870 "src/xnnpack/AlignedAllocator.h",
3871 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003872 deps = MICROKERNEL_BENCHMARK_DEPS,
3873)
3874
3875xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876 name = "f32_spmm_bench",
3877 srcs = [
3878 "bench/f32-spmm.cc",
3879 "bench/gemm.h",
3880 "src/xnnpack/AlignedAllocator.h",
3881 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882 deps = MICROKERNEL_BENCHMARK_DEPS,
3883)
3884
3885xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003886 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003887 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003888 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003889 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003890 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003891 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003892)
3893
3894xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003895 name = "f32_vscaleexpminusmax_bench",
3896 srcs = [
3897 "bench/f32-vscaleexpminusmax.cc",
3898 "src/xnnpack/AlignedAllocator.h",
3899 ] + MICROKERNEL_BENCHMARK_HDRS,
3900 deps = MICROKERNEL_BENCHMARK_DEPS,
3901)
3902
3903xnnpack_benchmark(
3904 name = "f32_vscaleextexp_bench",
3905 srcs = [
3906 "bench/f32-vscaleextexp.cc",
3907 "src/xnnpack/AlignedAllocator.h",
3908 ] + MICROKERNEL_BENCHMARK_HDRS,
3909 deps = MICROKERNEL_BENCHMARK_DEPS,
3910)
3911
3912xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003913 name = "f32_vsqrt_bench",
3914 srcs = [
3915 "bench/f32-vsqrt.cc",
3916 "src/xnnpack/AlignedAllocator.h",
3917 ] + MICROKERNEL_BENCHMARK_HDRS,
3918 deps = MICROKERNEL_BENCHMARK_DEPS,
3919)
3920
3921xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 name = "f32_im2col_gemm_bench",
3923 srcs = [
3924 "bench/f32-im2col-gemm.cc",
3925 "bench/conv.h",
3926 "src/xnnpack/AlignedAllocator.h",
3927 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003928 deps = MICROKERNEL_BENCHMARK_DEPS + [
3929 ":im2col",
3930 ":packing",
3931 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932)
3933
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003934xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003935 name = "rounding_bench",
3936 srcs = [
3937 "bench/rounding.cc",
3938 "src/xnnpack/math-stubs.h",
3939 "src/xnnpack/AlignedAllocator.h",
3940 ] + MICROKERNEL_BENCHMARK_HDRS,
3941 deps = MICROKERNEL_BENCHMARK_DEPS,
3942)
3943
Marat Dukhan08c4a432019-10-03 09:29:21 -07003944########################### Benchmarks for operators ###########################
3945
3946xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003947 name = "average_pooling_bench",
3948 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003949 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003950 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003951 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003952)
3953
3954xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003955 name = "bankers_rounding_bench",
3956 srcs = ["bench/bankers-rounding.cc"],
3957 copts = xnnpack_optional_tflite_copts(),
3958 tags = ["nowin32"],
3959 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3960)
3961
3962xnnpack_benchmark(
3963 name = "ceiling_bench",
3964 srcs = ["bench/ceiling.cc"],
3965 copts = xnnpack_optional_tflite_copts(),
3966 tags = ["nowin32"],
3967 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3968)
3969
3970xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 name = "channel_shuffle_bench",
3972 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003973 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974)
3975
3976xnnpack_benchmark(
3977 name = "convolution_bench",
3978 srcs = ["bench/convolution.cc"],
3979 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003980 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003981 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003982)
3983
3984xnnpack_benchmark(
3985 name = "deconvolution_bench",
3986 srcs = ["bench/deconvolution.cc"],
3987 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003988 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003989 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990)
3991
3992xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003993 name = "floor_bench",
3994 srcs = ["bench/floor.cc"],
3995 copts = xnnpack_optional_tflite_copts(),
3996 tags = ["nowin32"],
3997 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3998)
3999
4000xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001 name = "global_average_pooling_bench",
4002 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004003 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004004)
4005
4006xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004007 name = "hardswish_bench",
4008 srcs = ["bench/hardswish.cc"],
4009 copts = xnnpack_optional_tflite_copts(),
4010 tags = ["nowin32"],
4011 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4012)
4013
4014xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004015 name = "max_pooling_bench",
4016 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004017 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004018)
4019
4020xnnpack_benchmark(
4021 name = "sigmoid_bench",
4022 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004023 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004024 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004025 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004026)
4027
4028xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004029 name = "prelu_bench",
4030 srcs = ["bench/prelu.cc"],
4031 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004032 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004033 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004034)
4035
4036xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004037 name = "softmax_bench",
4038 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004039 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004040 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004041 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004042)
4043
Marat Dukhan87727142020-06-24 15:24:10 -07004044xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004045 name = "square_root_bench",
4046 srcs = ["bench/square-root.cc"],
4047 copts = xnnpack_optional_tflite_copts(),
4048 tags = ["nowin32"],
4049 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4050)
4051
4052xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004053 name = "truncation_bench",
4054 srcs = ["bench/truncation.cc"],
4055 deps = OPERATOR_BENCHMARK_DEPS,
4056)
4057
Marat Dukhanc068bb62019-10-04 13:24:39 -07004058############################# End-to-end benchmarks ############################
4059
4060cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004061 name = "fp32_mobilenet_v1",
4062 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004063 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004064 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004065 linkstatic = True,
4066 deps = [
4067 ":XNNPACK",
4068 "@pthreadpool",
4069 ],
4070)
4071
4072cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004073 name = "fp16_mobilenet_v1",
4074 srcs = ["models/fp16-mobilenet-v1.cc"],
4075 hdrs = ["models/models.h"],
4076 copts = xnnpack_std_cxxopts(),
4077 linkstatic = True,
4078 deps = [
4079 ":XNNPACK",
4080 "@FP16",
4081 "@pthreadpool",
4082 ],
4083)
4084
4085cc_library(
4086 name = "fp32_mobilenet_v2",
4087 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004088 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004089 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004090 linkstatic = True,
4091 deps = [
4092 ":XNNPACK",
4093 "@pthreadpool",
4094 ],
4095)
4096
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004097cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004098 name = "fp16_mobilenet_v2",
4099 srcs = ["models/fp16-mobilenet-v2.cc"],
4100 hdrs = ["models/models.h"],
4101 copts = xnnpack_std_cxxopts(),
4102 linkstatic = True,
4103 deps = [
4104 ":XNNPACK",
4105 "@FP16",
4106 "@pthreadpool",
4107 ],
4108)
4109
4110cc_library(
4111 name = "fp32_mobilenet_v3_large",
4112 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004113 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004114 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004115 linkstatic = True,
4116 deps = [
4117 ":XNNPACK",
4118 "@pthreadpool",
4119 ],
4120)
4121
4122cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004123 name = "fp32_mobilenet_v3_small",
4124 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004125 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004126 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004127 linkstatic = True,
4128 deps = [
4129 ":XNNPACK",
4130 "@pthreadpool",
4131 ],
4132)
4133
Marat Dukhanc068bb62019-10-04 13:24:39 -07004134xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004135 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004136 srcs = [
4137 "bench/f32-dwconv-e2e.cc",
4138 "bench/end2end.h",
4139 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004140 deps = MICROKERNEL_BENCHMARK_DEPS + [
4141 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004142 ":fp32_mobilenet_v1",
4143 ":fp32_mobilenet_v2",
4144 ":fp32_mobilenet_v3_large",
4145 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004146 ],
4147)
4148
4149xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004150 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004151 srcs = [
4152 "bench/f32-gemm-e2e.cc",
4153 "bench/end2end.h",
4154 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004155 deps = MICROKERNEL_BENCHMARK_DEPS + [
4156 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004157 ":fp32_mobilenet_v1",
4158 ":fp32_mobilenet_v2",
4159 ":fp32_mobilenet_v3_large",
4160 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004161 ],
4162)
4163
4164xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004165 name = "end2end_bench",
4166 srcs = ["bench/end2end.cc"],
4167 deps = [
4168 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004169 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004170 ":fp16_mobilenet_v1",
4171 ":fp16_mobilenet_v2",
4172 ":fp32_mobilenet_v1",
4173 ":fp32_mobilenet_v2",
4174 ":fp32_mobilenet_v3_large",
4175 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004176 "@pthreadpool",
4177 ],
4178)
4179
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004180#################### Accuracy evaluation for math functions ####################
4181
4182xnnpack_benchmark(
4183 name = "f32_exp_eval",
4184 srcs = [
4185 "eval/f32-exp.cc",
4186 "src/xnnpack/AlignedAllocator.h",
4187 ] + ACCURACY_EVAL_HDRS,
4188 deps = ACCURACY_EVAL_DEPS,
4189)
4190
Marat Dukhan515c9772019-10-17 18:07:57 -07004191xnnpack_benchmark(
4192 name = "f32_expminus_eval",
4193 srcs = [
4194 "eval/f32-expminus.cc",
4195 "src/xnnpack/AlignedAllocator.h",
4196 ] + ACCURACY_EVAL_HDRS,
4197 deps = ACCURACY_EVAL_DEPS,
4198)
4199
Marat Dukhan98ba4412019-10-23 02:14:28 -07004200xnnpack_benchmark(
4201 name = "f32_extexp_eval",
4202 srcs = [
4203 "eval/f32-extexp.cc",
4204 "src/xnnpack/AlignedAllocator.h",
4205 ] + ACCURACY_EVAL_HDRS,
4206 deps = ACCURACY_EVAL_DEPS,
4207)
4208
Marat Dukhan8853b822020-05-07 12:19:01 -07004209xnnpack_unit_test(
4210 name = "f32_roundne_eval",
4211 srcs = [
4212 "eval/f32-roundne.cc",
4213 "src/xnnpack/AlignedAllocator.h",
4214 "src/xnnpack/math-stubs.h",
4215 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004216 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004217 deps = MICROKERNEL_TEST_DEPS,
4218)
4219
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004220xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004221 name = "f32_roundd_eval",
4222 srcs = [
4223 "eval/f32-roundd.cc",
4224 "src/xnnpack/AlignedAllocator.h",
4225 "src/xnnpack/math-stubs.h",
4226 ] + MICROKERNEL_TEST_HDRS,
4227 automatic = False,
4228 deps = MICROKERNEL_TEST_DEPS,
4229)
4230
4231xnnpack_unit_test(
4232 name = "f32_roundu_eval",
4233 srcs = [
4234 "eval/f32-roundu.cc",
4235 "src/xnnpack/AlignedAllocator.h",
4236 "src/xnnpack/math-stubs.h",
4237 ] + MICROKERNEL_TEST_HDRS,
4238 automatic = False,
4239 deps = MICROKERNEL_TEST_DEPS,
4240)
4241
4242xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004243 name = "f32_roundz_eval",
4244 srcs = [
4245 "eval/f32-roundz.cc",
4246 "src/xnnpack/AlignedAllocator.h",
4247 "src/xnnpack/math-stubs.h",
4248 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004249 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004250 deps = MICROKERNEL_TEST_DEPS,
4251)
4252
Marat Dukhan346a9e52019-11-15 09:06:30 -08004253xnnpack_benchmark(
4254 name = "f32_sigmoid_eval",
4255 srcs = [
4256 "eval/f32-sigmoid.cc",
4257 "src/xnnpack/AlignedAllocator.h",
4258 ] + ACCURACY_EVAL_HDRS,
4259 deps = ACCURACY_EVAL_DEPS,
4260)
4261
Marat Dukhan84000762020-06-29 18:38:43 -07004262xnnpack_benchmark(
4263 name = "f32_sqrt_eval",
4264 srcs = [
4265 "eval/f32-sqrt.cc",
4266 "src/xnnpack/AlignedAllocator.h",
4267 ] + ACCURACY_EVAL_HDRS,
4268 deps = ACCURACY_EVAL_DEPS + [
4269 ":bench_utils",
4270 ],
4271)
4272
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273######################### Unit tests for micro-kernels #########################
4274
4275xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004276 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004278 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279 "test/gemm-microkernel-tester.h",
4280 "src/xnnpack/AlignedAllocator.h",
4281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004282 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283)
4284
4285xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004286 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004287 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004288 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004289 "test/spmm-microkernel-tester.h",
4290 "src/xnnpack/AlignedAllocator.h",
4291 ] + MICROKERNEL_TEST_HDRS,
4292 deps = MICROKERNEL_TEST_DEPS,
4293)
4294
4295xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004296 name = "f32_argmaxpool_test",
4297 srcs = [
4298 "test/f32-argmaxpool.cc",
4299 "test/argmaxpool-microkernel-tester.h",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + MICROKERNEL_TEST_HDRS,
4302 deps = MICROKERNEL_TEST_DEPS,
4303)
4304
4305xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004306 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004308 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004309 "test/avgpool-microkernel-tester.h",
4310 "src/xnnpack/AlignedAllocator.h",
4311 ] + MICROKERNEL_TEST_HDRS,
4312 deps = MICROKERNEL_TEST_DEPS,
4313)
4314
4315xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004316 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004317 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004318 "test/f32-ibilinear.cc",
4319 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004320 "src/xnnpack/AlignedAllocator.h",
4321 ] + MICROKERNEL_TEST_HDRS,
4322 deps = MICROKERNEL_TEST_DEPS,
4323)
4324
4325xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004326 name = "f16_clamp_test",
4327 srcs = [
4328 "test/f16-clamp.cc",
4329 "test/clamp-microkernel-tester.h",
4330 ] + MICROKERNEL_TEST_HDRS,
4331 deps = MICROKERNEL_TEST_DEPS,
4332)
4333
4334xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 name = "f32_clamp_test",
4336 srcs = [
4337 "test/f32-clamp.cc",
4338 "test/clamp-microkernel-tester.h",
4339 ] + MICROKERNEL_TEST_HDRS,
4340 deps = MICROKERNEL_TEST_DEPS,
4341)
4342
4343xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004344 name = "f32_igemm_test",
4345 srcs = [
4346 "test/f32-igemm.cc",
4347 "test/gemm-microkernel-tester.h",
4348 "src/xnnpack/AlignedAllocator.h",
4349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004350 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004351)
4352
4353xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004354 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004356 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 "test/gemm-microkernel-tester.h",
4358 "src/xnnpack/AlignedAllocator.h",
4359 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004360 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361)
4362
4363xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004364 name = "f32_igemm_minmax_test",
4365 srcs = [
4366 "test/f32-igemm-minmax.cc",
4367 "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 Dukhane207b7b2020-05-28 16:27:42 -07004371)
4372
4373xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004374 name = "f16_igemm_minmax_test",
4375 srcs = [
4376 "test/f16-igemm-minmax.cc",
4377 "test/gemm-microkernel-tester.h",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004380 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004381)
4382
4383xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_conv_hwc_test",
4385 srcs = [
4386 "test/f32-conv-hwc.cc",
4387 "test/conv-hwc-microkernel-tester.h",
4388 "src/xnnpack/AlignedAllocator.h",
4389 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004390 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391)
4392
4393xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004394 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004396 "test/f32-conv-hwc2chw.cc",
4397 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 "src/xnnpack/AlignedAllocator.h",
4399 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004400 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401)
4402
4403xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004404 name = "f32_dwconv_test",
4405 srcs = [
4406 "test/f32-dwconv.cc",
4407 "test/dwconv-microkernel-tester.h",
4408 "src/xnnpack/AlignedAllocator.h",
4409 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004410 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004411)
4412
4413xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004414 name = "f16_dwconv_minmax_test",
4415 srcs = [
4416 "test/f16-dwconv-minmax.cc",
4417 "test/dwconv-microkernel-tester.h",
4418 "src/xnnpack/AlignedAllocator.h",
4419 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004420 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004421)
4422
4423xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004424 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004425 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004426 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004427 "test/dwconv-microkernel-tester.h",
4428 "src/xnnpack/AlignedAllocator.h",
4429 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004430 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004431)
4432
4433xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004434 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004436 "test/f32-dwconv-chw.cc",
4437 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438 "src/xnnpack/AlignedAllocator.h",
4439 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004440 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441)
4442
4443xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004444 name = "f16_gavgpool_minmax_test",
4445 srcs = [
4446 "test/f16-gavgpool-minmax.cc",
4447 "test/gavgpool-microkernel-tester.h",
4448 "src/xnnpack/AlignedAllocator.h",
4449 ] + MICROKERNEL_TEST_HDRS,
4450 deps = MICROKERNEL_TEST_DEPS,
4451)
4452
4453xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004454 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004455 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004456 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004457 "test/gavgpool-microkernel-tester.h",
4458 "src/xnnpack/AlignedAllocator.h",
4459 ] + MICROKERNEL_TEST_HDRS,
4460 deps = MICROKERNEL_TEST_DEPS,
4461)
4462
4463xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004464 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004466 "test/f32-gavgpool-cw.cc",
4467 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004468 "src/xnnpack/AlignedAllocator.h",
4469 ] + MICROKERNEL_TEST_HDRS,
4470 deps = MICROKERNEL_TEST_DEPS,
4471)
4472
4473xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004474 name = "f32_gemm_test",
4475 srcs = [
4476 "test/f32-gemm.cc",
4477 "test/gemm-microkernel-tester.h",
4478 "src/xnnpack/AlignedAllocator.h",
4479 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004480 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004481)
4482
4483xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004484 name = "f32_gemm_relu_test",
4485 srcs = [
4486 "test/f32-gemm-relu.cc",
4487 "test/gemm-microkernel-tester.h",
4488 "src/xnnpack/AlignedAllocator.h",
4489 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004490 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004491)
4492
4493xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004494 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004496 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 "test/gemm-microkernel-tester.h",
4498 "src/xnnpack/AlignedAllocator.h",
4499 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004500 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501)
4502
4503xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004504 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004506 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507 "test/gemm-microkernel-tester.h",
4508 "src/xnnpack/AlignedAllocator.h",
4509 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004510 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511)
4512
4513xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004514 name = "f16_hswish_test",
4515 srcs = [
4516 "test/f16-hswish.cc",
4517 "test/hswish-microkernel-tester.h",
4518 ] + MICROKERNEL_TEST_HDRS,
4519 deps = MICROKERNEL_TEST_DEPS,
4520)
4521
4522xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004523 name = "f32_hswish_test",
4524 srcs = [
4525 "test/f32-hswish.cc",
4526 "test/hswish-microkernel-tester.h",
4527 ] + MICROKERNEL_TEST_HDRS,
4528 deps = MICROKERNEL_TEST_DEPS,
4529)
4530
4531xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004532 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004533 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004534 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535 "test/maxpool-microkernel-tester.h",
4536 ] + MICROKERNEL_TEST_HDRS,
4537 deps = MICROKERNEL_TEST_DEPS,
4538)
4539
4540xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004541 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004542 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004543 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544 "test/avgpool-microkernel-tester.h",
4545 "src/xnnpack/AlignedAllocator.h",
4546 ] + MICROKERNEL_TEST_HDRS,
4547 deps = MICROKERNEL_TEST_DEPS,
4548)
4549
4550xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004551 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004553 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 "test/gemm-microkernel-tester.h",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004557 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004558)
4559
4560xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004561 name = "f16_prelu_test",
4562 srcs = [
4563 "test/f16-prelu.cc",
4564 "test/prelu-microkernel-tester.h",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_TEST_HDRS,
4567 deps = MICROKERNEL_TEST_DEPS,
4568)
4569
4570xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 name = "f32_prelu_test",
4572 srcs = [
4573 "test/f32-prelu.cc",
4574 "test/prelu-microkernel-tester.h",
4575 "src/xnnpack/AlignedAllocator.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 deps = MICROKERNEL_TEST_DEPS,
4578)
4579
4580xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004581 name = "f32_raddexpminusmax_test",
4582 srcs = [
4583 "test/f32-raddexpminusmax.cc",
4584 "test/raddexpminusmax-microkernel-tester.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004590 name = "f32_raddextexp_test",
4591 srcs = [
4592 "test/f32-raddextexp.cc",
4593 "test/raddextexp-microkernel-tester.h",
4594 ] + MICROKERNEL_TEST_HDRS,
4595 deps = MICROKERNEL_TEST_DEPS,
4596)
4597
4598xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004599 name = "f32_raddstoreexpminusmax_test",
4600 srcs = [
4601 "test/f32-raddstoreexpminusmax.cc",
4602 "test/raddstoreexpminusmax-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004608 name = "f32_relu_test",
4609 srcs = [
4610 "test/f32-relu.cc",
4611 "test/vunary-microkernel-tester.h",
4612 ] + MICROKERNEL_TEST_HDRS,
4613 deps = MICROKERNEL_TEST_DEPS,
4614)
4615
4616xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004617 name = "f32_rmax_test",
4618 srcs = [
4619 "test/f32-rmax.cc",
4620 "test/rmax-microkernel-tester.h",
4621 ] + MICROKERNEL_TEST_HDRS,
4622 deps = MICROKERNEL_TEST_DEPS,
4623)
4624
4625xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004626 name = "f32_sigmoid_test",
4627 srcs = [
4628 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004629 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004630 ] + MICROKERNEL_TEST_HDRS,
4631 deps = MICROKERNEL_TEST_DEPS,
4632)
4633
4634xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004635 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004637 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004638 "test/spmm-microkernel-tester.h",
4639 "src/xnnpack/AlignedAllocator.h",
4640 ] + MICROKERNEL_TEST_HDRS,
4641 deps = MICROKERNEL_TEST_DEPS,
4642)
4643
4644xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004645 name = "f16_vadd_minmax_test",
4646 srcs = [
4647 "test/f16-vadd-minmax.cc",
4648 "test/vbinary-microkernel-tester.h",
4649 ] + MICROKERNEL_TEST_HDRS,
4650 deps = MICROKERNEL_TEST_DEPS,
4651)
4652
4653xnnpack_unit_test(
4654 name = "f16_vaddc_minmax_test",
4655 srcs = [
4656 "test/f16-vaddc-minmax.cc",
4657 "test/vbinaryc-microkernel-tester.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
4663 name = "f16_vdiv_minmax_test",
4664 srcs = [
4665 "test/f16-vdiv-minmax.cc",
4666 "test/vbinary-microkernel-tester.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
4672 name = "f16_vdivc_minmax_test",
4673 srcs = [
4674 "test/f16-vdivc-minmax.cc",
4675 "test/vbinaryc-microkernel-tester.h",
4676 ] + MICROKERNEL_TEST_HDRS,
4677 deps = MICROKERNEL_TEST_DEPS,
4678)
4679
4680xnnpack_unit_test(
4681 name = "f16_vrdivc_minmax_test",
4682 srcs = [
4683 "test/f16-vrdivc-minmax.cc",
4684 "test/vbinaryc-microkernel-tester.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
4690 name = "f16_vmax_test",
4691 srcs = [
4692 "test/f16-vmax.cc",
4693 "test/vbinary-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
4699 name = "f16_vmaxc_test",
4700 srcs = [
4701 "test/f16-vmaxc.cc",
4702 "test/vbinaryc-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
4708 name = "f16_vmin_test",
4709 srcs = [
4710 "test/f16-vmin.cc",
4711 "test/vbinary-microkernel-tester.h",
4712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
4717 name = "f16_vminc_test",
4718 srcs = [
4719 "test/f16-vminc.cc",
4720 "test/vbinaryc-microkernel-tester.h",
4721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
4726 name = "f16_vmul_minmax_test",
4727 srcs = [
4728 "test/f16-vmul-minmax.cc",
4729 "test/vbinary-microkernel-tester.h",
4730 ] + MICROKERNEL_TEST_HDRS,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
4735 name = "f16_vmulc_minmax_test",
4736 srcs = [
4737 "test/f16-vmulc-minmax.cc",
4738 "test/vbinaryc-microkernel-tester.h",
4739 ] + MICROKERNEL_TEST_HDRS,
4740 deps = MICROKERNEL_TEST_DEPS,
4741)
4742
4743xnnpack_unit_test(
4744 name = "f16_vsub_minmax_test",
4745 srcs = [
4746 "test/f16-vsub-minmax.cc",
4747 "test/vbinary-microkernel-tester.h",
4748 ] + MICROKERNEL_TEST_HDRS,
4749 deps = MICROKERNEL_TEST_DEPS,
4750)
4751
4752xnnpack_unit_test(
4753 name = "f16_vsubc_minmax_test",
4754 srcs = [
4755 "test/f16-vsubc-minmax.cc",
4756 "test/vbinaryc-microkernel-tester.h",
4757 ] + MICROKERNEL_TEST_HDRS,
4758 deps = MICROKERNEL_TEST_DEPS,
4759)
4760
4761xnnpack_unit_test(
4762 name = "f16_vrsubc_minmax_test",
4763 srcs = [
4764 "test/f16-vrsubc-minmax.cc",
4765 "test/vbinaryc-microkernel-tester.h",
4766 ] + MICROKERNEL_TEST_HDRS,
4767 deps = MICROKERNEL_TEST_DEPS,
4768)
4769
4770xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004771 name = "f32_vabs_test",
4772 srcs = [
4773 "test/f32-vabs.cc",
4774 "test/vunary-microkernel-tester.h",
4775 ] + MICROKERNEL_TEST_HDRS,
4776 deps = MICROKERNEL_TEST_DEPS,
4777)
4778
4779xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004780 name = "f32_vadd_test",
4781 srcs = [
4782 "test/f32-vadd.cc",
4783 "test/vbinary-microkernel-tester.h",
4784 ] + MICROKERNEL_TEST_HDRS,
4785 deps = MICROKERNEL_TEST_DEPS,
4786)
4787
4788xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004789 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004790 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004791 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004792 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004793 ] + MICROKERNEL_TEST_HDRS,
4794 deps = MICROKERNEL_TEST_DEPS,
4795)
4796
4797xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004798 name = "f32_vaddc_test",
4799 srcs = [
4800 "test/f32-vaddc.cc",
4801 "test/vbinaryc-microkernel-tester.h",
4802 ] + MICROKERNEL_TEST_HDRS,
4803 deps = MICROKERNEL_TEST_DEPS,
4804)
4805
4806xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004807 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004808 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004809 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004810 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004811 ] + MICROKERNEL_TEST_HDRS,
4812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
4815xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004816 name = "f32_vdiv_test",
4817 srcs = [
4818 "test/f32-vdiv.cc",
4819 "test/vbinary-microkernel-tester.h",
4820 ] + MICROKERNEL_TEST_HDRS,
4821 deps = MICROKERNEL_TEST_DEPS,
4822)
4823
4824xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004825 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004826 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004827 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004828 "test/vbinary-microkernel-tester.h",
4829 ] + MICROKERNEL_TEST_HDRS,
4830 deps = MICROKERNEL_TEST_DEPS,
4831)
4832
4833xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004834 name = "f32_vdivc_test",
4835 srcs = [
4836 "test/f32-vdivc.cc",
4837 "test/vbinaryc-microkernel-tester.h",
4838 ] + MICROKERNEL_TEST_HDRS,
4839 deps = MICROKERNEL_TEST_DEPS,
4840)
4841
4842xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004843 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004844 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004845 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004846 "test/vbinaryc-microkernel-tester.h",
4847 ] + MICROKERNEL_TEST_HDRS,
4848 deps = MICROKERNEL_TEST_DEPS,
4849)
4850
4851xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004852 name = "f32_vrdivc_test",
4853 srcs = [
4854 "test/f32-vrdivc.cc",
4855 "test/vbinaryc-microkernel-tester.h",
4856 ] + MICROKERNEL_TEST_HDRS,
4857 deps = MICROKERNEL_TEST_DEPS,
4858)
4859
4860xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004861 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004862 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004863 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004864 "test/vbinaryc-microkernel-tester.h",
4865 ] + MICROKERNEL_TEST_HDRS,
4866 deps = MICROKERNEL_TEST_DEPS,
4867)
4868
4869xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004870 name = "f32_vmax_test",
4871 srcs = [
4872 "test/f32-vmax.cc",
4873 "test/vbinary-microkernel-tester.h",
4874 ] + MICROKERNEL_TEST_HDRS,
4875 deps = MICROKERNEL_TEST_DEPS,
4876)
4877
4878xnnpack_unit_test(
4879 name = "f32_vmaxc_test",
4880 srcs = [
4881 "test/f32-vmaxc.cc",
4882 "test/vbinaryc-microkernel-tester.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 deps = MICROKERNEL_TEST_DEPS,
4885)
4886
4887xnnpack_unit_test(
4888 name = "f32_vmin_test",
4889 srcs = [
4890 "test/f32-vmin.cc",
4891 "test/vbinary-microkernel-tester.h",
4892 ] + MICROKERNEL_TEST_HDRS,
4893 deps = MICROKERNEL_TEST_DEPS,
4894)
4895
4896xnnpack_unit_test(
4897 name = "f32_vminc_test",
4898 srcs = [
4899 "test/f32-vminc.cc",
4900 "test/vbinaryc-microkernel-tester.h",
4901 ] + MICROKERNEL_TEST_HDRS,
4902 deps = MICROKERNEL_TEST_DEPS,
4903)
4904
4905xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004906 name = "f32_vmul_test",
4907 srcs = [
4908 "test/f32-vmul.cc",
4909 "test/vbinary-microkernel-tester.h",
4910 ] + MICROKERNEL_TEST_HDRS,
4911 deps = MICROKERNEL_TEST_DEPS,
4912)
4913
4914xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004915 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004916 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004917 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004918 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004919 ] + MICROKERNEL_TEST_HDRS,
4920 deps = MICROKERNEL_TEST_DEPS,
4921)
4922
4923xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004924 name = "f32_vmulc_test",
4925 srcs = [
4926 "test/f32-vmulc.cc",
4927 "test/vbinaryc-microkernel-tester.h",
4928 ] + MICROKERNEL_TEST_HDRS,
4929 deps = MICROKERNEL_TEST_DEPS,
4930)
4931
4932xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004933 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004934 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004935 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004936 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937 ] + MICROKERNEL_TEST_HDRS,
4938 deps = MICROKERNEL_TEST_DEPS,
4939)
4940
4941xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004942 name = "f16_vmulcaddc_minmax_test",
4943 srcs = [
4944 "test/f16-vmulcaddc-minmax.cc",
4945 "test/vmulcaddc-microkernel-tester.h",
4946 "src/xnnpack/AlignedAllocator.h",
4947 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004948 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004949)
4950
4951xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004952 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004953 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004954 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004955 "test/vmulcaddc-microkernel-tester.h",
4956 "src/xnnpack/AlignedAllocator.h",
4957 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004958 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004959)
4960
4961xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004962 name = "f32_vlrelu_test",
4963 srcs = [
4964 "test/f32-vlrelu.cc",
4965 "test/vunary-microkernel-tester.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004971 name = "f32_vneg_test",
4972 srcs = [
4973 "test/f32-vneg.cc",
4974 "test/vunary-microkernel-tester.h",
4975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004980 name = "f32_vrndne_test",
4981 srcs = [
4982 "test/f32-vrndne.cc",
4983 "test/vunary-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
4989 name = "f32_vrndz_test",
4990 srcs = [
4991 "test/f32-vrndz.cc",
4992 "test/vunary-microkernel-tester.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
4998 name = "f32_vrndu_test",
4999 srcs = [
5000 "test/f32-vrndu.cc",
5001 "test/vunary-microkernel-tester.h",
5002 ] + MICROKERNEL_TEST_HDRS,
5003 deps = MICROKERNEL_TEST_DEPS,
5004)
5005
5006xnnpack_unit_test(
5007 name = "f32_vrndd_test",
5008 srcs = [
5009 "test/f32-vrndd.cc",
5010 "test/vunary-microkernel-tester.h",
5011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005016 name = "f32_vscale_test",
5017 srcs = [
5018 "test/f32-vscale.cc",
5019 "test/vscale-microkernel-tester.h",
5020 ] + MICROKERNEL_TEST_HDRS,
5021 deps = MICROKERNEL_TEST_DEPS,
5022)
5023
5024xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005025 name = "f32_vscaleexpminusmax_test",
5026 srcs = [
5027 "test/f32-vscaleexpminusmax.cc",
5028 "test/vscaleexpminusmax-microkernel-tester.h",
5029 ] + MICROKERNEL_TEST_HDRS,
5030 deps = MICROKERNEL_TEST_DEPS,
5031)
5032
5033xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005034 name = "f32_vscaleextexp_test",
5035 srcs = [
5036 "test/f32-vscaleextexp.cc",
5037 "test/vscaleextexp-microkernel-tester.h",
5038 ] + MICROKERNEL_TEST_HDRS,
5039 deps = MICROKERNEL_TEST_DEPS,
5040)
5041
5042xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005043 name = "f32_vsqr_test",
5044 srcs = [
5045 "test/f32-vsqr.cc",
5046 "test/vunary-microkernel-tester.h",
5047 ] + MICROKERNEL_TEST_HDRS,
5048 deps = MICROKERNEL_TEST_DEPS,
5049)
5050
5051xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005052 name = "f32_vsqrdiff_test",
5053 srcs = [
5054 "test/f32-vsqrdiff.cc",
5055 "test/vbinary-microkernel-tester.h",
5056 ] + MICROKERNEL_TEST_HDRS,
5057 deps = MICROKERNEL_TEST_DEPS,
5058)
5059
5060xnnpack_unit_test(
5061 name = "f32_vsqrdiffc_test",
5062 srcs = [
5063 "test/f32-vsqrdiffc.cc",
5064 "test/vbinaryc-microkernel-tester.h",
5065 ] + MICROKERNEL_TEST_HDRS,
5066 deps = MICROKERNEL_TEST_DEPS,
5067)
5068
5069xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005070 name = "f32_vsqrt_test",
5071 srcs = [
5072 "test/f32-vsqrt.cc",
5073 "test/vunary-microkernel-tester.h",
5074 ] + MICROKERNEL_TEST_HDRS,
5075 deps = MICROKERNEL_TEST_DEPS,
5076)
5077
5078xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005079 name = "f32_vsub_test",
5080 srcs = [
5081 "test/f32-vsub.cc",
5082 "test/vbinary-microkernel-tester.h",
5083 ] + MICROKERNEL_TEST_HDRS,
5084 deps = MICROKERNEL_TEST_DEPS,
5085)
5086
5087xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005088 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005089 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005090 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005091 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005092 ] + MICROKERNEL_TEST_HDRS,
5093 deps = MICROKERNEL_TEST_DEPS,
5094)
5095
5096xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005097 name = "f32_vsubc_test",
5098 srcs = [
5099 "test/f32-vsubc.cc",
5100 "test/vbinaryc-microkernel-tester.h",
5101 ] + MICROKERNEL_TEST_HDRS,
5102 deps = MICROKERNEL_TEST_DEPS,
5103)
5104
5105xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005106 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005107 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005108 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005109 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005115 name = "f32_vrsubc_test",
5116 srcs = [
5117 "test/f32-vrsubc.cc",
5118 "test/vbinaryc-microkernel-tester.h",
5119 ] + MICROKERNEL_TEST_HDRS,
5120 deps = MICROKERNEL_TEST_DEPS,
5121)
5122
5123xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005124 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005125 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005126 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005127 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005128 ] + MICROKERNEL_TEST_HDRS,
5129 deps = MICROKERNEL_TEST_DEPS,
5130)
5131
5132xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005133 name = "qs8_gemm_minmax_test",
5134 srcs = [
5135 "test/qs8-gemm-minmax.cc",
5136 "test/gemm-microkernel-tester.h",
5137 "src/xnnpack/AlignedAllocator.h",
5138 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5140)
5141
5142xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005143 name = "qs8_igemm_minmax_test",
5144 srcs = [
5145 "test/qs8-igemm-minmax.cc",
5146 "test/gemm-microkernel-tester.h",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5149 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5150)
5151
5152xnnpack_unit_test(
5153 name = "qs8_requantization_test",
5154 srcs = [
5155 "src/xnnpack/requantization-stubs.h",
5156 "test/qs8-requantization.cc",
5157 "test/requantization-tester.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005163 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005165 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005166 "test/avgpool-microkernel-tester.h",
5167 "src/xnnpack/AlignedAllocator.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005173 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005175 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005176 "test/gemm-microkernel-tester.h",
5177 "src/xnnpack/AlignedAllocator.h",
5178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005179 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180)
5181
5182xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005183 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005184 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005185 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005186 "test/dwconv-microkernel-tester.h",
5187 "src/xnnpack/AlignedAllocator.h",
5188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005189 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005193 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005195 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 "test/gavgpool-microkernel-tester.h",
5197 "src/xnnpack/AlignedAllocator.h",
5198 ] + MICROKERNEL_TEST_HDRS,
5199 deps = MICROKERNEL_TEST_DEPS,
5200)
5201
5202xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005203 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005205 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 "test/gemm-microkernel-tester.h",
5207 "src/xnnpack/AlignedAllocator.h",
5208 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005209 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210)
5211
5212xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005213 name = "qu8_requantization_test",
5214 srcs = [
5215 "src/xnnpack/requantization-stubs.h",
5216 "test/qu8-requantization.cc",
5217 "test/requantization-tester.h",
5218 ] + MICROKERNEL_TEST_HDRS,
5219 deps = MICROKERNEL_TEST_DEPS,
5220)
5221
5222xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005223 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005225 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226 "test/vadd-microkernel-tester.h",
5227 ] + MICROKERNEL_TEST_HDRS,
5228 deps = MICROKERNEL_TEST_DEPS,
5229)
5230
5231xnnpack_unit_test(
5232 name = "u8_clamp_test",
5233 srcs = [
5234 "test/u8-clamp.cc",
5235 "test/clamp-microkernel-tester.h",
5236 ] + MICROKERNEL_TEST_HDRS,
5237 deps = MICROKERNEL_TEST_DEPS,
5238)
5239
5240xnnpack_unit_test(
5241 name = "u8_lut32norm_test",
5242 srcs = [
5243 "test/u8-lut32norm.cc",
5244 "test/lut-norm-microkernel-tester.h",
5245 ] + MICROKERNEL_TEST_HDRS,
5246 deps = MICROKERNEL_TEST_DEPS,
5247)
5248
5249xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005250 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005252 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 "test/maxpool-microkernel-tester.h",
5254 ] + MICROKERNEL_TEST_HDRS,
5255 deps = MICROKERNEL_TEST_DEPS,
5256)
5257
5258xnnpack_unit_test(
5259 name = "u8_rmax_test",
5260 srcs = [
5261 "test/u8-rmax.cc",
5262 "test/rmax-microkernel-tester.h",
5263 ] + MICROKERNEL_TEST_HDRS,
5264 deps = MICROKERNEL_TEST_DEPS,
5265)
5266
5267xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005268 name = "x32_fill_test",
5269 srcs = [
5270 "test/x32-fill.cc",
5271 "test/fill-microkernel-tester.h",
5272 ] + MICROKERNEL_TEST_HDRS,
5273 deps = MICROKERNEL_TEST_DEPS,
5274)
5275
5276xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005277 name = "x32_packx_test",
5278 srcs = [
5279 "test/x32-packx.cc",
5280 "test/pack-microkernel-tester.h",
5281 "src/xnnpack/AlignedAllocator.h",
5282 ] + MICROKERNEL_TEST_HDRS,
5283 deps = MICROKERNEL_TEST_DEPS,
5284)
5285
5286xnnpack_unit_test(
5287 name = "x32_pad_test",
5288 srcs = [
5289 "test/x32-pad.cc",
5290 "test/pad-microkernel-tester.h",
5291 ] + MICROKERNEL_TEST_HDRS,
5292 deps = MICROKERNEL_TEST_DEPS,
5293)
5294
5295xnnpack_unit_test(
5296 name = "x32_unpool_test",
5297 srcs = [
5298 "test/x32-unpool.cc",
5299 "test/unpool-microkernel-tester.h",
5300 ] + MICROKERNEL_TEST_HDRS,
5301 deps = MICROKERNEL_TEST_DEPS,
5302)
5303
5304xnnpack_unit_test(
5305 name = "x32_zip_test",
5306 srcs = [
5307 "test/x32-zip.cc",
5308 "test/zip-microkernel-tester.h",
5309 ] + MICROKERNEL_TEST_HDRS,
5310 deps = MICROKERNEL_TEST_DEPS,
5311)
5312
5313xnnpack_unit_test(
5314 name = "x8_lut_test",
5315 srcs = [
5316 "test/x8-lut.cc",
5317 "test/lut-microkernel-tester.h",
5318 ] + MICROKERNEL_TEST_HDRS,
5319 deps = MICROKERNEL_TEST_DEPS,
5320)
5321
5322xnnpack_unit_test(
5323 name = "x8_zip_test",
5324 srcs = [
5325 "test/x8-zip.cc",
5326 "test/zip-microkernel-tester.h",
5327 ] + MICROKERNEL_TEST_HDRS,
5328 deps = MICROKERNEL_TEST_DEPS,
5329)
5330
Marat Dukhan20c3b922020-03-10 03:45:06 -07005331########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332
5333xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005334 name = "operator_size_test",
5335 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005336 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337)
5338
Marat Dukhan20c3b922020-03-10 03:45:06 -07005339xnnpack_binary(
5340 name = "subgraph_size_test",
5341 srcs = ["test/subgraph-size.c"],
5342 deps = [":XNNPACK"],
5343)
5344
5345########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346
5347xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005348 name = "abs_nc_test",
5349 srcs = [
5350 "test/abs-nc.cc",
5351 "test/abs-operator-tester.h",
5352 ],
5353 deps = OPERATOR_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005357 name = "add_nd_test",
5358 srcs = [
5359 "test/add-nd.cc",
5360 "test/binary-elementwise-operator-tester.h",
5361 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005362 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005363)
5364
5365xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005366 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005368 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005369 "test/argmax-pooling-operator-tester.h",
5370 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005371 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372)
5373
5374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005375 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005377 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005378 "test/average-pooling-operator-tester.h",
5379 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381)
5382
5383xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005384 name = "bankers_rounding_nc_test",
5385 srcs = [
5386 "test/bankers-rounding-nc.cc",
5387 "test/bankers-rounding-operator-tester.h",
5388 ],
5389 deps = OPERATOR_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
5393 name = "ceiling_nc_test",
5394 srcs = [
5395 "test/ceiling-nc.cc",
5396 "test/ceiling-operator-tester.h",
5397 ],
5398 deps = OPERATOR_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005402 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005404 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405 "test/channel-shuffle-operator-tester.h",
5406 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005407 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408)
5409
5410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005411 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005413 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414 "test/clamp-operator-tester.h",
5415 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005420 name = "constant_pad_nd_test",
5421 srcs = [
5422 "test/constant-pad-nd.cc",
5423 "test/constant-pad-operator-tester.h",
5424 ],
5425 deps = OPERATOR_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005429 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005431 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005432 "test/convolution-operator-tester.h",
5433 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005434 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435)
5436
5437xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005438 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005440 "test/convolution-nchw.cc",
5441 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005443 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005444)
5445
5446xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005447 name = "copy_nc_test",
5448 srcs = [
5449 "test/copy-nc.cc",
5450 "test/copy-operator-tester.h",
5451 ],
5452 deps = OPERATOR_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005456 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005458 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005459 "test/deconvolution-operator-tester.h",
5460 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462)
5463
5464xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005465 name = "divide_nd_test",
5466 srcs = [
5467 "test/binary-elementwise-operator-tester.h",
5468 "test/divide-nd.cc",
5469 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005471)
5472
5473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005474 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005476 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477 "test/fully-connected-operator-tester.h",
5478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480)
5481
5482xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005483 name = "floor_nc_test",
5484 srcs = [
5485 "test/floor-nc.cc",
5486 "test/floor-operator-tester.h",
5487 ],
5488 deps = OPERATOR_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005492 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005493 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005494 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005496 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005497 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498)
5499
5500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005501 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005503 "test/global-average-pooling-ncw.cc",
5504 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005507)
5508
5509xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005510 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005511 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005512 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005513 "test/hardswish-operator-tester.h",
5514 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005515 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005516)
5517
5518xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005519 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005520 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005521 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005522 "test/leaky-relu-operator-tester.h",
5523 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005524 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005525)
5526
5527xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005528 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005529 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005530 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005531 "test/max-pooling-operator-tester.h",
5532 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005533 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005534)
5535
5536xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005537 name = "maximum_nd_test",
5538 srcs = [
5539 "test/binary-elementwise-operator-tester.h",
5540 "test/maximum-nd.cc",
5541 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005542 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005543)
5544
5545xnnpack_unit_test(
5546 name = "minimum_nd_test",
5547 srcs = [
5548 "test/binary-elementwise-operator-tester.h",
5549 "test/minimum-nd.cc",
5550 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005551 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005552)
5553
5554xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005555 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005556 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005557 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005558 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005559 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005560 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005561)
5562
5563xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005564 name = "negate_nc_test",
5565 srcs = [
5566 "test/negate-nc.cc",
5567 "test/negate-operator-tester.h",
5568 ],
5569 deps = OPERATOR_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005573 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005574 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005575 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576 "test/prelu-operator-tester.h",
5577 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005578 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005579)
5580
5581xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005582 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005583 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005584 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005585 "test/resize-bilinear-operator-tester.h",
5586 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005587 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005588)
5589
5590xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005591 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005592 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005593 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005594 "test/sigmoid-operator-tester.h",
5595 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005596 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597)
5598
5599xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005600 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005601 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005602 "test/softmax-nc.cc",
5603 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005604 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005605 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005606)
5607
5608xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005609 name = "square_nc_test",
5610 srcs = [
5611 "test/square-nc.cc",
5612 "test/square-operator-tester.h",
5613 ],
5614 deps = OPERATOR_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005618 name = "square_root_nc_test",
5619 srcs = [
5620 "test/square-root-nc.cc",
5621 "test/square-root-operator-tester.h",
5622 ],
5623 deps = OPERATOR_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005627 name = "squared_difference_nd_test",
5628 srcs = [
5629 "test/binary-elementwise-operator-tester.h",
5630 "test/squared-difference-nd.cc",
5631 ],
5632 deps = OPERATOR_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005636 name = "subtract_nd_test",
5637 srcs = [
5638 "test/binary-elementwise-operator-tester.h",
5639 "test/subtract-nd.cc",
5640 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005641 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005642)
5643
5644xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005645 name = "truncation_nc_test",
5646 srcs = [
5647 "test/truncation-nc.cc",
5648 "test/truncation-operator-tester.h",
5649 ],
5650 deps = OPERATOR_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005654 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005656 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005657 "test/unpooling-operator-tester.h",
5658 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005659 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005660)
5661
Chao Mei6ddfc602020-05-13 22:29:36 -07005662############################### Misc unit tests ###############################
5663
5664xnnpack_unit_test(
5665 name = "memory_planner_test",
5666 srcs = [
5667 "test/memory-planner-test.cc",
5668 ],
5669 deps = [
5670 ":XNNPACK",
5671 ":memory_planner",
5672 ],
5673)
5674
Marat Dukhan08c4a432019-10-03 09:29:21 -07005675############################# Build configurations #############################
5676
Marat Dukhanb8642352019-10-30 15:43:02 -07005677# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005678config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005679 name = "xnn_enable_assembly_explicit_true",
5680 define_values = {"xnn_enable_assembly": "true"},
5681)
5682
5683# Disables usage of assembly kernels.
5684config_setting(
5685 name = "xnn_enable_assembly_explicit_false",
5686 define_values = {"xnn_enable_assembly": "false"},
5687)
5688
Marat Dukhan9de90e02020-06-18 16:04:12 -07005689# Enables usage of sparse inference.
5690config_setting(
5691 name = "xnn_enable_sparse_explicit_true",
5692 define_values = {"xnn_enable_sparse": "true"},
5693)
5694
5695# Disables usage of sparse inference.
5696config_setting(
5697 name = "xnn_enable_sparse_explicit_false",
5698 define_values = {"xnn_enable_sparse": "false"},
5699)
5700
Marat Dukhan05702cf2020-03-26 15:41:33 -07005701# Disables usage of HMP-aware optimizations.
5702config_setting(
5703 name = "xnn_enable_hmp_explicit_false",
5704 define_values = {"xnn_enable_hmp": "false"},
5705)
5706
Chao Mei6ddfc602020-05-13 22:29:36 -07005707# Enable usage of optimized memory allocation
5708config_setting(
5709 name = "xnn_enable_memopt_explicit_true",
5710 define_values = {"xnnpack_enable_memopt": "true"},
5711)
5712
5713# Disable usage of optimized memory allocation
5714config_setting(
5715 name = "xnn_enable_memopt_explicit_false",
5716 define_values = {"xnnpack_enable_memopt": "false"},
5717)
5718
Marat Dukhanb8642352019-10-30 15:43:02 -07005719# Builds with -c dbg
5720config_setting(
5721 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005722 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005723 "compilation_mode": "dbg",
5724 },
5725)
5726
5727# Builds with -c opt
5728config_setting(
5729 name = "optimized_build",
5730 values = {
5731 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005732 },
5733)
5734
5735config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005736 name = "linux_k8",
5737 values = {"cpu": "k8"},
5738)
5739
5740config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005741 name = "linux_arm",
5742 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005743)
5744
5745config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005746 name = "linux_armeabi",
5747 values = {"cpu": "armeabi"},
5748)
5749
5750config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005751 name = "linux_armhf",
5752 values = {"cpu": "armhf"},
5753)
5754
5755config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005756 name = "linux_armv7a",
5757 values = {"cpu": "armv7a"},
5758)
5759
5760config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005761 name = "linux_aarch64",
5762 values = {"cpu": "aarch64"},
5763)
5764
5765config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005766 name = "android",
5767 values = {"crosstool_top": "//external:android/crosstool"},
5768)
5769
5770config_setting(
5771 name = "android_armv7",
5772 values = {
5773 "crosstool_top": "//external:android/crosstool",
5774 "cpu": "armeabi-v7a",
5775 },
5776)
5777
5778config_setting(
5779 name = "android_arm64",
5780 values = {
5781 "crosstool_top": "//external:android/crosstool",
5782 "cpu": "arm64-v8a",
5783 },
5784)
5785
5786config_setting(
5787 name = "android_x86",
5788 values = {
5789 "crosstool_top": "//external:android/crosstool",
5790 "cpu": "x86",
5791 },
5792)
5793
5794config_setting(
5795 name = "android_x86_64",
5796 values = {
5797 "crosstool_top": "//external:android/crosstool",
5798 "cpu": "x86_64",
5799 },
5800)
5801
5802config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005803 name = "windows_x86_64",
5804 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005805)
5806
5807config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005808 name = "windows_x86_64_clang",
5809 values = {
5810 "compiler": "clang-cl",
5811 "cpu": "x64_windows",
5812 },
5813)
5814
5815config_setting(
5816 name = "windows_x86_64_mingw",
5817 values = {
5818 "compiler": "mingw-gcc",
5819 "cpu": "x64_windows",
5820 },
5821)
5822
5823config_setting(
5824 name = "windows_x86_64_msys",
5825 values = {
5826 "compiler": "msys-gcc",
5827 "cpu": "x64_windows",
5828 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005829)
5830
5831config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005832 name = "macos_x86_64",
5833 values = {
5834 "apple_platform_type": "macos",
5835 "cpu": "darwin",
5836 },
5837)
5838
5839config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005840 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005841 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005842)
5843
5844config_setting(
5845 name = "emscripten_wasm",
5846 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005847 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005848 "cpu": "wasm",
5849 },
5850)
5851
5852config_setting(
5853 name = "emscripten_wasmsimd",
5854 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005855 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005857 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005858 },
5859)
5860
5861config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005862 name = "ios_armv7",
5863 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005864 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005865 "cpu": "ios_armv7",
5866 },
5867)
5868
5869config_setting(
5870 name = "ios_arm64",
5871 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005872 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005873 "cpu": "ios_arm64",
5874 },
5875)
5876
5877config_setting(
5878 name = "ios_arm64e",
5879 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005880 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005881 "cpu": "ios_arm64e",
5882 },
5883)
5884
5885config_setting(
5886 name = "ios_x86",
5887 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005888 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005889 "cpu": "ios_i386",
5890 },
5891)
5892
5893config_setting(
5894 name = "ios_x86_64",
5895 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005896 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005897 "cpu": "ios_x86_64",
5898 },
5899)
5900
5901config_setting(
5902 name = "watchos_armv7k",
5903 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005904 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005905 "cpu": "watchos_armv7k",
5906 },
5907)
5908
5909config_setting(
5910 name = "watchos_arm64_32",
5911 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005912 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005913 "cpu": "watchos_arm64_32",
5914 },
5915)
5916
5917config_setting(
5918 name = "watchos_x86",
5919 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005920 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005921 "cpu": "watchos_i386",
5922 },
5923)
5924
5925config_setting(
5926 name = "watchos_x86_64",
5927 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005928 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005929 "cpu": "watchos_x86_64",
5930 },
5931)
5932
5933config_setting(
5934 name = "tvos_arm64",
5935 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005936 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005937 "cpu": "tvos_arm64",
5938 },
5939)
5940
5941config_setting(
5942 name = "tvos_x86_64",
5943 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005944 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005945 "cpu": "tvos_x86_64",
5946 },
5947)