blob: c0889fb3af0c2b486952ac11925dececba25ec09 [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 Dukhancc8f34c2020-08-05 16:36:38 -0700851 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
852 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
853 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700854 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
855 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
856 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
857 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
858 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
859 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
860 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
861 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
862 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
863 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
864 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
865 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
866 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
867 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
868 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700869 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700870 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700871 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700872 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700873 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700874 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700875 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700876 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700877 "src/x32-zip/x2-wasmsimd.c",
878 "src/x32-zip/x3-wasmsimd.c",
879 "src/x32-zip/x4-wasmsimd.c",
880 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700881 "src/math/roundne-wasmsimd-addsub.c",
882 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700883 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700884 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700885 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700886 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700887 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700888 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
889 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700890]
891
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800892PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700893 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700894 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700895 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700896 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700897 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700898 "src/f32-spmm/gen/4x1-minmax-psimd.c",
899 "src/f32-spmm/gen/8x1-minmax-psimd.c",
900 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901]
902
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800903PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700904 "src/qs8-requantization/precise-psimd.c",
905 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700906 "src/qu8-requantization/precise-psimd.c",
907 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800908]
909
Marat Dukhan08c4a432019-10-03 09:29:21 -0700910# ISA-specific micro-kernels
911NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700912 "src/f32-argmaxpool/4x-neon-c4.c",
913 "src/f32-argmaxpool/9p8x-neon-c4.c",
914 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700915 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
916 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700917 "src/f32-clamp/gen/neon-x4.c",
918 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700919 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
920 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700921 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
922 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700923 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
924 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700925 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
926 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700927 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
928 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
929 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
930 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700931 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
932 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
933 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
934 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700935 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
936 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
937 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
938 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700939 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700940 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
941 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700942 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
945 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
946 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
947 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
948 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
949 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
950 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
951 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
952 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
953 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
954 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
955 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
956 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
957 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
958 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
959 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
960 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
961 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
962 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
963 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
964 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
965 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
966 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
967 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
968 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
969 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
970 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
971 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
972 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700973 "src/f32-hswish/gen/hswish-neon-x4.c",
974 "src/f32-hswish/gen/hswish-neon-x8.c",
975 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700976 "src/f32-ibilinear/gen/neon-c4.c",
977 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700978 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
979 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
980 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
981 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
982 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
983 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
984 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
985 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
986 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
987 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
988 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
989 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
990 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
991 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
992 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
993 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700994 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
995 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
996 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700997 "src/f32-ppmm/gen/4x8-minmax-neon.c",
998 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700999 "src/f32-prelu/gen/neon-1x4.c",
1000 "src/f32-prelu/gen/neon-1x8.c",
1001 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001002 "src/f32-prelu/gen/neon-2x4.c",
1003 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001004 "src/f32-prelu/gen/neon-2x16.c",
1005 "src/f32-prelu/gen/neon-4x4.c",
1006 "src/f32-prelu/gen/neon-4x8.c",
1007 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001008 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1014 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1015 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1016 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1017 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1018 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001032 "src/f32-relu/gen/neon-x4.c",
1033 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001034 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001035 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001036 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1037 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1038 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1039 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1040 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1041 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1042 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1043 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1044 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1045 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1046 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1047 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1048 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1049 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1050 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1051 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1052 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1053 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001054 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1055 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1056 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1057 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001058 "src/f32-vbinary/gen/vmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vmax-neon-x8.c",
1060 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1061 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1062 "src/f32-vbinary/gen/vmin-neon-x4.c",
1063 "src/f32-vbinary/gen/vmin-neon-x8.c",
1064 "src/f32-vbinary/gen/vminc-neon-x4.c",
1065 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001066 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1067 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1068 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1069 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1070 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1071 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001072 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1073 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1074 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1075 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001076 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1077 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1078 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1079 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001080 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1081 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001082 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1083 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001084 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1085 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1086 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1087 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1088 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1089 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1090 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1091 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001092 "src/f32-vunary/gen/vabs-neon-x4.c",
1093 "src/f32-vunary/gen/vabs-neon-x8.c",
1094 "src/f32-vunary/gen/vneg-neon-x4.c",
1095 "src/f32-vunary/gen/vneg-neon-x8.c",
1096 "src/f32-vunary/gen/vsqr-neon-x4.c",
1097 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001098 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1099 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1100 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1101 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1102 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1103 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1104 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1105 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001106 "src/qs8-requantization/precise-neon.c",
1107 "src/qs8-requantization/fp32-neon.c",
1108 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001109 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1110 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1111 "src/qu8-dwconv/up8x9-minmax-neon.c",
1112 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1113 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1114 "src/qu8-gemm/4x8-minmax-neon.c",
1115 "src/qu8-gemm/8x8-minmax-neon.c",
1116 "src/qu8-igemm/4x8-minmax-neon.c",
1117 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001118 "src/qu8-requantization/precise-neon.c",
1119 "src/qu8-requantization/fp32-neon.c",
1120 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001121 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001122 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001123 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001125 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001127 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001128 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001129 "src/x32-zip/x2-neon.c",
1130 "src/x32-zip/x3-neon.c",
1131 "src/x32-zip/x4-neon.c",
1132 "src/x32-zip/xm-neon.c",
1133 "src/x8-zip/x2-neon.c",
1134 "src/x8-zip/x3-neon.c",
1135 "src/x8-zip/x4-neon.c",
1136 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001137 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001138 "src/math/roundd-neon-addsub.c",
1139 "src/math/roundd-neon-cvt.c",
1140 "src/math/roundu-neon-addsub.c",
1141 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001142 "src/math/roundz-neon-addsub.c",
1143 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001144 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001145 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1146 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1147 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1148 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1149 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1150 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001151 "src/math/sqrt-neon-nr1rsqrts.c",
1152 "src/math/sqrt-neon-nr2rsqrts.c",
1153 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001154]
1155
1156NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001157 "src/f32-ibilinear/gen/neonfma-c4.c",
1158 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001159 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1160 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1161 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1162 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1163 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1164 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1165 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1166 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1167 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001168 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1169 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1170 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1171 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001172 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1173 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1174 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1175 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001176 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1177 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1178 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1179 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001180 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1181 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1182 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1183 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1184 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1185 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1186 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1187 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1188 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1189 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1190 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1191 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1192 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1193 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1194 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1195 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1196 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1197 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001198 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1199 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001200 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1221 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1222 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1223 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001278 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1279 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001280 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1281 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1282 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1283 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1284 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1285 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1286 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1287 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1288 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1289 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1290 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1291 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1292 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1293 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1294 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1295 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1296 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1297 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1298 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1299 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001300 "src/math/exp-neonfma-lut64-p2.c",
1301 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001302 "src/math/expminus-neonfma-lut2048-p1.c",
1303 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001304 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001305 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1306 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1307 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1308 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1309 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1310 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1311 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1312 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1313 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1314 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1315 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1316 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1317 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1318 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1319 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1320 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1321 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1322 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001323 "src/math/sqrt-neonfma-nr1fma.c",
1324 "src/math/sqrt-neonfma-nr2fma.c",
1325 "src/math/sqrt-neonfma-nr3fma.c",
1326 "src/math/sqrt-neonfma-nr2fma1adj.c",
1327 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001328]
1329
1330AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001331 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1332 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1333 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1334 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1335 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1336 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001337 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1338 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1339 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1340 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1341 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1342 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1343 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1344 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1345 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1346 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1347 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1348 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1349 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1350 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1351 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1352 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1353 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1354 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1355 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1356 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001357 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1358 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001359 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1360 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001361 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1362 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001363 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1364 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001365 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1366 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1367 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1368 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1369 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001370 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1385 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1386 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1387 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001388 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1389 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1390 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1391 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1392 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1393 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1394 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1395 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1396 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1397 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1398 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1399 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1400 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1401 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1402 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1403 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1404 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1405 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001406 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1407 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001408 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1409 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1410 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1411 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1412 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1413 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414]
1415
Marat Dukhan8853b822020-05-07 12:19:01 -07001416NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001417 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1418 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1419 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1420 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1421 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1422 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1423 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1424 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001425 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001426 "src/math/roundd-neonv8.c",
1427 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001428 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001429]
1430
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001432 "src/f16-clamp/gen/neonfp16arith-x16.c",
1433 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001434 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1435 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1436 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1437 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1438 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1439 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1440 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1441 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1442 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1443 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1444 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1445 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001446 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1447 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001448 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1449 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001450 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001451 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001452 "src/f16-relu/gen/neonfp16arith-x16.c",
1453 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001454 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1455 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1456 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1457 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1458 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1459 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1460 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1461 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1462 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1463 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1464 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1465 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1466 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1467 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1468 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1469 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1470 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1471 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1472 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1473 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1474 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1475 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1476 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1477 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1478 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1479 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1480 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1481 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001482 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1483 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001484 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1487 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1489 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1490 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1491 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001492 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1493 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1494 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1495 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001496 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1497 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1498 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1499 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1501 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1502 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1503 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1504 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1505 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1506 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1507 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001508 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1509 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1510 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1511 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1512 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1513 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1514 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1515 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001516]
1517
1518SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001519 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1520 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001521 "src/f32-clamp/gen/sse-x4.c",
1522 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001523 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1524 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001525 "src/f32-dwconv-chw/3x3p1-sse.c",
1526 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1528 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1529 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1530 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1531 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1532 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1533 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1534 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1535 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1536 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1537 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1538 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001539 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001540 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1541 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001542 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1543 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1544 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1545 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1546 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1547 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1548 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1549 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1550 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1551 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1552 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1553 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1554 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001555 "src/f32-hswish/gen/hswish-sse-x4.c",
1556 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001557 "src/f32-ibilinear/gen/sse-c4.c",
1558 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001559 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1560 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1561 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1562 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1563 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1564 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1565 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001566 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1567 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1568 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001569 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001570 "src/f32-prelu/gen/sse-2x4.c",
1571 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001572 "src/f32-relu/gen/sse-x4.c",
1573 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001574 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001575 "src/f32-spmm/gen/4x1-minmax-sse.c",
1576 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001577 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001578 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1582 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1583 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1584 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1585 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001586 "src/f32-vbinary/gen/vmax-sse-x4.c",
1587 "src/f32-vbinary/gen/vmax-sse-x8.c",
1588 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1589 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1590 "src/f32-vbinary/gen/vmin-sse-x4.c",
1591 "src/f32-vbinary/gen/vmin-sse-x8.c",
1592 "src/f32-vbinary/gen/vminc-sse-x4.c",
1593 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001594 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1595 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1596 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1597 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1598 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1599 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1600 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1601 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001602 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1603 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1604 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1605 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001606 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1607 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1608 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1609 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001610 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1611 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001612 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1613 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001614 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1615 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001616 "src/f32-vunary/gen/vabs-sse-x4.c",
1617 "src/f32-vunary/gen/vabs-sse-x8.c",
1618 "src/f32-vunary/gen/vneg-sse-x4.c",
1619 "src/f32-vunary/gen/vneg-sse-x8.c",
1620 "src/f32-vunary/gen/vsqr-sse-x4.c",
1621 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001622 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001623 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001624 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001625 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001626 "src/math/roundd-sse-addsub.c",
1627 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001628 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001629 "src/math/sqrt-sse-hh1mac.c",
1630 "src/math/sqrt-sse-nr1mac.c",
1631 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632]
1633
1634SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001635 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1636 "src/f32-argmaxpool/4x-sse2-c4.c",
1637 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001638 "src/f32-prelu/gen/sse2-2x4.c",
1639 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001640 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1641 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1642 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1643 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1644 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1645 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1646 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1647 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1648 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1649 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1650 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1651 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001652 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001653 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001654 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001655 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001656 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1657 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001658 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1659 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001660 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1661 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1662 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1663 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1664 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1665 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1666 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1667 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001668 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1669 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1670 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001671 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1672 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1673 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001674 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1675 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1676 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001677 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1678 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001679 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1680 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001681 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1682 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001683 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1684 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001685 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001686 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1687 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001688 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001689 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1690 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1691 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001692 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1693 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001694 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1695 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001696 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1697 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001698 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001699 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1700 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001701 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001702 "src/qs8-requantization/precise-sse2.c",
1703 "src/qs8-requantization/fp32-sse2.c",
1704 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001705 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1706 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1707 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1708 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1709 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1710 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1711 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1712 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001713 "src/qu8-requantization/precise-sse2.c",
1714 "src/qu8-requantization/fp32-sse2.c",
1715 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001716 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001717 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001718 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001719 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001720 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 "src/x32-zip/x2-sse2.c",
1722 "src/x32-zip/x3-sse2.c",
1723 "src/x32-zip/x4-sse2.c",
1724 "src/x32-zip/xm-sse2.c",
1725 "src/x8-zip/x2-sse2.c",
1726 "src/x8-zip/x3-sse2.c",
1727 "src/x8-zip/x4-sse2.c",
1728 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001729 "src/math/exp-sse2-p5.c",
1730 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001731 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001732 "src/math/roundd-sse2-cvt.c",
1733 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001734 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001735 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001736]
1737
1738SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001739 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1740 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1741 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001742 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1743 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1744 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001745 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1746 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1747 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001748 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1749 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001750 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1751 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001752 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1753 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001754 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1755 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001756 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001757 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1758 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001759 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001760 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1761 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1762 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001763 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1764 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001765 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1766 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001767 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1768 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001769 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001770 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1771 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001772 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001773 "src/qs8-requantization/precise-ssse3.c",
1774 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001775 "src/qu8-requantization/precise-ssse3.c",
1776 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777]
1778
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001779SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001780 "src/f32-prelu/gen/sse41-2x4.c",
1781 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001782 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001783 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001784 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001785 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001786 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1787 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001788 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1789 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001790 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1791 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1792 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1793 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1794 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1795 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1796 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1797 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001798 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1799 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1800 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001801 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1802 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1803 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001804 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1805 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1806 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001807 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1808 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001809 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1810 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001811 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1812 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001813 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1814 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001815 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001816 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1817 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001818 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001819 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1820 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1821 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001822 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1823 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001824 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1825 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001826 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1827 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001828 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001829 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1830 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001831 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001832 "src/qs8-requantization/fp32-sse4.c",
1833 "src/qs8-requantization/precise-sse4.c",
1834 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001835 "src/qu8-requantization/precise-sse4.c",
1836 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001837 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001838 "src/math/roundd-sse41.c",
1839 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001840 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001841]
1842
Marat Dukhan08c4a432019-10-03 09:29:21 -07001843AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001844 "src/f32-clamp/gen/avx-x8.c",
1845 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001846 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1847 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1848 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1849 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1850 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1851 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1852 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1853 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1854 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1855 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1856 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1857 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1858 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1859 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1860 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1861 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1862 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1863 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1864 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1865 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1866 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1867 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1868 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1869 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1870 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1871 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1872 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1873 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1874 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1875 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001876 "src/f32-hswish/gen/hswish-avx-x8.c",
1877 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001878 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1879 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1880 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1881 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1882 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1883 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1884 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1885 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1886 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001887 "src/f32-prelu/gen/avx-2x8.c",
1888 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001889 "src/f32-relu/gen/avx-x16.c",
1890 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001892 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1893 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1894 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1895 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1896 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1897 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1898 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1899 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001900 "src/f32-vbinary/gen/vmax-avx-x8.c",
1901 "src/f32-vbinary/gen/vmax-avx-x16.c",
1902 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1903 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1904 "src/f32-vbinary/gen/vmin-avx-x8.c",
1905 "src/f32-vbinary/gen/vmin-avx-x16.c",
1906 "src/f32-vbinary/gen/vminc-avx-x8.c",
1907 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001908 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1909 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1910 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1911 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1912 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1913 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1914 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1915 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001916 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1917 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1918 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1919 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001920 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1921 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1922 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1923 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001924 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1925 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001926 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001927 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1928 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1929 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1930 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1931 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1932 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1933 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1934 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001935 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1936 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001937 "src/f32-vunary/gen/vabs-avx-x8.c",
1938 "src/f32-vunary/gen/vabs-avx-x16.c",
1939 "src/f32-vunary/gen/vneg-avx-x8.c",
1940 "src/f32-vunary/gen/vneg-avx-x16.c",
1941 "src/f32-vunary/gen/vsqr-avx-x8.c",
1942 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001943]
1944
Marat Dukhan1566fee2020-08-02 21:55:41 -07001945XOP_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001946 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul16.c",
1947 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul16.c",
1948 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul16.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001949 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1950 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1951 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1952 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001953 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1954 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001955 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1956 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001957 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001958 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1959 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001960 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001961 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1962 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1963 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001964 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1965 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1966 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1967 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1968 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1969 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001970 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001971 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1972 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001973 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001974]
1975
Marat Dukhanfda12b82019-11-21 12:27:59 -08001976FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001977 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1978 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1979 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1980 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1981 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1982 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1983 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1984 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1985 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1986 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1987 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1988 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1989 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1990 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1991 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1992 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1993 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1994 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1995 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1996 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1997 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1998 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1999 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2000 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2001 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2002 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2003 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2004 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2005 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2006 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2007 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2008 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2009 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2010 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2011 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2012 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2013 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2014 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2015 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2016 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002017 "src/f32-hswish/gen/hswish-fma3-x8.c",
2018 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002019 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2020 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2021 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2022 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2023 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2024 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2025 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2026 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2027 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2028 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2029 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2030 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2031 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2032 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002033 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2034 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2035 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2036 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2037 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2038 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2039 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2040 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002041 "src/math/sqrt-fma3-nr1fma.c",
2042 "src/math/sqrt-fma3-nr2fma.c",
2043 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002044]
2045
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002046AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002047 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2048 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2049 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2050 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2051 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2052 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2053 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2054 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2055 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2056 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2057 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2058 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2059 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2060 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2061 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2062 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2063 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2064 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2065 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2066 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2067 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2068 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2069 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2070 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2071 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2072 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2073 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2074 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2075 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2076 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2077 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2078 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2079 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2080 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2081 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2082 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002083 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2084 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2085 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2086 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2087 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2088 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2089 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2090 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2091 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2092 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2093 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2094 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2095 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2096 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2097 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2098 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2099 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2100 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2101 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2102 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2103 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2104 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2105 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2106 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2107 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2108 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2109 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2110 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2111 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2112 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002113 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2114 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2115 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2116 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2117 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2118 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2119 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2120 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2121 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2122 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2123 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2124 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2125 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2126 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2127 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2128 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2129 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2130 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2131 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2132 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2133 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2134 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2135 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2136 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002137 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2138 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2139 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2140 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002141 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2142 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2143 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002144 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2145 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2146 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002147 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2148 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2149 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002150 "src/math/exp-avx2-p5.c",
2151 "src/math/exp-avx2-perm-p3.c",
2152 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002153 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002154 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002155 "src/math/sigmoid-avx2-rr2-p5-div.c",
2156 "src/math/sigmoid-avx2-rr1-p5-div.c",
2157 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2158 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2159 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2160 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002161]
2162
Marat Dukhan08c4a432019-10-03 09:29:21 -07002163AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002164 "src/f32-clamp/gen/avx512f-x16.c",
2165 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002166 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2167 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2168 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2169 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2170 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2171 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2172 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2173 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2174 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2175 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2176 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2177 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2178 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2179 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2180 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2181 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2182 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2183 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2184 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2185 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2186 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2187 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2188 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2189 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002190 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2191 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002192 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2193 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2194 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2195 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2196 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2197 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002198 "src/f32-prelu/gen/avx512f-2x16.c",
2199 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002200 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2201 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2202 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2203 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2204 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2205 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2206 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2207 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2208 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2209 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2210 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2211 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2212 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2213 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2214 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2215 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2216 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2217 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2218 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2219 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2220 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2221 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2222 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2223 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2224 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2225 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2226 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2227 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2228 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2229 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2230 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2231 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2232 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2233 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2234 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2235 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002236 "src/f32-relu/gen/avx512f-x16.c",
2237 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002239 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2240 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2241 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2242 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2243 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2244 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2245 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2246 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002247 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2248 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2249 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2250 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2251 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2252 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2253 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2254 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002255 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2256 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2257 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2258 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2259 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2260 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2261 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2262 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002263 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2264 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2265 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2266 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002267 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2268 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2269 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2270 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002271 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2272 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002273 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002274 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2275 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2276 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2277 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2278 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2279 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2280 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2281 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2282 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2283 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2284 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2285 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2286 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2287 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2288 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2289 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2290 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2291 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2292 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2293 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2294 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2295 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2296 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2297 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002298 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2299 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2300 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2301 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2302 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2303 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2304 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2305 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002306 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2307 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2308 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2309 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2310 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2311 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2312 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2313 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002314 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2315 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2316 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2317 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2318 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2319 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002320 "src/math/exp-avx512f-p5-scalef.c",
2321 "src/math/exp-avx512f-p5.c",
2322 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002323 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002324 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002325 "src/math/sqrt-avx512f-nr1fma.c",
2326 "src/math/sqrt-avx512f-nr2fma.c",
2327 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328]
2329
2330AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002331 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002332 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002333 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2334 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002335 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002336 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002337 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002338 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002339 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2340 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002341 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2342 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2343 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2344 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345]
2346
2347AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002348 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002349 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2350 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002351 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002352 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2353 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002354 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2355 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2356 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002357 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002358 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2359 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2360 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002361 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002362 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2363 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002364 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2366 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2367 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2368 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2369 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2370 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2371 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2372 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2373 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2374 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2375 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2376 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2377 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2378 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2379 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2380 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2381 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2382 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2384 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002385 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002386 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2387 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2388 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2389 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2390 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2391 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2392 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2393 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2394 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2395 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2396 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2397 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2398 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2399 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2400 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2401 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2402 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2403 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002404 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2405 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2406 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2407 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2408 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2409 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2410 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2411 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2412 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2413 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2414 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2415 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2416 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2417 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2418 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2419 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2420 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2421 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422]
2423
Marat Dukhan1b354632020-03-23 12:50:22 -07002424INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002425 "src/xnnpack/argmaxpool.h",
2426 "src/xnnpack/avgpool.h",
2427 "src/xnnpack/clamp.h",
2428 "src/xnnpack/common.h",
2429 "src/xnnpack/conv.h",
2430 "src/xnnpack/dwconv.h",
2431 "src/xnnpack/gavgpool.h",
2432 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002433 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002435 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002436 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002437 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002438 "src/xnnpack/lut.h",
2439 "src/xnnpack/math.h",
2440 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002441 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002442 "src/xnnpack/packx.h",
2443 "src/xnnpack/pad.h",
2444 "src/xnnpack/params.h",
2445 "src/xnnpack/pavgpool.h",
2446 "src/xnnpack/ppmm.h",
2447 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002448 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002449 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002450 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002451 "src/xnnpack/rmax.h",
2452 "src/xnnpack/scalar-utils.h",
2453 "src/xnnpack/spmm.h",
2454 "src/xnnpack/unpool.h",
2455 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002456 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002457 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002458 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002459 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002460 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002461 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002462 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002463]
2464
2465INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002466 "include/xnnpack.h",
2467 "src/xnnpack/allocator.h",
2468 "src/xnnpack/compute.h",
2469 "src/xnnpack/im2col.h",
2470 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002471 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002472 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002473 "src/xnnpack/operator.h",
2474 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002475 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002476 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002477 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002478 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002479]
2480
Marat Dukhan1b354632020-03-23 12:50:22 -07002481ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002482 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002483]
2484
Marat Dukhan1b354632020-03-23 12:50:22 -07002485MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002486 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002487 "include/xnnpack.h",
2488]
2489
Marat Dukhan1b354632020-03-23 12:50:22 -07002490MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002491 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002492 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002493 "src/xnnpack/requantization.h",
2494 "include/xnnpack.h",
2495]
2496
2497OPERATOR_TEST_PARAMS_HDRS = [
2498 "src/xnnpack/params.h",
2499 "src/xnnpack/common.h",
2500]
2501
2502WEIGHTS_PACK_HDRS = [
2503 "src/xnnpack/pack.h",
2504 "src/xnnpack/operator.h",
2505 "src/xnnpack/compute.h",
2506]
2507
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002508LOGGING_COPTS = select({
2509 # No logging in optimized mode
2510 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2511 # Full logging in debug mode
2512 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2513 # Error-only logging in default (fastbuild) mode
2514 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2515})
2516
Marat Dukhan3b59de22020-06-03 20:15:19 -07002517LOGGING_SRCS = select({
2518 # No logging in optimized mode
2519 ":optimized_build": [],
2520 "//conditions:default": [
2521 "src/operator-strings.c",
2522 "src/subgraph-strings.c",
2523 ],
2524})
2525
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002526LOGGING_HDRS = [
2527 "src/xnnpack/log.h",
2528]
2529
Marat Dukhan08c4a432019-10-03 09:29:21 -07002530xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002531 name = "tables",
2532 srcs = TABLE_SRCS,
2533 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002534 gcc_copts = xnnpack_gcc_std_copts(),
2535 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002536)
2537
2538xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002539 name = "scalar_ukernels",
2540 srcs = SCALAR_UKERNELS,
2541 hdrs = INTERNAL_HDRS,
2542 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002543 gcc_copts = xnnpack_gcc_std_copts(),
2544 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002545 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002546 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 "@FP16",
2548 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002549 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002550 ],
2551)
2552
2553xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002554 name = "scalar_ukernels_test_mode",
2555 srcs = SCALAR_UKERNELS,
2556 hdrs = INTERNAL_HDRS,
2557 aarch32_copts = ["-marm"],
2558 copts = [
2559 "-UNDEBUG",
2560 "-DXNN_TEST_MODE=1",
2561 ],
2562 gcc_copts = xnnpack_gcc_std_copts(),
2563 msvc_copts = xnnpack_msvc_std_copts(),
2564 deps = [
2565 ":tables",
2566 "@FP16",
2567 "@FXdiv",
2568 "@pthreadpool",
2569 ],
2570)
2571
2572xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002573 name = "wasm_ukernels",
2574 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002575 gcc_copts = xnnpack_gcc_std_copts(),
2576 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002577 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002578 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002579 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002580 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002581 "@FP16",
2582 "@FXdiv",
2583 "@pthreadpool",
2584 ],
2585)
2586
2587xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002588 name = "wasm_ukernels_test_mode",
2589 hdrs = INTERNAL_HDRS,
2590 copts = [
2591 "-UNDEBUG",
2592 "-DXNN_TEST_MODE=1",
2593 ],
2594 gcc_copts = xnnpack_gcc_std_copts(),
2595 msvc_copts = xnnpack_msvc_std_copts(),
2596 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002597 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002598 deps = [
2599 ":tables",
2600 "@FP16",
2601 "@FXdiv",
2602 "@pthreadpool",
2603 ],
2604)
2605
2606xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002607 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608 hdrs = INTERNAL_HDRS,
2609 aarch32_copts = [
2610 "-marm",
2611 "-mfpu=neon",
2612 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002613 gcc_copts = xnnpack_gcc_std_copts(),
2614 gcc_x86_copts = ["-msse2"],
2615 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002616 optimized_copts = [
2617 "-O3",
2618 "-ffast-math",
2619 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002620 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002622 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002623 "@FP16",
2624 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002625 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 ],
2627)
2628
2629xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002630 name = "psimd_fastmath_ukernels_test_mode",
2631 hdrs = INTERNAL_HDRS,
2632 aarch32_copts = [
2633 "-marm",
2634 "-mfpu=neon",
2635 ],
2636 copts = [
2637 "-UNDEBUG",
2638 "-DXNN_TEST_MODE=1",
2639 ],
2640 gcc_copts = xnnpack_gcc_std_copts(),
2641 gcc_x86_copts = ["-msse2"],
2642 msvc_copts = xnnpack_msvc_std_copts(),
2643 optimized_copts = [
2644 "-O3",
2645 "-ffast-math",
2646 ],
2647 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2648 deps = [
2649 ":tables",
2650 "@FP16",
2651 "@psimd",
2652 "@pthreadpool",
2653 ],
2654)
2655
2656xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002657 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002658 hdrs = INTERNAL_HDRS,
2659 aarch32_copts = [
2660 "-marm",
2661 "-mfpu=neon",
2662 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002663 gcc_copts = xnnpack_gcc_std_copts(),
2664 gcc_x86_copts = ["-msse2"],
2665 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002666 optimized_copts = [
2667 "-O3",
2668 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002669 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002670 deps = [
2671 ":tables",
2672 "@FP16",
2673 "@psimd",
2674 "@pthreadpool",
2675 ],
2676)
2677
2678xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002679 name = "psimd_accmath_ukernels_test_mode",
2680 hdrs = INTERNAL_HDRS,
2681 aarch32_copts = [
2682 "-marm",
2683 "-mfpu=neon",
2684 ],
2685 copts = [
2686 "-UNDEBUG",
2687 "-DXNN_TEST_MODE=1",
2688 ],
2689 gcc_copts = xnnpack_gcc_std_copts(),
2690 gcc_x86_copts = ["-msse2"],
2691 msvc_copts = xnnpack_msvc_std_copts(),
2692 optimized_copts = [
2693 "-O3",
2694 ],
2695 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2696 deps = [
2697 ":tables",
2698 "@FP16",
2699 "@psimd",
2700 "@pthreadpool",
2701 ],
2702)
2703
2704xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002705 name = "neon_ukernels",
2706 hdrs = INTERNAL_HDRS,
2707 aarch32_copts = [
2708 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002709 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002710 "-mfpu=neon",
2711 ],
2712 aarch32_srcs = NEON_UKERNELS,
2713 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002714 gcc_copts = xnnpack_gcc_std_copts(),
2715 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002716 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002717 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002718 "@FP16",
2719 "@pthreadpool",
2720 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002721)
2722
2723xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002724 name = "neon_ukernels_test_mode",
2725 hdrs = INTERNAL_HDRS,
2726 aarch32_copts = [
2727 "-marm",
2728 "-march=armv7-a",
2729 "-mfpu=neon",
2730 ],
2731 aarch32_srcs = NEON_UKERNELS,
2732 aarch64_srcs = NEON_UKERNELS,
2733 copts = [
2734 "-UNDEBUG",
2735 "-DXNN_TEST_MODE=1",
2736 ],
2737 gcc_copts = xnnpack_gcc_std_copts(),
2738 msvc_copts = xnnpack_msvc_std_copts(),
2739 deps = [
2740 ":tables",
2741 "@FP16",
2742 "@pthreadpool",
2743 ],
2744)
2745
2746xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 name = "neonfma_ukernels",
2748 hdrs = INTERNAL_HDRS,
2749 aarch32_copts = [
2750 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002751 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752 "-mfpu=neon-vfpv4",
2753 ],
2754 aarch32_srcs = NEONFMA_UKERNELS,
2755 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002756 apple_aarch32_copts = [
2757 "-mcpu=swift",
2758 "-mtune=generic",
2759 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002760 gcc_copts = xnnpack_gcc_std_copts(),
2761 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002762 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002763 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002764 "@FP16",
2765 "@pthreadpool",
2766 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767)
2768
2769xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002770 name = "neonfma_ukernels_test_mode",
2771 hdrs = INTERNAL_HDRS,
2772 aarch32_copts = [
2773 "-marm",
2774 "-march=armv7-a",
2775 "-mfpu=neon-vfpv4",
2776 ],
2777 aarch32_srcs = NEONFMA_UKERNELS,
2778 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002779 apple_aarch32_copts = [
2780 "-mcpu=swift",
2781 "-mtune=generic",
2782 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002783 copts = [
2784 "-UNDEBUG",
2785 "-DXNN_TEST_MODE=1",
2786 ],
2787 gcc_copts = xnnpack_gcc_std_copts(),
2788 msvc_copts = xnnpack_msvc_std_copts(),
2789 deps = [
2790 ":tables",
2791 "@FP16",
2792 "@pthreadpool",
2793 ],
2794)
2795
2796xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002797 name = "neonv8_ukernels",
2798 hdrs = INTERNAL_HDRS,
2799 aarch32_copts = [
2800 "-marm",
2801 "-march=armv8-a",
2802 "-mfpu=neon-fp-armv8",
2803 ],
2804 aarch32_srcs = NEONV8_UKERNELS,
2805 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002806 apple_aarch32_copts = [
2807 "-mcpu=cyclone",
2808 "-mtune=generic",
2809 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002810 gcc_copts = xnnpack_gcc_std_copts(),
2811 msvc_copts = xnnpack_msvc_std_copts(),
2812 deps = [
2813 ":tables",
2814 "@FP16",
2815 "@pthreadpool",
2816 ],
2817)
2818
2819xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002820 name = "neonv8_ukernels_test_mode",
2821 hdrs = INTERNAL_HDRS,
2822 aarch32_copts = [
2823 "-marm",
2824 "-march=armv8-a",
2825 "-mfpu=neon-fp-armv8",
2826 ],
2827 aarch32_srcs = NEONV8_UKERNELS,
2828 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002829 apple_aarch32_copts = [
2830 "-mcpu=cyclone",
2831 "-mtune=generic",
2832 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002833 copts = [
2834 "-UNDEBUG",
2835 "-DXNN_TEST_MODE=1",
2836 ],
2837 gcc_copts = xnnpack_gcc_std_copts(),
2838 msvc_copts = xnnpack_msvc_std_copts(),
2839 deps = [
2840 ":tables",
2841 "@FP16",
2842 "@pthreadpool",
2843 ],
2844)
2845
2846xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002847 name = "neonfp16arith_ukernels",
2848 hdrs = INTERNAL_HDRS,
2849 aarch64_copts = ["-march=armv8.2-a+fp16"],
2850 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002851 gcc_copts = xnnpack_gcc_std_copts(),
2852 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002853 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002854 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002855 "@FP16",
2856 "@pthreadpool",
2857 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002858)
2859
2860xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002861 name = "neonfp16arith_ukernels_test_mode",
2862 hdrs = INTERNAL_HDRS,
2863 aarch64_copts = ["-march=armv8.2-a+fp16"],
2864 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2865 copts = [
2866 "-UNDEBUG",
2867 "-DXNN_TEST_MODE=1",
2868 ],
2869 gcc_copts = xnnpack_gcc_std_copts(),
2870 msvc_copts = xnnpack_msvc_std_copts(),
2871 deps = [
2872 ":tables",
2873 "@FP16",
2874 "@pthreadpool",
2875 ],
2876)
2877
2878xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 name = "sse2_ukernels",
2880 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002881 gcc_copts = xnnpack_gcc_std_copts(),
2882 gcc_x86_copts = ["-msse2"],
2883 msvc_copts = xnnpack_msvc_std_copts(),
2884 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002886 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002887 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002888 "@FP16",
2889 "@pthreadpool",
2890 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891)
2892
2893xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002894 name = "sse2_ukernels_test_mode",
2895 hdrs = INTERNAL_HDRS,
2896 copts = [
2897 "-UNDEBUG",
2898 "-DXNN_TEST_MODE=1",
2899 ],
2900 gcc_copts = xnnpack_gcc_std_copts(),
2901 gcc_x86_copts = ["-msse2"],
2902 msvc_copts = xnnpack_msvc_std_copts(),
2903 msvc_x86_32_copts = ["/arch:SSE2"],
2904 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2905 deps = [
2906 ":tables",
2907 "@FP16",
2908 "@pthreadpool",
2909 ],
2910)
2911
2912xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002913 name = "ssse3_ukernels",
2914 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002915 gcc_copts = xnnpack_gcc_std_copts(),
2916 gcc_x86_copts = ["-mssse3"],
2917 msvc_copts = xnnpack_msvc_std_copts(),
2918 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002919 x86_srcs = SSSE3_UKERNELS,
2920 deps = [
2921 ":tables",
2922 "@FP16",
2923 "@pthreadpool",
2924 ],
2925)
2926
2927xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002928 name = "ssse3_ukernels_test_mode",
2929 hdrs = INTERNAL_HDRS,
2930 copts = [
2931 "-UNDEBUG",
2932 "-DXNN_TEST_MODE=1",
2933 ],
2934 gcc_copts = xnnpack_gcc_std_copts(),
2935 gcc_x86_copts = ["-mssse3"],
2936 msvc_copts = xnnpack_msvc_std_copts(),
2937 msvc_x86_32_copts = ["/arch:SSE2"],
2938 x86_srcs = SSSE3_UKERNELS,
2939 deps = [
2940 ":tables",
2941 "@FP16",
2942 "@pthreadpool",
2943 ],
2944)
2945
2946xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002947 name = "sse41_ukernels",
2948 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002949 gcc_copts = xnnpack_gcc_std_copts(),
2950 gcc_x86_copts = ["-msse4.1"],
2951 msvc_copts = xnnpack_msvc_std_copts(),
2952 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002953 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002954 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002955 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002956 "@FP16",
2957 "@pthreadpool",
2958 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002959)
2960
2961xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002962 name = "sse41_ukernels_test_mode",
2963 hdrs = INTERNAL_HDRS,
2964 copts = [
2965 "-UNDEBUG",
2966 "-DXNN_TEST_MODE=1",
2967 ],
2968 gcc_copts = xnnpack_gcc_std_copts(),
2969 gcc_x86_copts = ["-msse4.1"],
2970 msvc_copts = xnnpack_msvc_std_copts(),
2971 msvc_x86_32_copts = ["/arch:SSE2"],
2972 x86_srcs = SSE41_UKERNELS,
2973 deps = [
2974 ":tables",
2975 "@FP16",
2976 "@pthreadpool",
2977 ],
2978)
2979
2980xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981 name = "avx_ukernels",
2982 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002983 gcc_copts = xnnpack_gcc_std_copts(),
2984 gcc_x86_copts = ["-mavx"],
2985 msvc_copts = xnnpack_msvc_std_copts(),
2986 msvc_x86_32_copts = ["/arch:AVX"],
2987 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002989 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002990 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002991 "@FP16",
2992 "@pthreadpool",
2993 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994)
2995
2996xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002997 name = "avx_ukernels_test_mode",
2998 hdrs = INTERNAL_HDRS,
2999 copts = [
3000 "-UNDEBUG",
3001 "-DXNN_TEST_MODE=1",
3002 ],
3003 gcc_copts = xnnpack_gcc_std_copts(),
3004 gcc_x86_copts = ["-mavx"],
3005 msvc_copts = xnnpack_msvc_std_copts(),
3006 msvc_x86_32_copts = ["/arch:AVX"],
3007 msvc_x86_64_copts = ["/arch:AVX"],
3008 x86_srcs = AVX_UKERNELS,
3009 deps = [
3010 ":tables",
3011 "@FP16",
3012 "@pthreadpool",
3013 ],
3014)
3015
3016xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003017 name = "xop_ukernels",
3018 hdrs = INTERNAL_HDRS,
3019 gcc_copts = xnnpack_gcc_std_copts(),
3020 gcc_x86_copts = ["-mxop"],
3021 msvc_copts = xnnpack_msvc_std_copts(),
3022 msvc_x86_32_copts = ["/arch:AVX"],
3023 msvc_x86_64_copts = ["/arch:AVX"],
3024 x86_srcs = XOP_UKERNELS,
3025 deps = [
3026 ":tables",
3027 "@FP16",
3028 "@pthreadpool",
3029 ],
3030)
3031
3032xnnpack_cc_library(
3033 name = "xop_ukernels_test_mode",
3034 hdrs = INTERNAL_HDRS,
3035 copts = [
3036 "-UNDEBUG",
3037 "-DXNN_TEST_MODE=1",
3038 ],
3039 gcc_copts = xnnpack_gcc_std_copts(),
3040 gcc_x86_copts = ["-mxop"],
3041 msvc_copts = xnnpack_msvc_std_copts(),
3042 msvc_x86_32_copts = ["/arch:AVX"],
3043 msvc_x86_64_copts = ["/arch:AVX"],
3044 x86_srcs = XOP_UKERNELS,
3045 deps = [
3046 ":tables",
3047 "@FP16",
3048 "@pthreadpool",
3049 ],
3050)
3051
3052xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003053 name = "fma3_ukernels",
3054 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003055 gcc_copts = xnnpack_gcc_std_copts(),
3056 gcc_x86_copts = ["-mfma"],
3057 msvc_copts = xnnpack_msvc_std_copts(),
3058 msvc_x86_32_copts = ["/arch:AVX"],
3059 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003060 x86_srcs = FMA3_UKERNELS,
3061 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003062 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003063 "@FP16",
3064 "@pthreadpool",
3065 ],
3066)
3067
3068xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003069 name = "fma3_ukernels_test_mode",
3070 hdrs = INTERNAL_HDRS,
3071 copts = [
3072 "-UNDEBUG",
3073 "-DXNN_TEST_MODE=1",
3074 ],
3075 gcc_copts = xnnpack_gcc_std_copts(),
3076 gcc_x86_copts = ["-mfma"],
3077 msvc_copts = xnnpack_msvc_std_copts(),
3078 msvc_x86_32_copts = ["/arch:AVX"],
3079 msvc_x86_64_copts = ["/arch:AVX"],
3080 x86_srcs = FMA3_UKERNELS,
3081 deps = [
3082 ":tables",
3083 "@FP16",
3084 "@pthreadpool",
3085 ],
3086)
3087
3088xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003089 name = "avx2_ukernels",
3090 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003091 gcc_copts = xnnpack_gcc_std_copts(),
3092 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003093 "-mfma",
3094 "-mavx2",
3095 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003096 msvc_copts = xnnpack_msvc_std_copts(),
3097 msvc_x86_32_copts = ["/arch:AVX2"],
3098 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003099 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003100 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003101 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003102 "@FP16",
3103 "@pthreadpool",
3104 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003105)
3106
3107xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003108 name = "avx2_ukernels_test_mode",
3109 hdrs = INTERNAL_HDRS,
3110 copts = [
3111 "-UNDEBUG",
3112 "-DXNN_TEST_MODE=1",
3113 ],
3114 gcc_copts = xnnpack_gcc_std_copts(),
3115 gcc_x86_copts = [
3116 "-mfma",
3117 "-mavx2",
3118 ],
3119 msvc_copts = xnnpack_msvc_std_copts(),
3120 msvc_x86_32_copts = ["/arch:AVX2"],
3121 msvc_x86_64_copts = ["/arch:AVX2"],
3122 x86_srcs = AVX2_UKERNELS,
3123 deps = [
3124 ":tables",
3125 "@FP16",
3126 "@pthreadpool",
3127 ],
3128)
3129
3130xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 name = "avx512f_ukernels",
3132 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003133 gcc_copts = xnnpack_gcc_std_copts(),
3134 gcc_x86_copts = ["-mavx512f"],
3135 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3136 msvc_copts = xnnpack_msvc_std_copts(),
3137 msvc_x86_32_copts = ["/arch:AVX512"],
3138 msvc_x86_64_copts = ["/arch:AVX512"],
3139 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003141 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003142 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003143 "@FP16",
3144 "@pthreadpool",
3145 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146)
3147
3148xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003149 name = "avx512f_ukernels_test_mode",
3150 hdrs = INTERNAL_HDRS,
3151 copts = [
3152 "-UNDEBUG",
3153 "-DXNN_TEST_MODE=1",
3154 ],
3155 gcc_copts = xnnpack_gcc_std_copts(),
3156 gcc_x86_copts = ["-mavx512f"],
3157 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3158 msvc_copts = xnnpack_msvc_std_copts(),
3159 msvc_x86_32_copts = ["/arch:AVX512"],
3160 msvc_x86_64_copts = ["/arch:AVX512"],
3161 msys_copts = ["-fno-asynchronous-unwind-tables"],
3162 x86_srcs = AVX512F_UKERNELS,
3163 deps = [
3164 ":tables",
3165 "@FP16",
3166 "@pthreadpool",
3167 ],
3168)
3169
3170xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 name = "asm_ukernels",
3172 hdrs = ["src/xnnpack/assembly.h"],
3173 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003174 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175 aarch64_srcs = AARCH64_ASM_UKERNELS,
3176)
3177
Marat Dukhan3b59de22020-06-03 20:15:19 -07003178xnnpack_cc_library(
3179 name = "logging_utils",
3180 srcs = LOGGING_SRCS,
3181 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3182 copts = LOGGING_COPTS + [
3183 "-Isrc",
3184 "-Iinclude",
3185 ] + select({
3186 ":debug_build": [],
3187 "//conditions:default": xnnpack_min_size_copts(),
3188 }),
3189 gcc_copts = xnnpack_gcc_std_copts(),
3190 msvc_copts = xnnpack_msvc_std_copts(),
3191 visibility = xnnpack_visibility(),
3192 deps = [
3193 "@FP16",
3194 "@clog",
3195 "@pthreadpool",
3196 ],
3197)
3198
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199xnnpack_aggregate_library(
3200 name = "ukernels",
3201 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202 ":neon_ukernels",
3203 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003204 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205 ":asm_ukernels",
3206 ],
3207 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208 ":neon_ukernels",
3209 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003210 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003211 ":neonfp16arith_ukernels",
3212 ":asm_ukernels",
3213 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003214 generic_deps = [
3215 ":scalar_ukernels",
3216 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003217 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003218 ":psimd_fastmath_ukernels",
3219 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003221 wasm_deps = [
3222 ":wasm_ukernels",
3223 ],
3224 wasmsimd_deps = [
3225 ":wasm_ukernels",
3226 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003229 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003230 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003232 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003233 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003234 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235 ":avx512f_ukernels",
3236 ],
3237)
3238
Marat Dukhan33fcf782020-05-24 14:27:15 -07003239xnnpack_aggregate_library(
3240 name = "ukernels_test_mode",
3241 aarch32_deps = [
3242 ":neon_ukernels_test_mode",
3243 ":neonfma_ukernels_test_mode",
3244 ":neonv8_ukernels_test_mode",
3245 ":asm_ukernels",
3246 ],
3247 aarch64_deps = [
3248 ":neon_ukernels_test_mode",
3249 ":neonfma_ukernels_test_mode",
3250 ":neonv8_ukernels_test_mode",
3251 ":neonfp16arith_ukernels_test_mode",
3252 ":asm_ukernels",
3253 ],
3254 generic_deps = [
3255 ":scalar_ukernels_test_mode",
3256 ],
3257 psimd_deps = [
3258 ":psimd_fastmath_ukernels_test_mode",
3259 ":psimd_accmath_ukernels_test_mode",
3260 ],
3261 wasm_deps = [
3262 ":wasm_ukernels_test_mode",
3263 ],
3264 wasmsimd_deps = [
3265 ":wasm_ukernels_test_mode",
3266 ],
3267 x86_deps = [
3268 ":sse2_ukernels_test_mode",
3269 ":ssse3_ukernels_test_mode",
3270 ":sse41_ukernels_test_mode",
3271 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003272 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003273 ":fma3_ukernels_test_mode",
3274 ":avx2_ukernels_test_mode",
3275 ":avx512f_ukernels_test_mode",
3276 ],
3277)
3278
Marat Dukhan86df3842020-06-17 14:03:31 -07003279# This target must be compatible with Android NDK r17 and compile with Clang 6.
3280# It is similar to :ukernels target, but excludes microkernels which require
3281# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003282xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003283 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003284 aarch32_deps = [
3285 ":neon_ukernels",
3286 ":neonfma_ukernels",
3287 ":neonv8_ukernels",
3288 ":asm_ukernels",
3289 ],
3290 aarch64_deps = [
3291 ":neon_ukernels",
3292 ":neonfma_ukernels",
3293 ":neonv8_ukernels",
3294 ":asm_ukernels",
3295 ],
3296 generic_deps = [
3297 ":scalar_ukernels",
3298 ],
3299 psimd_deps = [
3300 ":psimd_fastmath_ukernels",
3301 ":psimd_accmath_ukernels",
3302 ],
3303 wasm_deps = [
3304 ":wasm_ukernels",
3305 ],
3306 wasmsimd_deps = [
3307 ":wasm_ukernels",
3308 ],
3309 x86_deps = [
3310 ":sse2_ukernels",
3311 ":ssse3_ukernels",
3312 ":sse41_ukernels",
3313 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003314 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003315 ":fma3_ukernels",
3316 ":avx2_ukernels",
3317 ":avx512f_ukernels",
3318 ],
3319)
3320
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321xnnpack_cc_library(
3322 name = "im2col",
3323 srcs = ["src/im2col.c"],
3324 hdrs = [
3325 "src/xnnpack/common.h",
3326 "src/xnnpack/im2col.h",
3327 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003328 gcc_copts = xnnpack_gcc_std_copts(),
3329 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003330)
3331
3332xnnpack_cc_library(
3333 name = "indirection",
3334 srcs = ["src/indirection.c"],
3335 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003336 gcc_copts = xnnpack_gcc_std_copts(),
3337 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003338 deps = [
3339 "@FP16",
3340 "@FXdiv",
3341 "@pthreadpool",
3342 ],
3343)
3344
3345xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003346 name = "indirection_test_mode",
3347 srcs = ["src/indirection.c"],
3348 hdrs = INTERNAL_HDRS,
3349 copts = [
3350 "-UNDEBUG",
3351 "-DXNN_TEST_MODE=1",
3352 ],
3353 gcc_copts = xnnpack_gcc_std_copts(),
3354 msvc_copts = xnnpack_msvc_std_copts(),
3355 deps = [
3356 "@FP16",
3357 "@FXdiv",
3358 "@pthreadpool",
3359 ],
3360)
3361
3362xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003363 name = "packing",
3364 srcs = ["src/packing.c"],
3365 hdrs = INTERNAL_HDRS,
3366 gcc_copts = xnnpack_gcc_std_copts(),
3367 msvc_copts = xnnpack_msvc_std_copts(),
3368 deps = [
3369 "@FP16",
3370 "@FXdiv",
3371 "@pthreadpool",
3372 ],
3373)
3374
3375xnnpack_cc_library(
3376 name = "packing_test_mode",
3377 srcs = ["src/packing.c"],
3378 hdrs = INTERNAL_HDRS,
3379 copts = [
3380 "-UNDEBUG",
3381 "-DXNN_TEST_MODE=1",
3382 ],
3383 gcc_copts = xnnpack_gcc_std_copts(),
3384 msvc_copts = xnnpack_msvc_std_copts(),
3385 deps = [
3386 "@FP16",
3387 "@FXdiv",
3388 "@pthreadpool",
3389 ],
3390)
3391
3392xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 name = "operator_run",
3394 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003395 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003396 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003397 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3398 "//conditions:default": [],
3399 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003400 gcc_copts = xnnpack_gcc_std_copts(),
3401 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003403 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404 "@FP16",
3405 "@FXdiv",
3406 "@clog",
3407 "@pthreadpool",
3408 ],
3409)
3410
Chao Mei6ddfc602020-05-13 22:29:36 -07003411xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003412 name = "operator_run_test_mode",
3413 srcs = ["src/operator-run.c"],
3414 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3415 copts = LOGGING_COPTS + [
3416 "-UNDEBUG",
3417 "-DXNN_TEST_MODE=1",
3418 ] + select({
3419 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3420 "//conditions:default": [],
3421 }),
3422 gcc_copts = xnnpack_gcc_std_copts(),
3423 msvc_copts = xnnpack_msvc_std_copts(),
3424 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003425 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003426 "@FP16",
3427 "@FXdiv",
3428 "@clog",
3429 "@pthreadpool",
3430 ],
3431)
3432
3433xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003434 name = "memory_planner",
3435 srcs = ["src/memory-planner.c"],
3436 hdrs = INTERNAL_HDRS,
3437 defines = select({
3438 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3439 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3440 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3441 }),
3442 gcc_copts = xnnpack_gcc_std_copts(),
3443 msvc_copts = xnnpack_msvc_std_copts(),
3444 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003445 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003446 "@pthreadpool",
3447 ],
3448)
3449
Marat Dukhan33fcf782020-05-24 14:27:15 -07003450xnnpack_cc_library(
3451 name = "memory_planner_test_mode",
3452 srcs = ["src/memory-planner.c"],
3453 hdrs = INTERNAL_HDRS,
3454 copts = [
3455 "-UNDEBUG",
3456 "-DXNN_TEST_MODE=1",
3457 ],
3458 defines = select({
3459 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3460 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3461 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3462 }),
3463 gcc_copts = xnnpack_gcc_std_copts(),
3464 msvc_copts = xnnpack_msvc_std_copts(),
3465 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003466 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 "@pthreadpool",
3468 ],
3469)
3470
Marat Dukhan08c4a432019-10-03 09:29:21 -07003471cc_library(
3472 name = "enable_assembly",
3473 defines = select({
3474 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3475 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003476 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 }),
3478)
3479
Marat Dukhan9de90e02020-06-18 16:04:12 -07003480cc_library(
3481 name = "enable_sparse",
3482 defines = select({
3483 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3484 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003485 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003486 }),
3487)
3488
Marat Dukhancf056b22019-10-07 10:26:29 -07003489xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 name = "operators",
3491 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003492 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003494 ],
3495 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003496 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497 "-Isrc",
3498 "-Iinclude",
3499 ] + select({
3500 ":debug_build": [],
3501 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003502 }) + select({
3503 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3504 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003506 gcc_copts = xnnpack_gcc_std_copts(),
3507 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003509 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003510 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003511 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003512 "@FP16",
3513 "@FXdiv",
3514 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003516 ],
3517)
3518
Marat Dukhan10a38082020-04-17 03:58:35 -07003519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003520 name = "operators_test_mode",
3521 srcs = OPERATOR_SRCS + [
3522 "src/memory.c",
3523 "src/operator-delete.c",
3524 ],
3525 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3526 copts = LOGGING_COPTS + [
3527 "-Isrc",
3528 "-Iinclude",
3529 "-UNDEBUG",
3530 "-DXNN_TEST_MODE=1",
3531 ] + select({
3532 ":debug_build": [],
3533 "//conditions:default": xnnpack_min_size_copts(),
3534 }) + select({
3535 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3536 "//conditions:default": [],
3537 }),
3538 gcc_copts = xnnpack_gcc_std_copts(),
3539 msvc_copts = xnnpack_msvc_std_copts(),
3540 deps = [
3541 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003542 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003543 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003544 "@FP16",
3545 "@FXdiv",
3546 "@clog",
3547 "@pthreadpool",
3548 ],
3549)
3550
3551xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003552 name = "XNNPACK",
3553 srcs = [
3554 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003555 "src/runtime.c",
3556 "src/subgraph.c",
3557 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003558 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003559 hdrs = ["include/xnnpack.h"],
3560 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003561 "-Isrc",
3562 "-Iinclude",
3563 ] + select({
3564 ":debug_build": [],
3565 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003566 }) + select({
3567 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3568 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003569 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003570 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003571 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003572 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003573 visibility = xnnpack_visibility(),
3574 deps = [
3575 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003576 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003577 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003578 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003579 ":operator_run",
3580 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003581 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003582 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003583 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003584 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003585 ] + select({
3586 ":emscripten": [],
3587 "//conditions:default": ["@cpuinfo"],
3588 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
Marat Dukhan10a38082020-04-17 03:58:35 -07003591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "XNNPACK_test_mode",
3593 srcs = [
3594 "src/init.c",
3595 "src/runtime.c",
3596 "src/subgraph.c",
3597 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003598 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003599 hdrs = ["include/xnnpack.h"],
3600 copts = LOGGING_COPTS + [
3601 "-Isrc",
3602 "-Iinclude",
3603 "-UNDEBUG",
3604 "-DXNN_TEST_MODE=1",
3605 ] + select({
3606 ":debug_build": [],
3607 "//conditions:default": xnnpack_min_size_copts(),
3608 }) + select({
3609 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3610 "//conditions:default": [],
3611 }),
3612 gcc_copts = xnnpack_gcc_std_copts(),
3613 includes = ["include"],
3614 msvc_copts = xnnpack_msvc_std_copts(),
3615 visibility = xnnpack_visibility(),
3616 deps = [
3617 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003618 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003619 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003620 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003621 ":operator_run_test_mode",
3622 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003623 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003624 "@clog",
3625 "@FP16",
3626 "@pthreadpool",
3627 ] + select({
3628 ":emscripten": [],
3629 "//conditions:default": ["@cpuinfo"],
3630 }),
3631)
3632
3633xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003634 name = "xnnpack_f32",
3635 srcs = [
3636 "src/init.c",
3637 "src/runtime.c",
3638 "src/subgraph.c",
3639 "src/tensor.c",
3640 ] + SUBGRAPH_SRCS,
3641 hdrs = ["include/xnnpack.h"],
3642 copts = LOGGING_COPTS + [
3643 "-Isrc",
3644 "-Iinclude",
3645 ] + select({
3646 ":debug_build": [],
3647 "//conditions:default": xnnpack_min_size_copts(),
3648 }) + select({
3649 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3650 "//conditions:default": [],
3651 }),
3652 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003653 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003654 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003655 "XNN_NO_U8_OPERATORS",
3656 "XNN_NO_X8_OPERATORS",
3657 "XNN_NO_F16_OPERATORS",
3658 "XNN_NO_X16_OPERATORS",
3659 ],
3660 gcc_copts = xnnpack_gcc_std_copts(),
3661 includes = ["include"],
3662 msvc_copts = xnnpack_msvc_std_copts(),
3663 visibility = xnnpack_visibility(),
3664 deps = [
3665 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003666 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003667 ":logging_utils",
3668 ":memory_planner",
3669 ":operator_run",
3670 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003671 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003672 "@clog",
3673 "@FP16",
3674 "@pthreadpool",
3675 ] + select({
3676 ":emscripten": [],
3677 "//conditions:default": ["@cpuinfo"],
3678 }),
3679)
3680
3681xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003682 name = "xnnpack_operators_nhwc_f32",
3683 srcs = [
3684 "src/init.c",
3685 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003686 hdrs = ["include/xnnpack.h"],
3687 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003688 "-Isrc",
3689 "-Iinclude",
3690 ] + select({
3691 ":debug_build": [],
3692 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003693 }) + select({
3694 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3695 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003696 }),
3697 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003698 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003699 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003700 "XNN_NO_U8_OPERATORS",
3701 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003702 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003703 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003704 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003706 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 visibility = xnnpack_visibility(),
3708 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003709 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003710 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 ":operator_run",
3712 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003713 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003714 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003716 ] + select({
3717 ":emscripten": [],
3718 "//conditions:default": ["@cpuinfo"],
3719 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003720)
3721
Marat Dukhancf056b22019-10-07 10:26:29 -07003722xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723 name = "bench_utils",
3724 srcs = ["bench/utils.cc"],
3725 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003726 deps = [
3727 "@com_google_benchmark//:benchmark",
3728 "@cpuinfo",
3729 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730)
3731
Frank Barchard7e955972019-10-11 10:34:25 -07003732######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003733
3734xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003735 name = "qs8_gemm_bench",
3736 srcs = [
3737 "bench/gemm.h",
3738 "bench/qs8-gemm.cc",
3739 "src/xnnpack/AlignedAllocator.h",
3740 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3741 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3742)
3743
3744xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003745 name = "qs8_requantization_bench",
3746 srcs = [
3747 "bench/qs8-requantization.cc",
3748 "src/xnnpack/requantization-stubs.h",
3749 "src/xnnpack/AlignedAllocator.h",
3750 ] + MICROKERNEL_BENCHMARK_HDRS,
3751 deps = MICROKERNEL_BENCHMARK_DEPS,
3752)
3753
3754xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003755 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 srcs = [
3757 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003758 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759 "src/xnnpack/AlignedAllocator.h",
3760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003761 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003762 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003763)
3764
3765xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003766 name = "qu8_requantization_bench",
3767 srcs = [
3768 "bench/qu8-requantization.cc",
3769 "src/xnnpack/requantization-stubs.h",
3770 "src/xnnpack/AlignedAllocator.h",
3771 ] + MICROKERNEL_BENCHMARK_HDRS,
3772 deps = MICROKERNEL_BENCHMARK_DEPS,
3773)
3774
3775xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003776 name = "f16_igemm_bench",
3777 srcs = [
3778 "bench/f16-igemm.cc",
3779 "bench/conv.h",
3780 "bench/google/conv.h",
3781 "src/xnnpack/AlignedAllocator.h",
3782 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003783 deps = MICROKERNEL_BENCHMARK_DEPS + [
3784 ":indirection",
3785 ":packing",
3786 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003787)
3788
3789xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790 name = "f16_gemm_bench",
3791 srcs = [
3792 "bench/f16-gemm.cc",
3793 "bench/gemm.h",
3794 "src/xnnpack/AlignedAllocator.h",
3795 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003796 deps = MICROKERNEL_BENCHMARK_DEPS + [
3797 ":packing",
3798 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799)
3800
3801xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003802 name = "f16_spmm_bench",
3803 srcs = [
3804 "bench/f16-spmm.cc",
3805 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003806 "src/xnnpack/AlignedAllocator.h",
3807 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003808 deps = MICROKERNEL_BENCHMARK_DEPS,
3809)
3810
3811xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003812 name = "f32_igemm_bench",
3813 srcs = [
3814 "bench/f32-igemm.cc",
3815 "bench/conv.h",
3816 "src/xnnpack/AlignedAllocator.h",
3817 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003818 deps = MICROKERNEL_BENCHMARK_DEPS + [
3819 ":indirection",
3820 ":packing",
3821 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822)
3823
3824xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003825 name = "f16_relu_bench",
3826 srcs = [
3827 "bench/f16-relu.cc",
3828 "src/xnnpack/AlignedAllocator.h",
3829 ] + MICROKERNEL_BENCHMARK_HDRS,
3830 deps = MICROKERNEL_BENCHMARK_DEPS,
3831)
3832
3833xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834 name = "f32_conv_hwc_bench",
3835 srcs = [
3836 "bench/f32-conv-hwc.cc",
3837 "bench/dconv.h",
3838 "src/xnnpack/AlignedAllocator.h",
3839 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003840 deps = MICROKERNEL_BENCHMARK_DEPS + [
3841 ":packing",
3842 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843)
3844
3845xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003846 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003847 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003848 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003849 "bench/dconv.h",
3850 "src/xnnpack/AlignedAllocator.h",
3851 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003852 deps = MICROKERNEL_BENCHMARK_DEPS + [
3853 ":packing",
3854 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003855)
3856
3857xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003858 name = "f16_dwconv_bench",
3859 srcs = [
3860 "bench/f16-dwconv.cc",
3861 "bench/dwconv.h",
3862 "bench/google/dwconv.h",
3863 "src/xnnpack/AlignedAllocator.h",
3864 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003865 deps = MICROKERNEL_BENCHMARK_DEPS + [
3866 ":indirection",
3867 ":packing",
3868 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003869)
3870
3871xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003872 name = "f32_dwconv_bench",
3873 srcs = [
3874 "bench/f32-dwconv.cc",
3875 "bench/dwconv.h",
3876 "src/xnnpack/AlignedAllocator.h",
3877 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003878 deps = MICROKERNEL_BENCHMARK_DEPS + [
3879 ":indirection",
3880 ":packing",
3881 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882)
3883
3884xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003885 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003886 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003887 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003888 "bench/dwconv.h",
3889 "src/xnnpack/AlignedAllocator.h",
3890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003891 deps = MICROKERNEL_BENCHMARK_DEPS + [
3892 ":indirection",
3893 ":packing",
3894 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003895)
3896
3897xnnpack_benchmark(
3898 name = "f32_gemm_bench",
3899 srcs = [
3900 "bench/f32-gemm.cc",
3901 "bench/gemm.h",
3902 "src/xnnpack/AlignedAllocator.h",
3903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003904 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003905 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906)
3907
3908xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003909 name = "f32_hswish_bench",
3910 srcs = [
3911 "bench/f32-hswish.cc",
3912 "src/xnnpack/AlignedAllocator.h",
3913 ] + MICROKERNEL_BENCHMARK_HDRS,
3914 deps = MICROKERNEL_BENCHMARK_DEPS,
3915)
3916
3917xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003918 name = "f32_raddexpminusmax_bench",
3919 srcs = [
3920 "bench/f32-raddexpminusmax.cc",
3921 "src/xnnpack/AlignedAllocator.h",
3922 ] + MICROKERNEL_BENCHMARK_HDRS,
3923 deps = MICROKERNEL_BENCHMARK_DEPS,
3924)
3925
3926xnnpack_benchmark(
3927 name = "f32_raddextexp_bench",
3928 srcs = [
3929 "bench/f32-raddextexp.cc",
3930 "src/xnnpack/AlignedAllocator.h",
3931 ] + MICROKERNEL_BENCHMARK_HDRS,
3932 deps = MICROKERNEL_BENCHMARK_DEPS,
3933)
3934
3935xnnpack_benchmark(
3936 name = "f32_raddstoreexpminusmax_bench",
3937 srcs = [
3938 "bench/f32-raddstoreexpminusmax.cc",
3939 "src/xnnpack/AlignedAllocator.h",
3940 ] + MICROKERNEL_BENCHMARK_HDRS,
3941 deps = MICROKERNEL_BENCHMARK_DEPS,
3942)
3943
3944xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003945 name = "f32_relu_bench",
3946 srcs = [
3947 "bench/f32-relu.cc",
3948 "src/xnnpack/AlignedAllocator.h",
3949 ] + MICROKERNEL_BENCHMARK_HDRS,
3950 deps = MICROKERNEL_BENCHMARK_DEPS,
3951)
3952
3953xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 name = "f32_rmax_bench",
3955 srcs = [
3956 "bench/f32-rmax.cc",
3957 "src/xnnpack/AlignedAllocator.h",
3958 ] + MICROKERNEL_BENCHMARK_HDRS,
3959 deps = MICROKERNEL_BENCHMARK_DEPS,
3960)
3961
3962xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003963 name = "f32_sigmoid_bench",
3964 srcs = [
3965 "bench/f32-sigmoid.cc",
3966 "src/xnnpack/AlignedAllocator.h",
3967 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003968 deps = MICROKERNEL_BENCHMARK_DEPS,
3969)
3970
3971xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003972 name = "f32_spmm_bench",
3973 srcs = [
3974 "bench/f32-spmm.cc",
3975 "bench/gemm.h",
3976 "src/xnnpack/AlignedAllocator.h",
3977 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003978 deps = MICROKERNEL_BENCHMARK_DEPS,
3979)
3980
3981xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003982 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003983 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003984 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003985 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003986 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003987 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003988)
3989
3990xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003991 name = "f32_vscaleexpminusmax_bench",
3992 srcs = [
3993 "bench/f32-vscaleexpminusmax.cc",
3994 "src/xnnpack/AlignedAllocator.h",
3995 ] + MICROKERNEL_BENCHMARK_HDRS,
3996 deps = MICROKERNEL_BENCHMARK_DEPS,
3997)
3998
3999xnnpack_benchmark(
4000 name = "f32_vscaleextexp_bench",
4001 srcs = [
4002 "bench/f32-vscaleextexp.cc",
4003 "src/xnnpack/AlignedAllocator.h",
4004 ] + MICROKERNEL_BENCHMARK_HDRS,
4005 deps = MICROKERNEL_BENCHMARK_DEPS,
4006)
4007
4008xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004009 name = "f32_vsqrt_bench",
4010 srcs = [
4011 "bench/f32-vsqrt.cc",
4012 "src/xnnpack/AlignedAllocator.h",
4013 ] + MICROKERNEL_BENCHMARK_HDRS,
4014 deps = MICROKERNEL_BENCHMARK_DEPS,
4015)
4016
4017xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004018 name = "f32_im2col_gemm_bench",
4019 srcs = [
4020 "bench/f32-im2col-gemm.cc",
4021 "bench/conv.h",
4022 "src/xnnpack/AlignedAllocator.h",
4023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004024 deps = MICROKERNEL_BENCHMARK_DEPS + [
4025 ":im2col",
4026 ":packing",
4027 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004028)
4029
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004030xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004031 name = "rounding_bench",
4032 srcs = [
4033 "bench/rounding.cc",
4034 "src/xnnpack/math-stubs.h",
4035 "src/xnnpack/AlignedAllocator.h",
4036 ] + MICROKERNEL_BENCHMARK_HDRS,
4037 deps = MICROKERNEL_BENCHMARK_DEPS,
4038)
4039
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040########################### Benchmarks for operators ###########################
4041
4042xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004043 name = "average_pooling_bench",
4044 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004045 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004046 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004047 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004048)
4049
4050xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004051 name = "bankers_rounding_bench",
4052 srcs = ["bench/bankers-rounding.cc"],
4053 copts = xnnpack_optional_tflite_copts(),
4054 tags = ["nowin32"],
4055 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4056)
4057
4058xnnpack_benchmark(
4059 name = "ceiling_bench",
4060 srcs = ["bench/ceiling.cc"],
4061 copts = xnnpack_optional_tflite_copts(),
4062 tags = ["nowin32"],
4063 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4064)
4065
4066xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004067 name = "channel_shuffle_bench",
4068 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004069 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004070)
4071
4072xnnpack_benchmark(
4073 name = "convolution_bench",
4074 srcs = ["bench/convolution.cc"],
4075 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004076 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004077 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078)
4079
4080xnnpack_benchmark(
4081 name = "deconvolution_bench",
4082 srcs = ["bench/deconvolution.cc"],
4083 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004084 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004085 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004086)
4087
4088xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004089 name = "floor_bench",
4090 srcs = ["bench/floor.cc"],
4091 copts = xnnpack_optional_tflite_copts(),
4092 tags = ["nowin32"],
4093 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4094)
4095
4096xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097 name = "global_average_pooling_bench",
4098 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004099 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004100)
4101
4102xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004103 name = "hardswish_bench",
4104 srcs = ["bench/hardswish.cc"],
4105 copts = xnnpack_optional_tflite_copts(),
4106 tags = ["nowin32"],
4107 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4108)
4109
4110xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004111 name = "max_pooling_bench",
4112 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004113 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004114)
4115
4116xnnpack_benchmark(
4117 name = "sigmoid_bench",
4118 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004119 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004120 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004121 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122)
4123
4124xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004125 name = "prelu_bench",
4126 srcs = ["bench/prelu.cc"],
4127 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004128 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004129 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004130)
4131
4132xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004133 name = "softmax_bench",
4134 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004135 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004136 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004137 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004138)
4139
Marat Dukhan87727142020-06-24 15:24:10 -07004140xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004141 name = "square_root_bench",
4142 srcs = ["bench/square-root.cc"],
4143 copts = xnnpack_optional_tflite_copts(),
4144 tags = ["nowin32"],
4145 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4146)
4147
4148xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004149 name = "truncation_bench",
4150 srcs = ["bench/truncation.cc"],
4151 deps = OPERATOR_BENCHMARK_DEPS,
4152)
4153
Marat Dukhanc068bb62019-10-04 13:24:39 -07004154############################# End-to-end benchmarks ############################
4155
4156cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004157 name = "fp32_mobilenet_v1",
4158 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004159 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004160 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004161 linkstatic = True,
4162 deps = [
4163 ":XNNPACK",
4164 "@pthreadpool",
4165 ],
4166)
4167
4168cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004169 name = "fp16_mobilenet_v1",
4170 srcs = ["models/fp16-mobilenet-v1.cc"],
4171 hdrs = ["models/models.h"],
4172 copts = xnnpack_std_cxxopts(),
4173 linkstatic = True,
4174 deps = [
4175 ":XNNPACK",
4176 "@FP16",
4177 "@pthreadpool",
4178 ],
4179)
4180
4181cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004182 name = "qs8_mobilenet_v1",
4183 srcs = ["models/qs8-mobilenet-v1.cc"],
4184 hdrs = ["models/models.h"],
4185 copts = xnnpack_std_cxxopts(),
4186 linkstatic = True,
4187 deps = [
4188 ":XNNPACK",
4189 "@pthreadpool",
4190 ],
4191)
4192
4193cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004194 name = "fp32_mobilenet_v2",
4195 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004196 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004197 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004198 linkstatic = True,
4199 deps = [
4200 ":XNNPACK",
4201 "@pthreadpool",
4202 ],
4203)
4204
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004205cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004206 name = "fp16_mobilenet_v2",
4207 srcs = ["models/fp16-mobilenet-v2.cc"],
4208 hdrs = ["models/models.h"],
4209 copts = xnnpack_std_cxxopts(),
4210 linkstatic = True,
4211 deps = [
4212 ":XNNPACK",
4213 "@FP16",
4214 "@pthreadpool",
4215 ],
4216)
4217
4218cc_library(
4219 name = "fp32_mobilenet_v3_large",
4220 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004221 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004222 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004223 linkstatic = True,
4224 deps = [
4225 ":XNNPACK",
4226 "@pthreadpool",
4227 ],
4228)
4229
4230cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004231 name = "fp32_mobilenet_v3_small",
4232 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004233 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004234 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004235 linkstatic = True,
4236 deps = [
4237 ":XNNPACK",
4238 "@pthreadpool",
4239 ],
4240)
4241
Marat Dukhanc068bb62019-10-04 13:24:39 -07004242xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004243 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004244 srcs = [
4245 "bench/f32-dwconv-e2e.cc",
4246 "bench/end2end.h",
4247 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004248 deps = MICROKERNEL_BENCHMARK_DEPS + [
4249 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004250 ":fp32_mobilenet_v1",
4251 ":fp32_mobilenet_v2",
4252 ":fp32_mobilenet_v3_large",
4253 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004254 ],
4255)
4256
4257xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004258 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004259 srcs = [
4260 "bench/f32-gemm-e2e.cc",
4261 "bench/end2end.h",
4262 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004263 deps = MICROKERNEL_BENCHMARK_DEPS + [
4264 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004265 ":fp32_mobilenet_v1",
4266 ":fp32_mobilenet_v2",
4267 ":fp32_mobilenet_v3_large",
4268 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004269 ],
4270)
4271
4272xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004273 name = "end2end_bench",
4274 srcs = ["bench/end2end.cc"],
4275 deps = [
4276 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004277 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004278 ":fp16_mobilenet_v1",
4279 ":fp16_mobilenet_v2",
4280 ":fp32_mobilenet_v1",
4281 ":fp32_mobilenet_v2",
4282 ":fp32_mobilenet_v3_large",
4283 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004284 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004285 "@pthreadpool",
4286 ],
4287)
4288
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004289#################### Accuracy evaluation for math functions ####################
4290
4291xnnpack_benchmark(
4292 name = "f32_exp_eval",
4293 srcs = [
4294 "eval/f32-exp.cc",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + ACCURACY_EVAL_HDRS,
4297 deps = ACCURACY_EVAL_DEPS,
4298)
4299
Marat Dukhan515c9772019-10-17 18:07:57 -07004300xnnpack_benchmark(
4301 name = "f32_expminus_eval",
4302 srcs = [
4303 "eval/f32-expminus.cc",
4304 "src/xnnpack/AlignedAllocator.h",
4305 ] + ACCURACY_EVAL_HDRS,
4306 deps = ACCURACY_EVAL_DEPS,
4307)
4308
Marat Dukhan98ba4412019-10-23 02:14:28 -07004309xnnpack_benchmark(
4310 name = "f32_extexp_eval",
4311 srcs = [
4312 "eval/f32-extexp.cc",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + ACCURACY_EVAL_HDRS,
4315 deps = ACCURACY_EVAL_DEPS,
4316)
4317
Marat Dukhan8853b822020-05-07 12:19:01 -07004318xnnpack_unit_test(
4319 name = "f32_roundne_eval",
4320 srcs = [
4321 "eval/f32-roundne.cc",
4322 "src/xnnpack/AlignedAllocator.h",
4323 "src/xnnpack/math-stubs.h",
4324 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004325 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004326 deps = MICROKERNEL_TEST_DEPS,
4327)
4328
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004329xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004330 name = "f32_roundd_eval",
4331 srcs = [
4332 "eval/f32-roundd.cc",
4333 "src/xnnpack/AlignedAllocator.h",
4334 "src/xnnpack/math-stubs.h",
4335 ] + MICROKERNEL_TEST_HDRS,
4336 automatic = False,
4337 deps = MICROKERNEL_TEST_DEPS,
4338)
4339
4340xnnpack_unit_test(
4341 name = "f32_roundu_eval",
4342 srcs = [
4343 "eval/f32-roundu.cc",
4344 "src/xnnpack/AlignedAllocator.h",
4345 "src/xnnpack/math-stubs.h",
4346 ] + MICROKERNEL_TEST_HDRS,
4347 automatic = False,
4348 deps = MICROKERNEL_TEST_DEPS,
4349)
4350
4351xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004352 name = "f32_roundz_eval",
4353 srcs = [
4354 "eval/f32-roundz.cc",
4355 "src/xnnpack/AlignedAllocator.h",
4356 "src/xnnpack/math-stubs.h",
4357 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004358 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004359 deps = MICROKERNEL_TEST_DEPS,
4360)
4361
Marat Dukhan346a9e52019-11-15 09:06:30 -08004362xnnpack_benchmark(
4363 name = "f32_sigmoid_eval",
4364 srcs = [
4365 "eval/f32-sigmoid.cc",
4366 "src/xnnpack/AlignedAllocator.h",
4367 ] + ACCURACY_EVAL_HDRS,
4368 deps = ACCURACY_EVAL_DEPS,
4369)
4370
Marat Dukhan84000762020-06-29 18:38:43 -07004371xnnpack_benchmark(
4372 name = "f32_sqrt_eval",
4373 srcs = [
4374 "eval/f32-sqrt.cc",
4375 "src/xnnpack/AlignedAllocator.h",
4376 ] + ACCURACY_EVAL_HDRS,
4377 deps = ACCURACY_EVAL_DEPS + [
4378 ":bench_utils",
4379 ],
4380)
4381
Marat Dukhan08c4a432019-10-03 09:29:21 -07004382######################### Unit tests for micro-kernels #########################
4383
4384xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004385 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004387 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 "test/gemm-microkernel-tester.h",
4389 "src/xnnpack/AlignedAllocator.h",
4390 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004391 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004392)
4393
4394xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004395 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004396 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004397 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004398 "test/spmm-microkernel-tester.h",
4399 "src/xnnpack/AlignedAllocator.h",
4400 ] + MICROKERNEL_TEST_HDRS,
4401 deps = MICROKERNEL_TEST_DEPS,
4402)
4403
4404xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004405 name = "f32_argmaxpool_test",
4406 srcs = [
4407 "test/f32-argmaxpool.cc",
4408 "test/argmaxpool-microkernel-tester.h",
4409 "src/xnnpack/AlignedAllocator.h",
4410 ] + MICROKERNEL_TEST_HDRS,
4411 deps = MICROKERNEL_TEST_DEPS,
4412)
4413
4414xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004415 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004417 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 "test/avgpool-microkernel-tester.h",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + MICROKERNEL_TEST_HDRS,
4421 deps = MICROKERNEL_TEST_DEPS,
4422)
4423
4424xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004425 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004426 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004427 "test/f32-ibilinear.cc",
4428 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004429 "src/xnnpack/AlignedAllocator.h",
4430 ] + MICROKERNEL_TEST_HDRS,
4431 deps = MICROKERNEL_TEST_DEPS,
4432)
4433
4434xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004435 name = "f16_clamp_test",
4436 srcs = [
4437 "test/f16-clamp.cc",
4438 "test/clamp-microkernel-tester.h",
4439 ] + MICROKERNEL_TEST_HDRS,
4440 deps = MICROKERNEL_TEST_DEPS,
4441)
4442
4443xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 name = "f32_clamp_test",
4445 srcs = [
4446 "test/f32-clamp.cc",
4447 "test/clamp-microkernel-tester.h",
4448 ] + MICROKERNEL_TEST_HDRS,
4449 deps = MICROKERNEL_TEST_DEPS,
4450)
4451
4452xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004453 name = "f32_igemm_test",
4454 srcs = [
4455 "test/f32-igemm.cc",
4456 "test/gemm-microkernel-tester.h",
4457 "src/xnnpack/AlignedAllocator.h",
4458 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004459 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004460)
4461
4462xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004463 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004465 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 "test/gemm-microkernel-tester.h",
4467 "src/xnnpack/AlignedAllocator.h",
4468 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004469 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470)
4471
4472xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004473 name = "f32_igemm_minmax_test",
4474 srcs = [
4475 "test/f32-igemm-minmax.cc",
4476 "test/gemm-microkernel-tester.h",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004479 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004480)
4481
4482xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004483 name = "f16_igemm_minmax_test",
4484 srcs = [
4485 "test/f16-igemm-minmax.cc",
4486 "test/gemm-microkernel-tester.h",
4487 "src/xnnpack/AlignedAllocator.h",
4488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004489 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004490)
4491
4492xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493 name = "f32_conv_hwc_test",
4494 srcs = [
4495 "test/f32-conv-hwc.cc",
4496 "test/conv-hwc-microkernel-tester.h",
4497 "src/xnnpack/AlignedAllocator.h",
4498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004499 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500)
4501
4502xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004503 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004505 "test/f32-conv-hwc2chw.cc",
4506 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507 "src/xnnpack/AlignedAllocator.h",
4508 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004509 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510)
4511
4512xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004513 name = "f32_dwconv_test",
4514 srcs = [
4515 "test/f32-dwconv.cc",
4516 "test/dwconv-microkernel-tester.h",
4517 "src/xnnpack/AlignedAllocator.h",
4518 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004519 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004520)
4521
4522xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004523 name = "f16_dwconv_minmax_test",
4524 srcs = [
4525 "test/f16-dwconv-minmax.cc",
4526 "test/dwconv-microkernel-tester.h",
4527 "src/xnnpack/AlignedAllocator.h",
4528 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004529 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004530)
4531
4532xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004533 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004535 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536 "test/dwconv-microkernel-tester.h",
4537 "src/xnnpack/AlignedAllocator.h",
4538 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004539 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004540)
4541
4542xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004543 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004545 "test/f32-dwconv-chw.cc",
4546 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004547 "src/xnnpack/AlignedAllocator.h",
4548 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004549 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550)
4551
4552xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004553 name = "f16_gavgpool_minmax_test",
4554 srcs = [
4555 "test/f16-gavgpool-minmax.cc",
4556 "test/gavgpool-microkernel-tester.h",
4557 "src/xnnpack/AlignedAllocator.h",
4558 ] + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004563 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004564 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004565 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566 "test/gavgpool-microkernel-tester.h",
4567 "src/xnnpack/AlignedAllocator.h",
4568 ] + MICROKERNEL_TEST_HDRS,
4569 deps = MICROKERNEL_TEST_DEPS,
4570)
4571
4572xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004573 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004575 "test/f32-gavgpool-cw.cc",
4576 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004577 "src/xnnpack/AlignedAllocator.h",
4578 ] + MICROKERNEL_TEST_HDRS,
4579 deps = MICROKERNEL_TEST_DEPS,
4580)
4581
4582xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004583 name = "f32_gemm_test",
4584 srcs = [
4585 "test/f32-gemm.cc",
4586 "test/gemm-microkernel-tester.h",
4587 "src/xnnpack/AlignedAllocator.h",
4588 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004589 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004590)
4591
4592xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004593 name = "f32_gemm_relu_test",
4594 srcs = [
4595 "test/f32-gemm-relu.cc",
4596 "test/gemm-microkernel-tester.h",
4597 "src/xnnpack/AlignedAllocator.h",
4598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004599 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004600)
4601
4602xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004603 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004605 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004606 "test/gemm-microkernel-tester.h",
4607 "src/xnnpack/AlignedAllocator.h",
4608 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004609 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610)
4611
4612xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004613 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004614 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004615 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004616 "test/gemm-microkernel-tester.h",
4617 "src/xnnpack/AlignedAllocator.h",
4618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004619 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620)
4621
4622xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004623 name = "f16_hswish_test",
4624 srcs = [
4625 "test/f16-hswish.cc",
4626 "test/hswish-microkernel-tester.h",
4627 ] + MICROKERNEL_TEST_HDRS,
4628 deps = MICROKERNEL_TEST_DEPS,
4629)
4630
4631xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 name = "f32_hswish_test",
4633 srcs = [
4634 "test/f32-hswish.cc",
4635 "test/hswish-microkernel-tester.h",
4636 ] + MICROKERNEL_TEST_HDRS,
4637 deps = MICROKERNEL_TEST_DEPS,
4638)
4639
4640xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004641 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004642 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004643 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004644 "test/maxpool-microkernel-tester.h",
4645 ] + MICROKERNEL_TEST_HDRS,
4646 deps = MICROKERNEL_TEST_DEPS,
4647)
4648
4649xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004650 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004652 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004653 "test/avgpool-microkernel-tester.h",
4654 "src/xnnpack/AlignedAllocator.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004660 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004661 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004662 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 "test/gemm-microkernel-tester.h",
4664 "src/xnnpack/AlignedAllocator.h",
4665 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004666 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667)
4668
4669xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004670 name = "f16_prelu_test",
4671 srcs = [
4672 "test/f16-prelu.cc",
4673 "test/prelu-microkernel-tester.h",
4674 "src/xnnpack/AlignedAllocator.h",
4675 ] + MICROKERNEL_TEST_HDRS,
4676 deps = MICROKERNEL_TEST_DEPS,
4677)
4678
4679xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004680 name = "f32_prelu_test",
4681 srcs = [
4682 "test/f32-prelu.cc",
4683 "test/prelu-microkernel-tester.h",
4684 "src/xnnpack/AlignedAllocator.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004690 name = "f32_raddexpminusmax_test",
4691 srcs = [
4692 "test/f32-raddexpminusmax.cc",
4693 "test/raddexpminusmax-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004699 name = "f32_raddextexp_test",
4700 srcs = [
4701 "test/f32-raddextexp.cc",
4702 "test/raddextexp-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004708 name = "f32_raddstoreexpminusmax_test",
4709 srcs = [
4710 "test/f32-raddstoreexpminusmax.cc",
4711 "test/raddstoreexpminusmax-microkernel-tester.h",
4712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004717 name = "f32_relu_test",
4718 srcs = [
4719 "test/f32-relu.cc",
4720 "test/vunary-microkernel-tester.h",
4721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004726 name = "f32_rmax_test",
4727 srcs = [
4728 "test/f32-rmax.cc",
4729 "test/rmax-microkernel-tester.h",
4730 ] + MICROKERNEL_TEST_HDRS,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004735 name = "f32_sigmoid_test",
4736 srcs = [
4737 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004738 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004739 ] + MICROKERNEL_TEST_HDRS,
4740 deps = MICROKERNEL_TEST_DEPS,
4741)
4742
4743xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004744 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004745 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004746 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747 "test/spmm-microkernel-tester.h",
4748 "src/xnnpack/AlignedAllocator.h",
4749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004754 name = "f16_vadd_minmax_test",
4755 srcs = [
4756 "test/f16-vadd-minmax.cc",
4757 "test/vbinary-microkernel-tester.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 deps = MICROKERNEL_TEST_DEPS,
4760)
4761
4762xnnpack_unit_test(
4763 name = "f16_vaddc_minmax_test",
4764 srcs = [
4765 "test/f16-vaddc-minmax.cc",
4766 "test/vbinaryc-microkernel-tester.h",
4767 ] + MICROKERNEL_TEST_HDRS,
4768 deps = MICROKERNEL_TEST_DEPS,
4769)
4770
4771xnnpack_unit_test(
4772 name = "f16_vdiv_minmax_test",
4773 srcs = [
4774 "test/f16-vdiv-minmax.cc",
4775 "test/vbinary-microkernel-tester.h",
4776 ] + MICROKERNEL_TEST_HDRS,
4777 deps = MICROKERNEL_TEST_DEPS,
4778)
4779
4780xnnpack_unit_test(
4781 name = "f16_vdivc_minmax_test",
4782 srcs = [
4783 "test/f16-vdivc-minmax.cc",
4784 "test/vbinaryc-microkernel-tester.h",
4785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
4790 name = "f16_vrdivc_minmax_test",
4791 srcs = [
4792 "test/f16-vrdivc-minmax.cc",
4793 "test/vbinaryc-microkernel-tester.h",
4794 ] + MICROKERNEL_TEST_HDRS,
4795 deps = MICROKERNEL_TEST_DEPS,
4796)
4797
4798xnnpack_unit_test(
4799 name = "f16_vmax_test",
4800 srcs = [
4801 "test/f16-vmax.cc",
4802 "test/vbinary-microkernel-tester.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
4808 name = "f16_vmaxc_test",
4809 srcs = [
4810 "test/f16-vmaxc.cc",
4811 "test/vbinaryc-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
4817 name = "f16_vmin_test",
4818 srcs = [
4819 "test/f16-vmin.cc",
4820 "test/vbinary-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
4826 name = "f16_vminc_test",
4827 srcs = [
4828 "test/f16-vminc.cc",
4829 "test/vbinaryc-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
4835 name = "f16_vmul_minmax_test",
4836 srcs = [
4837 "test/f16-vmul-minmax.cc",
4838 "test/vbinary-microkernel-tester.h",
4839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
4844 name = "f16_vmulc_minmax_test",
4845 srcs = [
4846 "test/f16-vmulc-minmax.cc",
4847 "test/vbinaryc-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
4853 name = "f16_vsub_minmax_test",
4854 srcs = [
4855 "test/f16-vsub-minmax.cc",
4856 "test/vbinary-microkernel-tester.h",
4857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
4862 name = "f16_vsubc_minmax_test",
4863 srcs = [
4864 "test/f16-vsubc-minmax.cc",
4865 "test/vbinaryc-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
4871 name = "f16_vrsubc_minmax_test",
4872 srcs = [
4873 "test/f16-vrsubc-minmax.cc",
4874 "test/vbinaryc-microkernel-tester.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004880 name = "f32_vabs_test",
4881 srcs = [
4882 "test/f32-vabs.cc",
4883 "test/vunary-microkernel-tester.h",
4884 ] + MICROKERNEL_TEST_HDRS,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004889 name = "f32_vadd_test",
4890 srcs = [
4891 "test/f32-vadd.cc",
4892 "test/vbinary-microkernel-tester.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004898 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004900 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004901 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004902 ] + MICROKERNEL_TEST_HDRS,
4903 deps = MICROKERNEL_TEST_DEPS,
4904)
4905
4906xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004907 name = "f32_vaddc_test",
4908 srcs = [
4909 "test/f32-vaddc.cc",
4910 "test/vbinaryc-microkernel-tester.h",
4911 ] + MICROKERNEL_TEST_HDRS,
4912 deps = MICROKERNEL_TEST_DEPS,
4913)
4914
4915xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004916 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004917 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004918 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004919 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004920 ] + MICROKERNEL_TEST_HDRS,
4921 deps = MICROKERNEL_TEST_DEPS,
4922)
4923
4924xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004925 name = "f32_vdiv_test",
4926 srcs = [
4927 "test/f32-vdiv.cc",
4928 "test/vbinary-microkernel-tester.h",
4929 ] + MICROKERNEL_TEST_HDRS,
4930 deps = MICROKERNEL_TEST_DEPS,
4931)
4932
4933xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004934 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004935 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004936 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004937 "test/vbinary-microkernel-tester.h",
4938 ] + MICROKERNEL_TEST_HDRS,
4939 deps = MICROKERNEL_TEST_DEPS,
4940)
4941
4942xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004943 name = "f32_vdivc_test",
4944 srcs = [
4945 "test/f32-vdivc.cc",
4946 "test/vbinaryc-microkernel-tester.h",
4947 ] + MICROKERNEL_TEST_HDRS,
4948 deps = MICROKERNEL_TEST_DEPS,
4949)
4950
4951xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004952 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004953 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004954 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004955 "test/vbinaryc-microkernel-tester.h",
4956 ] + MICROKERNEL_TEST_HDRS,
4957 deps = MICROKERNEL_TEST_DEPS,
4958)
4959
4960xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004961 name = "f32_vrdivc_test",
4962 srcs = [
4963 "test/f32-vrdivc.cc",
4964 "test/vbinaryc-microkernel-tester.h",
4965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004970 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004971 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004972 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004973 "test/vbinaryc-microkernel-tester.h",
4974 ] + MICROKERNEL_TEST_HDRS,
4975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
4978xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004979 name = "f32_vmax_test",
4980 srcs = [
4981 "test/f32-vmax.cc",
4982 "test/vbinary-microkernel-tester.h",
4983 ] + MICROKERNEL_TEST_HDRS,
4984 deps = MICROKERNEL_TEST_DEPS,
4985)
4986
4987xnnpack_unit_test(
4988 name = "f32_vmaxc_test",
4989 srcs = [
4990 "test/f32-vmaxc.cc",
4991 "test/vbinaryc-microkernel-tester.h",
4992 ] + MICROKERNEL_TEST_HDRS,
4993 deps = MICROKERNEL_TEST_DEPS,
4994)
4995
4996xnnpack_unit_test(
4997 name = "f32_vmin_test",
4998 srcs = [
4999 "test/f32-vmin.cc",
5000 "test/vbinary-microkernel-tester.h",
5001 ] + MICROKERNEL_TEST_HDRS,
5002 deps = MICROKERNEL_TEST_DEPS,
5003)
5004
5005xnnpack_unit_test(
5006 name = "f32_vminc_test",
5007 srcs = [
5008 "test/f32-vminc.cc",
5009 "test/vbinaryc-microkernel-tester.h",
5010 ] + MICROKERNEL_TEST_HDRS,
5011 deps = MICROKERNEL_TEST_DEPS,
5012)
5013
5014xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005015 name = "f32_vmul_test",
5016 srcs = [
5017 "test/f32-vmul.cc",
5018 "test/vbinary-microkernel-tester.h",
5019 ] + MICROKERNEL_TEST_HDRS,
5020 deps = MICROKERNEL_TEST_DEPS,
5021)
5022
5023xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005024 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005026 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005027 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005028 ] + MICROKERNEL_TEST_HDRS,
5029 deps = MICROKERNEL_TEST_DEPS,
5030)
5031
5032xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005033 name = "f32_vmulc_test",
5034 srcs = [
5035 "test/f32-vmulc.cc",
5036 "test/vbinaryc-microkernel-tester.h",
5037 ] + MICROKERNEL_TEST_HDRS,
5038 deps = MICROKERNEL_TEST_DEPS,
5039)
5040
5041xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005042 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005043 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005044 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005045 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 ] + MICROKERNEL_TEST_HDRS,
5047 deps = MICROKERNEL_TEST_DEPS,
5048)
5049
5050xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005051 name = "f16_vmulcaddc_minmax_test",
5052 srcs = [
5053 "test/f16-vmulcaddc-minmax.cc",
5054 "test/vmulcaddc-microkernel-tester.h",
5055 "src/xnnpack/AlignedAllocator.h",
5056 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005057 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005058)
5059
5060xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005061 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005063 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005064 "test/vmulcaddc-microkernel-tester.h",
5065 "src/xnnpack/AlignedAllocator.h",
5066 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005067 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005068)
5069
5070xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005071 name = "f32_vlrelu_test",
5072 srcs = [
5073 "test/f32-vlrelu.cc",
5074 "test/vunary-microkernel-tester.h",
5075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005080 name = "f32_vneg_test",
5081 srcs = [
5082 "test/f32-vneg.cc",
5083 "test/vunary-microkernel-tester.h",
5084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005089 name = "f32_vrndne_test",
5090 srcs = [
5091 "test/f32-vrndne.cc",
5092 "test/vunary-microkernel-tester.h",
5093 ] + MICROKERNEL_TEST_HDRS,
5094 deps = MICROKERNEL_TEST_DEPS,
5095)
5096
5097xnnpack_unit_test(
5098 name = "f32_vrndz_test",
5099 srcs = [
5100 "test/f32-vrndz.cc",
5101 "test/vunary-microkernel-tester.h",
5102 ] + MICROKERNEL_TEST_HDRS,
5103 deps = MICROKERNEL_TEST_DEPS,
5104)
5105
5106xnnpack_unit_test(
5107 name = "f32_vrndu_test",
5108 srcs = [
5109 "test/f32-vrndu.cc",
5110 "test/vunary-microkernel-tester.h",
5111 ] + MICROKERNEL_TEST_HDRS,
5112 deps = MICROKERNEL_TEST_DEPS,
5113)
5114
5115xnnpack_unit_test(
5116 name = "f32_vrndd_test",
5117 srcs = [
5118 "test/f32-vrndd.cc",
5119 "test/vunary-microkernel-tester.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 deps = MICROKERNEL_TEST_DEPS,
5122)
5123
5124xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005125 name = "f32_vscale_test",
5126 srcs = [
5127 "test/f32-vscale.cc",
5128 "test/vscale-microkernel-tester.h",
5129 ] + MICROKERNEL_TEST_HDRS,
5130 deps = MICROKERNEL_TEST_DEPS,
5131)
5132
5133xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005134 name = "f32_vscaleexpminusmax_test",
5135 srcs = [
5136 "test/f32-vscaleexpminusmax.cc",
5137 "test/vscaleexpminusmax-microkernel-tester.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005143 name = "f32_vscaleextexp_test",
5144 srcs = [
5145 "test/f32-vscaleextexp.cc",
5146 "test/vscaleextexp-microkernel-tester.h",
5147 ] + MICROKERNEL_TEST_HDRS,
5148 deps = MICROKERNEL_TEST_DEPS,
5149)
5150
5151xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005152 name = "f32_vsqr_test",
5153 srcs = [
5154 "test/f32-vsqr.cc",
5155 "test/vunary-microkernel-tester.h",
5156 ] + MICROKERNEL_TEST_HDRS,
5157 deps = MICROKERNEL_TEST_DEPS,
5158)
5159
5160xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005161 name = "f32_vsqrdiff_test",
5162 srcs = [
5163 "test/f32-vsqrdiff.cc",
5164 "test/vbinary-microkernel-tester.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
5170 name = "f32_vsqrdiffc_test",
5171 srcs = [
5172 "test/f32-vsqrdiffc.cc",
5173 "test/vbinaryc-microkernel-tester.h",
5174 ] + MICROKERNEL_TEST_HDRS,
5175 deps = MICROKERNEL_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005179 name = "f32_vsqrt_test",
5180 srcs = [
5181 "test/f32-vsqrt.cc",
5182 "test/vunary-microkernel-tester.h",
5183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005188 name = "f32_vsub_test",
5189 srcs = [
5190 "test/f32-vsub.cc",
5191 "test/vbinary-microkernel-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005197 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005198 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005199 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005200 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005206 name = "f32_vsubc_test",
5207 srcs = [
5208 "test/f32-vsubc.cc",
5209 "test/vbinaryc-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
5214xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005215 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005216 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005217 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005218 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005224 name = "f32_vrsubc_test",
5225 srcs = [
5226 "test/f32-vrsubc.cc",
5227 "test/vbinaryc-microkernel-tester.h",
5228 ] + MICROKERNEL_TEST_HDRS,
5229 deps = MICROKERNEL_TEST_DEPS,
5230)
5231
5232xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005233 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005234 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005235 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005236 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005242 name = "qs8_dwconv_minmax_test",
5243 srcs = [
5244 "test/qs8-dwconv-minmax.cc",
5245 "test/dwconv-microkernel-tester.h",
5246 "src/xnnpack/AlignedAllocator.h",
5247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5249)
5250
5251xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005252 name = "qs8_gavgpool_minmax_test",
5253 srcs = [
5254 "test/qs8-gavgpool-minmax.cc",
5255 "test/gavgpool-microkernel-tester.h",
5256 "src/xnnpack/AlignedAllocator.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005262 name = "qs8_gemm_minmax_test",
5263 srcs = [
5264 "test/qs8-gemm-minmax.cc",
5265 "test/gemm-microkernel-tester.h",
5266 "src/xnnpack/AlignedAllocator.h",
5267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5269)
5270
5271xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005272 name = "qs8_igemm_minmax_test",
5273 srcs = [
5274 "test/qs8-igemm-minmax.cc",
5275 "test/gemm-microkernel-tester.h",
5276 "src/xnnpack/AlignedAllocator.h",
5277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5279)
5280
5281xnnpack_unit_test(
5282 name = "qs8_requantization_test",
5283 srcs = [
5284 "src/xnnpack/requantization-stubs.h",
5285 "test/qs8-requantization.cc",
5286 "test/requantization-tester.h",
5287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005292 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005294 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 "test/avgpool-microkernel-tester.h",
5296 "src/xnnpack/AlignedAllocator.h",
5297 ] + MICROKERNEL_TEST_HDRS,
5298 deps = MICROKERNEL_TEST_DEPS,
5299)
5300
5301xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005302 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005304 "test/qu8-dwconv-minmax.cc",
5305 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306 "src/xnnpack/AlignedAllocator.h",
5307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309)
5310
5311xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005312 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005314 "test/qu8-igemm-minmax.cc",
5315 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 "src/xnnpack/AlignedAllocator.h",
5317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005318 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319)
5320
5321xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005322 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005324 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 "test/gavgpool-microkernel-tester.h",
5326 "src/xnnpack/AlignedAllocator.h",
5327 ] + MICROKERNEL_TEST_HDRS,
5328 deps = MICROKERNEL_TEST_DEPS,
5329)
5330
5331xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005332 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005334 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335 "test/gemm-microkernel-tester.h",
5336 "src/xnnpack/AlignedAllocator.h",
5337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339)
5340
5341xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005342 name = "qu8_requantization_test",
5343 srcs = [
5344 "src/xnnpack/requantization-stubs.h",
5345 "test/qu8-requantization.cc",
5346 "test/requantization-tester.h",
5347 ] + MICROKERNEL_TEST_HDRS,
5348 deps = MICROKERNEL_TEST_DEPS,
5349)
5350
5351xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005352 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005354 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355 "test/vadd-microkernel-tester.h",
5356 ] + MICROKERNEL_TEST_HDRS,
5357 deps = MICROKERNEL_TEST_DEPS,
5358)
5359
5360xnnpack_unit_test(
5361 name = "u8_clamp_test",
5362 srcs = [
5363 "test/u8-clamp.cc",
5364 "test/clamp-microkernel-tester.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
5370 name = "u8_lut32norm_test",
5371 srcs = [
5372 "test/u8-lut32norm.cc",
5373 "test/lut-norm-microkernel-tester.h",
5374 ] + MICROKERNEL_TEST_HDRS,
5375 deps = MICROKERNEL_TEST_DEPS,
5376)
5377
5378xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005379 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005381 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005382 "test/maxpool-microkernel-tester.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
5388 name = "u8_rmax_test",
5389 srcs = [
5390 "test/u8-rmax.cc",
5391 "test/rmax-microkernel-tester.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 deps = MICROKERNEL_TEST_DEPS,
5394)
5395
5396xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005397 name = "x32_fill_test",
5398 srcs = [
5399 "test/x32-fill.cc",
5400 "test/fill-microkernel-tester.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 name = "x32_packx_test",
5407 srcs = [
5408 "test/x32-packx.cc",
5409 "test/pack-microkernel-tester.h",
5410 "src/xnnpack/AlignedAllocator.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
5416 name = "x32_pad_test",
5417 srcs = [
5418 "test/x32-pad.cc",
5419 "test/pad-microkernel-tester.h",
5420 ] + MICROKERNEL_TEST_HDRS,
5421 deps = MICROKERNEL_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
5425 name = "x32_unpool_test",
5426 srcs = [
5427 "test/x32-unpool.cc",
5428 "test/unpool-microkernel-tester.h",
5429 ] + MICROKERNEL_TEST_HDRS,
5430 deps = MICROKERNEL_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
5434 name = "x32_zip_test",
5435 srcs = [
5436 "test/x32-zip.cc",
5437 "test/zip-microkernel-tester.h",
5438 ] + MICROKERNEL_TEST_HDRS,
5439 deps = MICROKERNEL_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
5443 name = "x8_lut_test",
5444 srcs = [
5445 "test/x8-lut.cc",
5446 "test/lut-microkernel-tester.h",
5447 ] + MICROKERNEL_TEST_HDRS,
5448 deps = MICROKERNEL_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
5452 name = "x8_zip_test",
5453 srcs = [
5454 "test/x8-zip.cc",
5455 "test/zip-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
Marat Dukhan20c3b922020-03-10 03:45:06 -07005460########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461
5462xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005463 name = "operator_size_test",
5464 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005465 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466)
5467
Marat Dukhan20c3b922020-03-10 03:45:06 -07005468xnnpack_binary(
5469 name = "subgraph_size_test",
5470 srcs = ["test/subgraph-size.c"],
5471 deps = [":XNNPACK"],
5472)
5473
5474########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475
5476xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005477 name = "abs_nc_test",
5478 srcs = [
5479 "test/abs-nc.cc",
5480 "test/abs-operator-tester.h",
5481 ],
5482 deps = OPERATOR_TEST_DEPS,
5483)
5484
5485xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005486 name = "add_nd_test",
5487 srcs = [
5488 "test/add-nd.cc",
5489 "test/binary-elementwise-operator-tester.h",
5490 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005491 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005492)
5493
5494xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005495 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005497 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 "test/argmax-pooling-operator-tester.h",
5499 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005500 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501)
5502
5503xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005504 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005506 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005507 "test/average-pooling-operator-tester.h",
5508 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005509 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510)
5511
5512xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005513 name = "bankers_rounding_nc_test",
5514 srcs = [
5515 "test/bankers-rounding-nc.cc",
5516 "test/bankers-rounding-operator-tester.h",
5517 ],
5518 deps = OPERATOR_TEST_DEPS,
5519)
5520
5521xnnpack_unit_test(
5522 name = "ceiling_nc_test",
5523 srcs = [
5524 "test/ceiling-nc.cc",
5525 "test/ceiling-operator-tester.h",
5526 ],
5527 deps = OPERATOR_TEST_DEPS,
5528)
5529
5530xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005531 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005532 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005533 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005534 "test/channel-shuffle-operator-tester.h",
5535 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005536 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537)
5538
5539xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005540 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005542 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 "test/clamp-operator-tester.h",
5544 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005545 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546)
5547
5548xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005549 name = "constant_pad_nd_test",
5550 srcs = [
5551 "test/constant-pad-nd.cc",
5552 "test/constant-pad-operator-tester.h",
5553 ],
5554 deps = OPERATOR_TEST_DEPS,
5555)
5556
5557xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005558 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005560 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561 "test/convolution-operator-tester.h",
5562 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005563 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005564)
5565
5566xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005567 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005568 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005569 "test/convolution-nchw.cc",
5570 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005572 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573)
5574
5575xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005576 name = "copy_nc_test",
5577 srcs = [
5578 "test/copy-nc.cc",
5579 "test/copy-operator-tester.h",
5580 ],
5581 deps = OPERATOR_TEST_DEPS,
5582)
5583
5584xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005585 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005586 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005587 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005588 "test/deconvolution-operator-tester.h",
5589 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005590 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591)
5592
5593xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005594 name = "divide_nd_test",
5595 srcs = [
5596 "test/binary-elementwise-operator-tester.h",
5597 "test/divide-nd.cc",
5598 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005599 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005600)
5601
5602xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005603 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005604 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005605 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005606 "test/fully-connected-operator-tester.h",
5607 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005608 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609)
5610
5611xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005612 name = "floor_nc_test",
5613 srcs = [
5614 "test/floor-nc.cc",
5615 "test/floor-operator-tester.h",
5616 ],
5617 deps = OPERATOR_TEST_DEPS,
5618)
5619
5620xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005621 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005622 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005623 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005624 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005625 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005626 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627)
5628
5629xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005630 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005631 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005632 "test/global-average-pooling-ncw.cc",
5633 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005635 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005636)
5637
5638xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005639 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005640 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005641 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005642 "test/hardswish-operator-tester.h",
5643 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005644 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005645)
5646
5647xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005648 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005650 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005651 "test/leaky-relu-operator-tester.h",
5652 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005653 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654)
5655
5656xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005657 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005659 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005660 "test/max-pooling-operator-tester.h",
5661 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005662 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663)
5664
5665xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005666 name = "maximum_nd_test",
5667 srcs = [
5668 "test/binary-elementwise-operator-tester.h",
5669 "test/maximum-nd.cc",
5670 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005671 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005672)
5673
5674xnnpack_unit_test(
5675 name = "minimum_nd_test",
5676 srcs = [
5677 "test/binary-elementwise-operator-tester.h",
5678 "test/minimum-nd.cc",
5679 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005680 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005681)
5682
5683xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005684 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005685 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005686 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005687 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005688 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005689 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005690)
5691
5692xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005693 name = "negate_nc_test",
5694 srcs = [
5695 "test/negate-nc.cc",
5696 "test/negate-operator-tester.h",
5697 ],
5698 deps = OPERATOR_TEST_DEPS,
5699)
5700
5701xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005702 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005703 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005704 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005705 "test/prelu-operator-tester.h",
5706 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005707 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005708)
5709
5710xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005711 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005712 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005713 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005714 "test/resize-bilinear-operator-tester.h",
5715 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005716 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005717)
5718
5719xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005720 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005721 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005722 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005723 "test/sigmoid-operator-tester.h",
5724 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005725 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726)
5727
5728xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005729 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005730 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005731 "test/softmax-nc.cc",
5732 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005734 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005735)
5736
5737xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005738 name = "square_nc_test",
5739 srcs = [
5740 "test/square-nc.cc",
5741 "test/square-operator-tester.h",
5742 ],
5743 deps = OPERATOR_TEST_DEPS,
5744)
5745
5746xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005747 name = "square_root_nc_test",
5748 srcs = [
5749 "test/square-root-nc.cc",
5750 "test/square-root-operator-tester.h",
5751 ],
5752 deps = OPERATOR_TEST_DEPS,
5753)
5754
5755xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005756 name = "squared_difference_nd_test",
5757 srcs = [
5758 "test/binary-elementwise-operator-tester.h",
5759 "test/squared-difference-nd.cc",
5760 ],
5761 deps = OPERATOR_TEST_DEPS,
5762)
5763
5764xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005765 name = "subtract_nd_test",
5766 srcs = [
5767 "test/binary-elementwise-operator-tester.h",
5768 "test/subtract-nd.cc",
5769 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005770 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005771)
5772
5773xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005774 name = "truncation_nc_test",
5775 srcs = [
5776 "test/truncation-nc.cc",
5777 "test/truncation-operator-tester.h",
5778 ],
5779 deps = OPERATOR_TEST_DEPS,
5780)
5781
5782xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005783 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005784 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005785 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005786 "test/unpooling-operator-tester.h",
5787 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005788 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005789)
5790
Chao Mei6ddfc602020-05-13 22:29:36 -07005791############################### Misc unit tests ###############################
5792
5793xnnpack_unit_test(
5794 name = "memory_planner_test",
5795 srcs = [
5796 "test/memory-planner-test.cc",
5797 ],
5798 deps = [
5799 ":XNNPACK",
5800 ":memory_planner",
5801 ],
5802)
5803
Marat Dukhan08c4a432019-10-03 09:29:21 -07005804############################# Build configurations #############################
5805
Marat Dukhanb8642352019-10-30 15:43:02 -07005806# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005807config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005808 name = "xnn_enable_assembly_explicit_true",
5809 define_values = {"xnn_enable_assembly": "true"},
5810)
5811
5812# Disables usage of assembly kernels.
5813config_setting(
5814 name = "xnn_enable_assembly_explicit_false",
5815 define_values = {"xnn_enable_assembly": "false"},
5816)
5817
Marat Dukhan9de90e02020-06-18 16:04:12 -07005818# Enables usage of sparse inference.
5819config_setting(
5820 name = "xnn_enable_sparse_explicit_true",
5821 define_values = {"xnn_enable_sparse": "true"},
5822)
5823
5824# Disables usage of sparse inference.
5825config_setting(
5826 name = "xnn_enable_sparse_explicit_false",
5827 define_values = {"xnn_enable_sparse": "false"},
5828)
5829
Marat Dukhan05702cf2020-03-26 15:41:33 -07005830# Disables usage of HMP-aware optimizations.
5831config_setting(
5832 name = "xnn_enable_hmp_explicit_false",
5833 define_values = {"xnn_enable_hmp": "false"},
5834)
5835
Chao Mei6ddfc602020-05-13 22:29:36 -07005836# Enable usage of optimized memory allocation
5837config_setting(
5838 name = "xnn_enable_memopt_explicit_true",
5839 define_values = {"xnnpack_enable_memopt": "true"},
5840)
5841
5842# Disable usage of optimized memory allocation
5843config_setting(
5844 name = "xnn_enable_memopt_explicit_false",
5845 define_values = {"xnnpack_enable_memopt": "false"},
5846)
5847
Marat Dukhanb8642352019-10-30 15:43:02 -07005848# Builds with -c dbg
5849config_setting(
5850 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005851 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005852 "compilation_mode": "dbg",
5853 },
5854)
5855
5856# Builds with -c opt
5857config_setting(
5858 name = "optimized_build",
5859 values = {
5860 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005861 },
5862)
5863
5864config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005865 name = "linux_k8",
5866 values = {"cpu": "k8"},
5867)
5868
5869config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005870 name = "linux_arm",
5871 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005872)
5873
5874config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005875 name = "linux_armeabi",
5876 values = {"cpu": "armeabi"},
5877)
5878
5879config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005880 name = "linux_armhf",
5881 values = {"cpu": "armhf"},
5882)
5883
5884config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005885 name = "linux_armv7a",
5886 values = {"cpu": "armv7a"},
5887)
5888
5889config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005890 name = "linux_aarch64",
5891 values = {"cpu": "aarch64"},
5892)
5893
5894config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005895 name = "android",
5896 values = {"crosstool_top": "//external:android/crosstool"},
5897)
5898
5899config_setting(
5900 name = "android_armv7",
5901 values = {
5902 "crosstool_top": "//external:android/crosstool",
5903 "cpu": "armeabi-v7a",
5904 },
5905)
5906
5907config_setting(
5908 name = "android_arm64",
5909 values = {
5910 "crosstool_top": "//external:android/crosstool",
5911 "cpu": "arm64-v8a",
5912 },
5913)
5914
5915config_setting(
5916 name = "android_x86",
5917 values = {
5918 "crosstool_top": "//external:android/crosstool",
5919 "cpu": "x86",
5920 },
5921)
5922
5923config_setting(
5924 name = "android_x86_64",
5925 values = {
5926 "crosstool_top": "//external:android/crosstool",
5927 "cpu": "x86_64",
5928 },
5929)
5930
5931config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005932 name = "windows_x86_64",
5933 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005934)
5935
5936config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005937 name = "windows_x86_64_clang",
5938 values = {
5939 "compiler": "clang-cl",
5940 "cpu": "x64_windows",
5941 },
5942)
5943
5944config_setting(
5945 name = "windows_x86_64_mingw",
5946 values = {
5947 "compiler": "mingw-gcc",
5948 "cpu": "x64_windows",
5949 },
5950)
5951
5952config_setting(
5953 name = "windows_x86_64_msys",
5954 values = {
5955 "compiler": "msys-gcc",
5956 "cpu": "x64_windows",
5957 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005958)
5959
5960config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005961 name = "macos_x86_64",
5962 values = {
5963 "apple_platform_type": "macos",
5964 "cpu": "darwin",
5965 },
5966)
5967
5968config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005969 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005970 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005971)
5972
5973config_setting(
5974 name = "emscripten_wasm",
5975 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005976 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005977 "cpu": "wasm",
5978 },
5979)
5980
5981config_setting(
5982 name = "emscripten_wasmsimd",
5983 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005984 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005985 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005986 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005987 },
5988)
5989
5990config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005991 name = "ios_armv7",
5992 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005993 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005994 "cpu": "ios_armv7",
5995 },
5996)
5997
5998config_setting(
5999 name = "ios_arm64",
6000 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006001 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006002 "cpu": "ios_arm64",
6003 },
6004)
6005
6006config_setting(
6007 name = "ios_arm64e",
6008 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006009 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006010 "cpu": "ios_arm64e",
6011 },
6012)
6013
6014config_setting(
6015 name = "ios_x86",
6016 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006017 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006018 "cpu": "ios_i386",
6019 },
6020)
6021
6022config_setting(
6023 name = "ios_x86_64",
6024 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006025 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006026 "cpu": "ios_x86_64",
6027 },
6028)
6029
6030config_setting(
6031 name = "watchos_armv7k",
6032 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006033 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006034 "cpu": "watchos_armv7k",
6035 },
6036)
6037
6038config_setting(
6039 name = "watchos_arm64_32",
6040 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006041 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006042 "cpu": "watchos_arm64_32",
6043 },
6044)
6045
6046config_setting(
6047 name = "watchos_x86",
6048 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006049 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006050 "cpu": "watchos_i386",
6051 },
6052)
6053
6054config_setting(
6055 name = "watchos_x86_64",
6056 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006057 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006058 "cpu": "watchos_x86_64",
6059 },
6060)
6061
6062config_setting(
6063 name = "tvos_arm64",
6064 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006065 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006066 "cpu": "tvos_arm64",
6067 },
6068)
6069
6070config_setting(
6071 name = "tvos_x86_64",
6072 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006073 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006074 "cpu": "tvos_x86_64",
6075 },
6076)