blob: 2355ff6533a88ce97ab78f81b4ab6ae8cc58070f [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700851 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700852 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700853 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700854 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700855 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700856 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700857 "src/x32-zip/x2-wasmsimd.c",
858 "src/x32-zip/x3-wasmsimd.c",
859 "src/x32-zip/x4-wasmsimd.c",
860 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700861 "src/math/roundne-wasmsimd-addsub.c",
862 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700863 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700864 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700865 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700866 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700867 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700868 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
869 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700870]
871
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800872PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700873 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700874 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700875 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700876 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700877 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700878 "src/f32-spmm/gen/4x1-minmax-psimd.c",
879 "src/f32-spmm/gen/8x1-minmax-psimd.c",
880 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700881]
882
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800883PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700884 "src/qs8-requantization/precise-psimd.c",
885 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700886 "src/qu8-requantization/precise-psimd.c",
887 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800888]
889
Marat Dukhan08c4a432019-10-03 09:29:21 -0700890# ISA-specific micro-kernels
891NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700892 "src/f32-argmaxpool/4x-neon-c4.c",
893 "src/f32-argmaxpool/9p8x-neon-c4.c",
894 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700895 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
896 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700897 "src/f32-clamp/gen/neon-x4.c",
898 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700899 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700903 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
904 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700907 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
908 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
909 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
910 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700911 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
912 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
913 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
914 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700915 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
916 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
917 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
918 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700919 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700920 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
921 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700922 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
923 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
925 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
927 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
929 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
930 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
931 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
932 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
933 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
934 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
935 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
936 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
938 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
939 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
940 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
942 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
944 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
945 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
948 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
949 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
950 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
951 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700953 "src/f32-hswish/gen/hswish-neon-x4.c",
954 "src/f32-hswish/gen/hswish-neon-x8.c",
955 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700956 "src/f32-ibilinear/gen/neon-c4.c",
957 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700958 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
959 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
960 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
962 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
963 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
964 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
965 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
966 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
967 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
968 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
969 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
970 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
971 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
972 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700974 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
975 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
976 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700977 "src/f32-ppmm/gen/4x8-minmax-neon.c",
978 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700979 "src/f32-prelu/gen/neon-1x4.c",
980 "src/f32-prelu/gen/neon-1x8.c",
981 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800982 "src/f32-prelu/gen/neon-2x4.c",
983 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700984 "src/f32-prelu/gen/neon-2x16.c",
985 "src/f32-prelu/gen/neon-4x4.c",
986 "src/f32-prelu/gen/neon-4x8.c",
987 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800988 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001012 "src/f32-relu/gen/neon-x4.c",
1013 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001014 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001015 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001016 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1017 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1022 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001034 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1035 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1036 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1037 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001038 "src/f32-vbinary/gen/vmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vmax-neon-x8.c",
1040 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1041 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1042 "src/f32-vbinary/gen/vmin-neon-x4.c",
1043 "src/f32-vbinary/gen/vmin-neon-x8.c",
1044 "src/f32-vbinary/gen/vminc-neon-x4.c",
1045 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001046 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1047 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1048 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001052 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1053 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1054 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001056 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1057 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1058 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001060 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1061 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001062 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1063 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001064 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1065 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1066 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1067 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1068 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1069 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1070 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1071 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001072 "src/f32-vunary/gen/vabs-neon-x4.c",
1073 "src/f32-vunary/gen/vabs-neon-x8.c",
1074 "src/f32-vunary/gen/vneg-neon-x4.c",
1075 "src/f32-vunary/gen/vneg-neon-x8.c",
1076 "src/f32-vunary/gen/vsqr-neon-x4.c",
1077 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001078 "src/qs8-requantization/precise-neon.c",
1079 "src/qs8-requantization/fp32-neon.c",
1080 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001081 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1082 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1083 "src/qu8-dwconv/up8x9-minmax-neon.c",
1084 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1085 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1086 "src/qu8-gemm/4x8-minmax-neon.c",
1087 "src/qu8-gemm/8x8-minmax-neon.c",
1088 "src/qu8-igemm/4x8-minmax-neon.c",
1089 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001090 "src/qu8-requantization/precise-neon.c",
1091 "src/qu8-requantization/fp32-neon.c",
1092 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001093 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001094 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001095 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001096 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001097 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001099 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001100 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001101 "src/x32-zip/x2-neon.c",
1102 "src/x32-zip/x3-neon.c",
1103 "src/x32-zip/x4-neon.c",
1104 "src/x32-zip/xm-neon.c",
1105 "src/x8-zip/x2-neon.c",
1106 "src/x8-zip/x3-neon.c",
1107 "src/x8-zip/x4-neon.c",
1108 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001109 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001110 "src/math/roundd-neon-addsub.c",
1111 "src/math/roundd-neon-cvt.c",
1112 "src/math/roundu-neon-addsub.c",
1113 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001114 "src/math/roundz-neon-addsub.c",
1115 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001116 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001117 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1118 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1119 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1120 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1121 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001123 "src/math/sqrt-neon-nr1rsqrts.c",
1124 "src/math/sqrt-neon-nr2rsqrts.c",
1125 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126]
1127
1128NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001129 "src/f32-ibilinear/gen/neonfma-c4.c",
1130 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001131 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1132 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1133 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1134 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1136 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1137 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1138 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001140 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1141 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1142 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001144 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1145 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1146 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001148 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1149 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1150 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001152 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1153 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1154 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1155 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1157 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1158 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1159 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1164 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1166 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1167 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1168 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001170 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1171 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001172 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001196 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001250 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1251 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001252 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1253 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1262 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1263 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001272 "src/math/exp-neonfma-lut64-p2.c",
1273 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001274 "src/math/expminus-neonfma-lut2048-p1.c",
1275 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001276 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001277 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1278 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1280 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1281 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1283 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1284 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1286 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1287 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1289 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1290 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1292 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1293 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001295 "src/math/sqrt-neonfma-nr1fma.c",
1296 "src/math/sqrt-neonfma-nr2fma.c",
1297 "src/math/sqrt-neonfma-nr3fma.c",
1298 "src/math/sqrt-neonfma-nr2fma1adj.c",
1299 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001300]
1301
1302AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001303 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1304 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1305 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1306 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1307 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1308 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001309 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1310 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1312 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1313 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1316 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1317 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1318 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1319 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1322 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1323 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1324 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1326 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1327 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001329 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1330 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001333 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1334 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001335 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1336 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001337 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1338 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1339 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1340 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1341 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001342 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001360 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1361 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1362 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1363 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1364 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1366 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1367 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1368 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1369 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1371 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1372 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1373 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1374 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1376 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1377 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001378 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1379 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001380 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1381 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1382 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1383 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1384 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1385 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001386]
1387
Marat Dukhan8853b822020-05-07 12:19:01 -07001388NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001389 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1390 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1391 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1392 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1393 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1394 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1395 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1396 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001397 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001398 "src/math/roundd-neonv8.c",
1399 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001400 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001401]
1402
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001404 "src/f16-clamp/gen/neonfp16arith-x16.c",
1405 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001406 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1407 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1408 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1409 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1410 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1411 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1412 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1413 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1414 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1415 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1416 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1417 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001418 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1419 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001420 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1421 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001422 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001423 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001424 "src/f16-relu/gen/neonfp16arith-x16.c",
1425 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001426 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1427 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1428 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1429 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1430 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1431 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1432 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1433 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1434 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1435 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1436 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1437 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1438 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1439 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1440 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1441 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1442 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1443 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1444 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1445 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1446 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1447 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1448 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1449 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1450 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1451 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1452 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1453 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001454 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1455 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001456 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1457 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001464 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1465 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001468 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1469 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001480 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1481 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1482 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1483 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1484 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1485 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1486 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1487 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001488]
1489
1490SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001491 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1492 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001493 "src/f32-clamp/gen/sse-x4.c",
1494 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001495 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1496 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001497 "src/f32-dwconv-chw/3x3p1-sse.c",
1498 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001499 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1500 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1501 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1502 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1503 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1504 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1505 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1506 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1507 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1508 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1509 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1510 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001511 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001512 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1513 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1515 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1516 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1517 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1518 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1519 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1520 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1521 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1522 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1523 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1524 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1526 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001527 "src/f32-hswish/gen/hswish-sse-x4.c",
1528 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001529 "src/f32-ibilinear/gen/sse-c4.c",
1530 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1532 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1533 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1534 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1535 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1536 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1537 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001538 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1539 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1540 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001542 "src/f32-prelu/gen/sse-2x4.c",
1543 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001544 "src/f32-relu/gen/sse-x4.c",
1545 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001547 "src/f32-spmm/gen/4x1-minmax-sse.c",
1548 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001549 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001550 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1551 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1552 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1554 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1556 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001558 "src/f32-vbinary/gen/vmax-sse-x4.c",
1559 "src/f32-vbinary/gen/vmax-sse-x8.c",
1560 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1561 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1562 "src/f32-vbinary/gen/vmin-sse-x4.c",
1563 "src/f32-vbinary/gen/vmin-sse-x8.c",
1564 "src/f32-vbinary/gen/vminc-sse-x4.c",
1565 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001566 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1567 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1568 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1569 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1570 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1571 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1572 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1573 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001574 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1575 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1576 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1577 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001578 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001582 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1583 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001584 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1585 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001586 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1587 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001588 "src/f32-vunary/gen/vabs-sse-x4.c",
1589 "src/f32-vunary/gen/vabs-sse-x8.c",
1590 "src/f32-vunary/gen/vneg-sse-x4.c",
1591 "src/f32-vunary/gen/vneg-sse-x8.c",
1592 "src/f32-vunary/gen/vsqr-sse-x4.c",
1593 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001594 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001596 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001597 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001598 "src/math/roundd-sse-addsub.c",
1599 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001600 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001601 "src/math/sqrt-sse-hh1mac.c",
1602 "src/math/sqrt-sse-nr1mac.c",
1603 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604]
1605
1606SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001607 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1608 "src/f32-argmaxpool/4x-sse2-c4.c",
1609 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001610 "src/f32-prelu/gen/sse2-2x4.c",
1611 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001612 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001624 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001625 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1629 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001630 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1631 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001632 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1633 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1634 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1636 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1637 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1638 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1639 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001640 "src/qs8-requantization/precise-sse2.c",
1641 "src/qs8-requantization/fp32-sse2.c",
1642 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001643 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1644 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001645 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1646 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001647 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1648 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001649 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1650 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001651 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1652 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001653 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1654 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001655 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1656 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
1657 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1658 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001659 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1660 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1661 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1662 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1663 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1664 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1665 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1666 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001667 "src/qu8-requantization/precise-sse2.c",
1668 "src/qu8-requantization/fp32-sse2.c",
1669 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001670 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001671 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001672 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001674 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675 "src/x32-zip/x2-sse2.c",
1676 "src/x32-zip/x3-sse2.c",
1677 "src/x32-zip/x4-sse2.c",
1678 "src/x32-zip/xm-sse2.c",
1679 "src/x8-zip/x2-sse2.c",
1680 "src/x8-zip/x3-sse2.c",
1681 "src/x8-zip/x4-sse2.c",
1682 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001683 "src/math/exp-sse2-p5.c",
1684 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001685 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001686 "src/math/roundd-sse2-cvt.c",
1687 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001688 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001689 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001690]
1691
1692SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001693 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1694 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001695 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1696 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001697 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1698 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001699 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1700 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001701 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1702 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001703 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1704 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001705 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1706 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
1707 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1708 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001709 "src/qs8-requantization/precise-ssse3.c",
1710 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001711 "src/qu8-requantization/precise-ssse3.c",
1712 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713]
1714
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001715SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001716 "src/f32-prelu/gen/sse41-2x4.c",
1717 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001718 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001719 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001720 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001721 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001722 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1723 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001724 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1725 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001726 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1727 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1728 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1729 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1730 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1731 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1732 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1733 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001734 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1735 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001736 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1737 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001738 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1739 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001740 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1741 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001742 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1743 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001744 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1745 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001746 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1747 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
1748 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1749 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001750 "src/qs8-requantization/fp32-sse4.c",
1751 "src/qs8-requantization/precise-sse4.c",
1752 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001753 "src/qu8-requantization/precise-sse4.c",
1754 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001755 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001756 "src/math/roundd-sse41.c",
1757 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001758 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001759]
1760
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001762 "src/f32-clamp/gen/avx-x8.c",
1763 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001764 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1765 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1766 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1767 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1768 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1769 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1770 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1771 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1772 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1773 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1774 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1775 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1776 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1777 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1778 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1779 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1780 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1781 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1782 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1783 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1784 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1785 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1786 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1787 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1788 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1789 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1790 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1791 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001794 "src/f32-hswish/gen/hswish-avx-x8.c",
1795 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001796 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1797 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1798 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1799 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1800 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1801 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1802 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1803 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1804 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001805 "src/f32-prelu/gen/avx-2x8.c",
1806 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001807 "src/f32-relu/gen/avx-x16.c",
1808 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001810 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1811 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1812 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1813 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1814 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1815 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1816 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1817 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001818 "src/f32-vbinary/gen/vmax-avx-x8.c",
1819 "src/f32-vbinary/gen/vmax-avx-x16.c",
1820 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1821 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1822 "src/f32-vbinary/gen/vmin-avx-x8.c",
1823 "src/f32-vbinary/gen/vmin-avx-x16.c",
1824 "src/f32-vbinary/gen/vminc-avx-x8.c",
1825 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001826 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1827 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1828 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1829 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1830 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1831 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1832 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1833 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001834 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1835 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1836 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1837 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001842 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1843 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001844 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001845 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1846 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1847 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1848 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1849 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1850 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1851 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1852 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001853 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1854 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001855 "src/f32-vunary/gen/vabs-avx-x8.c",
1856 "src/f32-vunary/gen/vabs-avx-x16.c",
1857 "src/f32-vunary/gen/vneg-avx-x8.c",
1858 "src/f32-vunary/gen/vneg-avx-x16.c",
1859 "src/f32-vunary/gen/vsqr-avx-x8.c",
1860 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001861]
1862
Marat Dukhanfda12b82019-11-21 12:27:59 -08001863FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001864 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1865 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1866 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1867 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1868 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1869 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1870 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1871 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1872 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1873 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1874 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1875 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1876 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1877 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1878 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1879 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1880 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1881 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1882 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1883 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1884 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1885 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1886 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1887 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1888 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1889 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1890 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1891 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1892 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1893 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1894 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1895 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1896 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1897 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1898 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1899 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1900 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1901 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1902 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1903 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001904 "src/f32-hswish/gen/hswish-fma3-x8.c",
1905 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001906 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1907 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1908 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1909 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1910 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1911 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1912 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1913 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1914 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1915 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1916 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1917 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1918 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1919 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001920 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1921 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1922 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1923 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1924 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1925 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1926 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1927 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001928 "src/math/sqrt-fma3-nr1fma.c",
1929 "src/math/sqrt-fma3-nr2fma.c",
1930 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001931]
1932
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001933AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001934 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1935 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1936 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1937 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1938 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1939 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1940 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1941 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1942 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1943 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1944 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1945 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1946 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1947 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1948 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1949 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1950 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1951 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1952 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1953 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1954 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1955 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1956 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1957 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1958 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1959 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1960 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1961 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1962 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1963 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1964 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1965 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1966 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1967 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1968 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1969 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001970 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1971 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1972 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1973 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1974 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1975 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1976 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1977 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1978 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1979 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1980 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1981 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1982 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1983 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1984 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1985 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1986 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1987 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1988 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1989 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1990 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1991 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1992 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1993 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1994 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1995 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1996 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1997 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1998 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1999 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002000 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2001 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2002 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2003 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2004 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2005 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2006 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2007 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2008 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2009 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2010 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2011 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2012 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2013 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2014 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2015 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2016 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2017 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2018 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2019 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2020 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2021 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2022 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2023 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002024 "src/math/exp-avx2-p5.c",
2025 "src/math/exp-avx2-perm-p3.c",
2026 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002027 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002028 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002029 "src/math/sigmoid-avx2-rr2-p5-div.c",
2030 "src/math/sigmoid-avx2-rr1-p5-div.c",
2031 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2032 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2033 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2034 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002035]
2036
Marat Dukhan08c4a432019-10-03 09:29:21 -07002037AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002038 "src/f32-clamp/gen/avx512f-x16.c",
2039 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002040 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2041 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2042 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2043 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2044 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2045 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2046 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2047 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2048 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2049 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2050 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2051 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2052 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2053 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2054 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2055 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2056 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2057 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2058 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2059 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2060 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2061 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2062 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2063 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002064 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2065 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002066 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2067 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2068 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2069 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2070 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2071 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002072 "src/f32-prelu/gen/avx512f-2x16.c",
2073 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002074 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2075 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2076 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2077 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2078 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2079 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2080 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2081 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2082 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2083 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2084 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2085 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2086 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2087 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2088 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2089 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2090 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2091 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2092 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2093 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2094 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2095 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2096 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2097 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2098 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2099 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2100 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2101 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2102 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2103 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2104 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2105 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2106 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2107 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2108 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2109 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002110 "src/f32-relu/gen/avx512f-x16.c",
2111 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002113 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2114 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2115 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2116 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2117 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2118 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2119 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2120 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002121 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2122 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2123 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2124 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2125 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2126 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2127 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2128 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002129 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2130 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2131 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2132 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2133 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2134 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2135 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2136 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002137 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2138 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2139 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2140 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002141 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2142 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2143 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2144 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002145 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2146 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002147 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002148 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2149 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2150 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2151 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2152 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2153 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2154 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2155 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2156 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2157 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2158 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2159 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2160 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2161 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2162 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2163 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2164 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2165 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2166 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2167 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2168 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2169 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2170 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2171 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002172 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2173 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2174 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2175 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2176 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2177 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2178 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2179 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002180 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2181 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2182 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2183 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2184 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2185 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2186 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2187 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002188 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2189 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2190 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2191 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2192 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2193 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002194 "src/math/exp-avx512f-p5-scalef.c",
2195 "src/math/exp-avx512f-p5.c",
2196 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002197 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002198 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002199 "src/math/sqrt-avx512f-nr1fma.c",
2200 "src/math/sqrt-avx512f-nr2fma.c",
2201 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002202]
2203
2204AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002205 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002206 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002207 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2208 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002209 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002210 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002211 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002212 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002213 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2214 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002215 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2216 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2217 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2218 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002219]
2220
2221AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002222 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002223 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2224 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002225 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002226 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2227 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002228 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2229 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2230 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002231 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002232 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2233 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2234 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002235 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002236 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2237 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002238 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002239 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2240 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2241 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2242 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2243 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2244 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2245 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2246 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2247 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2248 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2249 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2250 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2251 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2252 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2253 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2254 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2255 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2256 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002257 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2258 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002259 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002260 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2261 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2262 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2263 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2264 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2265 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2266 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2267 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2268 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2269 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2270 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2271 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2272 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2273 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2274 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2275 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2276 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2277 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002278 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2279 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2280 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2281 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2282 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2283 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2284 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2285 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2286 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2287 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2288 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2289 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2290 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2291 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2292 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2293 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2294 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2295 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296]
2297
Marat Dukhan1b354632020-03-23 12:50:22 -07002298INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002299 "src/xnnpack/argmaxpool.h",
2300 "src/xnnpack/avgpool.h",
2301 "src/xnnpack/clamp.h",
2302 "src/xnnpack/common.h",
2303 "src/xnnpack/conv.h",
2304 "src/xnnpack/dwconv.h",
2305 "src/xnnpack/gavgpool.h",
2306 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002307 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002308 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002309 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002310 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002311 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002312 "src/xnnpack/lut.h",
2313 "src/xnnpack/math.h",
2314 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002315 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002316 "src/xnnpack/packx.h",
2317 "src/xnnpack/pad.h",
2318 "src/xnnpack/params.h",
2319 "src/xnnpack/pavgpool.h",
2320 "src/xnnpack/ppmm.h",
2321 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002322 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002323 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002324 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325 "src/xnnpack/rmax.h",
2326 "src/xnnpack/scalar-utils.h",
2327 "src/xnnpack/spmm.h",
2328 "src/xnnpack/unpool.h",
2329 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002330 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002331 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002332 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002333 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002334 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002335 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002336 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002337]
2338
2339INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 "include/xnnpack.h",
2341 "src/xnnpack/allocator.h",
2342 "src/xnnpack/compute.h",
2343 "src/xnnpack/im2col.h",
2344 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002345 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002346 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347 "src/xnnpack/operator.h",
2348 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002349 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002350 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002351 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002352 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002353]
2354
Marat Dukhan1b354632020-03-23 12:50:22 -07002355ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002356 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357]
2358
Marat Dukhan1b354632020-03-23 12:50:22 -07002359MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002360 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002361 "include/xnnpack.h",
2362]
2363
Marat Dukhan1b354632020-03-23 12:50:22 -07002364MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002366 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002367 "src/xnnpack/requantization.h",
2368 "include/xnnpack.h",
2369]
2370
2371OPERATOR_TEST_PARAMS_HDRS = [
2372 "src/xnnpack/params.h",
2373 "src/xnnpack/common.h",
2374]
2375
2376WEIGHTS_PACK_HDRS = [
2377 "src/xnnpack/pack.h",
2378 "src/xnnpack/operator.h",
2379 "src/xnnpack/compute.h",
2380]
2381
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002382LOGGING_COPTS = select({
2383 # No logging in optimized mode
2384 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2385 # Full logging in debug mode
2386 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2387 # Error-only logging in default (fastbuild) mode
2388 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2389})
2390
Marat Dukhan3b59de22020-06-03 20:15:19 -07002391LOGGING_SRCS = select({
2392 # No logging in optimized mode
2393 ":optimized_build": [],
2394 "//conditions:default": [
2395 "src/operator-strings.c",
2396 "src/subgraph-strings.c",
2397 ],
2398})
2399
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002400LOGGING_HDRS = [
2401 "src/xnnpack/log.h",
2402]
2403
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002405 name = "tables",
2406 srcs = TABLE_SRCS,
2407 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002408 gcc_copts = xnnpack_gcc_std_copts(),
2409 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002410)
2411
2412xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002413 name = "scalar_ukernels",
2414 srcs = SCALAR_UKERNELS,
2415 hdrs = INTERNAL_HDRS,
2416 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002417 gcc_copts = xnnpack_gcc_std_copts(),
2418 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002419 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002420 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421 "@FP16",
2422 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002423 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002424 ],
2425)
2426
2427xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002428 name = "scalar_ukernels_test_mode",
2429 srcs = SCALAR_UKERNELS,
2430 hdrs = INTERNAL_HDRS,
2431 aarch32_copts = ["-marm"],
2432 copts = [
2433 "-UNDEBUG",
2434 "-DXNN_TEST_MODE=1",
2435 ],
2436 gcc_copts = xnnpack_gcc_std_copts(),
2437 msvc_copts = xnnpack_msvc_std_copts(),
2438 deps = [
2439 ":tables",
2440 "@FP16",
2441 "@FXdiv",
2442 "@pthreadpool",
2443 ],
2444)
2445
2446xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002447 name = "wasm_ukernels",
2448 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002449 gcc_copts = xnnpack_gcc_std_copts(),
2450 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002451 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002452 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002453 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002454 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002455 "@FP16",
2456 "@FXdiv",
2457 "@pthreadpool",
2458 ],
2459)
2460
2461xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002462 name = "wasm_ukernels_test_mode",
2463 hdrs = INTERNAL_HDRS,
2464 copts = [
2465 "-UNDEBUG",
2466 "-DXNN_TEST_MODE=1",
2467 ],
2468 gcc_copts = xnnpack_gcc_std_copts(),
2469 msvc_copts = xnnpack_msvc_std_copts(),
2470 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002471 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002472 deps = [
2473 ":tables",
2474 "@FP16",
2475 "@FXdiv",
2476 "@pthreadpool",
2477 ],
2478)
2479
2480xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002481 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002482 hdrs = INTERNAL_HDRS,
2483 aarch32_copts = [
2484 "-marm",
2485 "-mfpu=neon",
2486 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002487 gcc_copts = xnnpack_gcc_std_copts(),
2488 gcc_x86_copts = ["-msse2"],
2489 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002490 optimized_copts = [
2491 "-O3",
2492 "-ffast-math",
2493 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002494 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002495 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002496 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002497 "@FP16",
2498 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002499 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002500 ],
2501)
2502
2503xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002504 name = "psimd_fastmath_ukernels_test_mode",
2505 hdrs = INTERNAL_HDRS,
2506 aarch32_copts = [
2507 "-marm",
2508 "-mfpu=neon",
2509 ],
2510 copts = [
2511 "-UNDEBUG",
2512 "-DXNN_TEST_MODE=1",
2513 ],
2514 gcc_copts = xnnpack_gcc_std_copts(),
2515 gcc_x86_copts = ["-msse2"],
2516 msvc_copts = xnnpack_msvc_std_copts(),
2517 optimized_copts = [
2518 "-O3",
2519 "-ffast-math",
2520 ],
2521 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2522 deps = [
2523 ":tables",
2524 "@FP16",
2525 "@psimd",
2526 "@pthreadpool",
2527 ],
2528)
2529
2530xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002531 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002532 hdrs = INTERNAL_HDRS,
2533 aarch32_copts = [
2534 "-marm",
2535 "-mfpu=neon",
2536 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002537 gcc_copts = xnnpack_gcc_std_copts(),
2538 gcc_x86_copts = ["-msse2"],
2539 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002540 optimized_copts = [
2541 "-O3",
2542 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002543 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002544 deps = [
2545 ":tables",
2546 "@FP16",
2547 "@psimd",
2548 "@pthreadpool",
2549 ],
2550)
2551
2552xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002553 name = "psimd_accmath_ukernels_test_mode",
2554 hdrs = INTERNAL_HDRS,
2555 aarch32_copts = [
2556 "-marm",
2557 "-mfpu=neon",
2558 ],
2559 copts = [
2560 "-UNDEBUG",
2561 "-DXNN_TEST_MODE=1",
2562 ],
2563 gcc_copts = xnnpack_gcc_std_copts(),
2564 gcc_x86_copts = ["-msse2"],
2565 msvc_copts = xnnpack_msvc_std_copts(),
2566 optimized_copts = [
2567 "-O3",
2568 ],
2569 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2570 deps = [
2571 ":tables",
2572 "@FP16",
2573 "@psimd",
2574 "@pthreadpool",
2575 ],
2576)
2577
2578xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002579 name = "neon_ukernels",
2580 hdrs = INTERNAL_HDRS,
2581 aarch32_copts = [
2582 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002583 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002584 "-mfpu=neon",
2585 ],
2586 aarch32_srcs = NEON_UKERNELS,
2587 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002588 gcc_copts = xnnpack_gcc_std_copts(),
2589 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002590 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002591 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002592 "@FP16",
2593 "@pthreadpool",
2594 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595)
2596
2597xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002598 name = "neon_ukernels_test_mode",
2599 hdrs = INTERNAL_HDRS,
2600 aarch32_copts = [
2601 "-marm",
2602 "-march=armv7-a",
2603 "-mfpu=neon",
2604 ],
2605 aarch32_srcs = NEON_UKERNELS,
2606 aarch64_srcs = NEON_UKERNELS,
2607 copts = [
2608 "-UNDEBUG",
2609 "-DXNN_TEST_MODE=1",
2610 ],
2611 gcc_copts = xnnpack_gcc_std_copts(),
2612 msvc_copts = xnnpack_msvc_std_copts(),
2613 deps = [
2614 ":tables",
2615 "@FP16",
2616 "@pthreadpool",
2617 ],
2618)
2619
2620xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 name = "neonfma_ukernels",
2622 hdrs = INTERNAL_HDRS,
2623 aarch32_copts = [
2624 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002625 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 "-mfpu=neon-vfpv4",
2627 ],
2628 aarch32_srcs = NEONFMA_UKERNELS,
2629 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002630 apple_aarch32_copts = [
2631 "-mcpu=swift",
2632 "-mtune=generic",
2633 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002634 gcc_copts = xnnpack_gcc_std_copts(),
2635 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002636 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002637 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002638 "@FP16",
2639 "@pthreadpool",
2640 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002641)
2642
2643xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002644 name = "neonfma_ukernels_test_mode",
2645 hdrs = INTERNAL_HDRS,
2646 aarch32_copts = [
2647 "-marm",
2648 "-march=armv7-a",
2649 "-mfpu=neon-vfpv4",
2650 ],
2651 aarch32_srcs = NEONFMA_UKERNELS,
2652 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002653 apple_aarch32_copts = [
2654 "-mcpu=swift",
2655 "-mtune=generic",
2656 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002657 copts = [
2658 "-UNDEBUG",
2659 "-DXNN_TEST_MODE=1",
2660 ],
2661 gcc_copts = xnnpack_gcc_std_copts(),
2662 msvc_copts = xnnpack_msvc_std_copts(),
2663 deps = [
2664 ":tables",
2665 "@FP16",
2666 "@pthreadpool",
2667 ],
2668)
2669
2670xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002671 name = "neonv8_ukernels",
2672 hdrs = INTERNAL_HDRS,
2673 aarch32_copts = [
2674 "-marm",
2675 "-march=armv8-a",
2676 "-mfpu=neon-fp-armv8",
2677 ],
2678 aarch32_srcs = NEONV8_UKERNELS,
2679 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002680 apple_aarch32_copts = [
2681 "-mcpu=cyclone",
2682 "-mtune=generic",
2683 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002684 gcc_copts = xnnpack_gcc_std_copts(),
2685 msvc_copts = xnnpack_msvc_std_copts(),
2686 deps = [
2687 ":tables",
2688 "@FP16",
2689 "@pthreadpool",
2690 ],
2691)
2692
2693xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002694 name = "neonv8_ukernels_test_mode",
2695 hdrs = INTERNAL_HDRS,
2696 aarch32_copts = [
2697 "-marm",
2698 "-march=armv8-a",
2699 "-mfpu=neon-fp-armv8",
2700 ],
2701 aarch32_srcs = NEONV8_UKERNELS,
2702 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002703 apple_aarch32_copts = [
2704 "-mcpu=cyclone",
2705 "-mtune=generic",
2706 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002707 copts = [
2708 "-UNDEBUG",
2709 "-DXNN_TEST_MODE=1",
2710 ],
2711 gcc_copts = xnnpack_gcc_std_copts(),
2712 msvc_copts = xnnpack_msvc_std_copts(),
2713 deps = [
2714 ":tables",
2715 "@FP16",
2716 "@pthreadpool",
2717 ],
2718)
2719
2720xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002721 name = "neonfp16arith_ukernels",
2722 hdrs = INTERNAL_HDRS,
2723 aarch64_copts = ["-march=armv8.2-a+fp16"],
2724 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002725 gcc_copts = xnnpack_gcc_std_copts(),
2726 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002727 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002728 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002729 "@FP16",
2730 "@pthreadpool",
2731 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732)
2733
2734xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002735 name = "neonfp16arith_ukernels_test_mode",
2736 hdrs = INTERNAL_HDRS,
2737 aarch64_copts = ["-march=armv8.2-a+fp16"],
2738 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2739 copts = [
2740 "-UNDEBUG",
2741 "-DXNN_TEST_MODE=1",
2742 ],
2743 gcc_copts = xnnpack_gcc_std_copts(),
2744 msvc_copts = xnnpack_msvc_std_copts(),
2745 deps = [
2746 ":tables",
2747 "@FP16",
2748 "@pthreadpool",
2749 ],
2750)
2751
2752xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002753 name = "sse2_ukernels",
2754 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002755 gcc_copts = xnnpack_gcc_std_copts(),
2756 gcc_x86_copts = ["-msse2"],
2757 msvc_copts = xnnpack_msvc_std_copts(),
2758 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002760 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002761 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002762 "@FP16",
2763 "@pthreadpool",
2764 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765)
2766
2767xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002768 name = "sse2_ukernels_test_mode",
2769 hdrs = INTERNAL_HDRS,
2770 copts = [
2771 "-UNDEBUG",
2772 "-DXNN_TEST_MODE=1",
2773 ],
2774 gcc_copts = xnnpack_gcc_std_copts(),
2775 gcc_x86_copts = ["-msse2"],
2776 msvc_copts = xnnpack_msvc_std_copts(),
2777 msvc_x86_32_copts = ["/arch:SSE2"],
2778 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2779 deps = [
2780 ":tables",
2781 "@FP16",
2782 "@pthreadpool",
2783 ],
2784)
2785
2786xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002787 name = "ssse3_ukernels",
2788 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002789 gcc_copts = xnnpack_gcc_std_copts(),
2790 gcc_x86_copts = ["-mssse3"],
2791 msvc_copts = xnnpack_msvc_std_copts(),
2792 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002793 x86_srcs = SSSE3_UKERNELS,
2794 deps = [
2795 ":tables",
2796 "@FP16",
2797 "@pthreadpool",
2798 ],
2799)
2800
2801xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002802 name = "ssse3_ukernels_test_mode",
2803 hdrs = INTERNAL_HDRS,
2804 copts = [
2805 "-UNDEBUG",
2806 "-DXNN_TEST_MODE=1",
2807 ],
2808 gcc_copts = xnnpack_gcc_std_copts(),
2809 gcc_x86_copts = ["-mssse3"],
2810 msvc_copts = xnnpack_msvc_std_copts(),
2811 msvc_x86_32_copts = ["/arch:SSE2"],
2812 x86_srcs = SSSE3_UKERNELS,
2813 deps = [
2814 ":tables",
2815 "@FP16",
2816 "@pthreadpool",
2817 ],
2818)
2819
2820xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002821 name = "sse41_ukernels",
2822 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002823 gcc_copts = xnnpack_gcc_std_copts(),
2824 gcc_x86_copts = ["-msse4.1"],
2825 msvc_copts = xnnpack_msvc_std_copts(),
2826 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002827 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002828 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002829 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002830 "@FP16",
2831 "@pthreadpool",
2832 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002833)
2834
2835xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002836 name = "sse41_ukernels_test_mode",
2837 hdrs = INTERNAL_HDRS,
2838 copts = [
2839 "-UNDEBUG",
2840 "-DXNN_TEST_MODE=1",
2841 ],
2842 gcc_copts = xnnpack_gcc_std_copts(),
2843 gcc_x86_copts = ["-msse4.1"],
2844 msvc_copts = xnnpack_msvc_std_copts(),
2845 msvc_x86_32_copts = ["/arch:SSE2"],
2846 x86_srcs = SSE41_UKERNELS,
2847 deps = [
2848 ":tables",
2849 "@FP16",
2850 "@pthreadpool",
2851 ],
2852)
2853
2854xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002855 name = "avx_ukernels",
2856 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002857 gcc_copts = xnnpack_gcc_std_copts(),
2858 gcc_x86_copts = ["-mavx"],
2859 msvc_copts = xnnpack_msvc_std_copts(),
2860 msvc_x86_32_copts = ["/arch:AVX"],
2861 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002863 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002864 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002865 "@FP16",
2866 "@pthreadpool",
2867 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002868)
2869
2870xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002871 name = "avx_ukernels_test_mode",
2872 hdrs = INTERNAL_HDRS,
2873 copts = [
2874 "-UNDEBUG",
2875 "-DXNN_TEST_MODE=1",
2876 ],
2877 gcc_copts = xnnpack_gcc_std_copts(),
2878 gcc_x86_copts = ["-mavx"],
2879 msvc_copts = xnnpack_msvc_std_copts(),
2880 msvc_x86_32_copts = ["/arch:AVX"],
2881 msvc_x86_64_copts = ["/arch:AVX"],
2882 x86_srcs = AVX_UKERNELS,
2883 deps = [
2884 ":tables",
2885 "@FP16",
2886 "@pthreadpool",
2887 ],
2888)
2889
2890xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002891 name = "fma3_ukernels",
2892 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002893 gcc_copts = xnnpack_gcc_std_copts(),
2894 gcc_x86_copts = ["-mfma"],
2895 msvc_copts = xnnpack_msvc_std_copts(),
2896 msvc_x86_32_copts = ["/arch:AVX"],
2897 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002898 x86_srcs = FMA3_UKERNELS,
2899 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002900 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002901 "@FP16",
2902 "@pthreadpool",
2903 ],
2904)
2905
2906xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002907 name = "fma3_ukernels_test_mode",
2908 hdrs = INTERNAL_HDRS,
2909 copts = [
2910 "-UNDEBUG",
2911 "-DXNN_TEST_MODE=1",
2912 ],
2913 gcc_copts = xnnpack_gcc_std_copts(),
2914 gcc_x86_copts = ["-mfma"],
2915 msvc_copts = xnnpack_msvc_std_copts(),
2916 msvc_x86_32_copts = ["/arch:AVX"],
2917 msvc_x86_64_copts = ["/arch:AVX"],
2918 x86_srcs = FMA3_UKERNELS,
2919 deps = [
2920 ":tables",
2921 "@FP16",
2922 "@pthreadpool",
2923 ],
2924)
2925
2926xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002927 name = "avx2_ukernels",
2928 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002929 gcc_copts = xnnpack_gcc_std_copts(),
2930 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002931 "-mfma",
2932 "-mavx2",
2933 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002934 msvc_copts = xnnpack_msvc_std_copts(),
2935 msvc_x86_32_copts = ["/arch:AVX2"],
2936 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002937 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002938 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002939 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002940 "@FP16",
2941 "@pthreadpool",
2942 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002943)
2944
2945xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002946 name = "avx2_ukernels_test_mode",
2947 hdrs = INTERNAL_HDRS,
2948 copts = [
2949 "-UNDEBUG",
2950 "-DXNN_TEST_MODE=1",
2951 ],
2952 gcc_copts = xnnpack_gcc_std_copts(),
2953 gcc_x86_copts = [
2954 "-mfma",
2955 "-mavx2",
2956 ],
2957 msvc_copts = xnnpack_msvc_std_copts(),
2958 msvc_x86_32_copts = ["/arch:AVX2"],
2959 msvc_x86_64_copts = ["/arch:AVX2"],
2960 x86_srcs = AVX2_UKERNELS,
2961 deps = [
2962 ":tables",
2963 "@FP16",
2964 "@pthreadpool",
2965 ],
2966)
2967
2968xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969 name = "avx512f_ukernels",
2970 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002971 gcc_copts = xnnpack_gcc_std_copts(),
2972 gcc_x86_copts = ["-mavx512f"],
2973 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2974 msvc_copts = xnnpack_msvc_std_copts(),
2975 msvc_x86_32_copts = ["/arch:AVX512"],
2976 msvc_x86_64_copts = ["/arch:AVX512"],
2977 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002978 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002979 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002980 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002981 "@FP16",
2982 "@pthreadpool",
2983 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002984)
2985
2986xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002987 name = "avx512f_ukernels_test_mode",
2988 hdrs = INTERNAL_HDRS,
2989 copts = [
2990 "-UNDEBUG",
2991 "-DXNN_TEST_MODE=1",
2992 ],
2993 gcc_copts = xnnpack_gcc_std_copts(),
2994 gcc_x86_copts = ["-mavx512f"],
2995 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2996 msvc_copts = xnnpack_msvc_std_copts(),
2997 msvc_x86_32_copts = ["/arch:AVX512"],
2998 msvc_x86_64_copts = ["/arch:AVX512"],
2999 msys_copts = ["-fno-asynchronous-unwind-tables"],
3000 x86_srcs = AVX512F_UKERNELS,
3001 deps = [
3002 ":tables",
3003 "@FP16",
3004 "@pthreadpool",
3005 ],
3006)
3007
3008xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009 name = "asm_ukernels",
3010 hdrs = ["src/xnnpack/assembly.h"],
3011 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003012 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 aarch64_srcs = AARCH64_ASM_UKERNELS,
3014)
3015
Marat Dukhan3b59de22020-06-03 20:15:19 -07003016xnnpack_cc_library(
3017 name = "logging_utils",
3018 srcs = LOGGING_SRCS,
3019 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3020 copts = LOGGING_COPTS + [
3021 "-Isrc",
3022 "-Iinclude",
3023 ] + select({
3024 ":debug_build": [],
3025 "//conditions:default": xnnpack_min_size_copts(),
3026 }),
3027 gcc_copts = xnnpack_gcc_std_copts(),
3028 msvc_copts = xnnpack_msvc_std_copts(),
3029 visibility = xnnpack_visibility(),
3030 deps = [
3031 "@FP16",
3032 "@clog",
3033 "@pthreadpool",
3034 ],
3035)
3036
Marat Dukhan08c4a432019-10-03 09:29:21 -07003037xnnpack_aggregate_library(
3038 name = "ukernels",
3039 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003040 ":neon_ukernels",
3041 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003042 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 ":asm_ukernels",
3044 ],
3045 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 ":neon_ukernels",
3047 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003048 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 ":neonfp16arith_ukernels",
3050 ":asm_ukernels",
3051 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003052 generic_deps = [
3053 ":scalar_ukernels",
3054 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003055 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003056 ":psimd_fastmath_ukernels",
3057 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003058 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003059 wasm_deps = [
3060 ":wasm_ukernels",
3061 ],
3062 wasmsimd_deps = [
3063 ":wasm_ukernels",
3064 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003065 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003067 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003068 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003070 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003071 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003072 ":avx512f_ukernels",
3073 ],
3074)
3075
Marat Dukhan33fcf782020-05-24 14:27:15 -07003076xnnpack_aggregate_library(
3077 name = "ukernels_test_mode",
3078 aarch32_deps = [
3079 ":neon_ukernels_test_mode",
3080 ":neonfma_ukernels_test_mode",
3081 ":neonv8_ukernels_test_mode",
3082 ":asm_ukernels",
3083 ],
3084 aarch64_deps = [
3085 ":neon_ukernels_test_mode",
3086 ":neonfma_ukernels_test_mode",
3087 ":neonv8_ukernels_test_mode",
3088 ":neonfp16arith_ukernels_test_mode",
3089 ":asm_ukernels",
3090 ],
3091 generic_deps = [
3092 ":scalar_ukernels_test_mode",
3093 ],
3094 psimd_deps = [
3095 ":psimd_fastmath_ukernels_test_mode",
3096 ":psimd_accmath_ukernels_test_mode",
3097 ],
3098 wasm_deps = [
3099 ":wasm_ukernels_test_mode",
3100 ],
3101 wasmsimd_deps = [
3102 ":wasm_ukernels_test_mode",
3103 ],
3104 x86_deps = [
3105 ":sse2_ukernels_test_mode",
3106 ":ssse3_ukernels_test_mode",
3107 ":sse41_ukernels_test_mode",
3108 ":avx_ukernels_test_mode",
3109 ":fma3_ukernels_test_mode",
3110 ":avx2_ukernels_test_mode",
3111 ":avx512f_ukernels_test_mode",
3112 ],
3113)
3114
Marat Dukhan86df3842020-06-17 14:03:31 -07003115# This target must be compatible with Android NDK r17 and compile with Clang 6.
3116# It is similar to :ukernels target, but excludes microkernels which require
3117# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003118xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003119 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003120 aarch32_deps = [
3121 ":neon_ukernels",
3122 ":neonfma_ukernels",
3123 ":neonv8_ukernels",
3124 ":asm_ukernels",
3125 ],
3126 aarch64_deps = [
3127 ":neon_ukernels",
3128 ":neonfma_ukernels",
3129 ":neonv8_ukernels",
3130 ":asm_ukernels",
3131 ],
3132 generic_deps = [
3133 ":scalar_ukernels",
3134 ],
3135 psimd_deps = [
3136 ":psimd_fastmath_ukernels",
3137 ":psimd_accmath_ukernels",
3138 ],
3139 wasm_deps = [
3140 ":wasm_ukernels",
3141 ],
3142 wasmsimd_deps = [
3143 ":wasm_ukernels",
3144 ],
3145 x86_deps = [
3146 ":sse2_ukernels",
3147 ":ssse3_ukernels",
3148 ":sse41_ukernels",
3149 ":avx_ukernels",
3150 ":fma3_ukernels",
3151 ":avx2_ukernels",
3152 ":avx512f_ukernels",
3153 ],
3154)
3155
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156xnnpack_cc_library(
3157 name = "im2col",
3158 srcs = ["src/im2col.c"],
3159 hdrs = [
3160 "src/xnnpack/common.h",
3161 "src/xnnpack/im2col.h",
3162 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003163 gcc_copts = xnnpack_gcc_std_copts(),
3164 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165)
3166
3167xnnpack_cc_library(
3168 name = "indirection",
3169 srcs = ["src/indirection.c"],
3170 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003171 gcc_copts = xnnpack_gcc_std_copts(),
3172 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 deps = [
3174 "@FP16",
3175 "@FXdiv",
3176 "@pthreadpool",
3177 ],
3178)
3179
3180xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003181 name = "indirection_test_mode",
3182 srcs = ["src/indirection.c"],
3183 hdrs = INTERNAL_HDRS,
3184 copts = [
3185 "-UNDEBUG",
3186 "-DXNN_TEST_MODE=1",
3187 ],
3188 gcc_copts = xnnpack_gcc_std_copts(),
3189 msvc_copts = xnnpack_msvc_std_copts(),
3190 deps = [
3191 "@FP16",
3192 "@FXdiv",
3193 "@pthreadpool",
3194 ],
3195)
3196
3197xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003198 name = "packing",
3199 srcs = ["src/packing.c"],
3200 hdrs = INTERNAL_HDRS,
3201 gcc_copts = xnnpack_gcc_std_copts(),
3202 msvc_copts = xnnpack_msvc_std_copts(),
3203 deps = [
3204 "@FP16",
3205 "@FXdiv",
3206 "@pthreadpool",
3207 ],
3208)
3209
3210xnnpack_cc_library(
3211 name = "packing_test_mode",
3212 srcs = ["src/packing.c"],
3213 hdrs = INTERNAL_HDRS,
3214 copts = [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ],
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 msvc_copts = xnnpack_msvc_std_copts(),
3220 deps = [
3221 "@FP16",
3222 "@FXdiv",
3223 "@pthreadpool",
3224 ],
3225)
3226
3227xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228 name = "operator_run",
3229 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003230 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003231 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003232 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3233 "//conditions:default": [],
3234 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003235 gcc_copts = xnnpack_gcc_std_copts(),
3236 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003238 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 "@FP16",
3240 "@FXdiv",
3241 "@clog",
3242 "@pthreadpool",
3243 ],
3244)
3245
Chao Mei6ddfc602020-05-13 22:29:36 -07003246xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003247 name = "operator_run_test_mode",
3248 srcs = ["src/operator-run.c"],
3249 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3250 copts = LOGGING_COPTS + [
3251 "-UNDEBUG",
3252 "-DXNN_TEST_MODE=1",
3253 ] + select({
3254 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3255 "//conditions:default": [],
3256 }),
3257 gcc_copts = xnnpack_gcc_std_copts(),
3258 msvc_copts = xnnpack_msvc_std_copts(),
3259 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003260 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003261 "@FP16",
3262 "@FXdiv",
3263 "@clog",
3264 "@pthreadpool",
3265 ],
3266)
3267
3268xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003269 name = "memory_planner",
3270 srcs = ["src/memory-planner.c"],
3271 hdrs = INTERNAL_HDRS,
3272 defines = select({
3273 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3274 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3275 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3276 }),
3277 gcc_copts = xnnpack_gcc_std_copts(),
3278 msvc_copts = xnnpack_msvc_std_copts(),
3279 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003280 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003281 "@pthreadpool",
3282 ],
3283)
3284
Marat Dukhan33fcf782020-05-24 14:27:15 -07003285xnnpack_cc_library(
3286 name = "memory_planner_test_mode",
3287 srcs = ["src/memory-planner.c"],
3288 hdrs = INTERNAL_HDRS,
3289 copts = [
3290 "-UNDEBUG",
3291 "-DXNN_TEST_MODE=1",
3292 ],
3293 defines = select({
3294 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3295 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3296 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3297 }),
3298 gcc_copts = xnnpack_gcc_std_copts(),
3299 msvc_copts = xnnpack_msvc_std_copts(),
3300 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003301 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003302 "@pthreadpool",
3303 ],
3304)
3305
Marat Dukhan08c4a432019-10-03 09:29:21 -07003306cc_library(
3307 name = "enable_assembly",
3308 defines = select({
3309 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3310 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003311 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003312 }),
3313)
3314
Marat Dukhan9de90e02020-06-18 16:04:12 -07003315cc_library(
3316 name = "enable_sparse",
3317 defines = select({
3318 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3319 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003320 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003321 }),
3322)
3323
Marat Dukhancf056b22019-10-07 10:26:29 -07003324xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325 name = "operators",
3326 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003327 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003329 ],
3330 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003331 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332 "-Isrc",
3333 "-Iinclude",
3334 ] + select({
3335 ":debug_build": [],
3336 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003337 }) + select({
3338 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3339 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003340 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003341 gcc_copts = xnnpack_gcc_std_copts(),
3342 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003343 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003345 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003346 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 "@FP16",
3348 "@FXdiv",
3349 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003351 ],
3352)
3353
Marat Dukhan10a38082020-04-17 03:58:35 -07003354xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003355 name = "operators_test_mode",
3356 srcs = OPERATOR_SRCS + [
3357 "src/memory.c",
3358 "src/operator-delete.c",
3359 ],
3360 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3361 copts = LOGGING_COPTS + [
3362 "-Isrc",
3363 "-Iinclude",
3364 "-UNDEBUG",
3365 "-DXNN_TEST_MODE=1",
3366 ] + select({
3367 ":debug_build": [],
3368 "//conditions:default": xnnpack_min_size_copts(),
3369 }) + select({
3370 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3371 "//conditions:default": [],
3372 }),
3373 gcc_copts = xnnpack_gcc_std_copts(),
3374 msvc_copts = xnnpack_msvc_std_copts(),
3375 deps = [
3376 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003377 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003378 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003379 "@FP16",
3380 "@FXdiv",
3381 "@clog",
3382 "@pthreadpool",
3383 ],
3384)
3385
3386xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003387 name = "XNNPACK",
3388 srcs = [
3389 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003390 "src/runtime.c",
3391 "src/subgraph.c",
3392 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003393 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003394 hdrs = ["include/xnnpack.h"],
3395 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003396 "-Isrc",
3397 "-Iinclude",
3398 ] + select({
3399 ":debug_build": [],
3400 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003401 }) + select({
3402 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3403 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003404 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003405 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003406 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003407 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003408 visibility = xnnpack_visibility(),
3409 deps = [
3410 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003411 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003412 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003413 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003414 ":operator_run",
3415 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003416 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003417 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003418 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003419 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003420 ] + select({
3421 ":emscripten": [],
3422 "//conditions:default": ["@cpuinfo"],
3423 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003424)
3425
Marat Dukhan10a38082020-04-17 03:58:35 -07003426xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003427 name = "XNNPACK_test_mode",
3428 srcs = [
3429 "src/init.c",
3430 "src/runtime.c",
3431 "src/subgraph.c",
3432 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003433 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003434 hdrs = ["include/xnnpack.h"],
3435 copts = LOGGING_COPTS + [
3436 "-Isrc",
3437 "-Iinclude",
3438 "-UNDEBUG",
3439 "-DXNN_TEST_MODE=1",
3440 ] + select({
3441 ":debug_build": [],
3442 "//conditions:default": xnnpack_min_size_copts(),
3443 }) + select({
3444 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3445 "//conditions:default": [],
3446 }),
3447 gcc_copts = xnnpack_gcc_std_copts(),
3448 includes = ["include"],
3449 msvc_copts = xnnpack_msvc_std_copts(),
3450 visibility = xnnpack_visibility(),
3451 deps = [
3452 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003453 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003454 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003455 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003456 ":operator_run_test_mode",
3457 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003458 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003459 "@clog",
3460 "@FP16",
3461 "@pthreadpool",
3462 ] + select({
3463 ":emscripten": [],
3464 "//conditions:default": ["@cpuinfo"],
3465 }),
3466)
3467
3468xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003469 name = "xnnpack_f32",
3470 srcs = [
3471 "src/init.c",
3472 "src/runtime.c",
3473 "src/subgraph.c",
3474 "src/tensor.c",
3475 ] + SUBGRAPH_SRCS,
3476 hdrs = ["include/xnnpack.h"],
3477 copts = LOGGING_COPTS + [
3478 "-Isrc",
3479 "-Iinclude",
3480 ] + select({
3481 ":debug_build": [],
3482 "//conditions:default": xnnpack_min_size_copts(),
3483 }) + select({
3484 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3485 "//conditions:default": [],
3486 }),
3487 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003488 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003489 "XNN_NO_U8_OPERATORS",
3490 "XNN_NO_X8_OPERATORS",
3491 "XNN_NO_F16_OPERATORS",
3492 "XNN_NO_X16_OPERATORS",
3493 ],
3494 gcc_copts = xnnpack_gcc_std_copts(),
3495 includes = ["include"],
3496 msvc_copts = xnnpack_msvc_std_copts(),
3497 visibility = xnnpack_visibility(),
3498 deps = [
3499 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003500 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003501 ":logging_utils",
3502 ":memory_planner",
3503 ":operator_run",
3504 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003505 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003506 "@clog",
3507 "@FP16",
3508 "@pthreadpool",
3509 ] + select({
3510 ":emscripten": [],
3511 "//conditions:default": ["@cpuinfo"],
3512 }),
3513)
3514
3515xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003516 name = "xnnpack_operators_nhwc_f32",
3517 srcs = [
3518 "src/init.c",
3519 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003520 hdrs = ["include/xnnpack.h"],
3521 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003522 "-Isrc",
3523 "-Iinclude",
3524 ] + select({
3525 ":debug_build": [],
3526 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003527 }) + select({
3528 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3529 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003530 }),
3531 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003532 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003533 "XNN_NO_U8_OPERATORS",
3534 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003535 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003536 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003537 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003539 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003540 visibility = xnnpack_visibility(),
3541 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003542 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003543 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003544 ":operator_run",
3545 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003546 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003547 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003548 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003549 ] + select({
3550 ":emscripten": [],
3551 "//conditions:default": ["@cpuinfo"],
3552 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553)
3554
Marat Dukhancf056b22019-10-07 10:26:29 -07003555xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556 name = "bench_utils",
3557 srcs = ["bench/utils.cc"],
3558 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003559 deps = [
3560 "@com_google_benchmark//:benchmark",
3561 "@cpuinfo",
3562 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563)
3564
Frank Barchard7e955972019-10-11 10:34:25 -07003565######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566
3567xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003568 name = "qs8_gemm_bench",
3569 srcs = [
3570 "bench/gemm.h",
3571 "bench/qs8-gemm.cc",
3572 "src/xnnpack/AlignedAllocator.h",
3573 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3574 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3575)
3576
3577xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003578 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 srcs = [
3580 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003581 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 "src/xnnpack/AlignedAllocator.h",
3583 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003584 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003585 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586)
3587
3588xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003589 name = "qu8_requantization_bench",
3590 srcs = [
3591 "bench/qu8-requantization.cc",
3592 "src/xnnpack/requantization-stubs.h",
3593 "src/xnnpack/AlignedAllocator.h",
3594 ] + MICROKERNEL_BENCHMARK_HDRS,
3595 deps = MICROKERNEL_BENCHMARK_DEPS,
3596)
3597
3598xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003599 name = "f16_igemm_bench",
3600 srcs = [
3601 "bench/f16-igemm.cc",
3602 "bench/conv.h",
3603 "bench/google/conv.h",
3604 "src/xnnpack/AlignedAllocator.h",
3605 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003606 deps = MICROKERNEL_BENCHMARK_DEPS + [
3607 ":indirection",
3608 ":packing",
3609 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003610)
3611
3612xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613 name = "f16_gemm_bench",
3614 srcs = [
3615 "bench/f16-gemm.cc",
3616 "bench/gemm.h",
3617 "src/xnnpack/AlignedAllocator.h",
3618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003619 deps = MICROKERNEL_BENCHMARK_DEPS + [
3620 ":packing",
3621 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003622)
3623
3624xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003625 name = "f16_spmm_bench",
3626 srcs = [
3627 "bench/f16-spmm.cc",
3628 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003629 "src/xnnpack/AlignedAllocator.h",
3630 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003631 deps = MICROKERNEL_BENCHMARK_DEPS,
3632)
3633
3634xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003635 name = "f32_igemm_bench",
3636 srcs = [
3637 "bench/f32-igemm.cc",
3638 "bench/conv.h",
3639 "src/xnnpack/AlignedAllocator.h",
3640 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003641 deps = MICROKERNEL_BENCHMARK_DEPS + [
3642 ":indirection",
3643 ":packing",
3644 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645)
3646
3647xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003648 name = "f16_relu_bench",
3649 srcs = [
3650 "bench/f16-relu.cc",
3651 "src/xnnpack/AlignedAllocator.h",
3652 ] + MICROKERNEL_BENCHMARK_HDRS,
3653 deps = MICROKERNEL_BENCHMARK_DEPS,
3654)
3655
3656xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657 name = "f32_conv_hwc_bench",
3658 srcs = [
3659 "bench/f32-conv-hwc.cc",
3660 "bench/dconv.h",
3661 "src/xnnpack/AlignedAllocator.h",
3662 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003663 deps = MICROKERNEL_BENCHMARK_DEPS + [
3664 ":packing",
3665 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666)
3667
3668xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003669 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003670 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003671 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003672 "bench/dconv.h",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003675 deps = MICROKERNEL_BENCHMARK_DEPS + [
3676 ":packing",
3677 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003678)
3679
3680xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003681 name = "f16_dwconv_bench",
3682 srcs = [
3683 "bench/f16-dwconv.cc",
3684 "bench/dwconv.h",
3685 "bench/google/dwconv.h",
3686 "src/xnnpack/AlignedAllocator.h",
3687 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003688 deps = MICROKERNEL_BENCHMARK_DEPS + [
3689 ":indirection",
3690 ":packing",
3691 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003692)
3693
3694xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 name = "f32_dwconv_bench",
3696 srcs = [
3697 "bench/f32-dwconv.cc",
3698 "bench/dwconv.h",
3699 "src/xnnpack/AlignedAllocator.h",
3700 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003701 deps = MICROKERNEL_BENCHMARK_DEPS + [
3702 ":indirection",
3703 ":packing",
3704 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705)
3706
3707xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003708 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003710 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 "bench/dwconv.h",
3712 "src/xnnpack/AlignedAllocator.h",
3713 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003714 deps = MICROKERNEL_BENCHMARK_DEPS + [
3715 ":indirection",
3716 ":packing",
3717 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718)
3719
3720xnnpack_benchmark(
3721 name = "f32_gemm_bench",
3722 srcs = [
3723 "bench/f32-gemm.cc",
3724 "bench/gemm.h",
3725 "src/xnnpack/AlignedAllocator.h",
3726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003727 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003728 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003729)
3730
3731xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003732 name = "f32_hswish_bench",
3733 srcs = [
3734 "bench/f32-hswish.cc",
3735 "src/xnnpack/AlignedAllocator.h",
3736 ] + MICROKERNEL_BENCHMARK_HDRS,
3737 deps = MICROKERNEL_BENCHMARK_DEPS,
3738)
3739
3740xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003741 name = "f32_raddexpminusmax_bench",
3742 srcs = [
3743 "bench/f32-raddexpminusmax.cc",
3744 "src/xnnpack/AlignedAllocator.h",
3745 ] + MICROKERNEL_BENCHMARK_HDRS,
3746 deps = MICROKERNEL_BENCHMARK_DEPS,
3747)
3748
3749xnnpack_benchmark(
3750 name = "f32_raddextexp_bench",
3751 srcs = [
3752 "bench/f32-raddextexp.cc",
3753 "src/xnnpack/AlignedAllocator.h",
3754 ] + MICROKERNEL_BENCHMARK_HDRS,
3755 deps = MICROKERNEL_BENCHMARK_DEPS,
3756)
3757
3758xnnpack_benchmark(
3759 name = "f32_raddstoreexpminusmax_bench",
3760 srcs = [
3761 "bench/f32-raddstoreexpminusmax.cc",
3762 "src/xnnpack/AlignedAllocator.h",
3763 ] + MICROKERNEL_BENCHMARK_HDRS,
3764 deps = MICROKERNEL_BENCHMARK_DEPS,
3765)
3766
3767xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003768 name = "f32_relu_bench",
3769 srcs = [
3770 "bench/f32-relu.cc",
3771 "src/xnnpack/AlignedAllocator.h",
3772 ] + MICROKERNEL_BENCHMARK_HDRS,
3773 deps = MICROKERNEL_BENCHMARK_DEPS,
3774)
3775
3776xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003777 name = "f32_rmax_bench",
3778 srcs = [
3779 "bench/f32-rmax.cc",
3780 "src/xnnpack/AlignedAllocator.h",
3781 ] + MICROKERNEL_BENCHMARK_HDRS,
3782 deps = MICROKERNEL_BENCHMARK_DEPS,
3783)
3784
3785xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003786 name = "f32_sigmoid_bench",
3787 srcs = [
3788 "bench/f32-sigmoid.cc",
3789 "src/xnnpack/AlignedAllocator.h",
3790 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003791 deps = MICROKERNEL_BENCHMARK_DEPS,
3792)
3793
3794xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795 name = "f32_spmm_bench",
3796 srcs = [
3797 "bench/f32-spmm.cc",
3798 "bench/gemm.h",
3799 "src/xnnpack/AlignedAllocator.h",
3800 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003801 deps = MICROKERNEL_BENCHMARK_DEPS,
3802)
3803
3804xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003805 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003806 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003807 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003808 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003809 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003810 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003811)
3812
3813xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003814 name = "f32_vscaleexpminusmax_bench",
3815 srcs = [
3816 "bench/f32-vscaleexpminusmax.cc",
3817 "src/xnnpack/AlignedAllocator.h",
3818 ] + MICROKERNEL_BENCHMARK_HDRS,
3819 deps = MICROKERNEL_BENCHMARK_DEPS,
3820)
3821
3822xnnpack_benchmark(
3823 name = "f32_vscaleextexp_bench",
3824 srcs = [
3825 "bench/f32-vscaleextexp.cc",
3826 "src/xnnpack/AlignedAllocator.h",
3827 ] + MICROKERNEL_BENCHMARK_HDRS,
3828 deps = MICROKERNEL_BENCHMARK_DEPS,
3829)
3830
3831xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003832 name = "f32_vsqrt_bench",
3833 srcs = [
3834 "bench/f32-vsqrt.cc",
3835 "src/xnnpack/AlignedAllocator.h",
3836 ] + MICROKERNEL_BENCHMARK_HDRS,
3837 deps = MICROKERNEL_BENCHMARK_DEPS,
3838)
3839
3840xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003841 name = "f32_im2col_gemm_bench",
3842 srcs = [
3843 "bench/f32-im2col-gemm.cc",
3844 "bench/conv.h",
3845 "src/xnnpack/AlignedAllocator.h",
3846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003847 deps = MICROKERNEL_BENCHMARK_DEPS + [
3848 ":im2col",
3849 ":packing",
3850 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003851)
3852
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003853xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003854 name = "rounding_bench",
3855 srcs = [
3856 "bench/rounding.cc",
3857 "src/xnnpack/math-stubs.h",
3858 "src/xnnpack/AlignedAllocator.h",
3859 ] + MICROKERNEL_BENCHMARK_HDRS,
3860 deps = MICROKERNEL_BENCHMARK_DEPS,
3861)
3862
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863########################### Benchmarks for operators ###########################
3864
3865xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866 name = "average_pooling_bench",
3867 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003868 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003869 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003870 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003871)
3872
3873xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003874 name = "bankers_rounding_bench",
3875 srcs = ["bench/bankers-rounding.cc"],
3876 copts = xnnpack_optional_tflite_copts(),
3877 tags = ["nowin32"],
3878 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3879)
3880
3881xnnpack_benchmark(
3882 name = "ceiling_bench",
3883 srcs = ["bench/ceiling.cc"],
3884 copts = xnnpack_optional_tflite_copts(),
3885 tags = ["nowin32"],
3886 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3887)
3888
3889xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890 name = "channel_shuffle_bench",
3891 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003892 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003893)
3894
3895xnnpack_benchmark(
3896 name = "convolution_bench",
3897 srcs = ["bench/convolution.cc"],
3898 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003899 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003900 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901)
3902
3903xnnpack_benchmark(
3904 name = "deconvolution_bench",
3905 srcs = ["bench/deconvolution.cc"],
3906 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003907 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003908 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909)
3910
3911xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003912 name = "floor_bench",
3913 srcs = ["bench/floor.cc"],
3914 copts = xnnpack_optional_tflite_copts(),
3915 tags = ["nowin32"],
3916 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3917)
3918
3919xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003920 name = "global_average_pooling_bench",
3921 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003922 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003923)
3924
3925xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003926 name = "hardswish_bench",
3927 srcs = ["bench/hardswish.cc"],
3928 copts = xnnpack_optional_tflite_copts(),
3929 tags = ["nowin32"],
3930 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3931)
3932
3933xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934 name = "max_pooling_bench",
3935 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003936 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003937)
3938
3939xnnpack_benchmark(
3940 name = "sigmoid_bench",
3941 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003942 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003943 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003944 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003945)
3946
3947xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003948 name = "prelu_bench",
3949 srcs = ["bench/prelu.cc"],
3950 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003951 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003952 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003953)
3954
3955xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003956 name = "softmax_bench",
3957 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003958 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003959 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003960 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961)
3962
Marat Dukhan87727142020-06-24 15:24:10 -07003963xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003964 name = "square_root_bench",
3965 srcs = ["bench/square-root.cc"],
3966 copts = xnnpack_optional_tflite_copts(),
3967 tags = ["nowin32"],
3968 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3969)
3970
3971xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003972 name = "truncation_bench",
3973 srcs = ["bench/truncation.cc"],
3974 deps = OPERATOR_BENCHMARK_DEPS,
3975)
3976
Marat Dukhanc068bb62019-10-04 13:24:39 -07003977############################# End-to-end benchmarks ############################
3978
3979cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003980 name = "fp32_mobilenet_v1",
3981 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003982 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003983 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003984 linkstatic = True,
3985 deps = [
3986 ":XNNPACK",
3987 "@pthreadpool",
3988 ],
3989)
3990
3991cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003992 name = "fp16_mobilenet_v1",
3993 srcs = ["models/fp16-mobilenet-v1.cc"],
3994 hdrs = ["models/models.h"],
3995 copts = xnnpack_std_cxxopts(),
3996 linkstatic = True,
3997 deps = [
3998 ":XNNPACK",
3999 "@FP16",
4000 "@pthreadpool",
4001 ],
4002)
4003
4004cc_library(
4005 name = "fp32_mobilenet_v2",
4006 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004007 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004008 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004009 linkstatic = True,
4010 deps = [
4011 ":XNNPACK",
4012 "@pthreadpool",
4013 ],
4014)
4015
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004016cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004017 name = "fp16_mobilenet_v2",
4018 srcs = ["models/fp16-mobilenet-v2.cc"],
4019 hdrs = ["models/models.h"],
4020 copts = xnnpack_std_cxxopts(),
4021 linkstatic = True,
4022 deps = [
4023 ":XNNPACK",
4024 "@FP16",
4025 "@pthreadpool",
4026 ],
4027)
4028
4029cc_library(
4030 name = "fp32_mobilenet_v3_large",
4031 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004032 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004033 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004034 linkstatic = True,
4035 deps = [
4036 ":XNNPACK",
4037 "@pthreadpool",
4038 ],
4039)
4040
4041cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004042 name = "fp32_mobilenet_v3_small",
4043 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004044 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004045 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004046 linkstatic = True,
4047 deps = [
4048 ":XNNPACK",
4049 "@pthreadpool",
4050 ],
4051)
4052
Marat Dukhanc068bb62019-10-04 13:24:39 -07004053xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004054 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004055 srcs = [
4056 "bench/f32-dwconv-e2e.cc",
4057 "bench/end2end.h",
4058 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004059 deps = MICROKERNEL_BENCHMARK_DEPS + [
4060 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004061 ":fp32_mobilenet_v1",
4062 ":fp32_mobilenet_v2",
4063 ":fp32_mobilenet_v3_large",
4064 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004065 ],
4066)
4067
4068xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004069 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004070 srcs = [
4071 "bench/f32-gemm-e2e.cc",
4072 "bench/end2end.h",
4073 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004074 deps = MICROKERNEL_BENCHMARK_DEPS + [
4075 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004076 ":fp32_mobilenet_v1",
4077 ":fp32_mobilenet_v2",
4078 ":fp32_mobilenet_v3_large",
4079 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004080 ],
4081)
4082
4083xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004084 name = "end2end_bench",
4085 srcs = ["bench/end2end.cc"],
4086 deps = [
4087 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004088 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004089 ":fp16_mobilenet_v1",
4090 ":fp16_mobilenet_v2",
4091 ":fp32_mobilenet_v1",
4092 ":fp32_mobilenet_v2",
4093 ":fp32_mobilenet_v3_large",
4094 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004095 "@pthreadpool",
4096 ],
4097)
4098
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004099#################### Accuracy evaluation for math functions ####################
4100
4101xnnpack_benchmark(
4102 name = "f32_exp_eval",
4103 srcs = [
4104 "eval/f32-exp.cc",
4105 "src/xnnpack/AlignedAllocator.h",
4106 ] + ACCURACY_EVAL_HDRS,
4107 deps = ACCURACY_EVAL_DEPS,
4108)
4109
Marat Dukhan515c9772019-10-17 18:07:57 -07004110xnnpack_benchmark(
4111 name = "f32_expminus_eval",
4112 srcs = [
4113 "eval/f32-expminus.cc",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + ACCURACY_EVAL_HDRS,
4116 deps = ACCURACY_EVAL_DEPS,
4117)
4118
Marat Dukhan98ba4412019-10-23 02:14:28 -07004119xnnpack_benchmark(
4120 name = "f32_extexp_eval",
4121 srcs = [
4122 "eval/f32-extexp.cc",
4123 "src/xnnpack/AlignedAllocator.h",
4124 ] + ACCURACY_EVAL_HDRS,
4125 deps = ACCURACY_EVAL_DEPS,
4126)
4127
Marat Dukhan8853b822020-05-07 12:19:01 -07004128xnnpack_unit_test(
4129 name = "f32_roundne_eval",
4130 srcs = [
4131 "eval/f32-roundne.cc",
4132 "src/xnnpack/AlignedAllocator.h",
4133 "src/xnnpack/math-stubs.h",
4134 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004135 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004136 deps = MICROKERNEL_TEST_DEPS,
4137)
4138
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004139xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004140 name = "f32_roundd_eval",
4141 srcs = [
4142 "eval/f32-roundd.cc",
4143 "src/xnnpack/AlignedAllocator.h",
4144 "src/xnnpack/math-stubs.h",
4145 ] + MICROKERNEL_TEST_HDRS,
4146 automatic = False,
4147 deps = MICROKERNEL_TEST_DEPS,
4148)
4149
4150xnnpack_unit_test(
4151 name = "f32_roundu_eval",
4152 srcs = [
4153 "eval/f32-roundu.cc",
4154 "src/xnnpack/AlignedAllocator.h",
4155 "src/xnnpack/math-stubs.h",
4156 ] + MICROKERNEL_TEST_HDRS,
4157 automatic = False,
4158 deps = MICROKERNEL_TEST_DEPS,
4159)
4160
4161xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004162 name = "f32_roundz_eval",
4163 srcs = [
4164 "eval/f32-roundz.cc",
4165 "src/xnnpack/AlignedAllocator.h",
4166 "src/xnnpack/math-stubs.h",
4167 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004168 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004169 deps = MICROKERNEL_TEST_DEPS,
4170)
4171
Marat Dukhan346a9e52019-11-15 09:06:30 -08004172xnnpack_benchmark(
4173 name = "f32_sigmoid_eval",
4174 srcs = [
4175 "eval/f32-sigmoid.cc",
4176 "src/xnnpack/AlignedAllocator.h",
4177 ] + ACCURACY_EVAL_HDRS,
4178 deps = ACCURACY_EVAL_DEPS,
4179)
4180
Marat Dukhan84000762020-06-29 18:38:43 -07004181xnnpack_benchmark(
4182 name = "f32_sqrt_eval",
4183 srcs = [
4184 "eval/f32-sqrt.cc",
4185 "src/xnnpack/AlignedAllocator.h",
4186 ] + ACCURACY_EVAL_HDRS,
4187 deps = ACCURACY_EVAL_DEPS + [
4188 ":bench_utils",
4189 ],
4190)
4191
Marat Dukhan08c4a432019-10-03 09:29:21 -07004192######################### Unit tests for micro-kernels #########################
4193
4194xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004195 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004197 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 "test/gemm-microkernel-tester.h",
4199 "src/xnnpack/AlignedAllocator.h",
4200 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004201 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004202)
4203
4204xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004205 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004206 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004207 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004208 "test/spmm-microkernel-tester.h",
4209 "src/xnnpack/AlignedAllocator.h",
4210 ] + MICROKERNEL_TEST_HDRS,
4211 deps = MICROKERNEL_TEST_DEPS,
4212)
4213
4214xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 name = "f32_argmaxpool_test",
4216 srcs = [
4217 "test/f32-argmaxpool.cc",
4218 "test/argmaxpool-microkernel-tester.h",
4219 "src/xnnpack/AlignedAllocator.h",
4220 ] + MICROKERNEL_TEST_HDRS,
4221 deps = MICROKERNEL_TEST_DEPS,
4222)
4223
4224xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004225 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004227 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 "test/avgpool-microkernel-tester.h",
4229 "src/xnnpack/AlignedAllocator.h",
4230 ] + MICROKERNEL_TEST_HDRS,
4231 deps = MICROKERNEL_TEST_DEPS,
4232)
4233
4234xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004235 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004236 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004237 "test/f32-ibilinear.cc",
4238 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004239 "src/xnnpack/AlignedAllocator.h",
4240 ] + MICROKERNEL_TEST_HDRS,
4241 deps = MICROKERNEL_TEST_DEPS,
4242)
4243
4244xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004245 name = "f16_clamp_test",
4246 srcs = [
4247 "test/f16-clamp.cc",
4248 "test/clamp-microkernel-tester.h",
4249 ] + MICROKERNEL_TEST_HDRS,
4250 deps = MICROKERNEL_TEST_DEPS,
4251)
4252
4253xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004254 name = "f32_clamp_test",
4255 srcs = [
4256 "test/f32-clamp.cc",
4257 "test/clamp-microkernel-tester.h",
4258 ] + MICROKERNEL_TEST_HDRS,
4259 deps = MICROKERNEL_TEST_DEPS,
4260)
4261
4262xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004263 name = "f32_igemm_test",
4264 srcs = [
4265 "test/f32-igemm.cc",
4266 "test/gemm-microkernel-tester.h",
4267 "src/xnnpack/AlignedAllocator.h",
4268 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004269 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004270)
4271
4272xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004273 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004274 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004275 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 "test/gemm-microkernel-tester.h",
4277 "src/xnnpack/AlignedAllocator.h",
4278 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004279 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280)
4281
4282xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004283 name = "f32_igemm_minmax_test",
4284 srcs = [
4285 "test/f32-igemm-minmax.cc",
4286 "test/gemm-microkernel-tester.h",
4287 "src/xnnpack/AlignedAllocator.h",
4288 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004289 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004290)
4291
4292xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004293 name = "f16_igemm_minmax_test",
4294 srcs = [
4295 "test/f16-igemm-minmax.cc",
4296 "test/gemm-microkernel-tester.h",
4297 "src/xnnpack/AlignedAllocator.h",
4298 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004299 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004300)
4301
4302xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303 name = "f32_conv_hwc_test",
4304 srcs = [
4305 "test/f32-conv-hwc.cc",
4306 "test/conv-hwc-microkernel-tester.h",
4307 "src/xnnpack/AlignedAllocator.h",
4308 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004309 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310)
4311
4312xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004313 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004315 "test/f32-conv-hwc2chw.cc",
4316 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317 "src/xnnpack/AlignedAllocator.h",
4318 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004319 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004320)
4321
4322xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004323 name = "f32_dwconv_test",
4324 srcs = [
4325 "test/f32-dwconv.cc",
4326 "test/dwconv-microkernel-tester.h",
4327 "src/xnnpack/AlignedAllocator.h",
4328 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004329 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004330)
4331
4332xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004333 name = "f16_dwconv_minmax_test",
4334 srcs = [
4335 "test/f16-dwconv-minmax.cc",
4336 "test/dwconv-microkernel-tester.h",
4337 "src/xnnpack/AlignedAllocator.h",
4338 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004339 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004340)
4341
4342xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004343 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004344 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004345 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 "test/dwconv-microkernel-tester.h",
4347 "src/xnnpack/AlignedAllocator.h",
4348 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004349 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004350)
4351
4352xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004353 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004355 "test/f32-dwconv-chw.cc",
4356 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 "src/xnnpack/AlignedAllocator.h",
4358 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004359 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360)
4361
4362xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004363 name = "f16_gavgpool_minmax_test",
4364 srcs = [
4365 "test/f16-gavgpool-minmax.cc",
4366 "test/gavgpool-microkernel-tester.h",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + MICROKERNEL_TEST_HDRS,
4369 deps = MICROKERNEL_TEST_DEPS,
4370)
4371
4372xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004373 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004374 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004375 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 "test/gavgpool-microkernel-tester.h",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + MICROKERNEL_TEST_HDRS,
4379 deps = MICROKERNEL_TEST_DEPS,
4380)
4381
4382xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004383 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004385 "test/f32-gavgpool-cw.cc",
4386 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 "src/xnnpack/AlignedAllocator.h",
4388 ] + MICROKERNEL_TEST_HDRS,
4389 deps = MICROKERNEL_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004393 name = "f32_gemm_test",
4394 srcs = [
4395 "test/f32-gemm.cc",
4396 "test/gemm-microkernel-tester.h",
4397 "src/xnnpack/AlignedAllocator.h",
4398 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004399 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004400)
4401
4402xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004403 name = "f32_gemm_relu_test",
4404 srcs = [
4405 "test/f32-gemm-relu.cc",
4406 "test/gemm-microkernel-tester.h",
4407 "src/xnnpack/AlignedAllocator.h",
4408 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004409 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004410)
4411
4412xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004413 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004415 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 "test/gemm-microkernel-tester.h",
4417 "src/xnnpack/AlignedAllocator.h",
4418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004419 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004420)
4421
4422xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004423 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004425 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426 "test/gemm-microkernel-tester.h",
4427 "src/xnnpack/AlignedAllocator.h",
4428 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004429 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430)
4431
4432xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004433 name = "f16_hswish_test",
4434 srcs = [
4435 "test/f16-hswish.cc",
4436 "test/hswish-microkernel-tester.h",
4437 ] + MICROKERNEL_TEST_HDRS,
4438 deps = MICROKERNEL_TEST_DEPS,
4439)
4440
4441xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 name = "f32_hswish_test",
4443 srcs = [
4444 "test/f32-hswish.cc",
4445 "test/hswish-microkernel-tester.h",
4446 ] + MICROKERNEL_TEST_HDRS,
4447 deps = MICROKERNEL_TEST_DEPS,
4448)
4449
4450xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004451 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004453 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 "test/maxpool-microkernel-tester.h",
4455 ] + MICROKERNEL_TEST_HDRS,
4456 deps = MICROKERNEL_TEST_DEPS,
4457)
4458
4459xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004460 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004462 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004463 "test/avgpool-microkernel-tester.h",
4464 "src/xnnpack/AlignedAllocator.h",
4465 ] + MICROKERNEL_TEST_HDRS,
4466 deps = MICROKERNEL_TEST_DEPS,
4467)
4468
4469xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004470 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004472 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 "test/gemm-microkernel-tester.h",
4474 "src/xnnpack/AlignedAllocator.h",
4475 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004476 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477)
4478
4479xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004480 name = "f16_prelu_test",
4481 srcs = [
4482 "test/f16-prelu.cc",
4483 "test/prelu-microkernel-tester.h",
4484 "src/xnnpack/AlignedAllocator.h",
4485 ] + MICROKERNEL_TEST_HDRS,
4486 deps = MICROKERNEL_TEST_DEPS,
4487)
4488
4489xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490 name = "f32_prelu_test",
4491 srcs = [
4492 "test/f32-prelu.cc",
4493 "test/prelu-microkernel-tester.h",
4494 "src/xnnpack/AlignedAllocator.h",
4495 ] + MICROKERNEL_TEST_HDRS,
4496 deps = MICROKERNEL_TEST_DEPS,
4497)
4498
4499xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004500 name = "f32_raddexpminusmax_test",
4501 srcs = [
4502 "test/f32-raddexpminusmax.cc",
4503 "test/raddexpminusmax-microkernel-tester.h",
4504 ] + MICROKERNEL_TEST_HDRS,
4505 deps = MICROKERNEL_TEST_DEPS,
4506)
4507
4508xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004509 name = "f32_raddextexp_test",
4510 srcs = [
4511 "test/f32-raddextexp.cc",
4512 "test/raddextexp-microkernel-tester.h",
4513 ] + MICROKERNEL_TEST_HDRS,
4514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
4517xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004518 name = "f32_raddstoreexpminusmax_test",
4519 srcs = [
4520 "test/f32-raddstoreexpminusmax.cc",
4521 "test/raddstoreexpminusmax-microkernel-tester.h",
4522 ] + MICROKERNEL_TEST_HDRS,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004527 name = "f32_relu_test",
4528 srcs = [
4529 "test/f32-relu.cc",
4530 "test/vunary-microkernel-tester.h",
4531 ] + MICROKERNEL_TEST_HDRS,
4532 deps = MICROKERNEL_TEST_DEPS,
4533)
4534
4535xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536 name = "f32_rmax_test",
4537 srcs = [
4538 "test/f32-rmax.cc",
4539 "test/rmax-microkernel-tester.h",
4540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004545 name = "f32_sigmoid_test",
4546 srcs = [
4547 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004548 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004554 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004555 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004556 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557 "test/spmm-microkernel-tester.h",
4558 "src/xnnpack/AlignedAllocator.h",
4559 ] + MICROKERNEL_TEST_HDRS,
4560 deps = MICROKERNEL_TEST_DEPS,
4561)
4562
4563xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004564 name = "f16_vadd_minmax_test",
4565 srcs = [
4566 "test/f16-vadd-minmax.cc",
4567 "test/vbinary-microkernel-tester.h",
4568 ] + MICROKERNEL_TEST_HDRS,
4569 deps = MICROKERNEL_TEST_DEPS,
4570)
4571
4572xnnpack_unit_test(
4573 name = "f16_vaddc_minmax_test",
4574 srcs = [
4575 "test/f16-vaddc-minmax.cc",
4576 "test/vbinaryc-microkernel-tester.h",
4577 ] + MICROKERNEL_TEST_HDRS,
4578 deps = MICROKERNEL_TEST_DEPS,
4579)
4580
4581xnnpack_unit_test(
4582 name = "f16_vdiv_minmax_test",
4583 srcs = [
4584 "test/f16-vdiv-minmax.cc",
4585 "test/vbinary-microkernel-tester.h",
4586 ] + MICROKERNEL_TEST_HDRS,
4587 deps = MICROKERNEL_TEST_DEPS,
4588)
4589
4590xnnpack_unit_test(
4591 name = "f16_vdivc_minmax_test",
4592 srcs = [
4593 "test/f16-vdivc-minmax.cc",
4594 "test/vbinaryc-microkernel-tester.h",
4595 ] + MICROKERNEL_TEST_HDRS,
4596 deps = MICROKERNEL_TEST_DEPS,
4597)
4598
4599xnnpack_unit_test(
4600 name = "f16_vrdivc_minmax_test",
4601 srcs = [
4602 "test/f16-vrdivc-minmax.cc",
4603 "test/vbinaryc-microkernel-tester.h",
4604 ] + MICROKERNEL_TEST_HDRS,
4605 deps = MICROKERNEL_TEST_DEPS,
4606)
4607
4608xnnpack_unit_test(
4609 name = "f16_vmax_test",
4610 srcs = [
4611 "test/f16-vmax.cc",
4612 "test/vbinary-microkernel-tester.h",
4613 ] + MICROKERNEL_TEST_HDRS,
4614 deps = MICROKERNEL_TEST_DEPS,
4615)
4616
4617xnnpack_unit_test(
4618 name = "f16_vmaxc_test",
4619 srcs = [
4620 "test/f16-vmaxc.cc",
4621 "test/vbinaryc-microkernel-tester.h",
4622 ] + MICROKERNEL_TEST_HDRS,
4623 deps = MICROKERNEL_TEST_DEPS,
4624)
4625
4626xnnpack_unit_test(
4627 name = "f16_vmin_test",
4628 srcs = [
4629 "test/f16-vmin.cc",
4630 "test/vbinary-microkernel-tester.h",
4631 ] + MICROKERNEL_TEST_HDRS,
4632 deps = MICROKERNEL_TEST_DEPS,
4633)
4634
4635xnnpack_unit_test(
4636 name = "f16_vminc_test",
4637 srcs = [
4638 "test/f16-vminc.cc",
4639 "test/vbinaryc-microkernel-tester.h",
4640 ] + MICROKERNEL_TEST_HDRS,
4641 deps = MICROKERNEL_TEST_DEPS,
4642)
4643
4644xnnpack_unit_test(
4645 name = "f16_vmul_minmax_test",
4646 srcs = [
4647 "test/f16-vmul-minmax.cc",
4648 "test/vbinary-microkernel-tester.h",
4649 ] + MICROKERNEL_TEST_HDRS,
4650 deps = MICROKERNEL_TEST_DEPS,
4651)
4652
4653xnnpack_unit_test(
4654 name = "f16_vmulc_minmax_test",
4655 srcs = [
4656 "test/f16-vmulc-minmax.cc",
4657 "test/vbinaryc-microkernel-tester.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
4663 name = "f16_vsub_minmax_test",
4664 srcs = [
4665 "test/f16-vsub-minmax.cc",
4666 "test/vbinary-microkernel-tester.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
4672 name = "f16_vsubc_minmax_test",
4673 srcs = [
4674 "test/f16-vsubc-minmax.cc",
4675 "test/vbinaryc-microkernel-tester.h",
4676 ] + MICROKERNEL_TEST_HDRS,
4677 deps = MICROKERNEL_TEST_DEPS,
4678)
4679
4680xnnpack_unit_test(
4681 name = "f16_vrsubc_minmax_test",
4682 srcs = [
4683 "test/f16-vrsubc-minmax.cc",
4684 "test/vbinaryc-microkernel-tester.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004690 name = "f32_vabs_test",
4691 srcs = [
4692 "test/f32-vabs.cc",
4693 "test/vunary-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004699 name = "f32_vadd_test",
4700 srcs = [
4701 "test/f32-vadd.cc",
4702 "test/vbinary-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004708 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004709 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004710 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004711 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004717 name = "f32_vaddc_test",
4718 srcs = [
4719 "test/f32-vaddc.cc",
4720 "test/vbinaryc-microkernel-tester.h",
4721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004726 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004727 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004728 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004729 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004730 ] + MICROKERNEL_TEST_HDRS,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004735 name = "f32_vdiv_test",
4736 srcs = [
4737 "test/f32-vdiv.cc",
4738 "test/vbinary-microkernel-tester.h",
4739 ] + MICROKERNEL_TEST_HDRS,
4740 deps = MICROKERNEL_TEST_DEPS,
4741)
4742
4743xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004744 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004745 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004746 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004747 "test/vbinary-microkernel-tester.h",
4748 ] + MICROKERNEL_TEST_HDRS,
4749 deps = MICROKERNEL_TEST_DEPS,
4750)
4751
4752xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004753 name = "f32_vdivc_test",
4754 srcs = [
4755 "test/f32-vdivc.cc",
4756 "test/vbinaryc-microkernel-tester.h",
4757 ] + MICROKERNEL_TEST_HDRS,
4758 deps = MICROKERNEL_TEST_DEPS,
4759)
4760
4761xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004762 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004763 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004764 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004765 "test/vbinaryc-microkernel-tester.h",
4766 ] + MICROKERNEL_TEST_HDRS,
4767 deps = MICROKERNEL_TEST_DEPS,
4768)
4769
4770xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004771 name = "f32_vrdivc_test",
4772 srcs = [
4773 "test/f32-vrdivc.cc",
4774 "test/vbinaryc-microkernel-tester.h",
4775 ] + MICROKERNEL_TEST_HDRS,
4776 deps = MICROKERNEL_TEST_DEPS,
4777)
4778
4779xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004780 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004781 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004782 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004783 "test/vbinaryc-microkernel-tester.h",
4784 ] + MICROKERNEL_TEST_HDRS,
4785 deps = MICROKERNEL_TEST_DEPS,
4786)
4787
4788xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004789 name = "f32_vmax_test",
4790 srcs = [
4791 "test/f32-vmax.cc",
4792 "test/vbinary-microkernel-tester.h",
4793 ] + MICROKERNEL_TEST_HDRS,
4794 deps = MICROKERNEL_TEST_DEPS,
4795)
4796
4797xnnpack_unit_test(
4798 name = "f32_vmaxc_test",
4799 srcs = [
4800 "test/f32-vmaxc.cc",
4801 "test/vbinaryc-microkernel-tester.h",
4802 ] + MICROKERNEL_TEST_HDRS,
4803 deps = MICROKERNEL_TEST_DEPS,
4804)
4805
4806xnnpack_unit_test(
4807 name = "f32_vmin_test",
4808 srcs = [
4809 "test/f32-vmin.cc",
4810 "test/vbinary-microkernel-tester.h",
4811 ] + MICROKERNEL_TEST_HDRS,
4812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
4815xnnpack_unit_test(
4816 name = "f32_vminc_test",
4817 srcs = [
4818 "test/f32-vminc.cc",
4819 "test/vbinaryc-microkernel-tester.h",
4820 ] + MICROKERNEL_TEST_HDRS,
4821 deps = MICROKERNEL_TEST_DEPS,
4822)
4823
4824xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004825 name = "f32_vmul_test",
4826 srcs = [
4827 "test/f32-vmul.cc",
4828 "test/vbinary-microkernel-tester.h",
4829 ] + MICROKERNEL_TEST_HDRS,
4830 deps = MICROKERNEL_TEST_DEPS,
4831)
4832
4833xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004834 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004836 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004837 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004838 ] + MICROKERNEL_TEST_HDRS,
4839 deps = MICROKERNEL_TEST_DEPS,
4840)
4841
4842xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004843 name = "f32_vmulc_test",
4844 srcs = [
4845 "test/f32-vmulc.cc",
4846 "test/vbinaryc-microkernel-tester.h",
4847 ] + MICROKERNEL_TEST_HDRS,
4848 deps = MICROKERNEL_TEST_DEPS,
4849)
4850
4851xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004852 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004853 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004854 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004855 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004856 ] + MICROKERNEL_TEST_HDRS,
4857 deps = MICROKERNEL_TEST_DEPS,
4858)
4859
4860xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004861 name = "f16_vmulcaddc_minmax_test",
4862 srcs = [
4863 "test/f16-vmulcaddc-minmax.cc",
4864 "test/vmulcaddc-microkernel-tester.h",
4865 "src/xnnpack/AlignedAllocator.h",
4866 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004867 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004868)
4869
4870xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004871 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004872 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004873 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004874 "test/vmulcaddc-microkernel-tester.h",
4875 "src/xnnpack/AlignedAllocator.h",
4876 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004877 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878)
4879
4880xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004881 name = "f32_vlrelu_test",
4882 srcs = [
4883 "test/f32-vlrelu.cc",
4884 "test/vunary-microkernel-tester.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004890 name = "f32_vneg_test",
4891 srcs = [
4892 "test/f32-vneg.cc",
4893 "test/vunary-microkernel-tester.h",
4894 ] + MICROKERNEL_TEST_HDRS,
4895 deps = MICROKERNEL_TEST_DEPS,
4896)
4897
4898xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004899 name = "f32_vrndne_test",
4900 srcs = [
4901 "test/f32-vrndne.cc",
4902 "test/vunary-microkernel-tester.h",
4903 ] + MICROKERNEL_TEST_HDRS,
4904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
4907xnnpack_unit_test(
4908 name = "f32_vrndz_test",
4909 srcs = [
4910 "test/f32-vrndz.cc",
4911 "test/vunary-microkernel-tester.h",
4912 ] + MICROKERNEL_TEST_HDRS,
4913 deps = MICROKERNEL_TEST_DEPS,
4914)
4915
4916xnnpack_unit_test(
4917 name = "f32_vrndu_test",
4918 srcs = [
4919 "test/f32-vrndu.cc",
4920 "test/vunary-microkernel-tester.h",
4921 ] + MICROKERNEL_TEST_HDRS,
4922 deps = MICROKERNEL_TEST_DEPS,
4923)
4924
4925xnnpack_unit_test(
4926 name = "f32_vrndd_test",
4927 srcs = [
4928 "test/f32-vrndd.cc",
4929 "test/vunary-microkernel-tester.h",
4930 ] + MICROKERNEL_TEST_HDRS,
4931 deps = MICROKERNEL_TEST_DEPS,
4932)
4933
4934xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004935 name = "f32_vscale_test",
4936 srcs = [
4937 "test/f32-vscale.cc",
4938 "test/vscale-microkernel-tester.h",
4939 ] + MICROKERNEL_TEST_HDRS,
4940 deps = MICROKERNEL_TEST_DEPS,
4941)
4942
4943xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004944 name = "f32_vscaleexpminusmax_test",
4945 srcs = [
4946 "test/f32-vscaleexpminusmax.cc",
4947 "test/vscaleexpminusmax-microkernel-tester.h",
4948 ] + MICROKERNEL_TEST_HDRS,
4949 deps = MICROKERNEL_TEST_DEPS,
4950)
4951
4952xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004953 name = "f32_vscaleextexp_test",
4954 srcs = [
4955 "test/f32-vscaleextexp.cc",
4956 "test/vscaleextexp-microkernel-tester.h",
4957 ] + MICROKERNEL_TEST_HDRS,
4958 deps = MICROKERNEL_TEST_DEPS,
4959)
4960
4961xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004962 name = "f32_vsqr_test",
4963 srcs = [
4964 "test/f32-vsqr.cc",
4965 "test/vunary-microkernel-tester.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004971 name = "f32_vsqrdiff_test",
4972 srcs = [
4973 "test/f32-vsqrdiff.cc",
4974 "test/vbinary-microkernel-tester.h",
4975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
4980 name = "f32_vsqrdiffc_test",
4981 srcs = [
4982 "test/f32-vsqrdiffc.cc",
4983 "test/vbinaryc-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004989 name = "f32_vsqrt_test",
4990 srcs = [
4991 "test/f32-vsqrt.cc",
4992 "test/vunary-microkernel-tester.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004998 name = "f32_vsub_test",
4999 srcs = [
5000 "test/f32-vsub.cc",
5001 "test/vbinary-microkernel-tester.h",
5002 ] + MICROKERNEL_TEST_HDRS,
5003 deps = MICROKERNEL_TEST_DEPS,
5004)
5005
5006xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005007 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005008 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005009 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005010 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005016 name = "f32_vsubc_test",
5017 srcs = [
5018 "test/f32-vsubc.cc",
5019 "test/vbinaryc-microkernel-tester.h",
5020 ] + MICROKERNEL_TEST_HDRS,
5021 deps = MICROKERNEL_TEST_DEPS,
5022)
5023
5024xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005025 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005026 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005027 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005028 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005029 ] + MICROKERNEL_TEST_HDRS,
5030 deps = MICROKERNEL_TEST_DEPS,
5031)
5032
5033xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005034 name = "f32_vrsubc_test",
5035 srcs = [
5036 "test/f32-vrsubc.cc",
5037 "test/vbinaryc-microkernel-tester.h",
5038 ] + MICROKERNEL_TEST_HDRS,
5039 deps = MICROKERNEL_TEST_DEPS,
5040)
5041
5042xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005043 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005044 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005045 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005046 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005047 ] + MICROKERNEL_TEST_HDRS,
5048 deps = MICROKERNEL_TEST_DEPS,
5049)
5050
5051xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005052 name = "qs8_gemm_minmax_test",
5053 srcs = [
5054 "test/qs8-gemm-minmax.cc",
5055 "test/gemm-microkernel-tester.h",
5056 "src/xnnpack/AlignedAllocator.h",
5057 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5059)
5060
5061xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005062 name = "qs8_igemm_minmax_test",
5063 srcs = [
5064 "test/qs8-igemm-minmax.cc",
5065 "test/gemm-microkernel-tester.h",
5066 "src/xnnpack/AlignedAllocator.h",
5067 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5069)
5070
5071xnnpack_unit_test(
5072 name = "qs8_requantization_test",
5073 srcs = [
5074 "src/xnnpack/requantization-stubs.h",
5075 "test/qs8-requantization.cc",
5076 "test/requantization-tester.h",
5077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005082 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005084 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 "test/avgpool-microkernel-tester.h",
5086 "src/xnnpack/AlignedAllocator.h",
5087 ] + MICROKERNEL_TEST_HDRS,
5088 deps = MICROKERNEL_TEST_DEPS,
5089)
5090
5091xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005092 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005094 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 "test/gemm-microkernel-tester.h",
5096 "src/xnnpack/AlignedAllocator.h",
5097 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005098 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005099)
5100
5101xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005102 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005103 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005104 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005105 "test/dwconv-microkernel-tester.h",
5106 "src/xnnpack/AlignedAllocator.h",
5107 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005108 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005109)
5110
5111xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005112 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005113 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005114 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005115 "test/gavgpool-microkernel-tester.h",
5116 "src/xnnpack/AlignedAllocator.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005122 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005124 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 "test/gemm-microkernel-tester.h",
5126 "src/xnnpack/AlignedAllocator.h",
5127 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005128 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129)
5130
5131xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005132 name = "qu8_requantization_test",
5133 srcs = [
5134 "src/xnnpack/requantization-stubs.h",
5135 "test/qu8-requantization.cc",
5136 "test/requantization-tester.h",
5137 ] + MICROKERNEL_TEST_HDRS,
5138 deps = MICROKERNEL_TEST_DEPS,
5139)
5140
5141xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005142 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005144 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005145 "test/vadd-microkernel-tester.h",
5146 ] + MICROKERNEL_TEST_HDRS,
5147 deps = MICROKERNEL_TEST_DEPS,
5148)
5149
5150xnnpack_unit_test(
5151 name = "u8_clamp_test",
5152 srcs = [
5153 "test/u8-clamp.cc",
5154 "test/clamp-microkernel-tester.h",
5155 ] + MICROKERNEL_TEST_HDRS,
5156 deps = MICROKERNEL_TEST_DEPS,
5157)
5158
5159xnnpack_unit_test(
5160 name = "u8_lut32norm_test",
5161 srcs = [
5162 "test/u8-lut32norm.cc",
5163 "test/lut-norm-microkernel-tester.h",
5164 ] + MICROKERNEL_TEST_HDRS,
5165 deps = MICROKERNEL_TEST_DEPS,
5166)
5167
5168xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005169 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005170 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005171 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005172 "test/maxpool-microkernel-tester.h",
5173 ] + MICROKERNEL_TEST_HDRS,
5174 deps = MICROKERNEL_TEST_DEPS,
5175)
5176
5177xnnpack_unit_test(
5178 name = "u8_rmax_test",
5179 srcs = [
5180 "test/u8-rmax.cc",
5181 "test/rmax-microkernel-tester.h",
5182 ] + MICROKERNEL_TEST_HDRS,
5183 deps = MICROKERNEL_TEST_DEPS,
5184)
5185
5186xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005187 name = "x32_fill_test",
5188 srcs = [
5189 "test/x32-fill.cc",
5190 "test/fill-microkernel-tester.h",
5191 ] + MICROKERNEL_TEST_HDRS,
5192 deps = MICROKERNEL_TEST_DEPS,
5193)
5194
5195xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 name = "x32_packx_test",
5197 srcs = [
5198 "test/x32-packx.cc",
5199 "test/pack-microkernel-tester.h",
5200 "src/xnnpack/AlignedAllocator.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
5206 name = "x32_pad_test",
5207 srcs = [
5208 "test/x32-pad.cc",
5209 "test/pad-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
5214xnnpack_unit_test(
5215 name = "x32_unpool_test",
5216 srcs = [
5217 "test/x32-unpool.cc",
5218 "test/unpool-microkernel-tester.h",
5219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
5224 name = "x32_zip_test",
5225 srcs = [
5226 "test/x32-zip.cc",
5227 "test/zip-microkernel-tester.h",
5228 ] + MICROKERNEL_TEST_HDRS,
5229 deps = MICROKERNEL_TEST_DEPS,
5230)
5231
5232xnnpack_unit_test(
5233 name = "x8_lut_test",
5234 srcs = [
5235 "test/x8-lut.cc",
5236 "test/lut-microkernel-tester.h",
5237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
5242 name = "x8_zip_test",
5243 srcs = [
5244 "test/x8-zip.cc",
5245 "test/zip-microkernel-tester.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
Marat Dukhan20c3b922020-03-10 03:45:06 -07005250########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251
5252xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005253 name = "operator_size_test",
5254 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005255 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256)
5257
Marat Dukhan20c3b922020-03-10 03:45:06 -07005258xnnpack_binary(
5259 name = "subgraph_size_test",
5260 srcs = ["test/subgraph-size.c"],
5261 deps = [":XNNPACK"],
5262)
5263
5264########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265
5266xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005267 name = "abs_nc_test",
5268 srcs = [
5269 "test/abs-nc.cc",
5270 "test/abs-operator-tester.h",
5271 ],
5272 deps = OPERATOR_TEST_DEPS,
5273)
5274
5275xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005276 name = "add_nd_test",
5277 srcs = [
5278 "test/add-nd.cc",
5279 "test/binary-elementwise-operator-tester.h",
5280 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005281 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005282)
5283
5284xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005285 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005287 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 "test/argmax-pooling-operator-tester.h",
5289 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291)
5292
5293xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005294 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005296 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 "test/average-pooling-operator-tester.h",
5298 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005299 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300)
5301
5302xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005303 name = "bankers_rounding_nc_test",
5304 srcs = [
5305 "test/bankers-rounding-nc.cc",
5306 "test/bankers-rounding-operator-tester.h",
5307 ],
5308 deps = OPERATOR_TEST_DEPS,
5309)
5310
5311xnnpack_unit_test(
5312 name = "ceiling_nc_test",
5313 srcs = [
5314 "test/ceiling-nc.cc",
5315 "test/ceiling-operator-tester.h",
5316 ],
5317 deps = OPERATOR_TEST_DEPS,
5318)
5319
5320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005321 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005323 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005324 "test/channel-shuffle-operator-tester.h",
5325 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005326 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327)
5328
5329xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005330 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005332 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 "test/clamp-operator-tester.h",
5334 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005335 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336)
5337
5338xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005339 name = "constant_pad_nd_test",
5340 srcs = [
5341 "test/constant-pad-nd.cc",
5342 "test/constant-pad-operator-tester.h",
5343 ],
5344 deps = OPERATOR_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005348 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005350 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005351 "test/convolution-operator-tester.h",
5352 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005353 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354)
5355
5356xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005357 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005358 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005359 "test/convolution-nchw.cc",
5360 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005362 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363)
5364
5365xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005366 name = "copy_nc_test",
5367 srcs = [
5368 "test/copy-nc.cc",
5369 "test/copy-operator-tester.h",
5370 ],
5371 deps = OPERATOR_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005375 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005377 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005378 "test/deconvolution-operator-tester.h",
5379 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381)
5382
5383xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005384 name = "divide_nd_test",
5385 srcs = [
5386 "test/binary-elementwise-operator-tester.h",
5387 "test/divide-nd.cc",
5388 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005389 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005390)
5391
5392xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005393 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005395 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396 "test/fully-connected-operator-tester.h",
5397 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005398 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399)
5400
5401xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005402 name = "floor_nc_test",
5403 srcs = [
5404 "test/floor-nc.cc",
5405 "test/floor-operator-tester.h",
5406 ],
5407 deps = OPERATOR_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005411 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005413 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005415 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005420 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005421 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005422 "test/global-average-pooling-ncw.cc",
5423 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005424 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005425 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005426)
5427
5428xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005429 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005431 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005432 "test/hardswish-operator-tester.h",
5433 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005434 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435)
5436
5437xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005438 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005440 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441 "test/leaky-relu-operator-tester.h",
5442 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005443 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005444)
5445
5446xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005447 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005448 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005449 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005450 "test/max-pooling-operator-tester.h",
5451 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005452 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005453)
5454
5455xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005456 name = "maximum_nd_test",
5457 srcs = [
5458 "test/binary-elementwise-operator-tester.h",
5459 "test/maximum-nd.cc",
5460 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005462)
5463
5464xnnpack_unit_test(
5465 name = "minimum_nd_test",
5466 srcs = [
5467 "test/binary-elementwise-operator-tester.h",
5468 "test/minimum-nd.cc",
5469 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005471)
5472
5473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005474 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005475 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005476 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005477 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005479 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005480)
5481
5482xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005483 name = "negate_nc_test",
5484 srcs = [
5485 "test/negate-nc.cc",
5486 "test/negate-operator-tester.h",
5487 ],
5488 deps = OPERATOR_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005492 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005493 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005494 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 "test/prelu-operator-tester.h",
5496 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005497 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498)
5499
5500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005501 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005503 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005504 "test/resize-bilinear-operator-tester.h",
5505 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005507)
5508
5509xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005510 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005511 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005512 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005513 "test/sigmoid-operator-tester.h",
5514 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005515 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005516)
5517
5518xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005519 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005520 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005521 "test/softmax-nc.cc",
5522 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005523 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005524 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005525)
5526
5527xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005528 name = "square_nc_test",
5529 srcs = [
5530 "test/square-nc.cc",
5531 "test/square-operator-tester.h",
5532 ],
5533 deps = OPERATOR_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005537 name = "square_root_nc_test",
5538 srcs = [
5539 "test/square-root-nc.cc",
5540 "test/square-root-operator-tester.h",
5541 ],
5542 deps = OPERATOR_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005546 name = "squared_difference_nd_test",
5547 srcs = [
5548 "test/binary-elementwise-operator-tester.h",
5549 "test/squared-difference-nd.cc",
5550 ],
5551 deps = OPERATOR_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005555 name = "subtract_nd_test",
5556 srcs = [
5557 "test/binary-elementwise-operator-tester.h",
5558 "test/subtract-nd.cc",
5559 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005560 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005561)
5562
5563xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005564 name = "truncation_nc_test",
5565 srcs = [
5566 "test/truncation-nc.cc",
5567 "test/truncation-operator-tester.h",
5568 ],
5569 deps = OPERATOR_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005573 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005574 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005575 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576 "test/unpooling-operator-tester.h",
5577 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005578 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005579)
5580
Chao Mei6ddfc602020-05-13 22:29:36 -07005581############################### Misc unit tests ###############################
5582
5583xnnpack_unit_test(
5584 name = "memory_planner_test",
5585 srcs = [
5586 "test/memory-planner-test.cc",
5587 ],
5588 deps = [
5589 ":XNNPACK",
5590 ":memory_planner",
5591 ],
5592)
5593
Marat Dukhan08c4a432019-10-03 09:29:21 -07005594############################# Build configurations #############################
5595
Marat Dukhanb8642352019-10-30 15:43:02 -07005596# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005598 name = "xnn_enable_assembly_explicit_true",
5599 define_values = {"xnn_enable_assembly": "true"},
5600)
5601
5602# Disables usage of assembly kernels.
5603config_setting(
5604 name = "xnn_enable_assembly_explicit_false",
5605 define_values = {"xnn_enable_assembly": "false"},
5606)
5607
Marat Dukhan9de90e02020-06-18 16:04:12 -07005608# Enables usage of sparse inference.
5609config_setting(
5610 name = "xnn_enable_sparse_explicit_true",
5611 define_values = {"xnn_enable_sparse": "true"},
5612)
5613
5614# Disables usage of sparse inference.
5615config_setting(
5616 name = "xnn_enable_sparse_explicit_false",
5617 define_values = {"xnn_enable_sparse": "false"},
5618)
5619
Marat Dukhan05702cf2020-03-26 15:41:33 -07005620# Disables usage of HMP-aware optimizations.
5621config_setting(
5622 name = "xnn_enable_hmp_explicit_false",
5623 define_values = {"xnn_enable_hmp": "false"},
5624)
5625
Chao Mei6ddfc602020-05-13 22:29:36 -07005626# Enable usage of optimized memory allocation
5627config_setting(
5628 name = "xnn_enable_memopt_explicit_true",
5629 define_values = {"xnnpack_enable_memopt": "true"},
5630)
5631
5632# Disable usage of optimized memory allocation
5633config_setting(
5634 name = "xnn_enable_memopt_explicit_false",
5635 define_values = {"xnnpack_enable_memopt": "false"},
5636)
5637
Marat Dukhanb8642352019-10-30 15:43:02 -07005638# Builds with -c dbg
5639config_setting(
5640 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005642 "compilation_mode": "dbg",
5643 },
5644)
5645
5646# Builds with -c opt
5647config_setting(
5648 name = "optimized_build",
5649 values = {
5650 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005651 },
5652)
5653
5654config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005655 name = "linux_k8",
5656 values = {"cpu": "k8"},
5657)
5658
5659config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005660 name = "linux_arm",
5661 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005662)
5663
5664config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005665 name = "linux_armeabi",
5666 values = {"cpu": "armeabi"},
5667)
5668
5669config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005670 name = "linux_armhf",
5671 values = {"cpu": "armhf"},
5672)
5673
5674config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005675 name = "linux_armv7a",
5676 values = {"cpu": "armv7a"},
5677)
5678
5679config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005680 name = "linux_aarch64",
5681 values = {"cpu": "aarch64"},
5682)
5683
5684config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005685 name = "android",
5686 values = {"crosstool_top": "//external:android/crosstool"},
5687)
5688
5689config_setting(
5690 name = "android_armv7",
5691 values = {
5692 "crosstool_top": "//external:android/crosstool",
5693 "cpu": "armeabi-v7a",
5694 },
5695)
5696
5697config_setting(
5698 name = "android_arm64",
5699 values = {
5700 "crosstool_top": "//external:android/crosstool",
5701 "cpu": "arm64-v8a",
5702 },
5703)
5704
5705config_setting(
5706 name = "android_x86",
5707 values = {
5708 "crosstool_top": "//external:android/crosstool",
5709 "cpu": "x86",
5710 },
5711)
5712
5713config_setting(
5714 name = "android_x86_64",
5715 values = {
5716 "crosstool_top": "//external:android/crosstool",
5717 "cpu": "x86_64",
5718 },
5719)
5720
5721config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005722 name = "windows_x86_64",
5723 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005724)
5725
5726config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005727 name = "windows_x86_64_clang",
5728 values = {
5729 "compiler": "clang-cl",
5730 "cpu": "x64_windows",
5731 },
5732)
5733
5734config_setting(
5735 name = "windows_x86_64_mingw",
5736 values = {
5737 "compiler": "mingw-gcc",
5738 "cpu": "x64_windows",
5739 },
5740)
5741
5742config_setting(
5743 name = "windows_x86_64_msys",
5744 values = {
5745 "compiler": "msys-gcc",
5746 "cpu": "x64_windows",
5747 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005748)
5749
5750config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005751 name = "macos_x86_64",
5752 values = {
5753 "apple_platform_type": "macos",
5754 "cpu": "darwin",
5755 },
5756)
5757
5758config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005759 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005760 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005761)
5762
5763config_setting(
5764 name = "emscripten_wasm",
5765 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005766 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005767 "cpu": "wasm",
5768 },
5769)
5770
5771config_setting(
5772 name = "emscripten_wasmsimd",
5773 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005774 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005775 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005776 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005777 },
5778)
5779
5780config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005781 name = "ios_armv7",
5782 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005783 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005784 "cpu": "ios_armv7",
5785 },
5786)
5787
5788config_setting(
5789 name = "ios_arm64",
5790 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005791 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005792 "cpu": "ios_arm64",
5793 },
5794)
5795
5796config_setting(
5797 name = "ios_arm64e",
5798 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005799 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005800 "cpu": "ios_arm64e",
5801 },
5802)
5803
5804config_setting(
5805 name = "ios_x86",
5806 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005807 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005808 "cpu": "ios_i386",
5809 },
5810)
5811
5812config_setting(
5813 name = "ios_x86_64",
5814 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005815 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005816 "cpu": "ios_x86_64",
5817 },
5818)
5819
5820config_setting(
5821 name = "watchos_armv7k",
5822 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005823 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005824 "cpu": "watchos_armv7k",
5825 },
5826)
5827
5828config_setting(
5829 name = "watchos_arm64_32",
5830 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005831 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005832 "cpu": "watchos_arm64_32",
5833 },
5834)
5835
5836config_setting(
5837 name = "watchos_x86",
5838 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005839 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005840 "cpu": "watchos_i386",
5841 },
5842)
5843
5844config_setting(
5845 name = "watchos_x86_64",
5846 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005847 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005848 "cpu": "watchos_x86_64",
5849 },
5850)
5851
5852config_setting(
5853 name = "tvos_arm64",
5854 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005855 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005856 "cpu": "tvos_arm64",
5857 },
5858)
5859
5860config_setting(
5861 name = "tvos_x86_64",
5862 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005863 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005864 "cpu": "tvos_x86_64",
5865 },
5866)