blob: 7f5a2428696cd0dc543efc6a91822adcdb9b29be [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 Dukhane7edc802020-08-03 16:43:57 -07001649 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001650 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1651 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001652 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001653 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1654 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001655 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1656 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001657 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1658 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001659 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001660 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1661 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001662 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001663 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1664 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1665 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1666 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1667 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1668 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1669 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1670 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001671 "src/qu8-requantization/precise-sse2.c",
1672 "src/qu8-requantization/fp32-sse2.c",
1673 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001674 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001675 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001676 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001678 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679 "src/x32-zip/x2-sse2.c",
1680 "src/x32-zip/x3-sse2.c",
1681 "src/x32-zip/x4-sse2.c",
1682 "src/x32-zip/xm-sse2.c",
1683 "src/x8-zip/x2-sse2.c",
1684 "src/x8-zip/x3-sse2.c",
1685 "src/x8-zip/x4-sse2.c",
1686 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001687 "src/math/exp-sse2-p5.c",
1688 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001689 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001690 "src/math/roundd-sse2-cvt.c",
1691 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001692 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001693 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001694]
1695
1696SSSE3_UKERNELS = [
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001697 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1698 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001699 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1700 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001701 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1702 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001703 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001704 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1705 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001706 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001707 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1708 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001709 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1710 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001711 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1712 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001713 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001714 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1715 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001716 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001717 "src/qs8-requantization/precise-ssse3.c",
1718 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001719 "src/qu8-requantization/precise-ssse3.c",
1720 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721]
1722
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001723SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001724 "src/f32-prelu/gen/sse41-2x4.c",
1725 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001726 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001727 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001728 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001729 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001730 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1731 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001732 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1733 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001734 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1735 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1736 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1737 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1738 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1739 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1740 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1741 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001742 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1743 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001744 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1745 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001746 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1747 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001748 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001749 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1750 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001751 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001752 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1753 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001754 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1755 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001756 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1757 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001758 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001759 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1760 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001761 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001762 "src/qs8-requantization/fp32-sse4.c",
1763 "src/qs8-requantization/precise-sse4.c",
1764 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001765 "src/qu8-requantization/precise-sse4.c",
1766 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001767 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001768 "src/math/roundd-sse41.c",
1769 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001770 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001771]
1772
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001774 "src/f32-clamp/gen/avx-x8.c",
1775 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001776 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1777 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1778 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1779 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1780 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1781 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1782 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1783 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1784 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1785 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1786 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1787 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1788 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1789 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1790 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1791 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1794 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1795 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1796 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1797 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1798 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1799 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1800 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1802 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1803 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1804 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1805 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001806 "src/f32-hswish/gen/hswish-avx-x8.c",
1807 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001808 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1809 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1810 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1811 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1812 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1813 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1814 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1815 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1816 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001817 "src/f32-prelu/gen/avx-2x8.c",
1818 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001819 "src/f32-relu/gen/avx-x16.c",
1820 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001821 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001822 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1823 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1824 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1825 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1826 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1827 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1828 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1829 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001830 "src/f32-vbinary/gen/vmax-avx-x8.c",
1831 "src/f32-vbinary/gen/vmax-avx-x16.c",
1832 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1833 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1834 "src/f32-vbinary/gen/vmin-avx-x8.c",
1835 "src/f32-vbinary/gen/vmin-avx-x16.c",
1836 "src/f32-vbinary/gen/vminc-avx-x8.c",
1837 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1842 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1843 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1844 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1845 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001846 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1847 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1848 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1849 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001850 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1851 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1852 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1853 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001854 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1855 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001856 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001857 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1858 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1859 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1860 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1861 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1862 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1863 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1864 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001865 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1866 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001867 "src/f32-vunary/gen/vabs-avx-x8.c",
1868 "src/f32-vunary/gen/vabs-avx-x16.c",
1869 "src/f32-vunary/gen/vneg-avx-x8.c",
1870 "src/f32-vunary/gen/vneg-avx-x16.c",
1871 "src/f32-vunary/gen/vsqr-avx-x8.c",
1872 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873]
1874
Marat Dukhan1566fee2020-08-02 21:55:41 -07001875XOP_UKERNELS = [
1876 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1877 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1878 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1879 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
1880 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1881 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001882 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001883 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1884 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001885 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001886 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1887 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1888 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1889 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1890 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1891 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001892 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001893 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1894 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001895 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001896]
1897
Marat Dukhanfda12b82019-11-21 12:27:59 -08001898FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001899 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1900 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1901 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1902 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1903 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1904 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1905 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1906 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1907 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1908 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1909 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1910 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1911 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1912 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1913 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1914 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1915 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1916 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1917 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1918 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1919 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1920 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1921 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1922 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1923 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1924 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1925 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1926 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1927 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1928 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1929 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1930 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1931 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1932 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1933 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1934 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1935 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1936 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1937 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1938 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001939 "src/f32-hswish/gen/hswish-fma3-x8.c",
1940 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001941 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1942 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1943 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1944 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1945 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1946 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1947 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1948 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1949 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1950 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1951 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1952 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1953 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1954 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001955 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1956 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1957 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1958 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1959 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1960 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1961 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1962 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001963 "src/math/sqrt-fma3-nr1fma.c",
1964 "src/math/sqrt-fma3-nr2fma.c",
1965 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001966]
1967
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001968AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001969 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1970 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1971 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1972 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1973 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1974 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1975 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1976 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1977 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1978 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1979 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1980 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1981 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1982 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1983 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1984 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1985 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1986 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1987 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1988 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1989 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1990 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1991 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1992 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1993 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1994 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1995 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1996 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1997 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1998 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1999 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2000 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2001 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2002 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2003 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2004 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002005 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2006 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2007 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2008 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2009 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2010 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2011 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2012 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2013 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2014 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2015 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2016 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2017 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2018 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2019 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2020 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2021 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2022 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2023 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2024 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2025 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2026 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2027 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2028 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2029 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2030 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2031 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2032 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2033 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2034 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002035 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2036 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2037 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2038 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2039 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2040 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2041 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2042 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2043 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2044 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2045 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2046 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2047 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2048 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2049 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2050 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2051 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2052 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2053 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2054 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2055 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2056 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2057 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2058 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002059 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2060 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2061 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
2062 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2063 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2064 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002065 "src/math/exp-avx2-p5.c",
2066 "src/math/exp-avx2-perm-p3.c",
2067 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002068 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002069 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002070 "src/math/sigmoid-avx2-rr2-p5-div.c",
2071 "src/math/sigmoid-avx2-rr1-p5-div.c",
2072 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2073 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2074 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2075 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002076]
2077
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002079 "src/f32-clamp/gen/avx512f-x16.c",
2080 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002081 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2082 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2083 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2084 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2085 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2086 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2087 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2088 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2089 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2090 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2091 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2092 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2093 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2094 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2095 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2096 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2097 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2098 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2099 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2100 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2101 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2102 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2103 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2104 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002105 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2106 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002107 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2108 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2109 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2110 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2111 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2112 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002113 "src/f32-prelu/gen/avx512f-2x16.c",
2114 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002115 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2116 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2117 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2118 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2119 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2120 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2121 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2122 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2123 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2124 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2125 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2126 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2127 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2128 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2129 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2130 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2131 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2132 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2133 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2134 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2135 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2136 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2137 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2138 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2139 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2140 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2141 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2142 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2143 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2144 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2145 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2146 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2147 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2148 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2149 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2150 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002151 "src/f32-relu/gen/avx512f-x16.c",
2152 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002153 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002154 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2155 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2156 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2157 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2158 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2159 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2160 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2161 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002162 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2163 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2164 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2165 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2166 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2167 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2168 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2169 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002170 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2171 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2172 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2173 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2174 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2175 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2176 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2177 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002178 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2179 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2180 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2181 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002182 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2183 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2184 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2185 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002186 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2187 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002188 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002189 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2190 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2191 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2192 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2193 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2194 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2195 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2196 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2197 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2198 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2199 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2200 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2201 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2202 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2203 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2204 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2205 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2206 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2207 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2208 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2209 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2210 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2211 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2212 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002213 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2214 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2215 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2216 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2217 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2218 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2219 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2220 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002221 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2222 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2223 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2224 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2225 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2226 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2227 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2228 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002229 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2230 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2231 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2232 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2233 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2234 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002235 "src/math/exp-avx512f-p5-scalef.c",
2236 "src/math/exp-avx512f-p5.c",
2237 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002238 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002239 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002240 "src/math/sqrt-avx512f-nr1fma.c",
2241 "src/math/sqrt-avx512f-nr2fma.c",
2242 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243]
2244
2245AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002246 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002247 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002248 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2249 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002250 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002251 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002252 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002253 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002254 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2255 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002256 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2257 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2258 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2259 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002260]
2261
2262AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002263 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002264 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2265 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002266 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002267 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2268 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002269 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2270 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2271 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002272 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002273 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2274 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2275 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002276 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002277 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2278 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002279 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002280 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2281 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2282 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2283 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2284 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2285 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2286 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2287 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2288 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2289 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2290 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2291 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2292 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2293 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2294 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2295 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2296 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2297 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002298 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2299 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002300 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002301 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2302 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2303 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2304 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2305 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2306 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2307 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2308 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2309 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2310 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2311 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2312 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2313 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2314 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2315 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2316 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2317 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2318 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002319 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2320 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2321 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2322 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2323 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2324 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2325 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2326 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2327 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2328 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2329 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2330 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2331 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2332 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2333 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2334 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2335 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2336 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337]
2338
Marat Dukhan1b354632020-03-23 12:50:22 -07002339INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 "src/xnnpack/argmaxpool.h",
2341 "src/xnnpack/avgpool.h",
2342 "src/xnnpack/clamp.h",
2343 "src/xnnpack/common.h",
2344 "src/xnnpack/conv.h",
2345 "src/xnnpack/dwconv.h",
2346 "src/xnnpack/gavgpool.h",
2347 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002348 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002350 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002352 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353 "src/xnnpack/lut.h",
2354 "src/xnnpack/math.h",
2355 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002356 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357 "src/xnnpack/packx.h",
2358 "src/xnnpack/pad.h",
2359 "src/xnnpack/params.h",
2360 "src/xnnpack/pavgpool.h",
2361 "src/xnnpack/ppmm.h",
2362 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002363 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002364 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002365 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002366 "src/xnnpack/rmax.h",
2367 "src/xnnpack/scalar-utils.h",
2368 "src/xnnpack/spmm.h",
2369 "src/xnnpack/unpool.h",
2370 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002371 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002373 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002374 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002375 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002376 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002378]
2379
2380INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 "include/xnnpack.h",
2382 "src/xnnpack/allocator.h",
2383 "src/xnnpack/compute.h",
2384 "src/xnnpack/im2col.h",
2385 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002386 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002387 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388 "src/xnnpack/operator.h",
2389 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002390 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002392 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002393 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002394]
2395
Marat Dukhan1b354632020-03-23 12:50:22 -07002396ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002397 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398]
2399
Marat Dukhan1b354632020-03-23 12:50:22 -07002400MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002401 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402 "include/xnnpack.h",
2403]
2404
Marat Dukhan1b354632020-03-23 12:50:22 -07002405MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002406 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002407 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002408 "src/xnnpack/requantization.h",
2409 "include/xnnpack.h",
2410]
2411
2412OPERATOR_TEST_PARAMS_HDRS = [
2413 "src/xnnpack/params.h",
2414 "src/xnnpack/common.h",
2415]
2416
2417WEIGHTS_PACK_HDRS = [
2418 "src/xnnpack/pack.h",
2419 "src/xnnpack/operator.h",
2420 "src/xnnpack/compute.h",
2421]
2422
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002423LOGGING_COPTS = select({
2424 # No logging in optimized mode
2425 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2426 # Full logging in debug mode
2427 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2428 # Error-only logging in default (fastbuild) mode
2429 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2430})
2431
Marat Dukhan3b59de22020-06-03 20:15:19 -07002432LOGGING_SRCS = select({
2433 # No logging in optimized mode
2434 ":optimized_build": [],
2435 "//conditions:default": [
2436 "src/operator-strings.c",
2437 "src/subgraph-strings.c",
2438 ],
2439})
2440
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002441LOGGING_HDRS = [
2442 "src/xnnpack/log.h",
2443]
2444
Marat Dukhan08c4a432019-10-03 09:29:21 -07002445xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002446 name = "tables",
2447 srcs = TABLE_SRCS,
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 Dukhan3a77ea72019-12-23 12:10:24 -08002451)
2452
2453xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 name = "scalar_ukernels",
2455 srcs = SCALAR_UKERNELS,
2456 hdrs = INTERNAL_HDRS,
2457 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002458 gcc_copts = xnnpack_gcc_std_copts(),
2459 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002460 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002461 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002462 "@FP16",
2463 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002464 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465 ],
2466)
2467
2468xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002469 name = "scalar_ukernels_test_mode",
2470 srcs = SCALAR_UKERNELS,
2471 hdrs = INTERNAL_HDRS,
2472 aarch32_copts = ["-marm"],
2473 copts = [
2474 "-UNDEBUG",
2475 "-DXNN_TEST_MODE=1",
2476 ],
2477 gcc_copts = xnnpack_gcc_std_copts(),
2478 msvc_copts = xnnpack_msvc_std_copts(),
2479 deps = [
2480 ":tables",
2481 "@FP16",
2482 "@FXdiv",
2483 "@pthreadpool",
2484 ],
2485)
2486
2487xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002488 name = "wasm_ukernels",
2489 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002490 gcc_copts = xnnpack_gcc_std_copts(),
2491 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002492 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002493 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002494 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002495 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002496 "@FP16",
2497 "@FXdiv",
2498 "@pthreadpool",
2499 ],
2500)
2501
2502xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002503 name = "wasm_ukernels_test_mode",
2504 hdrs = INTERNAL_HDRS,
2505 copts = [
2506 "-UNDEBUG",
2507 "-DXNN_TEST_MODE=1",
2508 ],
2509 gcc_copts = xnnpack_gcc_std_copts(),
2510 msvc_copts = xnnpack_msvc_std_copts(),
2511 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002512 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002513 deps = [
2514 ":tables",
2515 "@FP16",
2516 "@FXdiv",
2517 "@pthreadpool",
2518 ],
2519)
2520
2521xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002522 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002523 hdrs = INTERNAL_HDRS,
2524 aarch32_copts = [
2525 "-marm",
2526 "-mfpu=neon",
2527 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002528 gcc_copts = xnnpack_gcc_std_copts(),
2529 gcc_x86_copts = ["-msse2"],
2530 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531 optimized_copts = [
2532 "-O3",
2533 "-ffast-math",
2534 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002535 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002537 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538 "@FP16",
2539 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002540 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002541 ],
2542)
2543
2544xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002545 name = "psimd_fastmath_ukernels_test_mode",
2546 hdrs = INTERNAL_HDRS,
2547 aarch32_copts = [
2548 "-marm",
2549 "-mfpu=neon",
2550 ],
2551 copts = [
2552 "-UNDEBUG",
2553 "-DXNN_TEST_MODE=1",
2554 ],
2555 gcc_copts = xnnpack_gcc_std_copts(),
2556 gcc_x86_copts = ["-msse2"],
2557 msvc_copts = xnnpack_msvc_std_copts(),
2558 optimized_copts = [
2559 "-O3",
2560 "-ffast-math",
2561 ],
2562 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2563 deps = [
2564 ":tables",
2565 "@FP16",
2566 "@psimd",
2567 "@pthreadpool",
2568 ],
2569)
2570
2571xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002572 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002573 hdrs = INTERNAL_HDRS,
2574 aarch32_copts = [
2575 "-marm",
2576 "-mfpu=neon",
2577 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002578 gcc_copts = xnnpack_gcc_std_copts(),
2579 gcc_x86_copts = ["-msse2"],
2580 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002581 optimized_copts = [
2582 "-O3",
2583 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002584 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002585 deps = [
2586 ":tables",
2587 "@FP16",
2588 "@psimd",
2589 "@pthreadpool",
2590 ],
2591)
2592
2593xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002594 name = "psimd_accmath_ukernels_test_mode",
2595 hdrs = INTERNAL_HDRS,
2596 aarch32_copts = [
2597 "-marm",
2598 "-mfpu=neon",
2599 ],
2600 copts = [
2601 "-UNDEBUG",
2602 "-DXNN_TEST_MODE=1",
2603 ],
2604 gcc_copts = xnnpack_gcc_std_copts(),
2605 gcc_x86_copts = ["-msse2"],
2606 msvc_copts = xnnpack_msvc_std_copts(),
2607 optimized_copts = [
2608 "-O3",
2609 ],
2610 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2611 deps = [
2612 ":tables",
2613 "@FP16",
2614 "@psimd",
2615 "@pthreadpool",
2616 ],
2617)
2618
2619xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620 name = "neon_ukernels",
2621 hdrs = INTERNAL_HDRS,
2622 aarch32_copts = [
2623 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002624 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002625 "-mfpu=neon",
2626 ],
2627 aarch32_srcs = NEON_UKERNELS,
2628 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002629 gcc_copts = xnnpack_gcc_std_copts(),
2630 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002631 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002632 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002633 "@FP16",
2634 "@pthreadpool",
2635 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002636)
2637
2638xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002639 name = "neon_ukernels_test_mode",
2640 hdrs = INTERNAL_HDRS,
2641 aarch32_copts = [
2642 "-marm",
2643 "-march=armv7-a",
2644 "-mfpu=neon",
2645 ],
2646 aarch32_srcs = NEON_UKERNELS,
2647 aarch64_srcs = NEON_UKERNELS,
2648 copts = [
2649 "-UNDEBUG",
2650 "-DXNN_TEST_MODE=1",
2651 ],
2652 gcc_copts = xnnpack_gcc_std_copts(),
2653 msvc_copts = xnnpack_msvc_std_copts(),
2654 deps = [
2655 ":tables",
2656 "@FP16",
2657 "@pthreadpool",
2658 ],
2659)
2660
2661xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 name = "neonfma_ukernels",
2663 hdrs = INTERNAL_HDRS,
2664 aarch32_copts = [
2665 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002666 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002667 "-mfpu=neon-vfpv4",
2668 ],
2669 aarch32_srcs = NEONFMA_UKERNELS,
2670 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002671 apple_aarch32_copts = [
2672 "-mcpu=swift",
2673 "-mtune=generic",
2674 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002675 gcc_copts = xnnpack_gcc_std_copts(),
2676 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002677 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002678 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002679 "@FP16",
2680 "@pthreadpool",
2681 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002682)
2683
2684xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002685 name = "neonfma_ukernels_test_mode",
2686 hdrs = INTERNAL_HDRS,
2687 aarch32_copts = [
2688 "-marm",
2689 "-march=armv7-a",
2690 "-mfpu=neon-vfpv4",
2691 ],
2692 aarch32_srcs = NEONFMA_UKERNELS,
2693 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002694 apple_aarch32_copts = [
2695 "-mcpu=swift",
2696 "-mtune=generic",
2697 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002698 copts = [
2699 "-UNDEBUG",
2700 "-DXNN_TEST_MODE=1",
2701 ],
2702 gcc_copts = xnnpack_gcc_std_copts(),
2703 msvc_copts = xnnpack_msvc_std_copts(),
2704 deps = [
2705 ":tables",
2706 "@FP16",
2707 "@pthreadpool",
2708 ],
2709)
2710
2711xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002712 name = "neonv8_ukernels",
2713 hdrs = INTERNAL_HDRS,
2714 aarch32_copts = [
2715 "-marm",
2716 "-march=armv8-a",
2717 "-mfpu=neon-fp-armv8",
2718 ],
2719 aarch32_srcs = NEONV8_UKERNELS,
2720 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002721 apple_aarch32_copts = [
2722 "-mcpu=cyclone",
2723 "-mtune=generic",
2724 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002725 gcc_copts = xnnpack_gcc_std_copts(),
2726 msvc_copts = xnnpack_msvc_std_copts(),
2727 deps = [
2728 ":tables",
2729 "@FP16",
2730 "@pthreadpool",
2731 ],
2732)
2733
2734xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002735 name = "neonv8_ukernels_test_mode",
2736 hdrs = INTERNAL_HDRS,
2737 aarch32_copts = [
2738 "-marm",
2739 "-march=armv8-a",
2740 "-mfpu=neon-fp-armv8",
2741 ],
2742 aarch32_srcs = NEONV8_UKERNELS,
2743 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002744 apple_aarch32_copts = [
2745 "-mcpu=cyclone",
2746 "-mtune=generic",
2747 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002748 copts = [
2749 "-UNDEBUG",
2750 "-DXNN_TEST_MODE=1",
2751 ],
2752 gcc_copts = xnnpack_gcc_std_copts(),
2753 msvc_copts = xnnpack_msvc_std_copts(),
2754 deps = [
2755 ":tables",
2756 "@FP16",
2757 "@pthreadpool",
2758 ],
2759)
2760
2761xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762 name = "neonfp16arith_ukernels",
2763 hdrs = INTERNAL_HDRS,
2764 aarch64_copts = ["-march=armv8.2-a+fp16"],
2765 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002766 gcc_copts = xnnpack_gcc_std_copts(),
2767 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002768 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002769 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002770 "@FP16",
2771 "@pthreadpool",
2772 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002773)
2774
2775xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002776 name = "neonfp16arith_ukernels_test_mode",
2777 hdrs = INTERNAL_HDRS,
2778 aarch64_copts = ["-march=armv8.2-a+fp16"],
2779 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2780 copts = [
2781 "-UNDEBUG",
2782 "-DXNN_TEST_MODE=1",
2783 ],
2784 gcc_copts = xnnpack_gcc_std_copts(),
2785 msvc_copts = xnnpack_msvc_std_copts(),
2786 deps = [
2787 ":tables",
2788 "@FP16",
2789 "@pthreadpool",
2790 ],
2791)
2792
2793xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002794 name = "sse2_ukernels",
2795 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002796 gcc_copts = xnnpack_gcc_std_copts(),
2797 gcc_x86_copts = ["-msse2"],
2798 msvc_copts = xnnpack_msvc_std_copts(),
2799 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002801 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002802 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002803 "@FP16",
2804 "@pthreadpool",
2805 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806)
2807
2808xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002809 name = "sse2_ukernels_test_mode",
2810 hdrs = INTERNAL_HDRS,
2811 copts = [
2812 "-UNDEBUG",
2813 "-DXNN_TEST_MODE=1",
2814 ],
2815 gcc_copts = xnnpack_gcc_std_copts(),
2816 gcc_x86_copts = ["-msse2"],
2817 msvc_copts = xnnpack_msvc_std_copts(),
2818 msvc_x86_32_copts = ["/arch:SSE2"],
2819 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2820 deps = [
2821 ":tables",
2822 "@FP16",
2823 "@pthreadpool",
2824 ],
2825)
2826
2827xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002828 name = "ssse3_ukernels",
2829 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002830 gcc_copts = xnnpack_gcc_std_copts(),
2831 gcc_x86_copts = ["-mssse3"],
2832 msvc_copts = xnnpack_msvc_std_copts(),
2833 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002834 x86_srcs = SSSE3_UKERNELS,
2835 deps = [
2836 ":tables",
2837 "@FP16",
2838 "@pthreadpool",
2839 ],
2840)
2841
2842xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002843 name = "ssse3_ukernels_test_mode",
2844 hdrs = INTERNAL_HDRS,
2845 copts = [
2846 "-UNDEBUG",
2847 "-DXNN_TEST_MODE=1",
2848 ],
2849 gcc_copts = xnnpack_gcc_std_copts(),
2850 gcc_x86_copts = ["-mssse3"],
2851 msvc_copts = xnnpack_msvc_std_copts(),
2852 msvc_x86_32_copts = ["/arch:SSE2"],
2853 x86_srcs = SSSE3_UKERNELS,
2854 deps = [
2855 ":tables",
2856 "@FP16",
2857 "@pthreadpool",
2858 ],
2859)
2860
2861xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002862 name = "sse41_ukernels",
2863 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002864 gcc_copts = xnnpack_gcc_std_copts(),
2865 gcc_x86_copts = ["-msse4.1"],
2866 msvc_copts = xnnpack_msvc_std_copts(),
2867 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002868 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002869 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002870 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002871 "@FP16",
2872 "@pthreadpool",
2873 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002874)
2875
2876xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002877 name = "sse41_ukernels_test_mode",
2878 hdrs = INTERNAL_HDRS,
2879 copts = [
2880 "-UNDEBUG",
2881 "-DXNN_TEST_MODE=1",
2882 ],
2883 gcc_copts = xnnpack_gcc_std_copts(),
2884 gcc_x86_copts = ["-msse4.1"],
2885 msvc_copts = xnnpack_msvc_std_copts(),
2886 msvc_x86_32_copts = ["/arch:SSE2"],
2887 x86_srcs = SSE41_UKERNELS,
2888 deps = [
2889 ":tables",
2890 "@FP16",
2891 "@pthreadpool",
2892 ],
2893)
2894
2895xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002896 name = "avx_ukernels",
2897 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002898 gcc_copts = xnnpack_gcc_std_copts(),
2899 gcc_x86_copts = ["-mavx"],
2900 msvc_copts = xnnpack_msvc_std_copts(),
2901 msvc_x86_32_copts = ["/arch:AVX"],
2902 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002904 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002905 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002906 "@FP16",
2907 "@pthreadpool",
2908 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002909)
2910
2911xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002912 name = "avx_ukernels_test_mode",
2913 hdrs = INTERNAL_HDRS,
2914 copts = [
2915 "-UNDEBUG",
2916 "-DXNN_TEST_MODE=1",
2917 ],
2918 gcc_copts = xnnpack_gcc_std_copts(),
2919 gcc_x86_copts = ["-mavx"],
2920 msvc_copts = xnnpack_msvc_std_copts(),
2921 msvc_x86_32_copts = ["/arch:AVX"],
2922 msvc_x86_64_copts = ["/arch:AVX"],
2923 x86_srcs = AVX_UKERNELS,
2924 deps = [
2925 ":tables",
2926 "@FP16",
2927 "@pthreadpool",
2928 ],
2929)
2930
2931xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002932 name = "xop_ukernels",
2933 hdrs = INTERNAL_HDRS,
2934 gcc_copts = xnnpack_gcc_std_copts(),
2935 gcc_x86_copts = ["-mxop"],
2936 msvc_copts = xnnpack_msvc_std_copts(),
2937 msvc_x86_32_copts = ["/arch:AVX"],
2938 msvc_x86_64_copts = ["/arch:AVX"],
2939 x86_srcs = XOP_UKERNELS,
2940 deps = [
2941 ":tables",
2942 "@FP16",
2943 "@pthreadpool",
2944 ],
2945)
2946
2947xnnpack_cc_library(
2948 name = "xop_ukernels_test_mode",
2949 hdrs = INTERNAL_HDRS,
2950 copts = [
2951 "-UNDEBUG",
2952 "-DXNN_TEST_MODE=1",
2953 ],
2954 gcc_copts = xnnpack_gcc_std_copts(),
2955 gcc_x86_copts = ["-mxop"],
2956 msvc_copts = xnnpack_msvc_std_copts(),
2957 msvc_x86_32_copts = ["/arch:AVX"],
2958 msvc_x86_64_copts = ["/arch:AVX"],
2959 x86_srcs = XOP_UKERNELS,
2960 deps = [
2961 ":tables",
2962 "@FP16",
2963 "@pthreadpool",
2964 ],
2965)
2966
2967xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002968 name = "fma3_ukernels",
2969 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002970 gcc_copts = xnnpack_gcc_std_copts(),
2971 gcc_x86_copts = ["-mfma"],
2972 msvc_copts = xnnpack_msvc_std_copts(),
2973 msvc_x86_32_copts = ["/arch:AVX"],
2974 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002975 x86_srcs = FMA3_UKERNELS,
2976 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002977 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002978 "@FP16",
2979 "@pthreadpool",
2980 ],
2981)
2982
2983xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002984 name = "fma3_ukernels_test_mode",
2985 hdrs = INTERNAL_HDRS,
2986 copts = [
2987 "-UNDEBUG",
2988 "-DXNN_TEST_MODE=1",
2989 ],
2990 gcc_copts = xnnpack_gcc_std_copts(),
2991 gcc_x86_copts = ["-mfma"],
2992 msvc_copts = xnnpack_msvc_std_copts(),
2993 msvc_x86_32_copts = ["/arch:AVX"],
2994 msvc_x86_64_copts = ["/arch:AVX"],
2995 x86_srcs = FMA3_UKERNELS,
2996 deps = [
2997 ":tables",
2998 "@FP16",
2999 "@pthreadpool",
3000 ],
3001)
3002
3003xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003004 name = "avx2_ukernels",
3005 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003006 gcc_copts = xnnpack_gcc_std_copts(),
3007 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003008 "-mfma",
3009 "-mavx2",
3010 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003011 msvc_copts = xnnpack_msvc_std_copts(),
3012 msvc_x86_32_copts = ["/arch:AVX2"],
3013 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003014 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003015 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003016 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003017 "@FP16",
3018 "@pthreadpool",
3019 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003020)
3021
3022xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003023 name = "avx2_ukernels_test_mode",
3024 hdrs = INTERNAL_HDRS,
3025 copts = [
3026 "-UNDEBUG",
3027 "-DXNN_TEST_MODE=1",
3028 ],
3029 gcc_copts = xnnpack_gcc_std_copts(),
3030 gcc_x86_copts = [
3031 "-mfma",
3032 "-mavx2",
3033 ],
3034 msvc_copts = xnnpack_msvc_std_copts(),
3035 msvc_x86_32_copts = ["/arch:AVX2"],
3036 msvc_x86_64_copts = ["/arch:AVX2"],
3037 x86_srcs = AVX2_UKERNELS,
3038 deps = [
3039 ":tables",
3040 "@FP16",
3041 "@pthreadpool",
3042 ],
3043)
3044
3045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 name = "avx512f_ukernels",
3047 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003048 gcc_copts = xnnpack_gcc_std_copts(),
3049 gcc_x86_copts = ["-mavx512f"],
3050 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3051 msvc_copts = xnnpack_msvc_std_copts(),
3052 msvc_x86_32_copts = ["/arch:AVX512"],
3053 msvc_x86_64_copts = ["/arch:AVX512"],
3054 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003055 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003056 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003057 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003058 "@FP16",
3059 "@pthreadpool",
3060 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061)
3062
3063xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003064 name = "avx512f_ukernels_test_mode",
3065 hdrs = INTERNAL_HDRS,
3066 copts = [
3067 "-UNDEBUG",
3068 "-DXNN_TEST_MODE=1",
3069 ],
3070 gcc_copts = xnnpack_gcc_std_copts(),
3071 gcc_x86_copts = ["-mavx512f"],
3072 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3073 msvc_copts = xnnpack_msvc_std_copts(),
3074 msvc_x86_32_copts = ["/arch:AVX512"],
3075 msvc_x86_64_copts = ["/arch:AVX512"],
3076 msys_copts = ["-fno-asynchronous-unwind-tables"],
3077 x86_srcs = AVX512F_UKERNELS,
3078 deps = [
3079 ":tables",
3080 "@FP16",
3081 "@pthreadpool",
3082 ],
3083)
3084
3085xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 name = "asm_ukernels",
3087 hdrs = ["src/xnnpack/assembly.h"],
3088 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003089 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 aarch64_srcs = AARCH64_ASM_UKERNELS,
3091)
3092
Marat Dukhan3b59de22020-06-03 20:15:19 -07003093xnnpack_cc_library(
3094 name = "logging_utils",
3095 srcs = LOGGING_SRCS,
3096 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3097 copts = LOGGING_COPTS + [
3098 "-Isrc",
3099 "-Iinclude",
3100 ] + select({
3101 ":debug_build": [],
3102 "//conditions:default": xnnpack_min_size_copts(),
3103 }),
3104 gcc_copts = xnnpack_gcc_std_copts(),
3105 msvc_copts = xnnpack_msvc_std_copts(),
3106 visibility = xnnpack_visibility(),
3107 deps = [
3108 "@FP16",
3109 "@clog",
3110 "@pthreadpool",
3111 ],
3112)
3113
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114xnnpack_aggregate_library(
3115 name = "ukernels",
3116 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117 ":neon_ukernels",
3118 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003119 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120 ":asm_ukernels",
3121 ],
3122 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123 ":neon_ukernels",
3124 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003125 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 ":neonfp16arith_ukernels",
3127 ":asm_ukernels",
3128 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003129 generic_deps = [
3130 ":scalar_ukernels",
3131 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003132 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003133 ":psimd_fastmath_ukernels",
3134 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003136 wasm_deps = [
3137 ":wasm_ukernels",
3138 ],
3139 wasmsimd_deps = [
3140 ":wasm_ukernels",
3141 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003143 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003144 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003145 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003147 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003148 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003149 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 ":avx512f_ukernels",
3151 ],
3152)
3153
Marat Dukhan33fcf782020-05-24 14:27:15 -07003154xnnpack_aggregate_library(
3155 name = "ukernels_test_mode",
3156 aarch32_deps = [
3157 ":neon_ukernels_test_mode",
3158 ":neonfma_ukernels_test_mode",
3159 ":neonv8_ukernels_test_mode",
3160 ":asm_ukernels",
3161 ],
3162 aarch64_deps = [
3163 ":neon_ukernels_test_mode",
3164 ":neonfma_ukernels_test_mode",
3165 ":neonv8_ukernels_test_mode",
3166 ":neonfp16arith_ukernels_test_mode",
3167 ":asm_ukernels",
3168 ],
3169 generic_deps = [
3170 ":scalar_ukernels_test_mode",
3171 ],
3172 psimd_deps = [
3173 ":psimd_fastmath_ukernels_test_mode",
3174 ":psimd_accmath_ukernels_test_mode",
3175 ],
3176 wasm_deps = [
3177 ":wasm_ukernels_test_mode",
3178 ],
3179 wasmsimd_deps = [
3180 ":wasm_ukernels_test_mode",
3181 ],
3182 x86_deps = [
3183 ":sse2_ukernels_test_mode",
3184 ":ssse3_ukernels_test_mode",
3185 ":sse41_ukernels_test_mode",
3186 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003187 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003188 ":fma3_ukernels_test_mode",
3189 ":avx2_ukernels_test_mode",
3190 ":avx512f_ukernels_test_mode",
3191 ],
3192)
3193
Marat Dukhan86df3842020-06-17 14:03:31 -07003194# This target must be compatible with Android NDK r17 and compile with Clang 6.
3195# It is similar to :ukernels target, but excludes microkernels which require
3196# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003197xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003198 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003199 aarch32_deps = [
3200 ":neon_ukernels",
3201 ":neonfma_ukernels",
3202 ":neonv8_ukernels",
3203 ":asm_ukernels",
3204 ],
3205 aarch64_deps = [
3206 ":neon_ukernels",
3207 ":neonfma_ukernels",
3208 ":neonv8_ukernels",
3209 ":asm_ukernels",
3210 ],
3211 generic_deps = [
3212 ":scalar_ukernels",
3213 ],
3214 psimd_deps = [
3215 ":psimd_fastmath_ukernels",
3216 ":psimd_accmath_ukernels",
3217 ],
3218 wasm_deps = [
3219 ":wasm_ukernels",
3220 ],
3221 wasmsimd_deps = [
3222 ":wasm_ukernels",
3223 ],
3224 x86_deps = [
3225 ":sse2_ukernels",
3226 ":ssse3_ukernels",
3227 ":sse41_ukernels",
3228 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003229 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003230 ":fma3_ukernels",
3231 ":avx2_ukernels",
3232 ":avx512f_ukernels",
3233 ],
3234)
3235
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236xnnpack_cc_library(
3237 name = "im2col",
3238 srcs = ["src/im2col.c"],
3239 hdrs = [
3240 "src/xnnpack/common.h",
3241 "src/xnnpack/im2col.h",
3242 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003243 gcc_copts = xnnpack_gcc_std_copts(),
3244 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245)
3246
3247xnnpack_cc_library(
3248 name = "indirection",
3249 srcs = ["src/indirection.c"],
3250 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003251 gcc_copts = xnnpack_gcc_std_copts(),
3252 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 deps = [
3254 "@FP16",
3255 "@FXdiv",
3256 "@pthreadpool",
3257 ],
3258)
3259
3260xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003261 name = "indirection_test_mode",
3262 srcs = ["src/indirection.c"],
3263 hdrs = INTERNAL_HDRS,
3264 copts = [
3265 "-UNDEBUG",
3266 "-DXNN_TEST_MODE=1",
3267 ],
3268 gcc_copts = xnnpack_gcc_std_copts(),
3269 msvc_copts = xnnpack_msvc_std_copts(),
3270 deps = [
3271 "@FP16",
3272 "@FXdiv",
3273 "@pthreadpool",
3274 ],
3275)
3276
3277xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003278 name = "packing",
3279 srcs = ["src/packing.c"],
3280 hdrs = INTERNAL_HDRS,
3281 gcc_copts = xnnpack_gcc_std_copts(),
3282 msvc_copts = xnnpack_msvc_std_copts(),
3283 deps = [
3284 "@FP16",
3285 "@FXdiv",
3286 "@pthreadpool",
3287 ],
3288)
3289
3290xnnpack_cc_library(
3291 name = "packing_test_mode",
3292 srcs = ["src/packing.c"],
3293 hdrs = INTERNAL_HDRS,
3294 copts = [
3295 "-UNDEBUG",
3296 "-DXNN_TEST_MODE=1",
3297 ],
3298 gcc_copts = xnnpack_gcc_std_copts(),
3299 msvc_copts = xnnpack_msvc_std_copts(),
3300 deps = [
3301 "@FP16",
3302 "@FXdiv",
3303 "@pthreadpool",
3304 ],
3305)
3306
3307xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003308 name = "operator_run",
3309 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003310 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003311 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003312 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3313 "//conditions:default": [],
3314 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003315 gcc_copts = xnnpack_gcc_std_copts(),
3316 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003318 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003319 "@FP16",
3320 "@FXdiv",
3321 "@clog",
3322 "@pthreadpool",
3323 ],
3324)
3325
Chao Mei6ddfc602020-05-13 22:29:36 -07003326xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003327 name = "operator_run_test_mode",
3328 srcs = ["src/operator-run.c"],
3329 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3330 copts = LOGGING_COPTS + [
3331 "-UNDEBUG",
3332 "-DXNN_TEST_MODE=1",
3333 ] + select({
3334 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3335 "//conditions:default": [],
3336 }),
3337 gcc_copts = xnnpack_gcc_std_copts(),
3338 msvc_copts = xnnpack_msvc_std_copts(),
3339 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003340 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003341 "@FP16",
3342 "@FXdiv",
3343 "@clog",
3344 "@pthreadpool",
3345 ],
3346)
3347
3348xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003349 name = "memory_planner",
3350 srcs = ["src/memory-planner.c"],
3351 hdrs = INTERNAL_HDRS,
3352 defines = select({
3353 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3354 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3355 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3356 }),
3357 gcc_copts = xnnpack_gcc_std_copts(),
3358 msvc_copts = xnnpack_msvc_std_copts(),
3359 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003360 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003361 "@pthreadpool",
3362 ],
3363)
3364
Marat Dukhan33fcf782020-05-24 14:27:15 -07003365xnnpack_cc_library(
3366 name = "memory_planner_test_mode",
3367 srcs = ["src/memory-planner.c"],
3368 hdrs = INTERNAL_HDRS,
3369 copts = [
3370 "-UNDEBUG",
3371 "-DXNN_TEST_MODE=1",
3372 ],
3373 defines = select({
3374 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3375 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3376 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3377 }),
3378 gcc_copts = xnnpack_gcc_std_copts(),
3379 msvc_copts = xnnpack_msvc_std_copts(),
3380 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003381 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003382 "@pthreadpool",
3383 ],
3384)
3385
Marat Dukhan08c4a432019-10-03 09:29:21 -07003386cc_library(
3387 name = "enable_assembly",
3388 defines = select({
3389 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3390 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003391 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 }),
3393)
3394
Marat Dukhan9de90e02020-06-18 16:04:12 -07003395cc_library(
3396 name = "enable_sparse",
3397 defines = select({
3398 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3399 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003400 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003401 }),
3402)
3403
Marat Dukhancf056b22019-10-07 10:26:29 -07003404xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 name = "operators",
3406 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003407 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003408 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003409 ],
3410 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003411 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412 "-Isrc",
3413 "-Iinclude",
3414 ] + select({
3415 ":debug_build": [],
3416 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003417 }) + select({
3418 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3419 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003421 gcc_copts = xnnpack_gcc_std_copts(),
3422 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003424 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003425 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003426 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 "@FP16",
3428 "@FXdiv",
3429 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003431 ],
3432)
3433
Marat Dukhan10a38082020-04-17 03:58:35 -07003434xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003435 name = "operators_test_mode",
3436 srcs = OPERATOR_SRCS + [
3437 "src/memory.c",
3438 "src/operator-delete.c",
3439 ],
3440 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3441 copts = LOGGING_COPTS + [
3442 "-Isrc",
3443 "-Iinclude",
3444 "-UNDEBUG",
3445 "-DXNN_TEST_MODE=1",
3446 ] + select({
3447 ":debug_build": [],
3448 "//conditions:default": xnnpack_min_size_copts(),
3449 }) + select({
3450 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3451 "//conditions:default": [],
3452 }),
3453 gcc_copts = xnnpack_gcc_std_copts(),
3454 msvc_copts = xnnpack_msvc_std_copts(),
3455 deps = [
3456 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003457 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003458 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003459 "@FP16",
3460 "@FXdiv",
3461 "@clog",
3462 "@pthreadpool",
3463 ],
3464)
3465
3466xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003467 name = "XNNPACK",
3468 srcs = [
3469 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003470 "src/runtime.c",
3471 "src/subgraph.c",
3472 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003473 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003474 hdrs = ["include/xnnpack.h"],
3475 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003476 "-Isrc",
3477 "-Iinclude",
3478 ] + select({
3479 ":debug_build": [],
3480 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003481 }) + select({
3482 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3483 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003484 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003485 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003486 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003487 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003488 visibility = xnnpack_visibility(),
3489 deps = [
3490 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003491 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003492 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003493 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003494 ":operator_run",
3495 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003496 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003497 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003498 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003499 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003500 ] + select({
3501 ":emscripten": [],
3502 "//conditions:default": ["@cpuinfo"],
3503 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504)
3505
Marat Dukhan10a38082020-04-17 03:58:35 -07003506xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003507 name = "XNNPACK_test_mode",
3508 srcs = [
3509 "src/init.c",
3510 "src/runtime.c",
3511 "src/subgraph.c",
3512 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003513 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003514 hdrs = ["include/xnnpack.h"],
3515 copts = LOGGING_COPTS + [
3516 "-Isrc",
3517 "-Iinclude",
3518 "-UNDEBUG",
3519 "-DXNN_TEST_MODE=1",
3520 ] + select({
3521 ":debug_build": [],
3522 "//conditions:default": xnnpack_min_size_copts(),
3523 }) + select({
3524 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3525 "//conditions:default": [],
3526 }),
3527 gcc_copts = xnnpack_gcc_std_copts(),
3528 includes = ["include"],
3529 msvc_copts = xnnpack_msvc_std_copts(),
3530 visibility = xnnpack_visibility(),
3531 deps = [
3532 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003533 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003534 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003535 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003536 ":operator_run_test_mode",
3537 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003538 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003539 "@clog",
3540 "@FP16",
3541 "@pthreadpool",
3542 ] + select({
3543 ":emscripten": [],
3544 "//conditions:default": ["@cpuinfo"],
3545 }),
3546)
3547
3548xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003549 name = "xnnpack_f32",
3550 srcs = [
3551 "src/init.c",
3552 "src/runtime.c",
3553 "src/subgraph.c",
3554 "src/tensor.c",
3555 ] + SUBGRAPH_SRCS,
3556 hdrs = ["include/xnnpack.h"],
3557 copts = LOGGING_COPTS + [
3558 "-Isrc",
3559 "-Iinclude",
3560 ] + select({
3561 ":debug_build": [],
3562 "//conditions:default": xnnpack_min_size_copts(),
3563 }) + select({
3564 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3565 "//conditions:default": [],
3566 }),
3567 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003568 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003569 "XNN_NO_U8_OPERATORS",
3570 "XNN_NO_X8_OPERATORS",
3571 "XNN_NO_F16_OPERATORS",
3572 "XNN_NO_X16_OPERATORS",
3573 ],
3574 gcc_copts = xnnpack_gcc_std_copts(),
3575 includes = ["include"],
3576 msvc_copts = xnnpack_msvc_std_copts(),
3577 visibility = xnnpack_visibility(),
3578 deps = [
3579 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003580 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003581 ":logging_utils",
3582 ":memory_planner",
3583 ":operator_run",
3584 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003585 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003586 "@clog",
3587 "@FP16",
3588 "@pthreadpool",
3589 ] + select({
3590 ":emscripten": [],
3591 "//conditions:default": ["@cpuinfo"],
3592 }),
3593)
3594
3595xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003596 name = "xnnpack_operators_nhwc_f32",
3597 srcs = [
3598 "src/init.c",
3599 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003600 hdrs = ["include/xnnpack.h"],
3601 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003602 "-Isrc",
3603 "-Iinclude",
3604 ] + select({
3605 ":debug_build": [],
3606 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003607 }) + select({
3608 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3609 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003610 }),
3611 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003612 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003613 "XNN_NO_U8_OPERATORS",
3614 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003615 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003616 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003617 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003619 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620 visibility = xnnpack_visibility(),
3621 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003622 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003623 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 ":operator_run",
3625 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003626 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003627 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003629 ] + select({
3630 ":emscripten": [],
3631 "//conditions:default": ["@cpuinfo"],
3632 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633)
3634
Marat Dukhancf056b22019-10-07 10:26:29 -07003635xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 name = "bench_utils",
3637 srcs = ["bench/utils.cc"],
3638 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003639 deps = [
3640 "@com_google_benchmark//:benchmark",
3641 "@cpuinfo",
3642 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643)
3644
Frank Barchard7e955972019-10-11 10:34:25 -07003645######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646
3647xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003648 name = "qs8_gemm_bench",
3649 srcs = [
3650 "bench/gemm.h",
3651 "bench/qs8-gemm.cc",
3652 "src/xnnpack/AlignedAllocator.h",
3653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3654 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3655)
3656
3657xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003658 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 srcs = [
3660 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003661 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 "src/xnnpack/AlignedAllocator.h",
3663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003664 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003665 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666)
3667
3668xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003669 name = "qu8_requantization_bench",
3670 srcs = [
3671 "bench/qu8-requantization.cc",
3672 "src/xnnpack/requantization-stubs.h",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + MICROKERNEL_BENCHMARK_HDRS,
3675 deps = MICROKERNEL_BENCHMARK_DEPS,
3676)
3677
3678xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003679 name = "f16_igemm_bench",
3680 srcs = [
3681 "bench/f16-igemm.cc",
3682 "bench/conv.h",
3683 "bench/google/conv.h",
3684 "src/xnnpack/AlignedAllocator.h",
3685 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003686 deps = MICROKERNEL_BENCHMARK_DEPS + [
3687 ":indirection",
3688 ":packing",
3689 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003690)
3691
3692xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693 name = "f16_gemm_bench",
3694 srcs = [
3695 "bench/f16-gemm.cc",
3696 "bench/gemm.h",
3697 "src/xnnpack/AlignedAllocator.h",
3698 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003699 deps = MICROKERNEL_BENCHMARK_DEPS + [
3700 ":packing",
3701 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702)
3703
3704xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003705 name = "f16_spmm_bench",
3706 srcs = [
3707 "bench/f16-spmm.cc",
3708 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003709 "src/xnnpack/AlignedAllocator.h",
3710 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003711 deps = MICROKERNEL_BENCHMARK_DEPS,
3712)
3713
3714xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 name = "f32_igemm_bench",
3716 srcs = [
3717 "bench/f32-igemm.cc",
3718 "bench/conv.h",
3719 "src/xnnpack/AlignedAllocator.h",
3720 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003721 deps = MICROKERNEL_BENCHMARK_DEPS + [
3722 ":indirection",
3723 ":packing",
3724 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725)
3726
3727xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003728 name = "f16_relu_bench",
3729 srcs = [
3730 "bench/f16-relu.cc",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + MICROKERNEL_BENCHMARK_HDRS,
3733 deps = MICROKERNEL_BENCHMARK_DEPS,
3734)
3735
3736xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 name = "f32_conv_hwc_bench",
3738 srcs = [
3739 "bench/f32-conv-hwc.cc",
3740 "bench/dconv.h",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003743 deps = MICROKERNEL_BENCHMARK_DEPS + [
3744 ":packing",
3745 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746)
3747
3748xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003749 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003750 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003751 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003752 "bench/dconv.h",
3753 "src/xnnpack/AlignedAllocator.h",
3754 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003755 deps = MICROKERNEL_BENCHMARK_DEPS + [
3756 ":packing",
3757 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003758)
3759
3760xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003761 name = "f16_dwconv_bench",
3762 srcs = [
3763 "bench/f16-dwconv.cc",
3764 "bench/dwconv.h",
3765 "bench/google/dwconv.h",
3766 "src/xnnpack/AlignedAllocator.h",
3767 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003768 deps = MICROKERNEL_BENCHMARK_DEPS + [
3769 ":indirection",
3770 ":packing",
3771 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003772)
3773
3774xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775 name = "f32_dwconv_bench",
3776 srcs = [
3777 "bench/f32-dwconv.cc",
3778 "bench/dwconv.h",
3779 "src/xnnpack/AlignedAllocator.h",
3780 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003781 deps = MICROKERNEL_BENCHMARK_DEPS + [
3782 ":indirection",
3783 ":packing",
3784 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003785)
3786
3787xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003788 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003790 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791 "bench/dwconv.h",
3792 "src/xnnpack/AlignedAllocator.h",
3793 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003794 deps = MICROKERNEL_BENCHMARK_DEPS + [
3795 ":indirection",
3796 ":packing",
3797 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798)
3799
3800xnnpack_benchmark(
3801 name = "f32_gemm_bench",
3802 srcs = [
3803 "bench/f32-gemm.cc",
3804 "bench/gemm.h",
3805 "src/xnnpack/AlignedAllocator.h",
3806 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003807 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003808 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003809)
3810
3811xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003812 name = "f32_hswish_bench",
3813 srcs = [
3814 "bench/f32-hswish.cc",
3815 "src/xnnpack/AlignedAllocator.h",
3816 ] + MICROKERNEL_BENCHMARK_HDRS,
3817 deps = MICROKERNEL_BENCHMARK_DEPS,
3818)
3819
3820xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003821 name = "f32_raddexpminusmax_bench",
3822 srcs = [
3823 "bench/f32-raddexpminusmax.cc",
3824 "src/xnnpack/AlignedAllocator.h",
3825 ] + MICROKERNEL_BENCHMARK_HDRS,
3826 deps = MICROKERNEL_BENCHMARK_DEPS,
3827)
3828
3829xnnpack_benchmark(
3830 name = "f32_raddextexp_bench",
3831 srcs = [
3832 "bench/f32-raddextexp.cc",
3833 "src/xnnpack/AlignedAllocator.h",
3834 ] + MICROKERNEL_BENCHMARK_HDRS,
3835 deps = MICROKERNEL_BENCHMARK_DEPS,
3836)
3837
3838xnnpack_benchmark(
3839 name = "f32_raddstoreexpminusmax_bench",
3840 srcs = [
3841 "bench/f32-raddstoreexpminusmax.cc",
3842 "src/xnnpack/AlignedAllocator.h",
3843 ] + MICROKERNEL_BENCHMARK_HDRS,
3844 deps = MICROKERNEL_BENCHMARK_DEPS,
3845)
3846
3847xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003848 name = "f32_relu_bench",
3849 srcs = [
3850 "bench/f32-relu.cc",
3851 "src/xnnpack/AlignedAllocator.h",
3852 ] + MICROKERNEL_BENCHMARK_HDRS,
3853 deps = MICROKERNEL_BENCHMARK_DEPS,
3854)
3855
3856xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003857 name = "f32_rmax_bench",
3858 srcs = [
3859 "bench/f32-rmax.cc",
3860 "src/xnnpack/AlignedAllocator.h",
3861 ] + MICROKERNEL_BENCHMARK_HDRS,
3862 deps = MICROKERNEL_BENCHMARK_DEPS,
3863)
3864
3865xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003866 name = "f32_sigmoid_bench",
3867 srcs = [
3868 "bench/f32-sigmoid.cc",
3869 "src/xnnpack/AlignedAllocator.h",
3870 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003871 deps = MICROKERNEL_BENCHMARK_DEPS,
3872)
3873
3874xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003875 name = "f32_spmm_bench",
3876 srcs = [
3877 "bench/f32-spmm.cc",
3878 "bench/gemm.h",
3879 "src/xnnpack/AlignedAllocator.h",
3880 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003881 deps = MICROKERNEL_BENCHMARK_DEPS,
3882)
3883
3884xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003885 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003886 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003887 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003888 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003889 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003890 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003891)
3892
3893xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003894 name = "f32_vscaleexpminusmax_bench",
3895 srcs = [
3896 "bench/f32-vscaleexpminusmax.cc",
3897 "src/xnnpack/AlignedAllocator.h",
3898 ] + MICROKERNEL_BENCHMARK_HDRS,
3899 deps = MICROKERNEL_BENCHMARK_DEPS,
3900)
3901
3902xnnpack_benchmark(
3903 name = "f32_vscaleextexp_bench",
3904 srcs = [
3905 "bench/f32-vscaleextexp.cc",
3906 "src/xnnpack/AlignedAllocator.h",
3907 ] + MICROKERNEL_BENCHMARK_HDRS,
3908 deps = MICROKERNEL_BENCHMARK_DEPS,
3909)
3910
3911xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003912 name = "f32_vsqrt_bench",
3913 srcs = [
3914 "bench/f32-vsqrt.cc",
3915 "src/xnnpack/AlignedAllocator.h",
3916 ] + MICROKERNEL_BENCHMARK_HDRS,
3917 deps = MICROKERNEL_BENCHMARK_DEPS,
3918)
3919
3920xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 name = "f32_im2col_gemm_bench",
3922 srcs = [
3923 "bench/f32-im2col-gemm.cc",
3924 "bench/conv.h",
3925 "src/xnnpack/AlignedAllocator.h",
3926 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003927 deps = MICROKERNEL_BENCHMARK_DEPS + [
3928 ":im2col",
3929 ":packing",
3930 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931)
3932
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003933xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003934 name = "rounding_bench",
3935 srcs = [
3936 "bench/rounding.cc",
3937 "src/xnnpack/math-stubs.h",
3938 "src/xnnpack/AlignedAllocator.h",
3939 ] + MICROKERNEL_BENCHMARK_HDRS,
3940 deps = MICROKERNEL_BENCHMARK_DEPS,
3941)
3942
Marat Dukhan08c4a432019-10-03 09:29:21 -07003943########################### Benchmarks for operators ###########################
3944
3945xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946 name = "average_pooling_bench",
3947 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003948 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003949 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003950 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951)
3952
3953xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003954 name = "bankers_rounding_bench",
3955 srcs = ["bench/bankers-rounding.cc"],
3956 copts = xnnpack_optional_tflite_copts(),
3957 tags = ["nowin32"],
3958 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3959)
3960
3961xnnpack_benchmark(
3962 name = "ceiling_bench",
3963 srcs = ["bench/ceiling.cc"],
3964 copts = xnnpack_optional_tflite_copts(),
3965 tags = ["nowin32"],
3966 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3967)
3968
3969xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970 name = "channel_shuffle_bench",
3971 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003972 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003973)
3974
3975xnnpack_benchmark(
3976 name = "convolution_bench",
3977 srcs = ["bench/convolution.cc"],
3978 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003979 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003980 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981)
3982
3983xnnpack_benchmark(
3984 name = "deconvolution_bench",
3985 srcs = ["bench/deconvolution.cc"],
3986 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003987 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003988 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003989)
3990
3991xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003992 name = "floor_bench",
3993 srcs = ["bench/floor.cc"],
3994 copts = xnnpack_optional_tflite_copts(),
3995 tags = ["nowin32"],
3996 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3997)
3998
3999xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000 name = "global_average_pooling_bench",
4001 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004002 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004003)
4004
4005xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004006 name = "hardswish_bench",
4007 srcs = ["bench/hardswish.cc"],
4008 copts = xnnpack_optional_tflite_copts(),
4009 tags = ["nowin32"],
4010 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4011)
4012
4013xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004014 name = "max_pooling_bench",
4015 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004016 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004017)
4018
4019xnnpack_benchmark(
4020 name = "sigmoid_bench",
4021 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004022 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004023 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004024 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004025)
4026
4027xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004028 name = "prelu_bench",
4029 srcs = ["bench/prelu.cc"],
4030 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004031 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004032 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004033)
4034
4035xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004036 name = "softmax_bench",
4037 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004038 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004039 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004040 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004041)
4042
Marat Dukhan87727142020-06-24 15:24:10 -07004043xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004044 name = "square_root_bench",
4045 srcs = ["bench/square-root.cc"],
4046 copts = xnnpack_optional_tflite_copts(),
4047 tags = ["nowin32"],
4048 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4049)
4050
4051xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004052 name = "truncation_bench",
4053 srcs = ["bench/truncation.cc"],
4054 deps = OPERATOR_BENCHMARK_DEPS,
4055)
4056
Marat Dukhanc068bb62019-10-04 13:24:39 -07004057############################# End-to-end benchmarks ############################
4058
4059cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004060 name = "fp32_mobilenet_v1",
4061 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004062 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004063 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004064 linkstatic = True,
4065 deps = [
4066 ":XNNPACK",
4067 "@pthreadpool",
4068 ],
4069)
4070
4071cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004072 name = "fp16_mobilenet_v1",
4073 srcs = ["models/fp16-mobilenet-v1.cc"],
4074 hdrs = ["models/models.h"],
4075 copts = xnnpack_std_cxxopts(),
4076 linkstatic = True,
4077 deps = [
4078 ":XNNPACK",
4079 "@FP16",
4080 "@pthreadpool",
4081 ],
4082)
4083
4084cc_library(
4085 name = "fp32_mobilenet_v2",
4086 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004087 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004088 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004089 linkstatic = True,
4090 deps = [
4091 ":XNNPACK",
4092 "@pthreadpool",
4093 ],
4094)
4095
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004096cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004097 name = "fp16_mobilenet_v2",
4098 srcs = ["models/fp16-mobilenet-v2.cc"],
4099 hdrs = ["models/models.h"],
4100 copts = xnnpack_std_cxxopts(),
4101 linkstatic = True,
4102 deps = [
4103 ":XNNPACK",
4104 "@FP16",
4105 "@pthreadpool",
4106 ],
4107)
4108
4109cc_library(
4110 name = "fp32_mobilenet_v3_large",
4111 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004112 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004113 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004114 linkstatic = True,
4115 deps = [
4116 ":XNNPACK",
4117 "@pthreadpool",
4118 ],
4119)
4120
4121cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004122 name = "fp32_mobilenet_v3_small",
4123 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004124 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004125 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004126 linkstatic = True,
4127 deps = [
4128 ":XNNPACK",
4129 "@pthreadpool",
4130 ],
4131)
4132
Marat Dukhanc068bb62019-10-04 13:24:39 -07004133xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004134 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004135 srcs = [
4136 "bench/f32-dwconv-e2e.cc",
4137 "bench/end2end.h",
4138 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004139 deps = MICROKERNEL_BENCHMARK_DEPS + [
4140 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004141 ":fp32_mobilenet_v1",
4142 ":fp32_mobilenet_v2",
4143 ":fp32_mobilenet_v3_large",
4144 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004145 ],
4146)
4147
4148xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004149 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004150 srcs = [
4151 "bench/f32-gemm-e2e.cc",
4152 "bench/end2end.h",
4153 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004154 deps = MICROKERNEL_BENCHMARK_DEPS + [
4155 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004156 ":fp32_mobilenet_v1",
4157 ":fp32_mobilenet_v2",
4158 ":fp32_mobilenet_v3_large",
4159 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004160 ],
4161)
4162
4163xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004164 name = "end2end_bench",
4165 srcs = ["bench/end2end.cc"],
4166 deps = [
4167 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004168 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004169 ":fp16_mobilenet_v1",
4170 ":fp16_mobilenet_v2",
4171 ":fp32_mobilenet_v1",
4172 ":fp32_mobilenet_v2",
4173 ":fp32_mobilenet_v3_large",
4174 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004175 "@pthreadpool",
4176 ],
4177)
4178
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004179#################### Accuracy evaluation for math functions ####################
4180
4181xnnpack_benchmark(
4182 name = "f32_exp_eval",
4183 srcs = [
4184 "eval/f32-exp.cc",
4185 "src/xnnpack/AlignedAllocator.h",
4186 ] + ACCURACY_EVAL_HDRS,
4187 deps = ACCURACY_EVAL_DEPS,
4188)
4189
Marat Dukhan515c9772019-10-17 18:07:57 -07004190xnnpack_benchmark(
4191 name = "f32_expminus_eval",
4192 srcs = [
4193 "eval/f32-expminus.cc",
4194 "src/xnnpack/AlignedAllocator.h",
4195 ] + ACCURACY_EVAL_HDRS,
4196 deps = ACCURACY_EVAL_DEPS,
4197)
4198
Marat Dukhan98ba4412019-10-23 02:14:28 -07004199xnnpack_benchmark(
4200 name = "f32_extexp_eval",
4201 srcs = [
4202 "eval/f32-extexp.cc",
4203 "src/xnnpack/AlignedAllocator.h",
4204 ] + ACCURACY_EVAL_HDRS,
4205 deps = ACCURACY_EVAL_DEPS,
4206)
4207
Marat Dukhan8853b822020-05-07 12:19:01 -07004208xnnpack_unit_test(
4209 name = "f32_roundne_eval",
4210 srcs = [
4211 "eval/f32-roundne.cc",
4212 "src/xnnpack/AlignedAllocator.h",
4213 "src/xnnpack/math-stubs.h",
4214 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004215 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004216 deps = MICROKERNEL_TEST_DEPS,
4217)
4218
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004219xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004220 name = "f32_roundd_eval",
4221 srcs = [
4222 "eval/f32-roundd.cc",
4223 "src/xnnpack/AlignedAllocator.h",
4224 "src/xnnpack/math-stubs.h",
4225 ] + MICROKERNEL_TEST_HDRS,
4226 automatic = False,
4227 deps = MICROKERNEL_TEST_DEPS,
4228)
4229
4230xnnpack_unit_test(
4231 name = "f32_roundu_eval",
4232 srcs = [
4233 "eval/f32-roundu.cc",
4234 "src/xnnpack/AlignedAllocator.h",
4235 "src/xnnpack/math-stubs.h",
4236 ] + MICROKERNEL_TEST_HDRS,
4237 automatic = False,
4238 deps = MICROKERNEL_TEST_DEPS,
4239)
4240
4241xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004242 name = "f32_roundz_eval",
4243 srcs = [
4244 "eval/f32-roundz.cc",
4245 "src/xnnpack/AlignedAllocator.h",
4246 "src/xnnpack/math-stubs.h",
4247 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004248 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004249 deps = MICROKERNEL_TEST_DEPS,
4250)
4251
Marat Dukhan346a9e52019-11-15 09:06:30 -08004252xnnpack_benchmark(
4253 name = "f32_sigmoid_eval",
4254 srcs = [
4255 "eval/f32-sigmoid.cc",
4256 "src/xnnpack/AlignedAllocator.h",
4257 ] + ACCURACY_EVAL_HDRS,
4258 deps = ACCURACY_EVAL_DEPS,
4259)
4260
Marat Dukhan84000762020-06-29 18:38:43 -07004261xnnpack_benchmark(
4262 name = "f32_sqrt_eval",
4263 srcs = [
4264 "eval/f32-sqrt.cc",
4265 "src/xnnpack/AlignedAllocator.h",
4266 ] + ACCURACY_EVAL_HDRS,
4267 deps = ACCURACY_EVAL_DEPS + [
4268 ":bench_utils",
4269 ],
4270)
4271
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272######################### Unit tests for micro-kernels #########################
4273
4274xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004275 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004277 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 "test/gemm-microkernel-tester.h",
4279 "src/xnnpack/AlignedAllocator.h",
4280 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004281 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004282)
4283
4284xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004285 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004286 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004287 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004288 "test/spmm-microkernel-tester.h",
4289 "src/xnnpack/AlignedAllocator.h",
4290 ] + MICROKERNEL_TEST_HDRS,
4291 deps = MICROKERNEL_TEST_DEPS,
4292)
4293
4294xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004295 name = "f32_argmaxpool_test",
4296 srcs = [
4297 "test/f32-argmaxpool.cc",
4298 "test/argmaxpool-microkernel-tester.h",
4299 "src/xnnpack/AlignedAllocator.h",
4300 ] + MICROKERNEL_TEST_HDRS,
4301 deps = MICROKERNEL_TEST_DEPS,
4302)
4303
4304xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004305 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004307 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308 "test/avgpool-microkernel-tester.h",
4309 "src/xnnpack/AlignedAllocator.h",
4310 ] + MICROKERNEL_TEST_HDRS,
4311 deps = MICROKERNEL_TEST_DEPS,
4312)
4313
4314xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004315 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004316 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004317 "test/f32-ibilinear.cc",
4318 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004319 "src/xnnpack/AlignedAllocator.h",
4320 ] + MICROKERNEL_TEST_HDRS,
4321 deps = MICROKERNEL_TEST_DEPS,
4322)
4323
4324xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004325 name = "f16_clamp_test",
4326 srcs = [
4327 "test/f16-clamp.cc",
4328 "test/clamp-microkernel-tester.h",
4329 ] + MICROKERNEL_TEST_HDRS,
4330 deps = MICROKERNEL_TEST_DEPS,
4331)
4332
4333xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004334 name = "f32_clamp_test",
4335 srcs = [
4336 "test/f32-clamp.cc",
4337 "test/clamp-microkernel-tester.h",
4338 ] + MICROKERNEL_TEST_HDRS,
4339 deps = MICROKERNEL_TEST_DEPS,
4340)
4341
4342xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004343 name = "f32_igemm_test",
4344 srcs = [
4345 "test/f32-igemm.cc",
4346 "test/gemm-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 Dukhan163a7e62020-04-09 04:19:26 -07004350)
4351
4352xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004353 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004355 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004356 "test/gemm-microkernel-tester.h",
4357 "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(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004363 name = "f32_igemm_minmax_test",
4364 srcs = [
4365 "test/f32-igemm-minmax.cc",
4366 "test/gemm-microkernel-tester.h",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004369 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004370)
4371
4372xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004373 name = "f16_igemm_minmax_test",
4374 srcs = [
4375 "test/f16-igemm-minmax.cc",
4376 "test/gemm-microkernel-tester.h",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004379 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004380)
4381
4382xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 name = "f32_conv_hwc_test",
4384 srcs = [
4385 "test/f32-conv-hwc.cc",
4386 "test/conv-hwc-microkernel-tester.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004389 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390)
4391
4392xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004393 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004395 "test/f32-conv-hwc2chw.cc",
4396 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004397 "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 Dukhan08c4a432019-10-03 09:29:21 -07004400)
4401
4402xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004403 name = "f32_dwconv_test",
4404 srcs = [
4405 "test/f32-dwconv.cc",
4406 "test/dwconv-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 Dukhan163a7e62020-04-09 04:19:26 -07004410)
4411
4412xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004413 name = "f16_dwconv_minmax_test",
4414 srcs = [
4415 "test/f16-dwconv-minmax.cc",
4416 "test/dwconv-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"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004420)
4421
4422xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004423 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004425 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426 "test/dwconv-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(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004433 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004434 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004435 "test/f32-dwconv-chw.cc",
4436 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004437 "src/xnnpack/AlignedAllocator.h",
4438 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004439 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440)
4441
4442xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004443 name = "f16_gavgpool_minmax_test",
4444 srcs = [
4445 "test/f16-gavgpool-minmax.cc",
4446 "test/gavgpool-microkernel-tester.h",
4447 "src/xnnpack/AlignedAllocator.h",
4448 ] + MICROKERNEL_TEST_HDRS,
4449 deps = MICROKERNEL_TEST_DEPS,
4450)
4451
4452xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004453 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004455 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456 "test/gavgpool-microkernel-tester.h",
4457 "src/xnnpack/AlignedAllocator.h",
4458 ] + MICROKERNEL_TEST_HDRS,
4459 deps = MICROKERNEL_TEST_DEPS,
4460)
4461
4462xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004463 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004465 "test/f32-gavgpool-cw.cc",
4466 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 "src/xnnpack/AlignedAllocator.h",
4468 ] + MICROKERNEL_TEST_HDRS,
4469 deps = MICROKERNEL_TEST_DEPS,
4470)
4471
4472xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004473 name = "f32_gemm_test",
4474 srcs = [
4475 "test/f32-gemm.cc",
4476 "test/gemm-microkernel-tester.h",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004479 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004480)
4481
4482xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004483 name = "f32_gemm_relu_test",
4484 srcs = [
4485 "test/f32-gemm-relu.cc",
4486 "test/gemm-microkernel-tester.h",
4487 "src/xnnpack/AlignedAllocator.h",
4488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004489 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004490)
4491
4492xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004493 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004495 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 "test/gemm-microkernel-tester.h",
4497 "src/xnnpack/AlignedAllocator.h",
4498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004499 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500)
4501
4502xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004503 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004505 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004506 "test/gemm-microkernel-tester.h",
4507 "src/xnnpack/AlignedAllocator.h",
4508 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004509 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510)
4511
4512xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004513 name = "f16_hswish_test",
4514 srcs = [
4515 "test/f16-hswish.cc",
4516 "test/hswish-microkernel-tester.h",
4517 ] + MICROKERNEL_TEST_HDRS,
4518 deps = MICROKERNEL_TEST_DEPS,
4519)
4520
4521xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522 name = "f32_hswish_test",
4523 srcs = [
4524 "test/f32-hswish.cc",
4525 "test/hswish-microkernel-tester.h",
4526 ] + MICROKERNEL_TEST_HDRS,
4527 deps = MICROKERNEL_TEST_DEPS,
4528)
4529
4530xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004531 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004533 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 "test/maxpool-microkernel-tester.h",
4535 ] + MICROKERNEL_TEST_HDRS,
4536 deps = MICROKERNEL_TEST_DEPS,
4537)
4538
4539xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004540 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004542 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 "test/avgpool-microkernel-tester.h",
4544 "src/xnnpack/AlignedAllocator.h",
4545 ] + MICROKERNEL_TEST_HDRS,
4546 deps = MICROKERNEL_TEST_DEPS,
4547)
4548
4549xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004550 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004552 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 "test/gemm-microkernel-tester.h",
4554 "src/xnnpack/AlignedAllocator.h",
4555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004556 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557)
4558
4559xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004560 name = "f16_prelu_test",
4561 srcs = [
4562 "test/f16-prelu.cc",
4563 "test/prelu-microkernel-tester.h",
4564 "src/xnnpack/AlignedAllocator.h",
4565 ] + MICROKERNEL_TEST_HDRS,
4566 deps = MICROKERNEL_TEST_DEPS,
4567)
4568
4569xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004570 name = "f32_prelu_test",
4571 srcs = [
4572 "test/f32-prelu.cc",
4573 "test/prelu-microkernel-tester.h",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_TEST_HDRS,
4576 deps = MICROKERNEL_TEST_DEPS,
4577)
4578
4579xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004580 name = "f32_raddexpminusmax_test",
4581 srcs = [
4582 "test/f32-raddexpminusmax.cc",
4583 "test/raddexpminusmax-microkernel-tester.h",
4584 ] + MICROKERNEL_TEST_HDRS,
4585 deps = MICROKERNEL_TEST_DEPS,
4586)
4587
4588xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004589 name = "f32_raddextexp_test",
4590 srcs = [
4591 "test/f32-raddextexp.cc",
4592 "test/raddextexp-microkernel-tester.h",
4593 ] + MICROKERNEL_TEST_HDRS,
4594 deps = MICROKERNEL_TEST_DEPS,
4595)
4596
4597xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004598 name = "f32_raddstoreexpminusmax_test",
4599 srcs = [
4600 "test/f32-raddstoreexpminusmax.cc",
4601 "test/raddstoreexpminusmax-microkernel-tester.h",
4602 ] + MICROKERNEL_TEST_HDRS,
4603 deps = MICROKERNEL_TEST_DEPS,
4604)
4605
4606xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004607 name = "f32_relu_test",
4608 srcs = [
4609 "test/f32-relu.cc",
4610 "test/vunary-microkernel-tester.h",
4611 ] + MICROKERNEL_TEST_HDRS,
4612 deps = MICROKERNEL_TEST_DEPS,
4613)
4614
4615xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004616 name = "f32_rmax_test",
4617 srcs = [
4618 "test/f32-rmax.cc",
4619 "test/rmax-microkernel-tester.h",
4620 ] + MICROKERNEL_TEST_HDRS,
4621 deps = MICROKERNEL_TEST_DEPS,
4622)
4623
4624xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004625 name = "f32_sigmoid_test",
4626 srcs = [
4627 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004628 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004629 ] + MICROKERNEL_TEST_HDRS,
4630 deps = MICROKERNEL_TEST_DEPS,
4631)
4632
4633xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004634 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004635 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004636 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004637 "test/spmm-microkernel-tester.h",
4638 "src/xnnpack/AlignedAllocator.h",
4639 ] + MICROKERNEL_TEST_HDRS,
4640 deps = MICROKERNEL_TEST_DEPS,
4641)
4642
4643xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004644 name = "f16_vadd_minmax_test",
4645 srcs = [
4646 "test/f16-vadd-minmax.cc",
4647 "test/vbinary-microkernel-tester.h",
4648 ] + MICROKERNEL_TEST_HDRS,
4649 deps = MICROKERNEL_TEST_DEPS,
4650)
4651
4652xnnpack_unit_test(
4653 name = "f16_vaddc_minmax_test",
4654 srcs = [
4655 "test/f16-vaddc-minmax.cc",
4656 "test/vbinaryc-microkernel-tester.h",
4657 ] + MICROKERNEL_TEST_HDRS,
4658 deps = MICROKERNEL_TEST_DEPS,
4659)
4660
4661xnnpack_unit_test(
4662 name = "f16_vdiv_minmax_test",
4663 srcs = [
4664 "test/f16-vdiv-minmax.cc",
4665 "test/vbinary-microkernel-tester.h",
4666 ] + MICROKERNEL_TEST_HDRS,
4667 deps = MICROKERNEL_TEST_DEPS,
4668)
4669
4670xnnpack_unit_test(
4671 name = "f16_vdivc_minmax_test",
4672 srcs = [
4673 "test/f16-vdivc-minmax.cc",
4674 "test/vbinaryc-microkernel-tester.h",
4675 ] + MICROKERNEL_TEST_HDRS,
4676 deps = MICROKERNEL_TEST_DEPS,
4677)
4678
4679xnnpack_unit_test(
4680 name = "f16_vrdivc_minmax_test",
4681 srcs = [
4682 "test/f16-vrdivc-minmax.cc",
4683 "test/vbinaryc-microkernel-tester.h",
4684 ] + MICROKERNEL_TEST_HDRS,
4685 deps = MICROKERNEL_TEST_DEPS,
4686)
4687
4688xnnpack_unit_test(
4689 name = "f16_vmax_test",
4690 srcs = [
4691 "test/f16-vmax.cc",
4692 "test/vbinary-microkernel-tester.h",
4693 ] + MICROKERNEL_TEST_HDRS,
4694 deps = MICROKERNEL_TEST_DEPS,
4695)
4696
4697xnnpack_unit_test(
4698 name = "f16_vmaxc_test",
4699 srcs = [
4700 "test/f16-vmaxc.cc",
4701 "test/vbinaryc-microkernel-tester.h",
4702 ] + MICROKERNEL_TEST_HDRS,
4703 deps = MICROKERNEL_TEST_DEPS,
4704)
4705
4706xnnpack_unit_test(
4707 name = "f16_vmin_test",
4708 srcs = [
4709 "test/f16-vmin.cc",
4710 "test/vbinary-microkernel-tester.h",
4711 ] + MICROKERNEL_TEST_HDRS,
4712 deps = MICROKERNEL_TEST_DEPS,
4713)
4714
4715xnnpack_unit_test(
4716 name = "f16_vminc_test",
4717 srcs = [
4718 "test/f16-vminc.cc",
4719 "test/vbinaryc-microkernel-tester.h",
4720 ] + MICROKERNEL_TEST_HDRS,
4721 deps = MICROKERNEL_TEST_DEPS,
4722)
4723
4724xnnpack_unit_test(
4725 name = "f16_vmul_minmax_test",
4726 srcs = [
4727 "test/f16-vmul-minmax.cc",
4728 "test/vbinary-microkernel-tester.h",
4729 ] + MICROKERNEL_TEST_HDRS,
4730 deps = MICROKERNEL_TEST_DEPS,
4731)
4732
4733xnnpack_unit_test(
4734 name = "f16_vmulc_minmax_test",
4735 srcs = [
4736 "test/f16-vmulc-minmax.cc",
4737 "test/vbinaryc-microkernel-tester.h",
4738 ] + MICROKERNEL_TEST_HDRS,
4739 deps = MICROKERNEL_TEST_DEPS,
4740)
4741
4742xnnpack_unit_test(
4743 name = "f16_vsub_minmax_test",
4744 srcs = [
4745 "test/f16-vsub-minmax.cc",
4746 "test/vbinary-microkernel-tester.h",
4747 ] + MICROKERNEL_TEST_HDRS,
4748 deps = MICROKERNEL_TEST_DEPS,
4749)
4750
4751xnnpack_unit_test(
4752 name = "f16_vsubc_minmax_test",
4753 srcs = [
4754 "test/f16-vsubc-minmax.cc",
4755 "test/vbinaryc-microkernel-tester.h",
4756 ] + MICROKERNEL_TEST_HDRS,
4757 deps = MICROKERNEL_TEST_DEPS,
4758)
4759
4760xnnpack_unit_test(
4761 name = "f16_vrsubc_minmax_test",
4762 srcs = [
4763 "test/f16-vrsubc-minmax.cc",
4764 "test/vbinaryc-microkernel-tester.h",
4765 ] + MICROKERNEL_TEST_HDRS,
4766 deps = MICROKERNEL_TEST_DEPS,
4767)
4768
4769xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004770 name = "f32_vabs_test",
4771 srcs = [
4772 "test/f32-vabs.cc",
4773 "test/vunary-microkernel-tester.h",
4774 ] + MICROKERNEL_TEST_HDRS,
4775 deps = MICROKERNEL_TEST_DEPS,
4776)
4777
4778xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004779 name = "f32_vadd_test",
4780 srcs = [
4781 "test/f32-vadd.cc",
4782 "test/vbinary-microkernel-tester.h",
4783 ] + MICROKERNEL_TEST_HDRS,
4784 deps = MICROKERNEL_TEST_DEPS,
4785)
4786
4787xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004788 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004790 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004791 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004792 ] + MICROKERNEL_TEST_HDRS,
4793 deps = MICROKERNEL_TEST_DEPS,
4794)
4795
4796xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004797 name = "f32_vaddc_test",
4798 srcs = [
4799 "test/f32-vaddc.cc",
4800 "test/vbinaryc-microkernel-tester.h",
4801 ] + MICROKERNEL_TEST_HDRS,
4802 deps = MICROKERNEL_TEST_DEPS,
4803)
4804
4805xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004806 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004807 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004808 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004809 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004810 ] + MICROKERNEL_TEST_HDRS,
4811 deps = MICROKERNEL_TEST_DEPS,
4812)
4813
4814xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004815 name = "f32_vdiv_test",
4816 srcs = [
4817 "test/f32-vdiv.cc",
4818 "test/vbinary-microkernel-tester.h",
4819 ] + MICROKERNEL_TEST_HDRS,
4820 deps = MICROKERNEL_TEST_DEPS,
4821)
4822
4823xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004824 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004825 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004826 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004827 "test/vbinary-microkernel-tester.h",
4828 ] + MICROKERNEL_TEST_HDRS,
4829 deps = MICROKERNEL_TEST_DEPS,
4830)
4831
4832xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004833 name = "f32_vdivc_test",
4834 srcs = [
4835 "test/f32-vdivc.cc",
4836 "test/vbinaryc-microkernel-tester.h",
4837 ] + MICROKERNEL_TEST_HDRS,
4838 deps = MICROKERNEL_TEST_DEPS,
4839)
4840
4841xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004842 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004843 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004844 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004845 "test/vbinaryc-microkernel-tester.h",
4846 ] + MICROKERNEL_TEST_HDRS,
4847 deps = MICROKERNEL_TEST_DEPS,
4848)
4849
4850xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004851 name = "f32_vrdivc_test",
4852 srcs = [
4853 "test/f32-vrdivc.cc",
4854 "test/vbinaryc-microkernel-tester.h",
4855 ] + MICROKERNEL_TEST_HDRS,
4856 deps = MICROKERNEL_TEST_DEPS,
4857)
4858
4859xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004860 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004861 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004862 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004863 "test/vbinaryc-microkernel-tester.h",
4864 ] + MICROKERNEL_TEST_HDRS,
4865 deps = MICROKERNEL_TEST_DEPS,
4866)
4867
4868xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004869 name = "f32_vmax_test",
4870 srcs = [
4871 "test/f32-vmax.cc",
4872 "test/vbinary-microkernel-tester.h",
4873 ] + MICROKERNEL_TEST_HDRS,
4874 deps = MICROKERNEL_TEST_DEPS,
4875)
4876
4877xnnpack_unit_test(
4878 name = "f32_vmaxc_test",
4879 srcs = [
4880 "test/f32-vmaxc.cc",
4881 "test/vbinaryc-microkernel-tester.h",
4882 ] + MICROKERNEL_TEST_HDRS,
4883 deps = MICROKERNEL_TEST_DEPS,
4884)
4885
4886xnnpack_unit_test(
4887 name = "f32_vmin_test",
4888 srcs = [
4889 "test/f32-vmin.cc",
4890 "test/vbinary-microkernel-tester.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 deps = MICROKERNEL_TEST_DEPS,
4893)
4894
4895xnnpack_unit_test(
4896 name = "f32_vminc_test",
4897 srcs = [
4898 "test/f32-vminc.cc",
4899 "test/vbinaryc-microkernel-tester.h",
4900 ] + MICROKERNEL_TEST_HDRS,
4901 deps = MICROKERNEL_TEST_DEPS,
4902)
4903
4904xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004905 name = "f32_vmul_test",
4906 srcs = [
4907 "test/f32-vmul.cc",
4908 "test/vbinary-microkernel-tester.h",
4909 ] + MICROKERNEL_TEST_HDRS,
4910 deps = MICROKERNEL_TEST_DEPS,
4911)
4912
4913xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004914 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004916 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004917 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004918 ] + MICROKERNEL_TEST_HDRS,
4919 deps = MICROKERNEL_TEST_DEPS,
4920)
4921
4922xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004923 name = "f32_vmulc_test",
4924 srcs = [
4925 "test/f32-vmulc.cc",
4926 "test/vbinaryc-microkernel-tester.h",
4927 ] + MICROKERNEL_TEST_HDRS,
4928 deps = MICROKERNEL_TEST_DEPS,
4929)
4930
4931xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004932 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004933 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004934 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004935 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004936 ] + MICROKERNEL_TEST_HDRS,
4937 deps = MICROKERNEL_TEST_DEPS,
4938)
4939
4940xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004941 name = "f16_vmulcaddc_minmax_test",
4942 srcs = [
4943 "test/f16-vmulcaddc-minmax.cc",
4944 "test/vmulcaddc-microkernel-tester.h",
4945 "src/xnnpack/AlignedAllocator.h",
4946 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004947 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004948)
4949
4950xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004951 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004952 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004953 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004954 "test/vmulcaddc-microkernel-tester.h",
4955 "src/xnnpack/AlignedAllocator.h",
4956 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004957 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958)
4959
4960xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004961 name = "f32_vlrelu_test",
4962 srcs = [
4963 "test/f32-vlrelu.cc",
4964 "test/vunary-microkernel-tester.h",
4965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004970 name = "f32_vneg_test",
4971 srcs = [
4972 "test/f32-vneg.cc",
4973 "test/vunary-microkernel-tester.h",
4974 ] + MICROKERNEL_TEST_HDRS,
4975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
4978xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004979 name = "f32_vrndne_test",
4980 srcs = [
4981 "test/f32-vrndne.cc",
4982 "test/vunary-microkernel-tester.h",
4983 ] + MICROKERNEL_TEST_HDRS,
4984 deps = MICROKERNEL_TEST_DEPS,
4985)
4986
4987xnnpack_unit_test(
4988 name = "f32_vrndz_test",
4989 srcs = [
4990 "test/f32-vrndz.cc",
4991 "test/vunary-microkernel-tester.h",
4992 ] + MICROKERNEL_TEST_HDRS,
4993 deps = MICROKERNEL_TEST_DEPS,
4994)
4995
4996xnnpack_unit_test(
4997 name = "f32_vrndu_test",
4998 srcs = [
4999 "test/f32-vrndu.cc",
5000 "test/vunary-microkernel-tester.h",
5001 ] + MICROKERNEL_TEST_HDRS,
5002 deps = MICROKERNEL_TEST_DEPS,
5003)
5004
5005xnnpack_unit_test(
5006 name = "f32_vrndd_test",
5007 srcs = [
5008 "test/f32-vrndd.cc",
5009 "test/vunary-microkernel-tester.h",
5010 ] + MICROKERNEL_TEST_HDRS,
5011 deps = MICROKERNEL_TEST_DEPS,
5012)
5013
5014xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005015 name = "f32_vscale_test",
5016 srcs = [
5017 "test/f32-vscale.cc",
5018 "test/vscale-microkernel-tester.h",
5019 ] + MICROKERNEL_TEST_HDRS,
5020 deps = MICROKERNEL_TEST_DEPS,
5021)
5022
5023xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005024 name = "f32_vscaleexpminusmax_test",
5025 srcs = [
5026 "test/f32-vscaleexpminusmax.cc",
5027 "test/vscaleexpminusmax-microkernel-tester.h",
5028 ] + MICROKERNEL_TEST_HDRS,
5029 deps = MICROKERNEL_TEST_DEPS,
5030)
5031
5032xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005033 name = "f32_vscaleextexp_test",
5034 srcs = [
5035 "test/f32-vscaleextexp.cc",
5036 "test/vscaleextexp-microkernel-tester.h",
5037 ] + MICROKERNEL_TEST_HDRS,
5038 deps = MICROKERNEL_TEST_DEPS,
5039)
5040
5041xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005042 name = "f32_vsqr_test",
5043 srcs = [
5044 "test/f32-vsqr.cc",
5045 "test/vunary-microkernel-tester.h",
5046 ] + MICROKERNEL_TEST_HDRS,
5047 deps = MICROKERNEL_TEST_DEPS,
5048)
5049
5050xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005051 name = "f32_vsqrdiff_test",
5052 srcs = [
5053 "test/f32-vsqrdiff.cc",
5054 "test/vbinary-microkernel-tester.h",
5055 ] + MICROKERNEL_TEST_HDRS,
5056 deps = MICROKERNEL_TEST_DEPS,
5057)
5058
5059xnnpack_unit_test(
5060 name = "f32_vsqrdiffc_test",
5061 srcs = [
5062 "test/f32-vsqrdiffc.cc",
5063 "test/vbinaryc-microkernel-tester.h",
5064 ] + MICROKERNEL_TEST_HDRS,
5065 deps = MICROKERNEL_TEST_DEPS,
5066)
5067
5068xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005069 name = "f32_vsqrt_test",
5070 srcs = [
5071 "test/f32-vsqrt.cc",
5072 "test/vunary-microkernel-tester.h",
5073 ] + MICROKERNEL_TEST_HDRS,
5074 deps = MICROKERNEL_TEST_DEPS,
5075)
5076
5077xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005078 name = "f32_vsub_test",
5079 srcs = [
5080 "test/f32-vsub.cc",
5081 "test/vbinary-microkernel-tester.h",
5082 ] + MICROKERNEL_TEST_HDRS,
5083 deps = MICROKERNEL_TEST_DEPS,
5084)
5085
5086xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005087 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005088 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005089 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005090 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005091 ] + MICROKERNEL_TEST_HDRS,
5092 deps = MICROKERNEL_TEST_DEPS,
5093)
5094
5095xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005096 name = "f32_vsubc_test",
5097 srcs = [
5098 "test/f32-vsubc.cc",
5099 "test/vbinaryc-microkernel-tester.h",
5100 ] + MICROKERNEL_TEST_HDRS,
5101 deps = MICROKERNEL_TEST_DEPS,
5102)
5103
5104xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005105 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005106 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005107 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005108 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005109 ] + MICROKERNEL_TEST_HDRS,
5110 deps = MICROKERNEL_TEST_DEPS,
5111)
5112
5113xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005114 name = "f32_vrsubc_test",
5115 srcs = [
5116 "test/f32-vrsubc.cc",
5117 "test/vbinaryc-microkernel-tester.h",
5118 ] + MICROKERNEL_TEST_HDRS,
5119 deps = MICROKERNEL_TEST_DEPS,
5120)
5121
5122xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005123 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005124 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005125 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005126 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005127 ] + MICROKERNEL_TEST_HDRS,
5128 deps = MICROKERNEL_TEST_DEPS,
5129)
5130
5131xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005132 name = "qs8_gemm_minmax_test",
5133 srcs = [
5134 "test/qs8-gemm-minmax.cc",
5135 "test/gemm-microkernel-tester.h",
5136 "src/xnnpack/AlignedAllocator.h",
5137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5138 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5139)
5140
5141xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005142 name = "qs8_igemm_minmax_test",
5143 srcs = [
5144 "test/qs8-igemm-minmax.cc",
5145 "test/gemm-microkernel-tester.h",
5146 "src/xnnpack/AlignedAllocator.h",
5147 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5148 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5149)
5150
5151xnnpack_unit_test(
5152 name = "qs8_requantization_test",
5153 srcs = [
5154 "src/xnnpack/requantization-stubs.h",
5155 "test/qs8-requantization.cc",
5156 "test/requantization-tester.h",
5157 ] + MICROKERNEL_TEST_HDRS,
5158 deps = MICROKERNEL_TEST_DEPS,
5159)
5160
5161xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005162 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005164 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 "test/avgpool-microkernel-tester.h",
5166 "src/xnnpack/AlignedAllocator.h",
5167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005172 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005174 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005175 "test/gemm-microkernel-tester.h",
5176 "src/xnnpack/AlignedAllocator.h",
5177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005178 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179)
5180
5181xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005182 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005184 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 "test/dwconv-microkernel-tester.h",
5186 "src/xnnpack/AlignedAllocator.h",
5187 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005188 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005189)
5190
5191xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005192 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005194 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 "test/gavgpool-microkernel-tester.h",
5196 "src/xnnpack/AlignedAllocator.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005202 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005204 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 "test/gemm-microkernel-tester.h",
5206 "src/xnnpack/AlignedAllocator.h",
5207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005208 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209)
5210
5211xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005212 name = "qu8_requantization_test",
5213 srcs = [
5214 "src/xnnpack/requantization-stubs.h",
5215 "test/qu8-requantization.cc",
5216 "test/requantization-tester.h",
5217 ] + MICROKERNEL_TEST_HDRS,
5218 deps = MICROKERNEL_TEST_DEPS,
5219)
5220
5221xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005222 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005224 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 "test/vadd-microkernel-tester.h",
5226 ] + MICROKERNEL_TEST_HDRS,
5227 deps = MICROKERNEL_TEST_DEPS,
5228)
5229
5230xnnpack_unit_test(
5231 name = "u8_clamp_test",
5232 srcs = [
5233 "test/u8-clamp.cc",
5234 "test/clamp-microkernel-tester.h",
5235 ] + MICROKERNEL_TEST_HDRS,
5236 deps = MICROKERNEL_TEST_DEPS,
5237)
5238
5239xnnpack_unit_test(
5240 name = "u8_lut32norm_test",
5241 srcs = [
5242 "test/u8-lut32norm.cc",
5243 "test/lut-norm-microkernel-tester.h",
5244 ] + MICROKERNEL_TEST_HDRS,
5245 deps = MICROKERNEL_TEST_DEPS,
5246)
5247
5248xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005249 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005251 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005252 "test/maxpool-microkernel-tester.h",
5253 ] + MICROKERNEL_TEST_HDRS,
5254 deps = MICROKERNEL_TEST_DEPS,
5255)
5256
5257xnnpack_unit_test(
5258 name = "u8_rmax_test",
5259 srcs = [
5260 "test/u8-rmax.cc",
5261 "test/rmax-microkernel-tester.h",
5262 ] + MICROKERNEL_TEST_HDRS,
5263 deps = MICROKERNEL_TEST_DEPS,
5264)
5265
5266xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005267 name = "x32_fill_test",
5268 srcs = [
5269 "test/x32-fill.cc",
5270 "test/fill-microkernel-tester.h",
5271 ] + MICROKERNEL_TEST_HDRS,
5272 deps = MICROKERNEL_TEST_DEPS,
5273)
5274
5275xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276 name = "x32_packx_test",
5277 srcs = [
5278 "test/x32-packx.cc",
5279 "test/pack-microkernel-tester.h",
5280 "src/xnnpack/AlignedAllocator.h",
5281 ] + MICROKERNEL_TEST_HDRS,
5282 deps = MICROKERNEL_TEST_DEPS,
5283)
5284
5285xnnpack_unit_test(
5286 name = "x32_pad_test",
5287 srcs = [
5288 "test/x32-pad.cc",
5289 "test/pad-microkernel-tester.h",
5290 ] + MICROKERNEL_TEST_HDRS,
5291 deps = MICROKERNEL_TEST_DEPS,
5292)
5293
5294xnnpack_unit_test(
5295 name = "x32_unpool_test",
5296 srcs = [
5297 "test/x32-unpool.cc",
5298 "test/unpool-microkernel-tester.h",
5299 ] + MICROKERNEL_TEST_HDRS,
5300 deps = MICROKERNEL_TEST_DEPS,
5301)
5302
5303xnnpack_unit_test(
5304 name = "x32_zip_test",
5305 srcs = [
5306 "test/x32-zip.cc",
5307 "test/zip-microkernel-tester.h",
5308 ] + MICROKERNEL_TEST_HDRS,
5309 deps = MICROKERNEL_TEST_DEPS,
5310)
5311
5312xnnpack_unit_test(
5313 name = "x8_lut_test",
5314 srcs = [
5315 "test/x8-lut.cc",
5316 "test/lut-microkernel-tester.h",
5317 ] + MICROKERNEL_TEST_HDRS,
5318 deps = MICROKERNEL_TEST_DEPS,
5319)
5320
5321xnnpack_unit_test(
5322 name = "x8_zip_test",
5323 srcs = [
5324 "test/x8-zip.cc",
5325 "test/zip-microkernel-tester.h",
5326 ] + MICROKERNEL_TEST_HDRS,
5327 deps = MICROKERNEL_TEST_DEPS,
5328)
5329
Marat Dukhan20c3b922020-03-10 03:45:06 -07005330########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331
5332xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005333 name = "operator_size_test",
5334 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005335 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336)
5337
Marat Dukhan20c3b922020-03-10 03:45:06 -07005338xnnpack_binary(
5339 name = "subgraph_size_test",
5340 srcs = ["test/subgraph-size.c"],
5341 deps = [":XNNPACK"],
5342)
5343
5344########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345
5346xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005347 name = "abs_nc_test",
5348 srcs = [
5349 "test/abs-nc.cc",
5350 "test/abs-operator-tester.h",
5351 ],
5352 deps = OPERATOR_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005356 name = "add_nd_test",
5357 srcs = [
5358 "test/add-nd.cc",
5359 "test/binary-elementwise-operator-tester.h",
5360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005361 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005362)
5363
5364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005365 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005367 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005368 "test/argmax-pooling-operator-tester.h",
5369 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005371)
5372
5373xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005374 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005376 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005377 "test/average-pooling-operator-tester.h",
5378 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380)
5381
5382xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005383 name = "bankers_rounding_nc_test",
5384 srcs = [
5385 "test/bankers-rounding-nc.cc",
5386 "test/bankers-rounding-operator-tester.h",
5387 ],
5388 deps = OPERATOR_TEST_DEPS,
5389)
5390
5391xnnpack_unit_test(
5392 name = "ceiling_nc_test",
5393 srcs = [
5394 "test/ceiling-nc.cc",
5395 "test/ceiling-operator-tester.h",
5396 ],
5397 deps = OPERATOR_TEST_DEPS,
5398)
5399
5400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005401 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005403 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 "test/channel-shuffle-operator-tester.h",
5405 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005407)
5408
5409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005410 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005412 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 "test/clamp-operator-tester.h",
5414 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416)
5417
5418xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005419 name = "constant_pad_nd_test",
5420 srcs = [
5421 "test/constant-pad-nd.cc",
5422 "test/constant-pad-operator-tester.h",
5423 ],
5424 deps = OPERATOR_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005428 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005430 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005431 "test/convolution-operator-tester.h",
5432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005434)
5435
5436xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005437 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005439 "test/convolution-nchw.cc",
5440 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005441 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005442 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443)
5444
5445xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005446 name = "copy_nc_test",
5447 srcs = [
5448 "test/copy-nc.cc",
5449 "test/copy-operator-tester.h",
5450 ],
5451 deps = OPERATOR_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005455 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005456 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005457 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005458 "test/deconvolution-operator-tester.h",
5459 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005460 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461)
5462
5463xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005464 name = "divide_nd_test",
5465 srcs = [
5466 "test/binary-elementwise-operator-tester.h",
5467 "test/divide-nd.cc",
5468 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005470)
5471
5472xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005473 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005475 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 "test/fully-connected-operator-tester.h",
5477 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479)
5480
5481xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005482 name = "floor_nc_test",
5483 srcs = [
5484 "test/floor-nc.cc",
5485 "test/floor-operator-tester.h",
5486 ],
5487 deps = OPERATOR_TEST_DEPS,
5488)
5489
5490xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005491 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005493 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005494 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005495 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005496 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497)
5498
5499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005500 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005502 "test/global-average-pooling-ncw.cc",
5503 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005506)
5507
5508xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005509 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005511 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005512 "test/hardswish-operator-tester.h",
5513 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005514 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005515)
5516
5517xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005518 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005519 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005520 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005521 "test/leaky-relu-operator-tester.h",
5522 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005523 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005524)
5525
5526xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005527 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005528 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005529 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005530 "test/max-pooling-operator-tester.h",
5531 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533)
5534
5535xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005536 name = "maximum_nd_test",
5537 srcs = [
5538 "test/binary-elementwise-operator-tester.h",
5539 "test/maximum-nd.cc",
5540 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005541 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005542)
5543
5544xnnpack_unit_test(
5545 name = "minimum_nd_test",
5546 srcs = [
5547 "test/binary-elementwise-operator-tester.h",
5548 "test/minimum-nd.cc",
5549 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005550 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005551)
5552
5553xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005554 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005555 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005556 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005557 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005558 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005559 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005560)
5561
5562xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005563 name = "negate_nc_test",
5564 srcs = [
5565 "test/negate-nc.cc",
5566 "test/negate-operator-tester.h",
5567 ],
5568 deps = OPERATOR_TEST_DEPS,
5569)
5570
5571xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005572 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005574 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005575 "test/prelu-operator-tester.h",
5576 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005577 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578)
5579
5580xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005581 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005582 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005583 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005584 "test/resize-bilinear-operator-tester.h",
5585 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005586 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005587)
5588
5589xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005590 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005591 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005592 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005593 "test/sigmoid-operator-tester.h",
5594 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005595 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005596)
5597
5598xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005599 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005601 "test/softmax-nc.cc",
5602 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005603 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005604 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005605)
5606
5607xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005608 name = "square_nc_test",
5609 srcs = [
5610 "test/square-nc.cc",
5611 "test/square-operator-tester.h",
5612 ],
5613 deps = OPERATOR_TEST_DEPS,
5614)
5615
5616xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005617 name = "square_root_nc_test",
5618 srcs = [
5619 "test/square-root-nc.cc",
5620 "test/square-root-operator-tester.h",
5621 ],
5622 deps = OPERATOR_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005626 name = "squared_difference_nd_test",
5627 srcs = [
5628 "test/binary-elementwise-operator-tester.h",
5629 "test/squared-difference-nd.cc",
5630 ],
5631 deps = OPERATOR_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005635 name = "subtract_nd_test",
5636 srcs = [
5637 "test/binary-elementwise-operator-tester.h",
5638 "test/subtract-nd.cc",
5639 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005640 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005641)
5642
5643xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005644 name = "truncation_nc_test",
5645 srcs = [
5646 "test/truncation-nc.cc",
5647 "test/truncation-operator-tester.h",
5648 ],
5649 deps = OPERATOR_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005653 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005655 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005656 "test/unpooling-operator-tester.h",
5657 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005658 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005659)
5660
Chao Mei6ddfc602020-05-13 22:29:36 -07005661############################### Misc unit tests ###############################
5662
5663xnnpack_unit_test(
5664 name = "memory_planner_test",
5665 srcs = [
5666 "test/memory-planner-test.cc",
5667 ],
5668 deps = [
5669 ":XNNPACK",
5670 ":memory_planner",
5671 ],
5672)
5673
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674############################# Build configurations #############################
5675
Marat Dukhanb8642352019-10-30 15:43:02 -07005676# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005678 name = "xnn_enable_assembly_explicit_true",
5679 define_values = {"xnn_enable_assembly": "true"},
5680)
5681
5682# Disables usage of assembly kernels.
5683config_setting(
5684 name = "xnn_enable_assembly_explicit_false",
5685 define_values = {"xnn_enable_assembly": "false"},
5686)
5687
Marat Dukhan9de90e02020-06-18 16:04:12 -07005688# Enables usage of sparse inference.
5689config_setting(
5690 name = "xnn_enable_sparse_explicit_true",
5691 define_values = {"xnn_enable_sparse": "true"},
5692)
5693
5694# Disables usage of sparse inference.
5695config_setting(
5696 name = "xnn_enable_sparse_explicit_false",
5697 define_values = {"xnn_enable_sparse": "false"},
5698)
5699
Marat Dukhan05702cf2020-03-26 15:41:33 -07005700# Disables usage of HMP-aware optimizations.
5701config_setting(
5702 name = "xnn_enable_hmp_explicit_false",
5703 define_values = {"xnn_enable_hmp": "false"},
5704)
5705
Chao Mei6ddfc602020-05-13 22:29:36 -07005706# Enable usage of optimized memory allocation
5707config_setting(
5708 name = "xnn_enable_memopt_explicit_true",
5709 define_values = {"xnnpack_enable_memopt": "true"},
5710)
5711
5712# Disable usage of optimized memory allocation
5713config_setting(
5714 name = "xnn_enable_memopt_explicit_false",
5715 define_values = {"xnnpack_enable_memopt": "false"},
5716)
5717
Marat Dukhanb8642352019-10-30 15:43:02 -07005718# Builds with -c dbg
5719config_setting(
5720 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005721 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005722 "compilation_mode": "dbg",
5723 },
5724)
5725
5726# Builds with -c opt
5727config_setting(
5728 name = "optimized_build",
5729 values = {
5730 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005731 },
5732)
5733
5734config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005735 name = "linux_k8",
5736 values = {"cpu": "k8"},
5737)
5738
5739config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005740 name = "linux_arm",
5741 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005742)
5743
5744config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005745 name = "linux_armeabi",
5746 values = {"cpu": "armeabi"},
5747)
5748
5749config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005750 name = "linux_armhf",
5751 values = {"cpu": "armhf"},
5752)
5753
5754config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005755 name = "linux_armv7a",
5756 values = {"cpu": "armv7a"},
5757)
5758
5759config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005760 name = "linux_aarch64",
5761 values = {"cpu": "aarch64"},
5762)
5763
5764config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005765 name = "android",
5766 values = {"crosstool_top": "//external:android/crosstool"},
5767)
5768
5769config_setting(
5770 name = "android_armv7",
5771 values = {
5772 "crosstool_top": "//external:android/crosstool",
5773 "cpu": "armeabi-v7a",
5774 },
5775)
5776
5777config_setting(
5778 name = "android_arm64",
5779 values = {
5780 "crosstool_top": "//external:android/crosstool",
5781 "cpu": "arm64-v8a",
5782 },
5783)
5784
5785config_setting(
5786 name = "android_x86",
5787 values = {
5788 "crosstool_top": "//external:android/crosstool",
5789 "cpu": "x86",
5790 },
5791)
5792
5793config_setting(
5794 name = "android_x86_64",
5795 values = {
5796 "crosstool_top": "//external:android/crosstool",
5797 "cpu": "x86_64",
5798 },
5799)
5800
5801config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005802 name = "windows_x86_64",
5803 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005804)
5805
5806config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005807 name = "windows_x86_64_clang",
5808 values = {
5809 "compiler": "clang-cl",
5810 "cpu": "x64_windows",
5811 },
5812)
5813
5814config_setting(
5815 name = "windows_x86_64_mingw",
5816 values = {
5817 "compiler": "mingw-gcc",
5818 "cpu": "x64_windows",
5819 },
5820)
5821
5822config_setting(
5823 name = "windows_x86_64_msys",
5824 values = {
5825 "compiler": "msys-gcc",
5826 "cpu": "x64_windows",
5827 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005828)
5829
5830config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005831 name = "macos_x86_64",
5832 values = {
5833 "apple_platform_type": "macos",
5834 "cpu": "darwin",
5835 },
5836)
5837
5838config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005839 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005840 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005841)
5842
5843config_setting(
5844 name = "emscripten_wasm",
5845 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005846 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005847 "cpu": "wasm",
5848 },
5849)
5850
5851config_setting(
5852 name = "emscripten_wasmsimd",
5853 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005854 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005856 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005857 },
5858)
5859
5860config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005861 name = "ios_armv7",
5862 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005863 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005864 "cpu": "ios_armv7",
5865 },
5866)
5867
5868config_setting(
5869 name = "ios_arm64",
5870 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005871 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005872 "cpu": "ios_arm64",
5873 },
5874)
5875
5876config_setting(
5877 name = "ios_arm64e",
5878 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005879 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005880 "cpu": "ios_arm64e",
5881 },
5882)
5883
5884config_setting(
5885 name = "ios_x86",
5886 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005887 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005888 "cpu": "ios_i386",
5889 },
5890)
5891
5892config_setting(
5893 name = "ios_x86_64",
5894 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005895 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005896 "cpu": "ios_x86_64",
5897 },
5898)
5899
5900config_setting(
5901 name = "watchos_armv7k",
5902 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005903 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005904 "cpu": "watchos_armv7k",
5905 },
5906)
5907
5908config_setting(
5909 name = "watchos_arm64_32",
5910 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005911 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005912 "cpu": "watchos_arm64_32",
5913 },
5914)
5915
5916config_setting(
5917 name = "watchos_x86",
5918 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005919 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005920 "cpu": "watchos_i386",
5921 },
5922)
5923
5924config_setting(
5925 name = "watchos_x86_64",
5926 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005927 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005928 "cpu": "watchos_x86_64",
5929 },
5930)
5931
5932config_setting(
5933 name = "tvos_arm64",
5934 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005935 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005936 "cpu": "tvos_arm64",
5937 },
5938)
5939
5940config_setting(
5941 name = "tvos_x86_64",
5942 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005943 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005944 "cpu": "tvos_x86_64",
5945 },
5946)