blob: 6fa565aacf8483f85d5d9d5d655a9f0481d5f1c4 [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 Dukhan595e1702020-07-31 10:12:52 -07001643 "src/qs8-gemm/gen/1x4c2-minmax-sse2.c",
1644 "src/qs8-gemm/gen/4x4c2-minmax-sse2.c",
Marat Dukhanf9480682020-07-31 14:50:24 -07001645 "src/qs8-igemm/gen/1x4c2-minmax-sse2.c",
1646 "src/qs8-igemm/gen/4x4c2-minmax-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001647 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1648 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1649 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1650 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1651 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1652 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1653 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1654 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001655 "src/qu8-requantization/precise-sse2.c",
1656 "src/qu8-requantization/fp32-sse2.c",
1657 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001658 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001659 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001660 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001661 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001662 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663 "src/x32-zip/x2-sse2.c",
1664 "src/x32-zip/x3-sse2.c",
1665 "src/x32-zip/x4-sse2.c",
1666 "src/x32-zip/xm-sse2.c",
1667 "src/x8-zip/x2-sse2.c",
1668 "src/x8-zip/x3-sse2.c",
1669 "src/x8-zip/x4-sse2.c",
1670 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001671 "src/math/exp-sse2-p5.c",
1672 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001673 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001674 "src/math/roundd-sse2-cvt.c",
1675 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001676 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001677 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001678]
1679
1680SSSE3_UKERNELS = [
Marat Dukhan595e1702020-07-31 10:12:52 -07001681 "src/qs8-gemm/gen/1x4c2-minmax-ssse3.c",
1682 "src/qs8-gemm/gen/4x4c2-minmax-ssse3.c",
Marat Dukhanf9480682020-07-31 14:50:24 -07001683 "src/qs8-igemm/gen/1x4c2-minmax-ssse3.c",
1684 "src/qs8-igemm/gen/4x4c2-minmax-ssse3.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001685 "src/qs8-requantization/precise-ssse3.c",
1686 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001687 "src/qu8-requantization/precise-ssse3.c",
1688 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001689]
1690
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001691SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001692 "src/f32-prelu/gen/sse41-2x4.c",
1693 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001694 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001695 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001696 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001697 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001698 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1699 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001700 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1701 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001702 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1703 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1704 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1705 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1706 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1707 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1708 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1709 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan595e1702020-07-31 10:12:52 -07001710 "src/qs8-gemm/gen/1x4c2-minmax-sse41.c",
1711 "src/qs8-gemm/gen/4x4c2-minmax-sse41.c",
Marat Dukhanf9480682020-07-31 14:50:24 -07001712 "src/qs8-igemm/gen/1x4c2-minmax-sse41.c",
1713 "src/qs8-igemm/gen/4x4c2-minmax-sse41.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001714 "src/qs8-requantization/fp32-sse4.c",
1715 "src/qs8-requantization/precise-sse4.c",
1716 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001717 "src/qu8-requantization/precise-sse4.c",
1718 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001719 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001720 "src/math/roundd-sse41.c",
1721 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001722 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001723]
1724
Marat Dukhan08c4a432019-10-03 09:29:21 -07001725AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001726 "src/f32-clamp/gen/avx-x8.c",
1727 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001728 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1729 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1730 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1731 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1732 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1733 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1734 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1735 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1736 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1737 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1738 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1739 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1740 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1741 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1742 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1743 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1744 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1745 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1746 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1747 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1748 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1749 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1750 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1751 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1752 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1753 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1754 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1755 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1756 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1757 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001758 "src/f32-hswish/gen/hswish-avx-x8.c",
1759 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001760 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1761 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1762 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1763 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1764 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1765 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1766 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1767 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1768 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001769 "src/f32-prelu/gen/avx-2x8.c",
1770 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001771 "src/f32-relu/gen/avx-x16.c",
1772 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001774 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1775 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1776 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1777 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1778 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1779 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1780 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1781 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001782 "src/f32-vbinary/gen/vmax-avx-x8.c",
1783 "src/f32-vbinary/gen/vmax-avx-x16.c",
1784 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1785 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1786 "src/f32-vbinary/gen/vmin-avx-x8.c",
1787 "src/f32-vbinary/gen/vmin-avx-x16.c",
1788 "src/f32-vbinary/gen/vminc-avx-x8.c",
1789 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001790 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1791 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1792 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1793 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1794 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1795 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1796 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1797 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001798 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1799 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1800 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1801 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001802 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1803 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1804 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1805 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001806 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1807 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001808 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001809 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1810 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1811 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1812 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1813 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1814 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1815 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1816 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001817 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1818 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001819 "src/f32-vunary/gen/vabs-avx-x8.c",
1820 "src/f32-vunary/gen/vabs-avx-x16.c",
1821 "src/f32-vunary/gen/vneg-avx-x8.c",
1822 "src/f32-vunary/gen/vneg-avx-x16.c",
1823 "src/f32-vunary/gen/vsqr-avx-x8.c",
1824 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825]
1826
Marat Dukhanfda12b82019-11-21 12:27:59 -08001827FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001828 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1829 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1830 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1831 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1832 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1833 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1834 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1835 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1836 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1837 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1838 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1839 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1840 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1841 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1842 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1843 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1844 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1845 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1846 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1847 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1848 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1849 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1850 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1856 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1857 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1858 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1859 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1860 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1861 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1862 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1863 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1864 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1865 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1866 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1867 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001868 "src/f32-hswish/gen/hswish-fma3-x8.c",
1869 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001870 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1871 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1872 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1873 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1874 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1875 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1876 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1877 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1878 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1879 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1880 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1881 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1882 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1883 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001884 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1885 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1886 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1887 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1888 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1889 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1890 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1891 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001892 "src/math/sqrt-fma3-nr1fma.c",
1893 "src/math/sqrt-fma3-nr2fma.c",
1894 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001895]
1896
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001897AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001898 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1899 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1900 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1901 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1902 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1903 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1904 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1905 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1906 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1907 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1908 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1909 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1910 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1911 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1912 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1913 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1914 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1915 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1916 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1917 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1918 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1919 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1920 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1921 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1922 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1923 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1924 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1925 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1926 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1927 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1928 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1929 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1930 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1931 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1932 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1933 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001934 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1935 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1936 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1937 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1938 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1939 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1940 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1941 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1942 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1943 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1944 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1952 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1953 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1954 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1955 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1956 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1957 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1958 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1959 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1960 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1961 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1962 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1963 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001964 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1965 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1966 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1967 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1968 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1969 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1970 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1971 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1972 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1973 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1974 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1975 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1976 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1977 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1978 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1979 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1980 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1981 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1982 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1983 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1984 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1985 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1986 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1987 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001988 "src/math/exp-avx2-p5.c",
1989 "src/math/exp-avx2-perm-p3.c",
1990 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001991 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001992 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001993 "src/math/sigmoid-avx2-rr2-p5-div.c",
1994 "src/math/sigmoid-avx2-rr1-p5-div.c",
1995 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1996 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1997 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1998 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001999]
2000
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002002 "src/f32-clamp/gen/avx512f-x16.c",
2003 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002004 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2005 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2006 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2007 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2008 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2009 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2010 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2011 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2012 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2013 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2014 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2015 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2016 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2017 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2018 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2019 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2020 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2021 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2022 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2023 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2024 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2025 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2026 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2027 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002028 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2029 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002030 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2031 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2032 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2033 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2034 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2035 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002036 "src/f32-prelu/gen/avx512f-2x16.c",
2037 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002038 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2039 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2040 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2041 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2042 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2043 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2044 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2045 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2046 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2047 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2048 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2049 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2050 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2051 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2052 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2053 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2054 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2055 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2056 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2057 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2058 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2059 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2060 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2061 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2062 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2063 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2064 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2065 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2066 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2067 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2068 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2069 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2070 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2071 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2072 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2073 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002074 "src/f32-relu/gen/avx512f-x16.c",
2075 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002076 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002077 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2078 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2079 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2080 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2081 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2082 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2083 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2084 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002085 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2086 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2087 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2088 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2089 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2090 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2091 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2092 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002093 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2094 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2095 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2096 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2097 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2098 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2099 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2100 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002101 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2102 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2103 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2104 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002105 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2106 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2107 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2108 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002109 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2110 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002111 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002112 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2113 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2114 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2115 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2116 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2117 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2118 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2119 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2120 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2121 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2122 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2123 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2124 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2125 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2126 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2127 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2128 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2129 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2130 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2131 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2132 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2133 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2134 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2135 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002136 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2137 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2138 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2139 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2140 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2141 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2142 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2143 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002144 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2145 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2146 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2147 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2148 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2149 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2150 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2151 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002152 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2153 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2154 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2155 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2156 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2157 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002158 "src/math/exp-avx512f-p5-scalef.c",
2159 "src/math/exp-avx512f-p5.c",
2160 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002161 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002162 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002163 "src/math/sqrt-avx512f-nr1fma.c",
2164 "src/math/sqrt-avx512f-nr2fma.c",
2165 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
2168AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002169 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002170 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002171 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2172 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002173 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002174 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002175 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002176 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002177 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2178 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002179 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2180 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2181 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2182 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183]
2184
2185AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002186 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002187 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2188 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002189 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002190 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2191 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002192 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2193 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2194 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002195 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002196 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2197 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2198 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002199 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002200 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2201 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002202 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002203 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2204 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2205 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2206 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2207 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2208 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2209 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2210 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2211 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2212 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2213 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2214 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2216 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2217 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2218 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2219 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2220 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002221 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2222 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002223 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002224 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2225 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2226 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2227 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2228 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2229 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2230 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2231 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2232 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2233 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2234 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2235 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2236 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2237 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2238 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2239 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2240 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2241 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002242 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2243 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2244 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2245 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2246 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2247 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2248 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2249 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2250 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2251 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2252 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2253 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2254 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2255 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2256 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2257 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2258 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2259 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002260]
2261
Marat Dukhan1b354632020-03-23 12:50:22 -07002262INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002263 "src/xnnpack/argmaxpool.h",
2264 "src/xnnpack/avgpool.h",
2265 "src/xnnpack/clamp.h",
2266 "src/xnnpack/common.h",
2267 "src/xnnpack/conv.h",
2268 "src/xnnpack/dwconv.h",
2269 "src/xnnpack/gavgpool.h",
2270 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002271 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002272 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002273 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002274 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002275 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276 "src/xnnpack/lut.h",
2277 "src/xnnpack/math.h",
2278 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002279 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002280 "src/xnnpack/packx.h",
2281 "src/xnnpack/pad.h",
2282 "src/xnnpack/params.h",
2283 "src/xnnpack/pavgpool.h",
2284 "src/xnnpack/ppmm.h",
2285 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002286 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002287 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002288 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "src/xnnpack/rmax.h",
2290 "src/xnnpack/scalar-utils.h",
2291 "src/xnnpack/spmm.h",
2292 "src/xnnpack/unpool.h",
2293 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002294 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002295 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002296 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002297 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002298 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002299 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002301]
2302
2303INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002304 "include/xnnpack.h",
2305 "src/xnnpack/allocator.h",
2306 "src/xnnpack/compute.h",
2307 "src/xnnpack/im2col.h",
2308 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002309 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002310 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002311 "src/xnnpack/operator.h",
2312 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002313 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002314 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002315 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002316 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002317]
2318
Marat Dukhan1b354632020-03-23 12:50:22 -07002319ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002320 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002321]
2322
Marat Dukhan1b354632020-03-23 12:50:22 -07002323MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002324 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325 "include/xnnpack.h",
2326]
2327
Marat Dukhan1b354632020-03-23 12:50:22 -07002328MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002330 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002331 "src/xnnpack/requantization.h",
2332 "include/xnnpack.h",
2333]
2334
2335OPERATOR_TEST_PARAMS_HDRS = [
2336 "src/xnnpack/params.h",
2337 "src/xnnpack/common.h",
2338]
2339
2340WEIGHTS_PACK_HDRS = [
2341 "src/xnnpack/pack.h",
2342 "src/xnnpack/operator.h",
2343 "src/xnnpack/compute.h",
2344]
2345
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002346LOGGING_COPTS = select({
2347 # No logging in optimized mode
2348 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2349 # Full logging in debug mode
2350 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2351 # Error-only logging in default (fastbuild) mode
2352 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2353})
2354
Marat Dukhan3b59de22020-06-03 20:15:19 -07002355LOGGING_SRCS = select({
2356 # No logging in optimized mode
2357 ":optimized_build": [],
2358 "//conditions:default": [
2359 "src/operator-strings.c",
2360 "src/subgraph-strings.c",
2361 ],
2362})
2363
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002364LOGGING_HDRS = [
2365 "src/xnnpack/log.h",
2366]
2367
Marat Dukhan08c4a432019-10-03 09:29:21 -07002368xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002369 name = "tables",
2370 srcs = TABLE_SRCS,
2371 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002372 gcc_copts = xnnpack_gcc_std_copts(),
2373 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002374)
2375
2376xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 name = "scalar_ukernels",
2378 srcs = SCALAR_UKERNELS,
2379 hdrs = INTERNAL_HDRS,
2380 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002381 gcc_copts = xnnpack_gcc_std_copts(),
2382 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002383 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002384 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002385 "@FP16",
2386 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002387 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388 ],
2389)
2390
2391xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002392 name = "scalar_ukernels_test_mode",
2393 srcs = SCALAR_UKERNELS,
2394 hdrs = INTERNAL_HDRS,
2395 aarch32_copts = ["-marm"],
2396 copts = [
2397 "-UNDEBUG",
2398 "-DXNN_TEST_MODE=1",
2399 ],
2400 gcc_copts = xnnpack_gcc_std_copts(),
2401 msvc_copts = xnnpack_msvc_std_copts(),
2402 deps = [
2403 ":tables",
2404 "@FP16",
2405 "@FXdiv",
2406 "@pthreadpool",
2407 ],
2408)
2409
2410xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002411 name = "wasm_ukernels",
2412 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002413 gcc_copts = xnnpack_gcc_std_copts(),
2414 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002415 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002416 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002417 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002418 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002419 "@FP16",
2420 "@FXdiv",
2421 "@pthreadpool",
2422 ],
2423)
2424
2425xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002426 name = "wasm_ukernels_test_mode",
2427 hdrs = INTERNAL_HDRS,
2428 copts = [
2429 "-UNDEBUG",
2430 "-DXNN_TEST_MODE=1",
2431 ],
2432 gcc_copts = xnnpack_gcc_std_copts(),
2433 msvc_copts = xnnpack_msvc_std_copts(),
2434 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002435 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002436 deps = [
2437 ":tables",
2438 "@FP16",
2439 "@FXdiv",
2440 "@pthreadpool",
2441 ],
2442)
2443
2444xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002445 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446 hdrs = INTERNAL_HDRS,
2447 aarch32_copts = [
2448 "-marm",
2449 "-mfpu=neon",
2450 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002451 gcc_copts = xnnpack_gcc_std_copts(),
2452 gcc_x86_copts = ["-msse2"],
2453 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 optimized_copts = [
2455 "-O3",
2456 "-ffast-math",
2457 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002458 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002459 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002460 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002461 "@FP16",
2462 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002463 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002464 ],
2465)
2466
2467xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002468 name = "psimd_fastmath_ukernels_test_mode",
2469 hdrs = INTERNAL_HDRS,
2470 aarch32_copts = [
2471 "-marm",
2472 "-mfpu=neon",
2473 ],
2474 copts = [
2475 "-UNDEBUG",
2476 "-DXNN_TEST_MODE=1",
2477 ],
2478 gcc_copts = xnnpack_gcc_std_copts(),
2479 gcc_x86_copts = ["-msse2"],
2480 msvc_copts = xnnpack_msvc_std_copts(),
2481 optimized_copts = [
2482 "-O3",
2483 "-ffast-math",
2484 ],
2485 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2486 deps = [
2487 ":tables",
2488 "@FP16",
2489 "@psimd",
2490 "@pthreadpool",
2491 ],
2492)
2493
2494xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002495 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002496 hdrs = INTERNAL_HDRS,
2497 aarch32_copts = [
2498 "-marm",
2499 "-mfpu=neon",
2500 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002501 gcc_copts = xnnpack_gcc_std_copts(),
2502 gcc_x86_copts = ["-msse2"],
2503 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002504 optimized_copts = [
2505 "-O3",
2506 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002507 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002508 deps = [
2509 ":tables",
2510 "@FP16",
2511 "@psimd",
2512 "@pthreadpool",
2513 ],
2514)
2515
2516xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002517 name = "psimd_accmath_ukernels_test_mode",
2518 hdrs = INTERNAL_HDRS,
2519 aarch32_copts = [
2520 "-marm",
2521 "-mfpu=neon",
2522 ],
2523 copts = [
2524 "-UNDEBUG",
2525 "-DXNN_TEST_MODE=1",
2526 ],
2527 gcc_copts = xnnpack_gcc_std_copts(),
2528 gcc_x86_copts = ["-msse2"],
2529 msvc_copts = xnnpack_msvc_std_copts(),
2530 optimized_copts = [
2531 "-O3",
2532 ],
2533 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2534 deps = [
2535 ":tables",
2536 "@FP16",
2537 "@psimd",
2538 "@pthreadpool",
2539 ],
2540)
2541
2542xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002543 name = "neon_ukernels",
2544 hdrs = INTERNAL_HDRS,
2545 aarch32_copts = [
2546 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002547 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002548 "-mfpu=neon",
2549 ],
2550 aarch32_srcs = NEON_UKERNELS,
2551 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002552 gcc_copts = xnnpack_gcc_std_copts(),
2553 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002554 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002555 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002556 "@FP16",
2557 "@pthreadpool",
2558 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002559)
2560
2561xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002562 name = "neon_ukernels_test_mode",
2563 hdrs = INTERNAL_HDRS,
2564 aarch32_copts = [
2565 "-marm",
2566 "-march=armv7-a",
2567 "-mfpu=neon",
2568 ],
2569 aarch32_srcs = NEON_UKERNELS,
2570 aarch64_srcs = NEON_UKERNELS,
2571 copts = [
2572 "-UNDEBUG",
2573 "-DXNN_TEST_MODE=1",
2574 ],
2575 gcc_copts = xnnpack_gcc_std_copts(),
2576 msvc_copts = xnnpack_msvc_std_copts(),
2577 deps = [
2578 ":tables",
2579 "@FP16",
2580 "@pthreadpool",
2581 ],
2582)
2583
2584xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002585 name = "neonfma_ukernels",
2586 hdrs = INTERNAL_HDRS,
2587 aarch32_copts = [
2588 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002589 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002590 "-mfpu=neon-vfpv4",
2591 ],
2592 aarch32_srcs = NEONFMA_UKERNELS,
2593 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002594 apple_aarch32_copts = [
2595 "-mcpu=swift",
2596 "-mtune=generic",
2597 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002598 gcc_copts = xnnpack_gcc_std_copts(),
2599 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002600 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002601 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002602 "@FP16",
2603 "@pthreadpool",
2604 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605)
2606
2607xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002608 name = "neonfma_ukernels_test_mode",
2609 hdrs = INTERNAL_HDRS,
2610 aarch32_copts = [
2611 "-marm",
2612 "-march=armv7-a",
2613 "-mfpu=neon-vfpv4",
2614 ],
2615 aarch32_srcs = NEONFMA_UKERNELS,
2616 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002617 apple_aarch32_copts = [
2618 "-mcpu=swift",
2619 "-mtune=generic",
2620 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002621 copts = [
2622 "-UNDEBUG",
2623 "-DXNN_TEST_MODE=1",
2624 ],
2625 gcc_copts = xnnpack_gcc_std_copts(),
2626 msvc_copts = xnnpack_msvc_std_copts(),
2627 deps = [
2628 ":tables",
2629 "@FP16",
2630 "@pthreadpool",
2631 ],
2632)
2633
2634xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002635 name = "neonv8_ukernels",
2636 hdrs = INTERNAL_HDRS,
2637 aarch32_copts = [
2638 "-marm",
2639 "-march=armv8-a",
2640 "-mfpu=neon-fp-armv8",
2641 ],
2642 aarch32_srcs = NEONV8_UKERNELS,
2643 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002644 apple_aarch32_copts = [
2645 "-mcpu=cyclone",
2646 "-mtune=generic",
2647 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002648 gcc_copts = xnnpack_gcc_std_copts(),
2649 msvc_copts = xnnpack_msvc_std_copts(),
2650 deps = [
2651 ":tables",
2652 "@FP16",
2653 "@pthreadpool",
2654 ],
2655)
2656
2657xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002658 name = "neonv8_ukernels_test_mode",
2659 hdrs = INTERNAL_HDRS,
2660 aarch32_copts = [
2661 "-marm",
2662 "-march=armv8-a",
2663 "-mfpu=neon-fp-armv8",
2664 ],
2665 aarch32_srcs = NEONV8_UKERNELS,
2666 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002667 apple_aarch32_copts = [
2668 "-mcpu=cyclone",
2669 "-mtune=generic",
2670 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002671 copts = [
2672 "-UNDEBUG",
2673 "-DXNN_TEST_MODE=1",
2674 ],
2675 gcc_copts = xnnpack_gcc_std_copts(),
2676 msvc_copts = xnnpack_msvc_std_copts(),
2677 deps = [
2678 ":tables",
2679 "@FP16",
2680 "@pthreadpool",
2681 ],
2682)
2683
2684xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 name = "neonfp16arith_ukernels",
2686 hdrs = INTERNAL_HDRS,
2687 aarch64_copts = ["-march=armv8.2-a+fp16"],
2688 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002689 gcc_copts = xnnpack_gcc_std_copts(),
2690 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002691 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002692 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002693 "@FP16",
2694 "@pthreadpool",
2695 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002696)
2697
2698xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002699 name = "neonfp16arith_ukernels_test_mode",
2700 hdrs = INTERNAL_HDRS,
2701 aarch64_copts = ["-march=armv8.2-a+fp16"],
2702 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2703 copts = [
2704 "-UNDEBUG",
2705 "-DXNN_TEST_MODE=1",
2706 ],
2707 gcc_copts = xnnpack_gcc_std_copts(),
2708 msvc_copts = xnnpack_msvc_std_copts(),
2709 deps = [
2710 ":tables",
2711 "@FP16",
2712 "@pthreadpool",
2713 ],
2714)
2715
2716xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002717 name = "sse2_ukernels",
2718 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002719 gcc_copts = xnnpack_gcc_std_copts(),
2720 gcc_x86_copts = ["-msse2"],
2721 msvc_copts = xnnpack_msvc_std_copts(),
2722 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002723 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002724 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002725 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002726 "@FP16",
2727 "@pthreadpool",
2728 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729)
2730
2731xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002732 name = "sse2_ukernels_test_mode",
2733 hdrs = INTERNAL_HDRS,
2734 copts = [
2735 "-UNDEBUG",
2736 "-DXNN_TEST_MODE=1",
2737 ],
2738 gcc_copts = xnnpack_gcc_std_copts(),
2739 gcc_x86_copts = ["-msse2"],
2740 msvc_copts = xnnpack_msvc_std_copts(),
2741 msvc_x86_32_copts = ["/arch:SSE2"],
2742 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2743 deps = [
2744 ":tables",
2745 "@FP16",
2746 "@pthreadpool",
2747 ],
2748)
2749
2750xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002751 name = "ssse3_ukernels",
2752 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002753 gcc_copts = xnnpack_gcc_std_copts(),
2754 gcc_x86_copts = ["-mssse3"],
2755 msvc_copts = xnnpack_msvc_std_copts(),
2756 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002757 x86_srcs = SSSE3_UKERNELS,
2758 deps = [
2759 ":tables",
2760 "@FP16",
2761 "@pthreadpool",
2762 ],
2763)
2764
2765xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002766 name = "ssse3_ukernels_test_mode",
2767 hdrs = INTERNAL_HDRS,
2768 copts = [
2769 "-UNDEBUG",
2770 "-DXNN_TEST_MODE=1",
2771 ],
2772 gcc_copts = xnnpack_gcc_std_copts(),
2773 gcc_x86_copts = ["-mssse3"],
2774 msvc_copts = xnnpack_msvc_std_copts(),
2775 msvc_x86_32_copts = ["/arch:SSE2"],
2776 x86_srcs = SSSE3_UKERNELS,
2777 deps = [
2778 ":tables",
2779 "@FP16",
2780 "@pthreadpool",
2781 ],
2782)
2783
2784xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002785 name = "sse41_ukernels",
2786 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002787 gcc_copts = xnnpack_gcc_std_copts(),
2788 gcc_x86_copts = ["-msse4.1"],
2789 msvc_copts = xnnpack_msvc_std_copts(),
2790 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002791 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002792 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002793 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002794 "@FP16",
2795 "@pthreadpool",
2796 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002797)
2798
2799xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002800 name = "sse41_ukernels_test_mode",
2801 hdrs = INTERNAL_HDRS,
2802 copts = [
2803 "-UNDEBUG",
2804 "-DXNN_TEST_MODE=1",
2805 ],
2806 gcc_copts = xnnpack_gcc_std_copts(),
2807 gcc_x86_copts = ["-msse4.1"],
2808 msvc_copts = xnnpack_msvc_std_copts(),
2809 msvc_x86_32_copts = ["/arch:SSE2"],
2810 x86_srcs = SSE41_UKERNELS,
2811 deps = [
2812 ":tables",
2813 "@FP16",
2814 "@pthreadpool",
2815 ],
2816)
2817
2818xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 name = "avx_ukernels",
2820 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002821 gcc_copts = xnnpack_gcc_std_copts(),
2822 gcc_x86_copts = ["-mavx"],
2823 msvc_copts = xnnpack_msvc_std_copts(),
2824 msvc_x86_32_copts = ["/arch:AVX"],
2825 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002828 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002829 "@FP16",
2830 "@pthreadpool",
2831 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002832)
2833
2834xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002835 name = "avx_ukernels_test_mode",
2836 hdrs = INTERNAL_HDRS,
2837 copts = [
2838 "-UNDEBUG",
2839 "-DXNN_TEST_MODE=1",
2840 ],
2841 gcc_copts = xnnpack_gcc_std_copts(),
2842 gcc_x86_copts = ["-mavx"],
2843 msvc_copts = xnnpack_msvc_std_copts(),
2844 msvc_x86_32_copts = ["/arch:AVX"],
2845 msvc_x86_64_copts = ["/arch:AVX"],
2846 x86_srcs = AVX_UKERNELS,
2847 deps = [
2848 ":tables",
2849 "@FP16",
2850 "@pthreadpool",
2851 ],
2852)
2853
2854xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002855 name = "fma3_ukernels",
2856 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002857 gcc_copts = xnnpack_gcc_std_copts(),
2858 gcc_x86_copts = ["-mfma"],
2859 msvc_copts = xnnpack_msvc_std_copts(),
2860 msvc_x86_32_copts = ["/arch:AVX"],
2861 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002862 x86_srcs = FMA3_UKERNELS,
2863 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002864 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002865 "@FP16",
2866 "@pthreadpool",
2867 ],
2868)
2869
2870xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002871 name = "fma3_ukernels_test_mode",
2872 hdrs = INTERNAL_HDRS,
2873 copts = [
2874 "-UNDEBUG",
2875 "-DXNN_TEST_MODE=1",
2876 ],
2877 gcc_copts = xnnpack_gcc_std_copts(),
2878 gcc_x86_copts = ["-mfma"],
2879 msvc_copts = xnnpack_msvc_std_copts(),
2880 msvc_x86_32_copts = ["/arch:AVX"],
2881 msvc_x86_64_copts = ["/arch:AVX"],
2882 x86_srcs = FMA3_UKERNELS,
2883 deps = [
2884 ":tables",
2885 "@FP16",
2886 "@pthreadpool",
2887 ],
2888)
2889
2890xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002891 name = "avx2_ukernels",
2892 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002893 gcc_copts = xnnpack_gcc_std_copts(),
2894 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002895 "-mfma",
2896 "-mavx2",
2897 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002898 msvc_copts = xnnpack_msvc_std_copts(),
2899 msvc_x86_32_copts = ["/arch:AVX2"],
2900 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002901 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002902 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002903 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002904 "@FP16",
2905 "@pthreadpool",
2906 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002907)
2908
2909xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002910 name = "avx2_ukernels_test_mode",
2911 hdrs = INTERNAL_HDRS,
2912 copts = [
2913 "-UNDEBUG",
2914 "-DXNN_TEST_MODE=1",
2915 ],
2916 gcc_copts = xnnpack_gcc_std_copts(),
2917 gcc_x86_copts = [
2918 "-mfma",
2919 "-mavx2",
2920 ],
2921 msvc_copts = xnnpack_msvc_std_copts(),
2922 msvc_x86_32_copts = ["/arch:AVX2"],
2923 msvc_x86_64_copts = ["/arch:AVX2"],
2924 x86_srcs = AVX2_UKERNELS,
2925 deps = [
2926 ":tables",
2927 "@FP16",
2928 "@pthreadpool",
2929 ],
2930)
2931
2932xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933 name = "avx512f_ukernels",
2934 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002935 gcc_copts = xnnpack_gcc_std_copts(),
2936 gcc_x86_copts = ["-mavx512f"],
2937 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2938 msvc_copts = xnnpack_msvc_std_copts(),
2939 msvc_x86_32_copts = ["/arch:AVX512"],
2940 msvc_x86_64_copts = ["/arch:AVX512"],
2941 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002943 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002944 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002945 "@FP16",
2946 "@pthreadpool",
2947 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948)
2949
2950xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002951 name = "avx512f_ukernels_test_mode",
2952 hdrs = INTERNAL_HDRS,
2953 copts = [
2954 "-UNDEBUG",
2955 "-DXNN_TEST_MODE=1",
2956 ],
2957 gcc_copts = xnnpack_gcc_std_copts(),
2958 gcc_x86_copts = ["-mavx512f"],
2959 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2960 msvc_copts = xnnpack_msvc_std_copts(),
2961 msvc_x86_32_copts = ["/arch:AVX512"],
2962 msvc_x86_64_copts = ["/arch:AVX512"],
2963 msys_copts = ["-fno-asynchronous-unwind-tables"],
2964 x86_srcs = AVX512F_UKERNELS,
2965 deps = [
2966 ":tables",
2967 "@FP16",
2968 "@pthreadpool",
2969 ],
2970)
2971
2972xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973 name = "asm_ukernels",
2974 hdrs = ["src/xnnpack/assembly.h"],
2975 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002976 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 aarch64_srcs = AARCH64_ASM_UKERNELS,
2978)
2979
Marat Dukhan3b59de22020-06-03 20:15:19 -07002980xnnpack_cc_library(
2981 name = "logging_utils",
2982 srcs = LOGGING_SRCS,
2983 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2984 copts = LOGGING_COPTS + [
2985 "-Isrc",
2986 "-Iinclude",
2987 ] + select({
2988 ":debug_build": [],
2989 "//conditions:default": xnnpack_min_size_copts(),
2990 }),
2991 gcc_copts = xnnpack_gcc_std_copts(),
2992 msvc_copts = xnnpack_msvc_std_copts(),
2993 visibility = xnnpack_visibility(),
2994 deps = [
2995 "@FP16",
2996 "@clog",
2997 "@pthreadpool",
2998 ],
2999)
3000
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001xnnpack_aggregate_library(
3002 name = "ukernels",
3003 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004 ":neon_ukernels",
3005 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003006 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 ":asm_ukernels",
3008 ],
3009 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 ":neon_ukernels",
3011 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003012 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 ":neonfp16arith_ukernels",
3014 ":asm_ukernels",
3015 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003016 generic_deps = [
3017 ":scalar_ukernels",
3018 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003019 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003020 ":psimd_fastmath_ukernels",
3021 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003023 wasm_deps = [
3024 ":wasm_ukernels",
3025 ],
3026 wasmsimd_deps = [
3027 ":wasm_ukernels",
3028 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003031 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003032 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003034 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003035 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003036 ":avx512f_ukernels",
3037 ],
3038)
3039
Marat Dukhan33fcf782020-05-24 14:27:15 -07003040xnnpack_aggregate_library(
3041 name = "ukernels_test_mode",
3042 aarch32_deps = [
3043 ":neon_ukernels_test_mode",
3044 ":neonfma_ukernels_test_mode",
3045 ":neonv8_ukernels_test_mode",
3046 ":asm_ukernels",
3047 ],
3048 aarch64_deps = [
3049 ":neon_ukernels_test_mode",
3050 ":neonfma_ukernels_test_mode",
3051 ":neonv8_ukernels_test_mode",
3052 ":neonfp16arith_ukernels_test_mode",
3053 ":asm_ukernels",
3054 ],
3055 generic_deps = [
3056 ":scalar_ukernels_test_mode",
3057 ],
3058 psimd_deps = [
3059 ":psimd_fastmath_ukernels_test_mode",
3060 ":psimd_accmath_ukernels_test_mode",
3061 ],
3062 wasm_deps = [
3063 ":wasm_ukernels_test_mode",
3064 ],
3065 wasmsimd_deps = [
3066 ":wasm_ukernels_test_mode",
3067 ],
3068 x86_deps = [
3069 ":sse2_ukernels_test_mode",
3070 ":ssse3_ukernels_test_mode",
3071 ":sse41_ukernels_test_mode",
3072 ":avx_ukernels_test_mode",
3073 ":fma3_ukernels_test_mode",
3074 ":avx2_ukernels_test_mode",
3075 ":avx512f_ukernels_test_mode",
3076 ],
3077)
3078
Marat Dukhan86df3842020-06-17 14:03:31 -07003079# This target must be compatible with Android NDK r17 and compile with Clang 6.
3080# It is similar to :ukernels target, but excludes microkernels which require
3081# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003082xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003083 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003084 aarch32_deps = [
3085 ":neon_ukernels",
3086 ":neonfma_ukernels",
3087 ":neonv8_ukernels",
3088 ":asm_ukernels",
3089 ],
3090 aarch64_deps = [
3091 ":neon_ukernels",
3092 ":neonfma_ukernels",
3093 ":neonv8_ukernels",
3094 ":asm_ukernels",
3095 ],
3096 generic_deps = [
3097 ":scalar_ukernels",
3098 ],
3099 psimd_deps = [
3100 ":psimd_fastmath_ukernels",
3101 ":psimd_accmath_ukernels",
3102 ],
3103 wasm_deps = [
3104 ":wasm_ukernels",
3105 ],
3106 wasmsimd_deps = [
3107 ":wasm_ukernels",
3108 ],
3109 x86_deps = [
3110 ":sse2_ukernels",
3111 ":ssse3_ukernels",
3112 ":sse41_ukernels",
3113 ":avx_ukernels",
3114 ":fma3_ukernels",
3115 ":avx2_ukernels",
3116 ":avx512f_ukernels",
3117 ],
3118)
3119
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120xnnpack_cc_library(
3121 name = "im2col",
3122 srcs = ["src/im2col.c"],
3123 hdrs = [
3124 "src/xnnpack/common.h",
3125 "src/xnnpack/im2col.h",
3126 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003127 gcc_copts = xnnpack_gcc_std_copts(),
3128 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129)
3130
3131xnnpack_cc_library(
3132 name = "indirection",
3133 srcs = ["src/indirection.c"],
3134 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003135 gcc_copts = xnnpack_gcc_std_copts(),
3136 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 deps = [
3138 "@FP16",
3139 "@FXdiv",
3140 "@pthreadpool",
3141 ],
3142)
3143
3144xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003145 name = "indirection_test_mode",
3146 srcs = ["src/indirection.c"],
3147 hdrs = INTERNAL_HDRS,
3148 copts = [
3149 "-UNDEBUG",
3150 "-DXNN_TEST_MODE=1",
3151 ],
3152 gcc_copts = xnnpack_gcc_std_copts(),
3153 msvc_copts = xnnpack_msvc_std_copts(),
3154 deps = [
3155 "@FP16",
3156 "@FXdiv",
3157 "@pthreadpool",
3158 ],
3159)
3160
3161xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003162 name = "packing",
3163 srcs = ["src/packing.c"],
3164 hdrs = INTERNAL_HDRS,
3165 gcc_copts = xnnpack_gcc_std_copts(),
3166 msvc_copts = xnnpack_msvc_std_copts(),
3167 deps = [
3168 "@FP16",
3169 "@FXdiv",
3170 "@pthreadpool",
3171 ],
3172)
3173
3174xnnpack_cc_library(
3175 name = "packing_test_mode",
3176 srcs = ["src/packing.c"],
3177 hdrs = INTERNAL_HDRS,
3178 copts = [
3179 "-UNDEBUG",
3180 "-DXNN_TEST_MODE=1",
3181 ],
3182 gcc_copts = xnnpack_gcc_std_copts(),
3183 msvc_copts = xnnpack_msvc_std_copts(),
3184 deps = [
3185 "@FP16",
3186 "@FXdiv",
3187 "@pthreadpool",
3188 ],
3189)
3190
3191xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192 name = "operator_run",
3193 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003194 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003195 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003196 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3197 "//conditions:default": [],
3198 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003199 gcc_copts = xnnpack_gcc_std_copts(),
3200 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003202 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 "@FP16",
3204 "@FXdiv",
3205 "@clog",
3206 "@pthreadpool",
3207 ],
3208)
3209
Chao Mei6ddfc602020-05-13 22:29:36 -07003210xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003211 name = "operator_run_test_mode",
3212 srcs = ["src/operator-run.c"],
3213 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3214 copts = LOGGING_COPTS + [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ] + select({
3218 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3219 "//conditions:default": [],
3220 }),
3221 gcc_copts = xnnpack_gcc_std_copts(),
3222 msvc_copts = xnnpack_msvc_std_copts(),
3223 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003224 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003225 "@FP16",
3226 "@FXdiv",
3227 "@clog",
3228 "@pthreadpool",
3229 ],
3230)
3231
3232xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003233 name = "memory_planner",
3234 srcs = ["src/memory-planner.c"],
3235 hdrs = INTERNAL_HDRS,
3236 defines = select({
3237 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3238 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3239 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3240 }),
3241 gcc_copts = xnnpack_gcc_std_copts(),
3242 msvc_copts = xnnpack_msvc_std_copts(),
3243 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003244 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003245 "@pthreadpool",
3246 ],
3247)
3248
Marat Dukhan33fcf782020-05-24 14:27:15 -07003249xnnpack_cc_library(
3250 name = "memory_planner_test_mode",
3251 srcs = ["src/memory-planner.c"],
3252 hdrs = INTERNAL_HDRS,
3253 copts = [
3254 "-UNDEBUG",
3255 "-DXNN_TEST_MODE=1",
3256 ],
3257 defines = select({
3258 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3259 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3260 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3261 }),
3262 gcc_copts = xnnpack_gcc_std_copts(),
3263 msvc_copts = xnnpack_msvc_std_copts(),
3264 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003265 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003266 "@pthreadpool",
3267 ],
3268)
3269
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270cc_library(
3271 name = "enable_assembly",
3272 defines = select({
3273 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3274 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003275 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 }),
3277)
3278
Marat Dukhan9de90e02020-06-18 16:04:12 -07003279cc_library(
3280 name = "enable_sparse",
3281 defines = select({
3282 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3283 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003284 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003285 }),
3286)
3287
Marat Dukhancf056b22019-10-07 10:26:29 -07003288xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289 name = "operators",
3290 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003291 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003292 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003293 ],
3294 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003295 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 "-Isrc",
3297 "-Iinclude",
3298 ] + select({
3299 ":debug_build": [],
3300 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003301 }) + select({
3302 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3303 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003305 gcc_copts = xnnpack_gcc_std_copts(),
3306 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003307 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003308 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003309 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003310 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 "@FP16",
3312 "@FXdiv",
3313 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003314 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003315 ],
3316)
3317
Marat Dukhan10a38082020-04-17 03:58:35 -07003318xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003319 name = "operators_test_mode",
3320 srcs = OPERATOR_SRCS + [
3321 "src/memory.c",
3322 "src/operator-delete.c",
3323 ],
3324 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3325 copts = LOGGING_COPTS + [
3326 "-Isrc",
3327 "-Iinclude",
3328 "-UNDEBUG",
3329 "-DXNN_TEST_MODE=1",
3330 ] + select({
3331 ":debug_build": [],
3332 "//conditions:default": xnnpack_min_size_copts(),
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 = [
3340 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003341 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003342 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003343 "@FP16",
3344 "@FXdiv",
3345 "@clog",
3346 "@pthreadpool",
3347 ],
3348)
3349
3350xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003351 name = "XNNPACK",
3352 srcs = [
3353 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003354 "src/runtime.c",
3355 "src/subgraph.c",
3356 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003357 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003358 hdrs = ["include/xnnpack.h"],
3359 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003360 "-Isrc",
3361 "-Iinclude",
3362 ] + select({
3363 ":debug_build": [],
3364 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003365 }) + select({
3366 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3367 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003368 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003369 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003370 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003371 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003372 visibility = xnnpack_visibility(),
3373 deps = [
3374 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003375 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003376 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003377 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003378 ":operator_run",
3379 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003380 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003381 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003382 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003383 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003384 ] + select({
3385 ":emscripten": [],
3386 "//conditions:default": ["@cpuinfo"],
3387 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388)
3389
Marat Dukhan10a38082020-04-17 03:58:35 -07003390xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003391 name = "XNNPACK_test_mode",
3392 srcs = [
3393 "src/init.c",
3394 "src/runtime.c",
3395 "src/subgraph.c",
3396 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003397 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003398 hdrs = ["include/xnnpack.h"],
3399 copts = LOGGING_COPTS + [
3400 "-Isrc",
3401 "-Iinclude",
3402 "-UNDEBUG",
3403 "-DXNN_TEST_MODE=1",
3404 ] + select({
3405 ":debug_build": [],
3406 "//conditions:default": xnnpack_min_size_copts(),
3407 }) + select({
3408 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3409 "//conditions:default": [],
3410 }),
3411 gcc_copts = xnnpack_gcc_std_copts(),
3412 includes = ["include"],
3413 msvc_copts = xnnpack_msvc_std_copts(),
3414 visibility = xnnpack_visibility(),
3415 deps = [
3416 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003417 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003418 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003419 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003420 ":operator_run_test_mode",
3421 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003422 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003423 "@clog",
3424 "@FP16",
3425 "@pthreadpool",
3426 ] + select({
3427 ":emscripten": [],
3428 "//conditions:default": ["@cpuinfo"],
3429 }),
3430)
3431
3432xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003433 name = "xnnpack_f32",
3434 srcs = [
3435 "src/init.c",
3436 "src/runtime.c",
3437 "src/subgraph.c",
3438 "src/tensor.c",
3439 ] + SUBGRAPH_SRCS,
3440 hdrs = ["include/xnnpack.h"],
3441 copts = LOGGING_COPTS + [
3442 "-Isrc",
3443 "-Iinclude",
3444 ] + select({
3445 ":debug_build": [],
3446 "//conditions:default": xnnpack_min_size_copts(),
3447 }) + select({
3448 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3449 "//conditions:default": [],
3450 }),
3451 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003452 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003453 "XNN_NO_U8_OPERATORS",
3454 "XNN_NO_X8_OPERATORS",
3455 "XNN_NO_F16_OPERATORS",
3456 "XNN_NO_X16_OPERATORS",
3457 ],
3458 gcc_copts = xnnpack_gcc_std_copts(),
3459 includes = ["include"],
3460 msvc_copts = xnnpack_msvc_std_copts(),
3461 visibility = xnnpack_visibility(),
3462 deps = [
3463 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003464 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003465 ":logging_utils",
3466 ":memory_planner",
3467 ":operator_run",
3468 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003469 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003470 "@clog",
3471 "@FP16",
3472 "@pthreadpool",
3473 ] + select({
3474 ":emscripten": [],
3475 "//conditions:default": ["@cpuinfo"],
3476 }),
3477)
3478
3479xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003480 name = "xnnpack_operators_nhwc_f32",
3481 srcs = [
3482 "src/init.c",
3483 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003484 hdrs = ["include/xnnpack.h"],
3485 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003486 "-Isrc",
3487 "-Iinclude",
3488 ] + select({
3489 ":debug_build": [],
3490 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003491 }) + select({
3492 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3493 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003494 }),
3495 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003496 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003497 "XNN_NO_U8_OPERATORS",
3498 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003499 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003500 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003501 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003503 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504 visibility = xnnpack_visibility(),
3505 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003506 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003507 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 ":operator_run",
3509 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003510 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003511 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003512 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003513 ] + select({
3514 ":emscripten": [],
3515 "//conditions:default": ["@cpuinfo"],
3516 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517)
3518
Marat Dukhancf056b22019-10-07 10:26:29 -07003519xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520 name = "bench_utils",
3521 srcs = ["bench/utils.cc"],
3522 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003523 deps = [
3524 "@com_google_benchmark//:benchmark",
3525 "@cpuinfo",
3526 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527)
3528
Frank Barchard7e955972019-10-11 10:34:25 -07003529######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530
3531xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003532 name = "qs8_gemm_bench",
3533 srcs = [
3534 "bench/gemm.h",
3535 "bench/qs8-gemm.cc",
3536 "src/xnnpack/AlignedAllocator.h",
3537 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3538 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3539)
3540
3541xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003542 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003543 srcs = [
3544 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003545 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 "src/xnnpack/AlignedAllocator.h",
3547 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003548 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003549 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003550)
3551
3552xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003553 name = "qu8_requantization_bench",
3554 srcs = [
3555 "bench/qu8-requantization.cc",
3556 "src/xnnpack/requantization-stubs.h",
3557 "src/xnnpack/AlignedAllocator.h",
3558 ] + MICROKERNEL_BENCHMARK_HDRS,
3559 deps = MICROKERNEL_BENCHMARK_DEPS,
3560)
3561
3562xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003563 name = "f16_igemm_bench",
3564 srcs = [
3565 "bench/f16-igemm.cc",
3566 "bench/conv.h",
3567 "bench/google/conv.h",
3568 "src/xnnpack/AlignedAllocator.h",
3569 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003570 deps = MICROKERNEL_BENCHMARK_DEPS + [
3571 ":indirection",
3572 ":packing",
3573 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003574)
3575
3576xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 name = "f16_gemm_bench",
3578 srcs = [
3579 "bench/f16-gemm.cc",
3580 "bench/gemm.h",
3581 "src/xnnpack/AlignedAllocator.h",
3582 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003583 deps = MICROKERNEL_BENCHMARK_DEPS + [
3584 ":packing",
3585 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586)
3587
3588xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003589 name = "f16_spmm_bench",
3590 srcs = [
3591 "bench/f16-spmm.cc",
3592 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003593 "src/xnnpack/AlignedAllocator.h",
3594 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003595 deps = MICROKERNEL_BENCHMARK_DEPS,
3596)
3597
3598xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003599 name = "f32_igemm_bench",
3600 srcs = [
3601 "bench/f32-igemm.cc",
3602 "bench/conv.h",
3603 "src/xnnpack/AlignedAllocator.h",
3604 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003605 deps = MICROKERNEL_BENCHMARK_DEPS + [
3606 ":indirection",
3607 ":packing",
3608 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609)
3610
3611xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003612 name = "f16_relu_bench",
3613 srcs = [
3614 "bench/f16-relu.cc",
3615 "src/xnnpack/AlignedAllocator.h",
3616 ] + MICROKERNEL_BENCHMARK_HDRS,
3617 deps = MICROKERNEL_BENCHMARK_DEPS,
3618)
3619
3620xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621 name = "f32_conv_hwc_bench",
3622 srcs = [
3623 "bench/f32-conv-hwc.cc",
3624 "bench/dconv.h",
3625 "src/xnnpack/AlignedAllocator.h",
3626 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003627 deps = MICROKERNEL_BENCHMARK_DEPS + [
3628 ":packing",
3629 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630)
3631
3632xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003633 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003634 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003635 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003636 "bench/dconv.h",
3637 "src/xnnpack/AlignedAllocator.h",
3638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003639 deps = MICROKERNEL_BENCHMARK_DEPS + [
3640 ":packing",
3641 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003642)
3643
3644xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003645 name = "f16_dwconv_bench",
3646 srcs = [
3647 "bench/f16-dwconv.cc",
3648 "bench/dwconv.h",
3649 "bench/google/dwconv.h",
3650 "src/xnnpack/AlignedAllocator.h",
3651 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003652 deps = MICROKERNEL_BENCHMARK_DEPS + [
3653 ":indirection",
3654 ":packing",
3655 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003656)
3657
3658xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 name = "f32_dwconv_bench",
3660 srcs = [
3661 "bench/f32-dwconv.cc",
3662 "bench/dwconv.h",
3663 "src/xnnpack/AlignedAllocator.h",
3664 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003665 deps = MICROKERNEL_BENCHMARK_DEPS + [
3666 ":indirection",
3667 ":packing",
3668 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669)
3670
3671xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003672 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003674 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003675 "bench/dwconv.h",
3676 "src/xnnpack/AlignedAllocator.h",
3677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003678 deps = MICROKERNEL_BENCHMARK_DEPS + [
3679 ":indirection",
3680 ":packing",
3681 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682)
3683
3684xnnpack_benchmark(
3685 name = "f32_gemm_bench",
3686 srcs = [
3687 "bench/f32-gemm.cc",
3688 "bench/gemm.h",
3689 "src/xnnpack/AlignedAllocator.h",
3690 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003691 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003692 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693)
3694
3695xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003696 name = "f32_hswish_bench",
3697 srcs = [
3698 "bench/f32-hswish.cc",
3699 "src/xnnpack/AlignedAllocator.h",
3700 ] + MICROKERNEL_BENCHMARK_HDRS,
3701 deps = MICROKERNEL_BENCHMARK_DEPS,
3702)
3703
3704xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003705 name = "f32_raddexpminusmax_bench",
3706 srcs = [
3707 "bench/f32-raddexpminusmax.cc",
3708 "src/xnnpack/AlignedAllocator.h",
3709 ] + MICROKERNEL_BENCHMARK_HDRS,
3710 deps = MICROKERNEL_BENCHMARK_DEPS,
3711)
3712
3713xnnpack_benchmark(
3714 name = "f32_raddextexp_bench",
3715 srcs = [
3716 "bench/f32-raddextexp.cc",
3717 "src/xnnpack/AlignedAllocator.h",
3718 ] + MICROKERNEL_BENCHMARK_HDRS,
3719 deps = MICROKERNEL_BENCHMARK_DEPS,
3720)
3721
3722xnnpack_benchmark(
3723 name = "f32_raddstoreexpminusmax_bench",
3724 srcs = [
3725 "bench/f32-raddstoreexpminusmax.cc",
3726 "src/xnnpack/AlignedAllocator.h",
3727 ] + MICROKERNEL_BENCHMARK_HDRS,
3728 deps = MICROKERNEL_BENCHMARK_DEPS,
3729)
3730
3731xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003732 name = "f32_relu_bench",
3733 srcs = [
3734 "bench/f32-relu.cc",
3735 "src/xnnpack/AlignedAllocator.h",
3736 ] + MICROKERNEL_BENCHMARK_HDRS,
3737 deps = MICROKERNEL_BENCHMARK_DEPS,
3738)
3739
3740xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003741 name = "f32_rmax_bench",
3742 srcs = [
3743 "bench/f32-rmax.cc",
3744 "src/xnnpack/AlignedAllocator.h",
3745 ] + MICROKERNEL_BENCHMARK_HDRS,
3746 deps = MICROKERNEL_BENCHMARK_DEPS,
3747)
3748
3749xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003750 name = "f32_sigmoid_bench",
3751 srcs = [
3752 "bench/f32-sigmoid.cc",
3753 "src/xnnpack/AlignedAllocator.h",
3754 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003755 deps = MICROKERNEL_BENCHMARK_DEPS,
3756)
3757
3758xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003759 name = "f32_spmm_bench",
3760 srcs = [
3761 "bench/f32-spmm.cc",
3762 "bench/gemm.h",
3763 "src/xnnpack/AlignedAllocator.h",
3764 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003765 deps = MICROKERNEL_BENCHMARK_DEPS,
3766)
3767
3768xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003769 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003770 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003771 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003772 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003773 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003774 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003775)
3776
3777xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003778 name = "f32_vscaleexpminusmax_bench",
3779 srcs = [
3780 "bench/f32-vscaleexpminusmax.cc",
3781 "src/xnnpack/AlignedAllocator.h",
3782 ] + MICROKERNEL_BENCHMARK_HDRS,
3783 deps = MICROKERNEL_BENCHMARK_DEPS,
3784)
3785
3786xnnpack_benchmark(
3787 name = "f32_vscaleextexp_bench",
3788 srcs = [
3789 "bench/f32-vscaleextexp.cc",
3790 "src/xnnpack/AlignedAllocator.h",
3791 ] + MICROKERNEL_BENCHMARK_HDRS,
3792 deps = MICROKERNEL_BENCHMARK_DEPS,
3793)
3794
3795xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003796 name = "f32_vsqrt_bench",
3797 srcs = [
3798 "bench/f32-vsqrt.cc",
3799 "src/xnnpack/AlignedAllocator.h",
3800 ] + MICROKERNEL_BENCHMARK_HDRS,
3801 deps = MICROKERNEL_BENCHMARK_DEPS,
3802)
3803
3804xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 name = "f32_im2col_gemm_bench",
3806 srcs = [
3807 "bench/f32-im2col-gemm.cc",
3808 "bench/conv.h",
3809 "src/xnnpack/AlignedAllocator.h",
3810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003811 deps = MICROKERNEL_BENCHMARK_DEPS + [
3812 ":im2col",
3813 ":packing",
3814 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003815)
3816
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003817xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003818 name = "rounding_bench",
3819 srcs = [
3820 "bench/rounding.cc",
3821 "src/xnnpack/math-stubs.h",
3822 "src/xnnpack/AlignedAllocator.h",
3823 ] + MICROKERNEL_BENCHMARK_HDRS,
3824 deps = MICROKERNEL_BENCHMARK_DEPS,
3825)
3826
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827########################### Benchmarks for operators ###########################
3828
3829xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830 name = "average_pooling_bench",
3831 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003832 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003833 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003834 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003835)
3836
3837xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003838 name = "bankers_rounding_bench",
3839 srcs = ["bench/bankers-rounding.cc"],
3840 copts = xnnpack_optional_tflite_copts(),
3841 tags = ["nowin32"],
3842 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3843)
3844
3845xnnpack_benchmark(
3846 name = "ceiling_bench",
3847 srcs = ["bench/ceiling.cc"],
3848 copts = xnnpack_optional_tflite_copts(),
3849 tags = ["nowin32"],
3850 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3851)
3852
3853xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854 name = "channel_shuffle_bench",
3855 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003856 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003857)
3858
3859xnnpack_benchmark(
3860 name = "convolution_bench",
3861 srcs = ["bench/convolution.cc"],
3862 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003863 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003864 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865)
3866
3867xnnpack_benchmark(
3868 name = "deconvolution_bench",
3869 srcs = ["bench/deconvolution.cc"],
3870 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003871 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003872 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873)
3874
3875xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003876 name = "floor_bench",
3877 srcs = ["bench/floor.cc"],
3878 copts = xnnpack_optional_tflite_copts(),
3879 tags = ["nowin32"],
3880 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3881)
3882
3883xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003884 name = "global_average_pooling_bench",
3885 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003886 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003887)
3888
3889xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003890 name = "hardswish_bench",
3891 srcs = ["bench/hardswish.cc"],
3892 copts = xnnpack_optional_tflite_copts(),
3893 tags = ["nowin32"],
3894 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3895)
3896
3897xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898 name = "max_pooling_bench",
3899 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003900 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901)
3902
3903xnnpack_benchmark(
3904 name = "sigmoid_bench",
3905 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003906 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003907 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003908 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909)
3910
3911xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003912 name = "prelu_bench",
3913 srcs = ["bench/prelu.cc"],
3914 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003915 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003916 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003917)
3918
3919xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003920 name = "softmax_bench",
3921 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003922 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003923 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003924 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925)
3926
Marat Dukhan87727142020-06-24 15:24:10 -07003927xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003928 name = "square_root_bench",
3929 srcs = ["bench/square-root.cc"],
3930 copts = xnnpack_optional_tflite_copts(),
3931 tags = ["nowin32"],
3932 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3933)
3934
3935xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003936 name = "truncation_bench",
3937 srcs = ["bench/truncation.cc"],
3938 deps = OPERATOR_BENCHMARK_DEPS,
3939)
3940
Marat Dukhanc068bb62019-10-04 13:24:39 -07003941############################# End-to-end benchmarks ############################
3942
3943cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003944 name = "fp32_mobilenet_v1",
3945 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003946 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003947 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003948 linkstatic = True,
3949 deps = [
3950 ":XNNPACK",
3951 "@pthreadpool",
3952 ],
3953)
3954
3955cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003956 name = "fp16_mobilenet_v1",
3957 srcs = ["models/fp16-mobilenet-v1.cc"],
3958 hdrs = ["models/models.h"],
3959 copts = xnnpack_std_cxxopts(),
3960 linkstatic = True,
3961 deps = [
3962 ":XNNPACK",
3963 "@FP16",
3964 "@pthreadpool",
3965 ],
3966)
3967
3968cc_library(
3969 name = "fp32_mobilenet_v2",
3970 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003971 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003972 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003973 linkstatic = True,
3974 deps = [
3975 ":XNNPACK",
3976 "@pthreadpool",
3977 ],
3978)
3979
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003980cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003981 name = "fp16_mobilenet_v2",
3982 srcs = ["models/fp16-mobilenet-v2.cc"],
3983 hdrs = ["models/models.h"],
3984 copts = xnnpack_std_cxxopts(),
3985 linkstatic = True,
3986 deps = [
3987 ":XNNPACK",
3988 "@FP16",
3989 "@pthreadpool",
3990 ],
3991)
3992
3993cc_library(
3994 name = "fp32_mobilenet_v3_large",
3995 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003996 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003997 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003998 linkstatic = True,
3999 deps = [
4000 ":XNNPACK",
4001 "@pthreadpool",
4002 ],
4003)
4004
4005cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004006 name = "fp32_mobilenet_v3_small",
4007 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004008 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004009 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004010 linkstatic = True,
4011 deps = [
4012 ":XNNPACK",
4013 "@pthreadpool",
4014 ],
4015)
4016
Marat Dukhanc068bb62019-10-04 13:24:39 -07004017xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004018 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004019 srcs = [
4020 "bench/f32-dwconv-e2e.cc",
4021 "bench/end2end.h",
4022 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004023 deps = MICROKERNEL_BENCHMARK_DEPS + [
4024 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004025 ":fp32_mobilenet_v1",
4026 ":fp32_mobilenet_v2",
4027 ":fp32_mobilenet_v3_large",
4028 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004029 ],
4030)
4031
4032xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004033 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004034 srcs = [
4035 "bench/f32-gemm-e2e.cc",
4036 "bench/end2end.h",
4037 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004038 deps = MICROKERNEL_BENCHMARK_DEPS + [
4039 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004040 ":fp32_mobilenet_v1",
4041 ":fp32_mobilenet_v2",
4042 ":fp32_mobilenet_v3_large",
4043 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004044 ],
4045)
4046
4047xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004048 name = "end2end_bench",
4049 srcs = ["bench/end2end.cc"],
4050 deps = [
4051 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004052 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004053 ":fp16_mobilenet_v1",
4054 ":fp16_mobilenet_v2",
4055 ":fp32_mobilenet_v1",
4056 ":fp32_mobilenet_v2",
4057 ":fp32_mobilenet_v3_large",
4058 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004059 "@pthreadpool",
4060 ],
4061)
4062
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004063#################### Accuracy evaluation for math functions ####################
4064
4065xnnpack_benchmark(
4066 name = "f32_exp_eval",
4067 srcs = [
4068 "eval/f32-exp.cc",
4069 "src/xnnpack/AlignedAllocator.h",
4070 ] + ACCURACY_EVAL_HDRS,
4071 deps = ACCURACY_EVAL_DEPS,
4072)
4073
Marat Dukhan515c9772019-10-17 18:07:57 -07004074xnnpack_benchmark(
4075 name = "f32_expminus_eval",
4076 srcs = [
4077 "eval/f32-expminus.cc",
4078 "src/xnnpack/AlignedAllocator.h",
4079 ] + ACCURACY_EVAL_HDRS,
4080 deps = ACCURACY_EVAL_DEPS,
4081)
4082
Marat Dukhan98ba4412019-10-23 02:14:28 -07004083xnnpack_benchmark(
4084 name = "f32_extexp_eval",
4085 srcs = [
4086 "eval/f32-extexp.cc",
4087 "src/xnnpack/AlignedAllocator.h",
4088 ] + ACCURACY_EVAL_HDRS,
4089 deps = ACCURACY_EVAL_DEPS,
4090)
4091
Marat Dukhan8853b822020-05-07 12:19:01 -07004092xnnpack_unit_test(
4093 name = "f32_roundne_eval",
4094 srcs = [
4095 "eval/f32-roundne.cc",
4096 "src/xnnpack/AlignedAllocator.h",
4097 "src/xnnpack/math-stubs.h",
4098 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004099 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004100 deps = MICROKERNEL_TEST_DEPS,
4101)
4102
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004103xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004104 name = "f32_roundd_eval",
4105 srcs = [
4106 "eval/f32-roundd.cc",
4107 "src/xnnpack/AlignedAllocator.h",
4108 "src/xnnpack/math-stubs.h",
4109 ] + MICROKERNEL_TEST_HDRS,
4110 automatic = False,
4111 deps = MICROKERNEL_TEST_DEPS,
4112)
4113
4114xnnpack_unit_test(
4115 name = "f32_roundu_eval",
4116 srcs = [
4117 "eval/f32-roundu.cc",
4118 "src/xnnpack/AlignedAllocator.h",
4119 "src/xnnpack/math-stubs.h",
4120 ] + MICROKERNEL_TEST_HDRS,
4121 automatic = False,
4122 deps = MICROKERNEL_TEST_DEPS,
4123)
4124
4125xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004126 name = "f32_roundz_eval",
4127 srcs = [
4128 "eval/f32-roundz.cc",
4129 "src/xnnpack/AlignedAllocator.h",
4130 "src/xnnpack/math-stubs.h",
4131 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004132 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004133 deps = MICROKERNEL_TEST_DEPS,
4134)
4135
Marat Dukhan346a9e52019-11-15 09:06:30 -08004136xnnpack_benchmark(
4137 name = "f32_sigmoid_eval",
4138 srcs = [
4139 "eval/f32-sigmoid.cc",
4140 "src/xnnpack/AlignedAllocator.h",
4141 ] + ACCURACY_EVAL_HDRS,
4142 deps = ACCURACY_EVAL_DEPS,
4143)
4144
Marat Dukhan84000762020-06-29 18:38:43 -07004145xnnpack_benchmark(
4146 name = "f32_sqrt_eval",
4147 srcs = [
4148 "eval/f32-sqrt.cc",
4149 "src/xnnpack/AlignedAllocator.h",
4150 ] + ACCURACY_EVAL_HDRS,
4151 deps = ACCURACY_EVAL_DEPS + [
4152 ":bench_utils",
4153 ],
4154)
4155
Marat Dukhan08c4a432019-10-03 09:29:21 -07004156######################### Unit tests for micro-kernels #########################
4157
4158xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004159 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004160 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004161 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004162 "test/gemm-microkernel-tester.h",
4163 "src/xnnpack/AlignedAllocator.h",
4164 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004165 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004166)
4167
4168xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004169 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004170 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004171 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004172 "test/spmm-microkernel-tester.h",
4173 "src/xnnpack/AlignedAllocator.h",
4174 ] + MICROKERNEL_TEST_HDRS,
4175 deps = MICROKERNEL_TEST_DEPS,
4176)
4177
4178xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004179 name = "f32_argmaxpool_test",
4180 srcs = [
4181 "test/f32-argmaxpool.cc",
4182 "test/argmaxpool-microkernel-tester.h",
4183 "src/xnnpack/AlignedAllocator.h",
4184 ] + MICROKERNEL_TEST_HDRS,
4185 deps = MICROKERNEL_TEST_DEPS,
4186)
4187
4188xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004189 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004190 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004191 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004192 "test/avgpool-microkernel-tester.h",
4193 "src/xnnpack/AlignedAllocator.h",
4194 ] + MICROKERNEL_TEST_HDRS,
4195 deps = MICROKERNEL_TEST_DEPS,
4196)
4197
4198xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004199 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004200 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004201 "test/f32-ibilinear.cc",
4202 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004203 "src/xnnpack/AlignedAllocator.h",
4204 ] + MICROKERNEL_TEST_HDRS,
4205 deps = MICROKERNEL_TEST_DEPS,
4206)
4207
4208xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004209 name = "f16_clamp_test",
4210 srcs = [
4211 "test/f16-clamp.cc",
4212 "test/clamp-microkernel-tester.h",
4213 ] + MICROKERNEL_TEST_HDRS,
4214 deps = MICROKERNEL_TEST_DEPS,
4215)
4216
4217xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 name = "f32_clamp_test",
4219 srcs = [
4220 "test/f32-clamp.cc",
4221 "test/clamp-microkernel-tester.h",
4222 ] + MICROKERNEL_TEST_HDRS,
4223 deps = MICROKERNEL_TEST_DEPS,
4224)
4225
4226xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004227 name = "f32_igemm_test",
4228 srcs = [
4229 "test/f32-igemm.cc",
4230 "test/gemm-microkernel-tester.h",
4231 "src/xnnpack/AlignedAllocator.h",
4232 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004233 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004234)
4235
4236xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004237 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004238 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004239 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240 "test/gemm-microkernel-tester.h",
4241 "src/xnnpack/AlignedAllocator.h",
4242 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004243 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004244)
4245
4246xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004247 name = "f32_igemm_minmax_test",
4248 srcs = [
4249 "test/f32-igemm-minmax.cc",
4250 "test/gemm-microkernel-tester.h",
4251 "src/xnnpack/AlignedAllocator.h",
4252 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004253 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004254)
4255
4256xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004257 name = "f16_igemm_minmax_test",
4258 srcs = [
4259 "test/f16-igemm-minmax.cc",
4260 "test/gemm-microkernel-tester.h",
4261 "src/xnnpack/AlignedAllocator.h",
4262 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004263 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004264)
4265
4266xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267 name = "f32_conv_hwc_test",
4268 srcs = [
4269 "test/f32-conv-hwc.cc",
4270 "test/conv-hwc-microkernel-tester.h",
4271 "src/xnnpack/AlignedAllocator.h",
4272 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004273 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004274)
4275
4276xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004277 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004279 "test/f32-conv-hwc2chw.cc",
4280 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281 "src/xnnpack/AlignedAllocator.h",
4282 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004283 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004284)
4285
4286xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004287 name = "f32_dwconv_test",
4288 srcs = [
4289 "test/f32-dwconv.cc",
4290 "test/dwconv-microkernel-tester.h",
4291 "src/xnnpack/AlignedAllocator.h",
4292 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004293 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004294)
4295
4296xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004297 name = "f16_dwconv_minmax_test",
4298 srcs = [
4299 "test/f16-dwconv-minmax.cc",
4300 "test/dwconv-microkernel-tester.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004303 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004304)
4305
4306xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004307 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004309 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 "test/dwconv-microkernel-tester.h",
4311 "src/xnnpack/AlignedAllocator.h",
4312 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004313 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314)
4315
4316xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004317 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004319 "test/f32-dwconv-chw.cc",
4320 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004321 "src/xnnpack/AlignedAllocator.h",
4322 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004323 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324)
4325
4326xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004327 name = "f16_gavgpool_minmax_test",
4328 srcs = [
4329 "test/f16-gavgpool-minmax.cc",
4330 "test/gavgpool-microkernel-tester.h",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + MICROKERNEL_TEST_HDRS,
4333 deps = MICROKERNEL_TEST_DEPS,
4334)
4335
4336xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004337 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004339 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340 "test/gavgpool-microkernel-tester.h",
4341 "src/xnnpack/AlignedAllocator.h",
4342 ] + MICROKERNEL_TEST_HDRS,
4343 deps = MICROKERNEL_TEST_DEPS,
4344)
4345
4346xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004347 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004349 "test/f32-gavgpool-cw.cc",
4350 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351 "src/xnnpack/AlignedAllocator.h",
4352 ] + MICROKERNEL_TEST_HDRS,
4353 deps = MICROKERNEL_TEST_DEPS,
4354)
4355
4356xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004357 name = "f32_gemm_test",
4358 srcs = [
4359 "test/f32-gemm.cc",
4360 "test/gemm-microkernel-tester.h",
4361 "src/xnnpack/AlignedAllocator.h",
4362 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004363 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004364)
4365
4366xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004367 name = "f32_gemm_relu_test",
4368 srcs = [
4369 "test/f32-gemm-relu.cc",
4370 "test/gemm-microkernel-tester.h",
4371 "src/xnnpack/AlignedAllocator.h",
4372 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004373 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004374)
4375
4376xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004377 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004379 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380 "test/gemm-microkernel-tester.h",
4381 "src/xnnpack/AlignedAllocator.h",
4382 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004383 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384)
4385
4386xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004387 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004389 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 "test/gemm-microkernel-tester.h",
4391 "src/xnnpack/AlignedAllocator.h",
4392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004393 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394)
4395
4396xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004397 name = "f16_hswish_test",
4398 srcs = [
4399 "test/f16-hswish.cc",
4400 "test/hswish-microkernel-tester.h",
4401 ] + MICROKERNEL_TEST_HDRS,
4402 deps = MICROKERNEL_TEST_DEPS,
4403)
4404
4405xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004406 name = "f32_hswish_test",
4407 srcs = [
4408 "test/f32-hswish.cc",
4409 "test/hswish-microkernel-tester.h",
4410 ] + MICROKERNEL_TEST_HDRS,
4411 deps = MICROKERNEL_TEST_DEPS,
4412)
4413
4414xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004415 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004417 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004418 "test/maxpool-microkernel-tester.h",
4419 ] + MICROKERNEL_TEST_HDRS,
4420 deps = MICROKERNEL_TEST_DEPS,
4421)
4422
4423xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004424 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004425 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004426 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004427 "test/avgpool-microkernel-tester.h",
4428 "src/xnnpack/AlignedAllocator.h",
4429 ] + MICROKERNEL_TEST_HDRS,
4430 deps = MICROKERNEL_TEST_DEPS,
4431)
4432
4433xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004434 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004436 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004437 "test/gemm-microkernel-tester.h",
4438 "src/xnnpack/AlignedAllocator.h",
4439 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004440 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441)
4442
4443xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004444 name = "f16_prelu_test",
4445 srcs = [
4446 "test/f16-prelu.cc",
4447 "test/prelu-microkernel-tester.h",
4448 "src/xnnpack/AlignedAllocator.h",
4449 ] + MICROKERNEL_TEST_HDRS,
4450 deps = MICROKERNEL_TEST_DEPS,
4451)
4452
4453xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 name = "f32_prelu_test",
4455 srcs = [
4456 "test/f32-prelu.cc",
4457 "test/prelu-microkernel-tester.h",
4458 "src/xnnpack/AlignedAllocator.h",
4459 ] + MICROKERNEL_TEST_HDRS,
4460 deps = MICROKERNEL_TEST_DEPS,
4461)
4462
4463xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004464 name = "f32_raddexpminusmax_test",
4465 srcs = [
4466 "test/f32-raddexpminusmax.cc",
4467 "test/raddexpminusmax-microkernel-tester.h",
4468 ] + MICROKERNEL_TEST_HDRS,
4469 deps = MICROKERNEL_TEST_DEPS,
4470)
4471
4472xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004473 name = "f32_raddextexp_test",
4474 srcs = [
4475 "test/f32-raddextexp.cc",
4476 "test/raddextexp-microkernel-tester.h",
4477 ] + MICROKERNEL_TEST_HDRS,
4478 deps = MICROKERNEL_TEST_DEPS,
4479)
4480
4481xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004482 name = "f32_raddstoreexpminusmax_test",
4483 srcs = [
4484 "test/f32-raddstoreexpminusmax.cc",
4485 "test/raddstoreexpminusmax-microkernel-tester.h",
4486 ] + MICROKERNEL_TEST_HDRS,
4487 deps = MICROKERNEL_TEST_DEPS,
4488)
4489
4490xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004491 name = "f32_relu_test",
4492 srcs = [
4493 "test/f32-relu.cc",
4494 "test/vunary-microkernel-tester.h",
4495 ] + MICROKERNEL_TEST_HDRS,
4496 deps = MICROKERNEL_TEST_DEPS,
4497)
4498
4499xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500 name = "f32_rmax_test",
4501 srcs = [
4502 "test/f32-rmax.cc",
4503 "test/rmax-microkernel-tester.h",
4504 ] + MICROKERNEL_TEST_HDRS,
4505 deps = MICROKERNEL_TEST_DEPS,
4506)
4507
4508xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004509 name = "f32_sigmoid_test",
4510 srcs = [
4511 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004512 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004513 ] + MICROKERNEL_TEST_HDRS,
4514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
4517xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004518 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004519 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004520 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521 "test/spmm-microkernel-tester.h",
4522 "src/xnnpack/AlignedAllocator.h",
4523 ] + MICROKERNEL_TEST_HDRS,
4524 deps = MICROKERNEL_TEST_DEPS,
4525)
4526
4527xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004528 name = "f16_vadd_minmax_test",
4529 srcs = [
4530 "test/f16-vadd-minmax.cc",
4531 "test/vbinary-microkernel-tester.h",
4532 ] + MICROKERNEL_TEST_HDRS,
4533 deps = MICROKERNEL_TEST_DEPS,
4534)
4535
4536xnnpack_unit_test(
4537 name = "f16_vaddc_minmax_test",
4538 srcs = [
4539 "test/f16-vaddc-minmax.cc",
4540 "test/vbinaryc-microkernel-tester.h",
4541 ] + MICROKERNEL_TEST_HDRS,
4542 deps = MICROKERNEL_TEST_DEPS,
4543)
4544
4545xnnpack_unit_test(
4546 name = "f16_vdiv_minmax_test",
4547 srcs = [
4548 "test/f16-vdiv-minmax.cc",
4549 "test/vbinary-microkernel-tester.h",
4550 ] + MICROKERNEL_TEST_HDRS,
4551 deps = MICROKERNEL_TEST_DEPS,
4552)
4553
4554xnnpack_unit_test(
4555 name = "f16_vdivc_minmax_test",
4556 srcs = [
4557 "test/f16-vdivc-minmax.cc",
4558 "test/vbinaryc-microkernel-tester.h",
4559 ] + MICROKERNEL_TEST_HDRS,
4560 deps = MICROKERNEL_TEST_DEPS,
4561)
4562
4563xnnpack_unit_test(
4564 name = "f16_vrdivc_minmax_test",
4565 srcs = [
4566 "test/f16-vrdivc-minmax.cc",
4567 "test/vbinaryc-microkernel-tester.h",
4568 ] + MICROKERNEL_TEST_HDRS,
4569 deps = MICROKERNEL_TEST_DEPS,
4570)
4571
4572xnnpack_unit_test(
4573 name = "f16_vmax_test",
4574 srcs = [
4575 "test/f16-vmax.cc",
4576 "test/vbinary-microkernel-tester.h",
4577 ] + MICROKERNEL_TEST_HDRS,
4578 deps = MICROKERNEL_TEST_DEPS,
4579)
4580
4581xnnpack_unit_test(
4582 name = "f16_vmaxc_test",
4583 srcs = [
4584 "test/f16-vmaxc.cc",
4585 "test/vbinaryc-microkernel-tester.h",
4586 ] + MICROKERNEL_TEST_HDRS,
4587 deps = MICROKERNEL_TEST_DEPS,
4588)
4589
4590xnnpack_unit_test(
4591 name = "f16_vmin_test",
4592 srcs = [
4593 "test/f16-vmin.cc",
4594 "test/vbinary-microkernel-tester.h",
4595 ] + MICROKERNEL_TEST_HDRS,
4596 deps = MICROKERNEL_TEST_DEPS,
4597)
4598
4599xnnpack_unit_test(
4600 name = "f16_vminc_test",
4601 srcs = [
4602 "test/f16-vminc.cc",
4603 "test/vbinaryc-microkernel-tester.h",
4604 ] + MICROKERNEL_TEST_HDRS,
4605 deps = MICROKERNEL_TEST_DEPS,
4606)
4607
4608xnnpack_unit_test(
4609 name = "f16_vmul_minmax_test",
4610 srcs = [
4611 "test/f16-vmul-minmax.cc",
4612 "test/vbinary-microkernel-tester.h",
4613 ] + MICROKERNEL_TEST_HDRS,
4614 deps = MICROKERNEL_TEST_DEPS,
4615)
4616
4617xnnpack_unit_test(
4618 name = "f16_vmulc_minmax_test",
4619 srcs = [
4620 "test/f16-vmulc-minmax.cc",
4621 "test/vbinaryc-microkernel-tester.h",
4622 ] + MICROKERNEL_TEST_HDRS,
4623 deps = MICROKERNEL_TEST_DEPS,
4624)
4625
4626xnnpack_unit_test(
4627 name = "f16_vsub_minmax_test",
4628 srcs = [
4629 "test/f16-vsub-minmax.cc",
4630 "test/vbinary-microkernel-tester.h",
4631 ] + MICROKERNEL_TEST_HDRS,
4632 deps = MICROKERNEL_TEST_DEPS,
4633)
4634
4635xnnpack_unit_test(
4636 name = "f16_vsubc_minmax_test",
4637 srcs = [
4638 "test/f16-vsubc-minmax.cc",
4639 "test/vbinaryc-microkernel-tester.h",
4640 ] + MICROKERNEL_TEST_HDRS,
4641 deps = MICROKERNEL_TEST_DEPS,
4642)
4643
4644xnnpack_unit_test(
4645 name = "f16_vrsubc_minmax_test",
4646 srcs = [
4647 "test/f16-vrsubc-minmax.cc",
4648 "test/vbinaryc-microkernel-tester.h",
4649 ] + MICROKERNEL_TEST_HDRS,
4650 deps = MICROKERNEL_TEST_DEPS,
4651)
4652
4653xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004654 name = "f32_vabs_test",
4655 srcs = [
4656 "test/f32-vabs.cc",
4657 "test/vunary-microkernel-tester.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004663 name = "f32_vadd_test",
4664 srcs = [
4665 "test/f32-vadd.cc",
4666 "test/vbinary-microkernel-tester.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004672 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004674 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004675 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004676 ] + MICROKERNEL_TEST_HDRS,
4677 deps = MICROKERNEL_TEST_DEPS,
4678)
4679
4680xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004681 name = "f32_vaddc_test",
4682 srcs = [
4683 "test/f32-vaddc.cc",
4684 "test/vbinaryc-microkernel-tester.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004690 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004691 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004692 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004693 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004699 name = "f32_vdiv_test",
4700 srcs = [
4701 "test/f32-vdiv.cc",
4702 "test/vbinary-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004708 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004709 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004710 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004711 "test/vbinary-microkernel-tester.h",
4712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004717 name = "f32_vdivc_test",
4718 srcs = [
4719 "test/f32-vdivc.cc",
4720 "test/vbinaryc-microkernel-tester.h",
4721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004726 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004727 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004728 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004729 "test/vbinaryc-microkernel-tester.h",
4730 ] + MICROKERNEL_TEST_HDRS,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004735 name = "f32_vrdivc_test",
4736 srcs = [
4737 "test/f32-vrdivc.cc",
4738 "test/vbinaryc-microkernel-tester.h",
4739 ] + MICROKERNEL_TEST_HDRS,
4740 deps = MICROKERNEL_TEST_DEPS,
4741)
4742
4743xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004744 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004745 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004746 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004747 "test/vbinaryc-microkernel-tester.h",
4748 ] + MICROKERNEL_TEST_HDRS,
4749 deps = MICROKERNEL_TEST_DEPS,
4750)
4751
4752xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004753 name = "f32_vmax_test",
4754 srcs = [
4755 "test/f32-vmax.cc",
4756 "test/vbinary-microkernel-tester.h",
4757 ] + MICROKERNEL_TEST_HDRS,
4758 deps = MICROKERNEL_TEST_DEPS,
4759)
4760
4761xnnpack_unit_test(
4762 name = "f32_vmaxc_test",
4763 srcs = [
4764 "test/f32-vmaxc.cc",
4765 "test/vbinaryc-microkernel-tester.h",
4766 ] + MICROKERNEL_TEST_HDRS,
4767 deps = MICROKERNEL_TEST_DEPS,
4768)
4769
4770xnnpack_unit_test(
4771 name = "f32_vmin_test",
4772 srcs = [
4773 "test/f32-vmin.cc",
4774 "test/vbinary-microkernel-tester.h",
4775 ] + MICROKERNEL_TEST_HDRS,
4776 deps = MICROKERNEL_TEST_DEPS,
4777)
4778
4779xnnpack_unit_test(
4780 name = "f32_vminc_test",
4781 srcs = [
4782 "test/f32-vminc.cc",
4783 "test/vbinaryc-microkernel-tester.h",
4784 ] + MICROKERNEL_TEST_HDRS,
4785 deps = MICROKERNEL_TEST_DEPS,
4786)
4787
4788xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004789 name = "f32_vmul_test",
4790 srcs = [
4791 "test/f32-vmul.cc",
4792 "test/vbinary-microkernel-tester.h",
4793 ] + MICROKERNEL_TEST_HDRS,
4794 deps = MICROKERNEL_TEST_DEPS,
4795)
4796
4797xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004798 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004800 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004801 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004802 ] + MICROKERNEL_TEST_HDRS,
4803 deps = MICROKERNEL_TEST_DEPS,
4804)
4805
4806xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004807 name = "f32_vmulc_test",
4808 srcs = [
4809 "test/f32-vmulc.cc",
4810 "test/vbinaryc-microkernel-tester.h",
4811 ] + MICROKERNEL_TEST_HDRS,
4812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
4815xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004816 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004817 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004818 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004819 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004820 ] + MICROKERNEL_TEST_HDRS,
4821 deps = MICROKERNEL_TEST_DEPS,
4822)
4823
4824xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004825 name = "f16_vmulcaddc_minmax_test",
4826 srcs = [
4827 "test/f16-vmulcaddc-minmax.cc",
4828 "test/vmulcaddc-microkernel-tester.h",
4829 "src/xnnpack/AlignedAllocator.h",
4830 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004831 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004832)
4833
4834xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004835 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004837 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004838 "test/vmulcaddc-microkernel-tester.h",
4839 "src/xnnpack/AlignedAllocator.h",
4840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004841 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004842)
4843
4844xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004845 name = "f32_vlrelu_test",
4846 srcs = [
4847 "test/f32-vlrelu.cc",
4848 "test/vunary-microkernel-tester.h",
4849 ] + MICROKERNEL_TEST_HDRS,
4850 deps = MICROKERNEL_TEST_DEPS,
4851)
4852
4853xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004854 name = "f32_vneg_test",
4855 srcs = [
4856 "test/f32-vneg.cc",
4857 "test/vunary-microkernel-tester.h",
4858 ] + MICROKERNEL_TEST_HDRS,
4859 deps = MICROKERNEL_TEST_DEPS,
4860)
4861
4862xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004863 name = "f32_vrndne_test",
4864 srcs = [
4865 "test/f32-vrndne.cc",
4866 "test/vunary-microkernel-tester.h",
4867 ] + MICROKERNEL_TEST_HDRS,
4868 deps = MICROKERNEL_TEST_DEPS,
4869)
4870
4871xnnpack_unit_test(
4872 name = "f32_vrndz_test",
4873 srcs = [
4874 "test/f32-vrndz.cc",
4875 "test/vunary-microkernel-tester.h",
4876 ] + MICROKERNEL_TEST_HDRS,
4877 deps = MICROKERNEL_TEST_DEPS,
4878)
4879
4880xnnpack_unit_test(
4881 name = "f32_vrndu_test",
4882 srcs = [
4883 "test/f32-vrndu.cc",
4884 "test/vunary-microkernel-tester.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
4890 name = "f32_vrndd_test",
4891 srcs = [
4892 "test/f32-vrndd.cc",
4893 "test/vunary-microkernel-tester.h",
4894 ] + MICROKERNEL_TEST_HDRS,
4895 deps = MICROKERNEL_TEST_DEPS,
4896)
4897
4898xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004899 name = "f32_vscale_test",
4900 srcs = [
4901 "test/f32-vscale.cc",
4902 "test/vscale-microkernel-tester.h",
4903 ] + MICROKERNEL_TEST_HDRS,
4904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
4907xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004908 name = "f32_vscaleexpminusmax_test",
4909 srcs = [
4910 "test/f32-vscaleexpminusmax.cc",
4911 "test/vscaleexpminusmax-microkernel-tester.h",
4912 ] + MICROKERNEL_TEST_HDRS,
4913 deps = MICROKERNEL_TEST_DEPS,
4914)
4915
4916xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004917 name = "f32_vscaleextexp_test",
4918 srcs = [
4919 "test/f32-vscaleextexp.cc",
4920 "test/vscaleextexp-microkernel-tester.h",
4921 ] + MICROKERNEL_TEST_HDRS,
4922 deps = MICROKERNEL_TEST_DEPS,
4923)
4924
4925xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004926 name = "f32_vsqr_test",
4927 srcs = [
4928 "test/f32-vsqr.cc",
4929 "test/vunary-microkernel-tester.h",
4930 ] + MICROKERNEL_TEST_HDRS,
4931 deps = MICROKERNEL_TEST_DEPS,
4932)
4933
4934xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004935 name = "f32_vsqrdiff_test",
4936 srcs = [
4937 "test/f32-vsqrdiff.cc",
4938 "test/vbinary-microkernel-tester.h",
4939 ] + MICROKERNEL_TEST_HDRS,
4940 deps = MICROKERNEL_TEST_DEPS,
4941)
4942
4943xnnpack_unit_test(
4944 name = "f32_vsqrdiffc_test",
4945 srcs = [
4946 "test/f32-vsqrdiffc.cc",
4947 "test/vbinaryc-microkernel-tester.h",
4948 ] + MICROKERNEL_TEST_HDRS,
4949 deps = MICROKERNEL_TEST_DEPS,
4950)
4951
4952xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004953 name = "f32_vsqrt_test",
4954 srcs = [
4955 "test/f32-vsqrt.cc",
4956 "test/vunary-microkernel-tester.h",
4957 ] + MICROKERNEL_TEST_HDRS,
4958 deps = MICROKERNEL_TEST_DEPS,
4959)
4960
4961xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004962 name = "f32_vsub_test",
4963 srcs = [
4964 "test/f32-vsub.cc",
4965 "test/vbinary-microkernel-tester.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004971 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004972 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004973 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004974 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004980 name = "f32_vsubc_test",
4981 srcs = [
4982 "test/f32-vsubc.cc",
4983 "test/vbinaryc-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004989 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004990 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004991 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004992 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004998 name = "f32_vrsubc_test",
4999 srcs = [
5000 "test/f32-vrsubc.cc",
5001 "test/vbinaryc-microkernel-tester.h",
5002 ] + MICROKERNEL_TEST_HDRS,
5003 deps = MICROKERNEL_TEST_DEPS,
5004)
5005
5006xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005007 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005008 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005009 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005010 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005016 name = "qs8_gemm_minmax_test",
5017 srcs = [
5018 "test/qs8-gemm-minmax.cc",
5019 "test/gemm-microkernel-tester.h",
5020 "src/xnnpack/AlignedAllocator.h",
5021 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5023)
5024
5025xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005026 name = "qs8_igemm_minmax_test",
5027 srcs = [
5028 "test/qs8-igemm-minmax.cc",
5029 "test/gemm-microkernel-tester.h",
5030 "src/xnnpack/AlignedAllocator.h",
5031 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5033)
5034
5035xnnpack_unit_test(
5036 name = "qs8_requantization_test",
5037 srcs = [
5038 "src/xnnpack/requantization-stubs.h",
5039 "test/qs8-requantization.cc",
5040 "test/requantization-tester.h",
5041 ] + MICROKERNEL_TEST_HDRS,
5042 deps = MICROKERNEL_TEST_DEPS,
5043)
5044
5045xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005046 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005048 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 "test/avgpool-microkernel-tester.h",
5050 "src/xnnpack/AlignedAllocator.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005056 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005058 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059 "test/gemm-microkernel-tester.h",
5060 "src/xnnpack/AlignedAllocator.h",
5061 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005062 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063)
5064
5065xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005066 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005068 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005069 "test/dwconv-microkernel-tester.h",
5070 "src/xnnpack/AlignedAllocator.h",
5071 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005072 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073)
5074
5075xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005076 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005077 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005078 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005079 "test/gavgpool-microkernel-tester.h",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005086 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005088 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089 "test/gemm-microkernel-tester.h",
5090 "src/xnnpack/AlignedAllocator.h",
5091 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005092 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093)
5094
5095xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005096 name = "qu8_requantization_test",
5097 srcs = [
5098 "src/xnnpack/requantization-stubs.h",
5099 "test/qu8-requantization.cc",
5100 "test/requantization-tester.h",
5101 ] + MICROKERNEL_TEST_HDRS,
5102 deps = MICROKERNEL_TEST_DEPS,
5103)
5104
5105xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005106 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005107 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005108 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005109 "test/vadd-microkernel-tester.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
5115 name = "u8_clamp_test",
5116 srcs = [
5117 "test/u8-clamp.cc",
5118 "test/clamp-microkernel-tester.h",
5119 ] + MICROKERNEL_TEST_HDRS,
5120 deps = MICROKERNEL_TEST_DEPS,
5121)
5122
5123xnnpack_unit_test(
5124 name = "u8_lut32norm_test",
5125 srcs = [
5126 "test/u8-lut32norm.cc",
5127 "test/lut-norm-microkernel-tester.h",
5128 ] + MICROKERNEL_TEST_HDRS,
5129 deps = MICROKERNEL_TEST_DEPS,
5130)
5131
5132xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005133 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005135 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005136 "test/maxpool-microkernel-tester.h",
5137 ] + MICROKERNEL_TEST_HDRS,
5138 deps = MICROKERNEL_TEST_DEPS,
5139)
5140
5141xnnpack_unit_test(
5142 name = "u8_rmax_test",
5143 srcs = [
5144 "test/u8-rmax.cc",
5145 "test/rmax-microkernel-tester.h",
5146 ] + MICROKERNEL_TEST_HDRS,
5147 deps = MICROKERNEL_TEST_DEPS,
5148)
5149
5150xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005151 name = "x32_fill_test",
5152 srcs = [
5153 "test/x32-fill.cc",
5154 "test/fill-microkernel-tester.h",
5155 ] + MICROKERNEL_TEST_HDRS,
5156 deps = MICROKERNEL_TEST_DEPS,
5157)
5158
5159xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005160 name = "x32_packx_test",
5161 srcs = [
5162 "test/x32-packx.cc",
5163 "test/pack-microkernel-tester.h",
5164 "src/xnnpack/AlignedAllocator.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
5170 name = "x32_pad_test",
5171 srcs = [
5172 "test/x32-pad.cc",
5173 "test/pad-microkernel-tester.h",
5174 ] + MICROKERNEL_TEST_HDRS,
5175 deps = MICROKERNEL_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
5179 name = "x32_unpool_test",
5180 srcs = [
5181 "test/x32-unpool.cc",
5182 "test/unpool-microkernel-tester.h",
5183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
5188 name = "x32_zip_test",
5189 srcs = [
5190 "test/x32-zip.cc",
5191 "test/zip-microkernel-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
5197 name = "x8_lut_test",
5198 srcs = [
5199 "test/x8-lut.cc",
5200 "test/lut-microkernel-tester.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
5206 name = "x8_zip_test",
5207 srcs = [
5208 "test/x8-zip.cc",
5209 "test/zip-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
Marat Dukhan20c3b922020-03-10 03:45:06 -07005214########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215
5216xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005217 name = "operator_size_test",
5218 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005219 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005220)
5221
Marat Dukhan20c3b922020-03-10 03:45:06 -07005222xnnpack_binary(
5223 name = "subgraph_size_test",
5224 srcs = ["test/subgraph-size.c"],
5225 deps = [":XNNPACK"],
5226)
5227
5228########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229
5230xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005231 name = "abs_nc_test",
5232 srcs = [
5233 "test/abs-nc.cc",
5234 "test/abs-operator-tester.h",
5235 ],
5236 deps = OPERATOR_TEST_DEPS,
5237)
5238
5239xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005240 name = "add_nd_test",
5241 srcs = [
5242 "test/add-nd.cc",
5243 "test/binary-elementwise-operator-tester.h",
5244 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005245 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005246)
5247
5248xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005249 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005251 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005252 "test/argmax-pooling-operator-tester.h",
5253 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255)
5256
5257xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005258 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005260 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005261 "test/average-pooling-operator-tester.h",
5262 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005263 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264)
5265
5266xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005267 name = "bankers_rounding_nc_test",
5268 srcs = [
5269 "test/bankers-rounding-nc.cc",
5270 "test/bankers-rounding-operator-tester.h",
5271 ],
5272 deps = OPERATOR_TEST_DEPS,
5273)
5274
5275xnnpack_unit_test(
5276 name = "ceiling_nc_test",
5277 srcs = [
5278 "test/ceiling-nc.cc",
5279 "test/ceiling-operator-tester.h",
5280 ],
5281 deps = OPERATOR_TEST_DEPS,
5282)
5283
5284xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005285 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005287 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 "test/channel-shuffle-operator-tester.h",
5289 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291)
5292
5293xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005294 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005296 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 "test/clamp-operator-tester.h",
5298 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005299 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300)
5301
5302xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005303 name = "constant_pad_nd_test",
5304 srcs = [
5305 "test/constant-pad-nd.cc",
5306 "test/constant-pad-operator-tester.h",
5307 ],
5308 deps = OPERATOR_TEST_DEPS,
5309)
5310
5311xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005312 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005313 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005314 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005315 "test/convolution-operator-tester.h",
5316 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005317 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318)
5319
5320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005321 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005323 "test/convolution-nchw.cc",
5324 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005326 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327)
5328
5329xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005330 name = "copy_nc_test",
5331 srcs = [
5332 "test/copy-nc.cc",
5333 "test/copy-operator-tester.h",
5334 ],
5335 deps = OPERATOR_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005339 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005340 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005341 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342 "test/deconvolution-operator-tester.h",
5343 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005344 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345)
5346
5347xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005348 name = "divide_nd_test",
5349 srcs = [
5350 "test/binary-elementwise-operator-tester.h",
5351 "test/divide-nd.cc",
5352 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005353 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005354)
5355
5356xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005357 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005358 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005359 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005360 "test/fully-connected-operator-tester.h",
5361 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005362 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363)
5364
5365xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005366 name = "floor_nc_test",
5367 srcs = [
5368 "test/floor-nc.cc",
5369 "test/floor-operator-tester.h",
5370 ],
5371 deps = OPERATOR_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005375 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005377 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005378 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005379 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381)
5382
5383xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005384 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005386 "test/global-average-pooling-ncw.cc",
5387 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005389 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005390)
5391
5392xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005393 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005395 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005396 "test/hardswish-operator-tester.h",
5397 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005398 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399)
5400
5401xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005402 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005404 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405 "test/leaky-relu-operator-tester.h",
5406 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005407 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408)
5409
5410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005411 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005413 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414 "test/max-pooling-operator-tester.h",
5415 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417)
5418
5419xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005420 name = "maximum_nd_test",
5421 srcs = [
5422 "test/binary-elementwise-operator-tester.h",
5423 "test/maximum-nd.cc",
5424 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005425 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005426)
5427
5428xnnpack_unit_test(
5429 name = "minimum_nd_test",
5430 srcs = [
5431 "test/binary-elementwise-operator-tester.h",
5432 "test/minimum-nd.cc",
5433 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005434 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005435)
5436
5437xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005438 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005439 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005440 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005441 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005442 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005443 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005444)
5445
5446xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005447 name = "negate_nc_test",
5448 srcs = [
5449 "test/negate-nc.cc",
5450 "test/negate-operator-tester.h",
5451 ],
5452 deps = OPERATOR_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005456 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005458 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005459 "test/prelu-operator-tester.h",
5460 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462)
5463
5464xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005465 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005466 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005467 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005468 "test/resize-bilinear-operator-tester.h",
5469 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005471)
5472
5473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005474 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005476 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477 "test/sigmoid-operator-tester.h",
5478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480)
5481
5482xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005483 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005484 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005485 "test/softmax-nc.cc",
5486 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005487 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005488 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005489)
5490
5491xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005492 name = "square_nc_test",
5493 srcs = [
5494 "test/square-nc.cc",
5495 "test/square-operator-tester.h",
5496 ],
5497 deps = OPERATOR_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005501 name = "square_root_nc_test",
5502 srcs = [
5503 "test/square-root-nc.cc",
5504 "test/square-root-operator-tester.h",
5505 ],
5506 deps = OPERATOR_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005510 name = "squared_difference_nd_test",
5511 srcs = [
5512 "test/binary-elementwise-operator-tester.h",
5513 "test/squared-difference-nd.cc",
5514 ],
5515 deps = OPERATOR_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005519 name = "subtract_nd_test",
5520 srcs = [
5521 "test/binary-elementwise-operator-tester.h",
5522 "test/subtract-nd.cc",
5523 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005524 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005525)
5526
5527xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005528 name = "truncation_nc_test",
5529 srcs = [
5530 "test/truncation-nc.cc",
5531 "test/truncation-operator-tester.h",
5532 ],
5533 deps = OPERATOR_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005537 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005538 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005539 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 "test/unpooling-operator-tester.h",
5541 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005542 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543)
5544
Chao Mei6ddfc602020-05-13 22:29:36 -07005545############################### Misc unit tests ###############################
5546
5547xnnpack_unit_test(
5548 name = "memory_planner_test",
5549 srcs = [
5550 "test/memory-planner-test.cc",
5551 ],
5552 deps = [
5553 ":XNNPACK",
5554 ":memory_planner",
5555 ],
5556)
5557
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558############################# Build configurations #############################
5559
Marat Dukhanb8642352019-10-30 15:43:02 -07005560# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005562 name = "xnn_enable_assembly_explicit_true",
5563 define_values = {"xnn_enable_assembly": "true"},
5564)
5565
5566# Disables usage of assembly kernels.
5567config_setting(
5568 name = "xnn_enable_assembly_explicit_false",
5569 define_values = {"xnn_enable_assembly": "false"},
5570)
5571
Marat Dukhan9de90e02020-06-18 16:04:12 -07005572# Enables usage of sparse inference.
5573config_setting(
5574 name = "xnn_enable_sparse_explicit_true",
5575 define_values = {"xnn_enable_sparse": "true"},
5576)
5577
5578# Disables usage of sparse inference.
5579config_setting(
5580 name = "xnn_enable_sparse_explicit_false",
5581 define_values = {"xnn_enable_sparse": "false"},
5582)
5583
Marat Dukhan05702cf2020-03-26 15:41:33 -07005584# Disables usage of HMP-aware optimizations.
5585config_setting(
5586 name = "xnn_enable_hmp_explicit_false",
5587 define_values = {"xnn_enable_hmp": "false"},
5588)
5589
Chao Mei6ddfc602020-05-13 22:29:36 -07005590# Enable usage of optimized memory allocation
5591config_setting(
5592 name = "xnn_enable_memopt_explicit_true",
5593 define_values = {"xnnpack_enable_memopt": "true"},
5594)
5595
5596# Disable usage of optimized memory allocation
5597config_setting(
5598 name = "xnn_enable_memopt_explicit_false",
5599 define_values = {"xnnpack_enable_memopt": "false"},
5600)
5601
Marat Dukhanb8642352019-10-30 15:43:02 -07005602# Builds with -c dbg
5603config_setting(
5604 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005605 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005606 "compilation_mode": "dbg",
5607 },
5608)
5609
5610# Builds with -c opt
5611config_setting(
5612 name = "optimized_build",
5613 values = {
5614 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005615 },
5616)
5617
5618config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005619 name = "linux_k8",
5620 values = {"cpu": "k8"},
5621)
5622
5623config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005624 name = "linux_arm",
5625 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005626)
5627
5628config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005629 name = "linux_armeabi",
5630 values = {"cpu": "armeabi"},
5631)
5632
5633config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005634 name = "linux_armhf",
5635 values = {"cpu": "armhf"},
5636)
5637
5638config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005639 name = "linux_armv7a",
5640 values = {"cpu": "armv7a"},
5641)
5642
5643config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005644 name = "linux_aarch64",
5645 values = {"cpu": "aarch64"},
5646)
5647
5648config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649 name = "android",
5650 values = {"crosstool_top": "//external:android/crosstool"},
5651)
5652
5653config_setting(
5654 name = "android_armv7",
5655 values = {
5656 "crosstool_top": "//external:android/crosstool",
5657 "cpu": "armeabi-v7a",
5658 },
5659)
5660
5661config_setting(
5662 name = "android_arm64",
5663 values = {
5664 "crosstool_top": "//external:android/crosstool",
5665 "cpu": "arm64-v8a",
5666 },
5667)
5668
5669config_setting(
5670 name = "android_x86",
5671 values = {
5672 "crosstool_top": "//external:android/crosstool",
5673 "cpu": "x86",
5674 },
5675)
5676
5677config_setting(
5678 name = "android_x86_64",
5679 values = {
5680 "crosstool_top": "//external:android/crosstool",
5681 "cpu": "x86_64",
5682 },
5683)
5684
5685config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005686 name = "windows_x86_64",
5687 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005688)
5689
5690config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005691 name = "windows_x86_64_clang",
5692 values = {
5693 "compiler": "clang-cl",
5694 "cpu": "x64_windows",
5695 },
5696)
5697
5698config_setting(
5699 name = "windows_x86_64_mingw",
5700 values = {
5701 "compiler": "mingw-gcc",
5702 "cpu": "x64_windows",
5703 },
5704)
5705
5706config_setting(
5707 name = "windows_x86_64_msys",
5708 values = {
5709 "compiler": "msys-gcc",
5710 "cpu": "x64_windows",
5711 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005712)
5713
5714config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005715 name = "macos_x86_64",
5716 values = {
5717 "apple_platform_type": "macos",
5718 "cpu": "darwin",
5719 },
5720)
5721
5722config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005723 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005724 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005725)
5726
5727config_setting(
5728 name = "emscripten_wasm",
5729 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005730 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005731 "cpu": "wasm",
5732 },
5733)
5734
5735config_setting(
5736 name = "emscripten_wasmsimd",
5737 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005738 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005739 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005740 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005741 },
5742)
5743
5744config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005745 name = "ios_armv7",
5746 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005747 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005748 "cpu": "ios_armv7",
5749 },
5750)
5751
5752config_setting(
5753 name = "ios_arm64",
5754 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005755 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005756 "cpu": "ios_arm64",
5757 },
5758)
5759
5760config_setting(
5761 name = "ios_arm64e",
5762 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005763 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005764 "cpu": "ios_arm64e",
5765 },
5766)
5767
5768config_setting(
5769 name = "ios_x86",
5770 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005771 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005772 "cpu": "ios_i386",
5773 },
5774)
5775
5776config_setting(
5777 name = "ios_x86_64",
5778 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005779 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005780 "cpu": "ios_x86_64",
5781 },
5782)
5783
5784config_setting(
5785 name = "watchos_armv7k",
5786 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005787 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005788 "cpu": "watchos_armv7k",
5789 },
5790)
5791
5792config_setting(
5793 name = "watchos_arm64_32",
5794 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005795 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005796 "cpu": "watchos_arm64_32",
5797 },
5798)
5799
5800config_setting(
5801 name = "watchos_x86",
5802 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005803 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005804 "cpu": "watchos_i386",
5805 },
5806)
5807
5808config_setting(
5809 name = "watchos_x86_64",
5810 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005811 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005812 "cpu": "watchos_x86_64",
5813 },
5814)
5815
5816config_setting(
5817 name = "tvos_arm64",
5818 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005819 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005820 "cpu": "tvos_arm64",
5821 },
5822)
5823
5824config_setting(
5825 name = "tvos_x86_64",
5826 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005827 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005828 "cpu": "tvos_x86_64",
5829 },
5830)