blob: 4ee6119a1d58b0bd5f6c92d6fa365bd966b9a6fa [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",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070039 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070047 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070048 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/argmax-pooling-nhwc.c",
54 "src/operators/average-pooling-nhwc.c",
55 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070056 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070057 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070058 "src/operators/convolution-nchw.c",
59 "src/operators/convolution-nhwc.c",
60 "src/operators/deconvolution-nhwc.c",
61 "src/operators/fully-connected-nc.c",
62 "src/operators/global-average-pooling-ncw.c",
63 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/leaky-relu-nc.c",
65 "src/operators/max-pooling-nhwc.c",
66 "src/operators/prelu-nc.c",
67 "src/operators/resize-bilinear-nhwc.c",
68 "src/operators/sigmoid-nc.c",
69 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070070 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070072]
73
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070074SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070075 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070076 "src/subgraph/add2.c",
77 "src/subgraph/argmax-pooling-2d.c",
78 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/bankers-rounding.c",
80 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070081 "src/subgraph/clamp.c",
82 "src/subgraph/convolution-2d.c",
83 "src/subgraph/deconvolution-2d.c",
84 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070085 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070086 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070087 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070088 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070089 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070090 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070091 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070092 "src/subgraph/maximum2.c",
93 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070094 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070095 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070096 "src/subgraph/prelu.c",
97 "src/subgraph/sigmoid.c",
98 "src/subgraph/softmax.c",
99 "src/subgraph/static-constant-pad.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700100 "src/subgraph/square.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700101 "src/subgraph/squared-difference.c",
102 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700103 "src/subgraph/unpooling-2d.c",
104]
105
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800106TABLE_SRCS = [
107 "src/tables/exp2-k-over-64.c",
108 "src/tables/exp2-k-over-2048.c",
109]
110
Marat Dukhan08c4a432019-10-03 09:29:21 -0700111SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800112 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800113 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800114 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700115 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
116 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700117 "src/f32-clamp/gen/scalar-x1.c",
118 "src/f32-clamp/gen/scalar-x2.c",
119 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800120 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800121 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700122 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
123 "src/f32-dwconv-chw/3x3p1-scalar.c",
124 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
125 "src/f32-dwconv-chw/5x5p2-scalar.c",
126 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
128 "src/f32-dwconv/gen/up1x4-scalar.c",
129 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
130 "src/f32-dwconv/gen/up1x9-scalar.c",
131 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x25-scalar.c",
133 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
134 "src/f32-dwconv/gen/up2x4-scalar.c",
135 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
136 "src/f32-dwconv/gen/up2x9-scalar.c",
137 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700139 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700143 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
147 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700149 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700151 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700152 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
153 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700154 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
155 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
156 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700157 "src/f32-gemm/gen/1x4-scalar.c",
158 "src/f32-gemm/gen/2x4-scalar.c",
159 "src/f32-gemm/gen/4x2-scalar.c",
160 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700161 "src/f32-gemm/gen/1x4-relu-scalar.c",
162 "src/f32-gemm/gen/2x4-relu-scalar.c",
163 "src/f32-gemm/gen/4x2-relu-scalar.c",
164 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700165 "src/f32-gemm/gen/1x4-minmax-scalar.c",
166 "src/f32-gemm/gen/2x4-minmax-scalar.c",
167 "src/f32-gemm/gen/4x2-minmax-scalar.c",
168 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800169 "src/f32-hswish/gen/scalar-x1.c",
170 "src/f32-hswish/gen/scalar-x2.c",
171 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700172 "src/f32-ibilinear/gen/scalar-c1.c",
173 "src/f32-ibilinear/gen/scalar-c2.c",
174 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700175 "src/f32-igemm/gen/1x4-scalar.c",
176 "src/f32-igemm/gen/2x4-scalar.c",
177 "src/f32-igemm/gen/4x2-scalar.c",
178 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700179 "src/f32-igemm/gen/1x4-relu-scalar.c",
180 "src/f32-igemm/gen/2x4-relu-scalar.c",
181 "src/f32-igemm/gen/4x2-relu-scalar.c",
182 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700183 "src/f32-igemm/gen/1x4-minmax-scalar.c",
184 "src/f32-igemm/gen/2x4-minmax-scalar.c",
185 "src/f32-igemm/gen/4x2-minmax-scalar.c",
186 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700187 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
188 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
189 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700190 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
191 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
192 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
193 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800194 "src/f32-prelu/gen/scalar-2x1.c",
195 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800196 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
197 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700208 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800209 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
210 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
211 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
212 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
213 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
214 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
215 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
216 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
217 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700218 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
219 "src/f32-spmm/gen/1x1-minmax-scalar.c",
220 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
221 "src/f32-spmm/gen/2x1-minmax-scalar.c",
222 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
223 "src/f32-spmm/gen/4x1-minmax-scalar.c",
224 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
225 "src/f32-spmm/gen/8x1-minmax-scalar.c",
226 "src/f32-spmm/gen/8x2-minmax-scalar.c",
227 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700228 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
229 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
230 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
231 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
232 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
233 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
234 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
235 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
236 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
237 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
238 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
239 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800240 "src/f32-vbinary/gen/vmax-scalar-x1.c",
241 "src/f32-vbinary/gen/vmax-scalar-x2.c",
242 "src/f32-vbinary/gen/vmax-scalar-x4.c",
243 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
244 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
245 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
246 "src/f32-vbinary/gen/vmin-scalar-x1.c",
247 "src/f32-vbinary/gen/vmin-scalar-x2.c",
248 "src/f32-vbinary/gen/vmin-scalar-x4.c",
249 "src/f32-vbinary/gen/vminc-scalar-x1.c",
250 "src/f32-vbinary/gen/vminc-scalar-x2.c",
251 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700252 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
253 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
254 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
255 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
256 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
257 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
258 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
259 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
260 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700261 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x1.c",
262 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x2.c",
263 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700264 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
265 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
266 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700267 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
268 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
269 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
270 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
271 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
272 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700273 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
274 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
275 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
276 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700279 "src/f32-vlrelu/gen/scalar-x1.c",
280 "src/f32-vlrelu/gen/scalar-x2.c",
281 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700282 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
283 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
284 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700285 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
286 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
287 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
288 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
289 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
290 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
291 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
292 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
293 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
294 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
295 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
296 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700297 "src/f32-vunary/gen/vabs-scalar-x1.c",
298 "src/f32-vunary/gen/vabs-scalar-x2.c",
299 "src/f32-vunary/gen/vabs-scalar-x4.c",
300 "src/f32-vunary/gen/vneg-scalar-x1.c",
301 "src/f32-vunary/gen/vneg-scalar-x2.c",
302 "src/f32-vunary/gen/vneg-scalar-x4.c",
303 "src/f32-vunary/gen/vsqr-scalar-x1.c",
304 "src/f32-vunary/gen/vsqr-scalar-x2.c",
305 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800306 "src/math/expminus-scalar-lut2048-p1.c",
307 "src/math/expminus-scalar-lut64-p2.c",
308 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700309 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700310 "src/math/roundne-scalar-nearbyint.c",
311 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700312 "src/math/roundd-scalar-addsub.c",
313 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700314 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700315 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700316 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700317 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700318 "src/math/roundz-scalar-addsub.c",
319 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700320 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800321 "src/math/sigmoid-scalar-lut2048-p1-div.c",
322 "src/math/sigmoid-scalar-lut64-p2-div.c",
323 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700324 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
325 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700326 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700327 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
328 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700329 "src/q8-gemm/2x2-minmax-scalar.c",
330 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700331 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700332 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700333 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700334 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700336 "src/x32-fill/scalar-float.c",
337 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700338 "src/x32-packx/x2-scalar.c",
339 "src/x32-packx/x3-scalar.c",
340 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700341 "src/x32-pad/scalar-int.c",
342 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700343 "src/x32-unpool/scalar.c",
344 "src/x32-zip/x2-scalar.c",
345 "src/x32-zip/x3-scalar.c",
346 "src/x32-zip/x4-scalar.c",
347 "src/x32-zip/xm-scalar.c",
348 "src/x8-lut/scalar.c",
349 "src/x8-zip/x2-scalar.c",
350 "src/x8-zip/x3-scalar.c",
351 "src/x8-zip/x4-scalar.c",
352 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700353 "src/requantization/precise-scalar.c",
354 "src/requantization/fp32-scalar.c",
355 "src/requantization/q31-scalar.c",
356 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700357]
358
Marat Dukhan436ebe62019-12-04 15:10:12 -0800359WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700360 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
361 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700362 "src/f32-clamp/gen/wasm-x1.c",
363 "src/f32-clamp/gen/wasm-x2.c",
364 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700365 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
366 "src/f32-dwconv/gen/up1x4-wasm.c",
367 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
368 "src/f32-dwconv/gen/up1x9-wasm.c",
369 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
370 "src/f32-dwconv/gen/up1x25-wasm.c",
371 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
372 "src/f32-dwconv/gen/up2x4-wasm.c",
373 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
374 "src/f32-dwconv/gen/up2x9-wasm.c",
375 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
376 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700377 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
378 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
379 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
380 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700381 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
382 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700383 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
384 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
385 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
386 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700387 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
388 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700389 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
390 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700391 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
392 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
393 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700394 "src/f32-gemm/gen/1x4-wasm.c",
395 "src/f32-gemm/gen/2x4-wasm.c",
396 "src/f32-gemm/gen/4x2-wasm.c",
397 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700398 "src/f32-gemm/gen/1x4-relu-wasm.c",
399 "src/f32-gemm/gen/2x4-relu-wasm.c",
400 "src/f32-gemm/gen/4x2-relu-wasm.c",
401 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700402 "src/f32-gemm/gen/1x4-minmax-wasm.c",
403 "src/f32-gemm/gen/2x4-minmax-wasm.c",
404 "src/f32-gemm/gen/4x2-minmax-wasm.c",
405 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800406 "src/f32-hswish/gen/wasm-x1.c",
407 "src/f32-hswish/gen/wasm-x2.c",
408 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700409 "src/f32-igemm/gen/1x4-wasm.c",
410 "src/f32-igemm/gen/2x4-wasm.c",
411 "src/f32-igemm/gen/4x2-wasm.c",
412 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700413 "src/f32-igemm/gen/1x4-relu-wasm.c",
414 "src/f32-igemm/gen/2x4-relu-wasm.c",
415 "src/f32-igemm/gen/4x2-relu-wasm.c",
416 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700417 "src/f32-igemm/gen/1x4-minmax-wasm.c",
418 "src/f32-igemm/gen/2x4-minmax-wasm.c",
419 "src/f32-igemm/gen/4x2-minmax-wasm.c",
420 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700421 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
422 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
423 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700424 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
425 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
426 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
427 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
428 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
429 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
430 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
431 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
432 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
433 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
434 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
435 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800436 "src/f32-vbinary/gen/vmax-wasm-x1.c",
437 "src/f32-vbinary/gen/vmax-wasm-x2.c",
438 "src/f32-vbinary/gen/vmax-wasm-x4.c",
439 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
440 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
441 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
442 "src/f32-vbinary/gen/vmin-wasm-x1.c",
443 "src/f32-vbinary/gen/vmin-wasm-x2.c",
444 "src/f32-vbinary/gen/vmin-wasm-x4.c",
445 "src/f32-vbinary/gen/vminc-wasm-x1.c",
446 "src/f32-vbinary/gen/vminc-wasm-x2.c",
447 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700448 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
449 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
450 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
451 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
452 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
453 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
454 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
455 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
456 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
457 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
458 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
459 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
460 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
461 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
462 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
463 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
464 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
465 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700466 "src/f32-vlrelu/gen/wasm-x1.c",
467 "src/f32-vlrelu/gen/wasm-x2.c",
468 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700469 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
470 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
471 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800472]
473
Marat Dukhan290055c2020-06-09 12:24:29 -0700474WASMSIMD_UKERNELS = [
475 "src/requantization/fp32-wasmsimd.c",
476]
477
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800478PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800479 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800480 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800481 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700482 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
483 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700484 "src/f32-clamp/gen/psimd-x4.c",
485 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700486 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700487 "src/f32-dwconv-chw/3x3p1-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700488 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
489 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
490 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
491 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
492 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
493 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
494 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
495 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
496 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
497 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
498 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
499 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700500 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700501 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
502 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700503 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
504 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
505 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
506 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
507 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
508 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
509 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
510 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
511 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
512 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
513 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
514 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
515 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
516 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
517 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
518 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
519 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
520 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
521 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800522 "src/f32-hswish/gen/psimd-x4.c",
523 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700524 "src/f32-ibilinear/gen/psimd-c4.c",
525 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700526 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
527 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
528 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
529 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
530 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
531 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
532 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
533 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
534 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
535 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700536 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700537 "src/f32-spmm/gen/4x1-minmax-psimd.c",
538 "src/f32-spmm/gen/8x1-minmax-psimd.c",
539 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
541 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700542 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800543 "src/f32-prelu/gen/psimd-2x4.c",
544 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800545 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
547 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
548 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
549 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
550 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
551 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
552 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
553 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800554 "src/f32-vbinary/gen/vmax-psimd-x4.c",
555 "src/f32-vbinary/gen/vmax-psimd-x8.c",
556 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
557 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
558 "src/f32-vbinary/gen/vmin-psimd-x4.c",
559 "src/f32-vbinary/gen/vmin-psimd-x8.c",
560 "src/f32-vbinary/gen/vminc-psimd-x4.c",
561 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700562 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
563 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
564 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
565 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
566 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
567 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700568 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x4.c",
569 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700570 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
571 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700572 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
573 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
574 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
575 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700576 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
577 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
578 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
579 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700580 "src/f32-vlrelu/gen/psimd-x4.c",
581 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700582 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
583 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700584 "src/f32-vunary/gen/vabs-psimd-x4.c",
585 "src/f32-vunary/gen/vabs-psimd-x8.c",
586 "src/f32-vunary/gen/vneg-psimd-x4.c",
587 "src/f32-vunary/gen/vneg-psimd-x8.c",
588 "src/f32-vunary/gen/vsqr-psimd-x4.c",
589 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700590 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700592 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700593 "src/x32-unpool/psimd.c",
594 "src/x32-zip/x2-psimd.c",
595 "src/x32-zip/x3-psimd.c",
596 "src/x32-zip/x4-psimd.c",
597 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700598 "src/requantization/precise-psimd.c",
599 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700600]
601
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800602PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800603 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
604 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
605 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
606 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
607 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
608 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
609 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
610 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
611 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
612 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
613 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
614 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800615 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
616 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
617 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
618 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
619 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
620 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700621 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
622 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
623 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
624 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
625 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
626 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
627 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
628 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700629 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700630 "src/math/roundd-psimd-addsub.c",
631 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700632 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800633 "src/math/sigmoid-psimd-p5-div.c",
634]
635
Marat Dukhan08c4a432019-10-03 09:29:21 -0700636# ISA-specific micro-kernels
637NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700638 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
639 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700640 "src/f32-clamp/gen/neon-x4.c",
641 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700642 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
643 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700644 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
645 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700646 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
647 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700648 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
649 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700650 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
651 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
652 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
653 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700654 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
655 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
656 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
657 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700658 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
659 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
660 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
661 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700662 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700663 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
664 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700665 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
666 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
667 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
668 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
669 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
670 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
671 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
672 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
673 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
674 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
675 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
676 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
677 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
678 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
679 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
680 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
681 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
682 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
683 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
684 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
685 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
686 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
687 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
688 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
689 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
690 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
691 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
692 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
693 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
694 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
695 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800696 "src/f32-hswish/gen/neon-x4.c",
697 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700698 "src/f32-ibilinear/gen/neon-c4.c",
699 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700700 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
701 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
702 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
703 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
704 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
705 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
706 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
707 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
708 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
709 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
710 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
711 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
712 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
713 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
714 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
715 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700716 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
717 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
718 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700719 "src/f32-ppmm/gen/4x8-minmax-neon.c",
720 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800721 "src/f32-prelu/gen/neon-2x4.c",
722 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800723 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
724 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
725 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
726 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
727 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
728 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
729 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
730 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
731 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
732 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
733 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
734 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
735 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
736 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
737 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
738 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
739 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
740 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
741 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
742 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
743 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
744 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
745 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
746 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700747 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800748 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800749 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
750 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
751 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
752 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
753 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
754 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
755 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
756 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
757 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
758 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
759 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
760 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
761 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
762 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
763 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
764 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
765 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
766 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700767 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
768 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
769 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
770 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800771 "src/f32-vbinary/gen/vmax-neon-x4.c",
772 "src/f32-vbinary/gen/vmax-neon-x8.c",
773 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
774 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
775 "src/f32-vbinary/gen/vmin-neon-x4.c",
776 "src/f32-vbinary/gen/vmin-neon-x8.c",
777 "src/f32-vbinary/gen/vminc-neon-x4.c",
778 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700779 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
780 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
781 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
782 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700783 "src/f32-vbinary/gen/vrsqrdiffc-neon-x4.c",
784 "src/f32-vbinary/gen/vrsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700785 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
786 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700787 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
788 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
789 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
790 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700791 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
792 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
793 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
794 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700795 "src/f32-vlrelu/gen/neon-x4.c",
796 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700797 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
798 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700799 "src/f32-vrnd/gen/vrndne-neon-x4.c",
800 "src/f32-vrnd/gen/vrndne-neon-x8.c",
801 "src/f32-vrnd/gen/vrndz-neon-x4.c",
802 "src/f32-vrnd/gen/vrndz-neon-x8.c",
803 "src/f32-vrnd/gen/vrndu-neon-x4.c",
804 "src/f32-vrnd/gen/vrndu-neon-x8.c",
805 "src/f32-vrnd/gen/vrndd-neon-x4.c",
806 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700807 "src/f32-vunary/gen/vabs-neon-x4.c",
808 "src/f32-vunary/gen/vabs-neon-x8.c",
809 "src/f32-vunary/gen/vneg-neon-x4.c",
810 "src/f32-vunary/gen/vneg-neon-x8.c",
811 "src/f32-vunary/gen/vsqr-neon-x4.c",
812 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700813 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
814 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700815 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700816 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
817 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700818 "src/q8-gemm/4x8-minmax-neon.c",
819 "src/q8-gemm/8x8-minmax-neon.c",
820 "src/q8-igemm/4x8-minmax-neon.c",
821 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700822 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700823 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700824 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700825 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700826 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700827 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700828 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700829 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700830 "src/x32-zip/x2-neon.c",
831 "src/x32-zip/x3-neon.c",
832 "src/x32-zip/x4-neon.c",
833 "src/x32-zip/xm-neon.c",
834 "src/x8-zip/x2-neon.c",
835 "src/x8-zip/x3-neon.c",
836 "src/x8-zip/x4-neon.c",
837 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700838 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700839 "src/math/roundd-neon-addsub.c",
840 "src/math/roundd-neon-cvt.c",
841 "src/math/roundu-neon-addsub.c",
842 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700843 "src/math/roundz-neon-addsub.c",
844 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800845 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800846 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
847 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
848 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
849 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
850 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
851 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700852 "src/requantization/precise-neon.c",
853 "src/requantization/fp32-neon.c",
854 "src/requantization/q31-neon.c",
855 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856]
857
858NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700859 "src/f32-ibilinear/gen/neonfma-c4.c",
860 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700861 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
862 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
863 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
864 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
865 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
866 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
867 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
868 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
869 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700870 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
871 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
872 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
873 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700874 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
875 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
876 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
877 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700878 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
879 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
880 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
881 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700882 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
883 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
884 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
885 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
886 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
887 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
888 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
889 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
890 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
891 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
892 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
893 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
894 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
895 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
896 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
897 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
898 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
899 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800900 "src/f32-hswish/gen/neonfma-x4.c",
901 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700902 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
903 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800904 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
905 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
906 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
907 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
916 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
917 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
918 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
919 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
920 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
921 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
922 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
923 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
924 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
925 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
926 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
927 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800928 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
929 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
930 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
931 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
932 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
933 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
934 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
935 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
936 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
937 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
938 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
939 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
940 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
941 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
942 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
943 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
944 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
945 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
946 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
947 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
948 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
949 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
950 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
951 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
952 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
953 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
954 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
955 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
956 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
957 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
958 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
959 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
960 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
961 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
962 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
963 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
964 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
965 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
966 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
967 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
968 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
969 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
970 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
971 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
972 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
973 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
974 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
975 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
976 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
977 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
978 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
979 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
980 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
981 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700982 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
983 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800984 "src/math/exp-neonfma-lut64-p2.c",
985 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800986 "src/math/expminus-neonfma-lut2048-p1.c",
987 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800988 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800989 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
990 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
991 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
992 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
993 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
994 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
995 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
996 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
997 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
998 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
999 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1000 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1001 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1002 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1003 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1004 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1005 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1006 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001007]
1008
1009AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001010 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1011 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1012 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1013 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1014 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1015 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001016 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1017 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1018 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1019 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1020 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1021 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1022 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1023 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1024 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1025 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1026 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1027 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1028 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1029 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1030 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1031 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1032 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1033 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1034 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1035 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001036 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1037 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001038 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1039 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001040 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1041 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001042 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1043 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001044 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1045 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1046 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1047 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1048 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001049 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1050 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1051 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1052 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1053 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1054 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1055 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1056 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1057 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1058 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1059 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1060 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1061 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1062 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1063 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1064 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1065 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1066 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001067 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1068 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1069 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1070 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1071 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1072 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1073 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1074 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1075 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1076 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1077 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1078 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1079 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1080 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1081 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1082 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1083 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1084 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001085 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1086 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1087 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1088 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1089 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1090 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001091]
1092
Marat Dukhan8853b822020-05-07 12:19:01 -07001093NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001094 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1095 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1096 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1097 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1098 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1099 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1100 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1101 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001102 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001103 "src/math/roundd-neonv8.c",
1104 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001105 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001106]
1107
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001109 "src/f16-clamp/gen/neonfp16arith-x16.c",
1110 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001111 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1112 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1113 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1114 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1115 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1116 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1117 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1118 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1119 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1120 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1121 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1122 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001123 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1124 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001125 "src/f16-hswish/gen/neonfp16arith-x16.c",
1126 "src/f16-hswish/gen/neonfp16arith-x8.c",
1127 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1128 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001129 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1130 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1131 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1132 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1133 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1134 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1135 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1136 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1137 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1138 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1139 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1140 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1141 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1142 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1143 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1144 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1145 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1146 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1147 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1148 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1149 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1150 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1151 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1152 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1153 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1154 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1155 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1156 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001157 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1158 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001159 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1160 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1161 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1162 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1163 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1164 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1165 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1166 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001167 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1168 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1169 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1170 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001171 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1172 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1173 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1174 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1175 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1176 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1177 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1178 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1179 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1180 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1181 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1182 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001183 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1184 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1185 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1186 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1187 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1188 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1189 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1190 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001191]
1192
1193SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001194 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1195 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001196 "src/f32-clamp/gen/sse-x4.c",
1197 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001198 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1199 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001200 "src/f32-dwconv-chw/3x3p1-sse.c",
1201 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001202 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1203 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1204 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1205 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1206 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1207 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1208 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1209 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1210 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1211 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1212 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1213 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001214 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001215 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1216 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001217 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1218 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1219 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1220 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1221 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1222 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1223 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1224 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1225 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1226 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1227 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1228 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1229 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001230 "src/f32-hswish/gen/sse-x4.c",
1231 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001232 "src/f32-ibilinear/gen/sse-c4.c",
1233 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001234 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1235 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1236 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1237 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1238 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1239 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1240 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001241 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1242 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1243 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001244 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001246 "src/f32-spmm/gen/4x1-minmax-sse.c",
1247 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001248 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001249 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1250 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1251 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1252 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1253 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1254 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1255 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1256 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001257 "src/f32-vbinary/gen/vmax-sse-x4.c",
1258 "src/f32-vbinary/gen/vmax-sse-x8.c",
1259 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1260 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1261 "src/f32-vbinary/gen/vmin-sse-x4.c",
1262 "src/f32-vbinary/gen/vmin-sse-x8.c",
1263 "src/f32-vbinary/gen/vminc-sse-x4.c",
1264 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1266 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1267 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1268 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1269 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1270 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001271 "src/f32-vbinary/gen/vrsqrdiffc-sse-x4.c",
1272 "src/f32-vbinary/gen/vrsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001273 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1274 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001275 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1276 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1277 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1278 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001279 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1280 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1281 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1282 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001283 "src/f32-vlrelu/gen/sse-x4.c",
1284 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001285 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1286 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001287 "src/f32-vunary/gen/vabs-sse-x4.c",
1288 "src/f32-vunary/gen/vabs-sse-x8.c",
1289 "src/f32-vunary/gen/vneg-sse-x4.c",
1290 "src/f32-vunary/gen/vneg-sse-x8.c",
1291 "src/f32-vunary/gen/vsqr-sse-x4.c",
1292 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001293 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001294 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001295 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001296 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001297 "src/math/roundd-sse-addsub.c",
1298 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001299 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001300]
1301
1302SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001303 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1304 "src/f32-argmaxpool/4x-sse2-c4.c",
1305 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001306 "src/f32-prelu/gen/sse2-2x4.c",
1307 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001308 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1309 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1310 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1311 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1312 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1313 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1314 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1315 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1316 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1317 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1318 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1319 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001320 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001321 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001322 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001323 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001324 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1325 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001326 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1327 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1328 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1329 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1330 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1331 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1332 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1333 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001334 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1335 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001336 "src/q8-igemm/4x4c2-minmax-sse2.c",
1337 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001338 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1339 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001340 "src/q8-gemm/2x4c8-minmax-sse2.c",
1341 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001342 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001343 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001344 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001345 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001346 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001347 "src/x32-zip/x2-sse2.c",
1348 "src/x32-zip/x3-sse2.c",
1349 "src/x32-zip/x4-sse2.c",
1350 "src/x32-zip/xm-sse2.c",
1351 "src/x8-zip/x2-sse2.c",
1352 "src/x8-zip/x3-sse2.c",
1353 "src/x8-zip/x4-sse2.c",
1354 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001355 "src/math/exp-sse2-p5.c",
1356 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001357 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001358 "src/math/roundd-sse2-cvt.c",
1359 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001360 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001361 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001362 "src/requantization/precise-sse2.c",
1363 "src/requantization/fp32-sse2.c",
1364 "src/requantization/q31-sse2.c",
1365 "src/requantization/gemmlowp-sse2.c",
1366]
1367
1368SSSE3_UKERNELS = [
1369 "src/requantization/precise-ssse3.c",
1370 "src/requantization/q31-ssse3.c",
1371 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001372]
1373
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001374SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001375 "src/f32-prelu/gen/sse41-2x4.c",
1376 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001377 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001378 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001379 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001380 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001381 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1382 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001383 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1384 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1385 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1386 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1387 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1388 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1389 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1390 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001391 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001392 "src/math/roundd-sse41.c",
1393 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001394 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001395 "src/requantization/precise-sse4.c",
1396 "src/requantization/q31-sse4.c",
1397 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001398]
1399
Marat Dukhan08c4a432019-10-03 09:29:21 -07001400AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001401 "src/f32-clamp/gen/avx-x8.c",
1402 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1404 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1405 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1406 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1407 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1408 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1409 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1410 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1411 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1412 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1413 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1414 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1415 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1416 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1417 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1418 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1419 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1420 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1421 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1422 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1423 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1424 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1425 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1426 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1427 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1428 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1429 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1430 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1431 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1432 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001433 "src/f32-hswish/gen/avx-x8.c",
1434 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001435 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1436 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1437 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1438 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1439 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1440 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1441 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1442 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1443 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001444 "src/f32-prelu/gen/avx-2x8.c",
1445 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001446 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001447 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1448 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1449 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1450 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1451 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1452 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1453 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1454 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001455 "src/f32-vbinary/gen/vmax-avx-x8.c",
1456 "src/f32-vbinary/gen/vmax-avx-x16.c",
1457 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1458 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1459 "src/f32-vbinary/gen/vmin-avx-x8.c",
1460 "src/f32-vbinary/gen/vmin-avx-x16.c",
1461 "src/f32-vbinary/gen/vminc-avx-x8.c",
1462 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001463 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1464 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1465 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1466 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1467 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1468 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001469 "src/f32-vbinary/gen/vrsqrdiffc-avx-x8.c",
1470 "src/f32-vbinary/gen/vrsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001471 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1472 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001473 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1474 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1475 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1476 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001477 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1478 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1479 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1480 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001481 "src/f32-vlrelu/gen/avx-x8.c",
1482 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001483 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001484 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1485 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1486 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1487 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1488 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1489 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1490 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1491 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001492 "src/f32-vunary/gen/vabs-avx-x8.c",
1493 "src/f32-vunary/gen/vabs-avx-x16.c",
1494 "src/f32-vunary/gen/vneg-avx-x8.c",
1495 "src/f32-vunary/gen/vneg-avx-x16.c",
1496 "src/f32-vunary/gen/vsqr-avx-x8.c",
1497 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001498]
1499
Marat Dukhanfda12b82019-11-21 12:27:59 -08001500FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1502 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1503 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1504 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1505 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1506 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1507 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1508 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1509 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1510 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1511 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1512 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1513 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1514 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1515 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1516 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1517 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1518 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1519 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1520 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1521 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1522 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1523 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1524 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1525 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1526 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1527 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1528 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1529 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1530 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1531 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1532 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1533 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1534 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1535 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1536 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1537 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1538 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1539 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1540 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001541 "src/f32-hswish/gen/fma3-x8.c",
1542 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001543 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1544 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1545 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1546 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1547 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1548 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1549 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1550 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1551 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1552 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1553 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1554 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1555 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1556 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001557]
1558
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001559AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001560 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1561 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1562 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1563 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1564 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1565 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1566 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1567 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1568 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1569 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1570 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1571 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1572 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1573 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1574 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1575 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1576 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1577 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1578 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1579 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1580 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1581 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1582 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1583 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1584 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1585 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1586 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1587 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1588 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1589 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1590 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1591 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1592 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1593 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1595 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001596 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1597 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1598 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1599 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1600 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1601 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1602 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1603 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1604 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1605 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1606 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1607 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1608 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1609 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1610 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1611 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1612 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1613 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1614 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1615 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1616 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1617 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1618 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1619 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1620 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1621 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1622 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1623 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1624 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1625 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001626 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1627 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1628 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1629 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1630 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1631 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1632 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1633 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1634 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1635 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1636 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1637 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1638 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1639 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1640 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1641 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1642 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1643 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1644 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1645 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1646 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1647 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1648 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1649 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001650 "src/math/exp-avx2-p5.c",
1651 "src/math/exp-avx2-perm-p3.c",
1652 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001653 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001654 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001655 "src/math/sigmoid-avx2-rr2-p5-div.c",
1656 "src/math/sigmoid-avx2-rr1-p5-div.c",
1657 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1658 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1659 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1660 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001661]
1662
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001664 "src/f32-clamp/gen/avx512f-x16.c",
1665 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001666 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1667 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1668 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1669 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1670 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1671 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1672 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1673 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1674 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1675 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1676 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1677 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1678 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1679 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1680 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1681 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1682 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1683 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1684 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1685 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1686 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1687 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1688 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1689 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001690 "src/f32-hswish/gen/avx512f-x16.c",
1691 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001692 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1693 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1694 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1695 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1696 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1697 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001698 "src/f32-prelu/gen/avx512f-2x16.c",
1699 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001700 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1701 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1702 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1703 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1704 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1705 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1706 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1707 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1708 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1709 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1710 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1711 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1712 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1713 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1714 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1715 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1716 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1717 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1718 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1719 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1720 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1721 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1722 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1723 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1724 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1725 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1726 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1727 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1728 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1729 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1730 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1731 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1732 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1733 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1734 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1735 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001736 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001737 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1738 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1739 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1740 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1741 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1742 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1743 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1744 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001745 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1746 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1747 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1748 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1749 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1750 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1751 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1752 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001753 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1754 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1755 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1756 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1757 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1758 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001759 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x16.c",
1760 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001761 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1762 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001763 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
1764 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
1765 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
1766 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001767 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1768 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1769 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1770 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001771 "src/f32-vlrelu/gen/avx512f-x16.c",
1772 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001773 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001774 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1775 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1776 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1777 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1778 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1779 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1780 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1781 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1782 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1783 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1784 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1785 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1786 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1787 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1788 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1789 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1790 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1791 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1792 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1793 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1794 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1795 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1796 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1797 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001798 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
1799 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
1800 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
1801 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
1802 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
1803 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
1804 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
1805 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001806 "src/f32-vunary/gen/vabs-avx512f-x16.c",
1807 "src/f32-vunary/gen/vabs-avx512f-x32.c",
1808 "src/f32-vunary/gen/vneg-avx512f-x16.c",
1809 "src/f32-vunary/gen/vneg-avx512f-x32.c",
1810 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
1811 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001812 "src/math/exp-avx512f-p5-scalef.c",
1813 "src/math/exp-avx512f-p5.c",
1814 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001815 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001816 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001817]
1818
1819AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001820 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001821 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001822 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1823 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1824 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1825 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1826 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1827 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1828 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1829 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1830 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1831 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832]
1833
1834AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001835 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001836 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1837 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001838 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001839 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1840 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001841 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1842 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1843 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001844 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001845 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1846 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1847 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001848 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001849 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1850 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001851 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001852 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1853 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1854 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1855 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1856 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1857 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1858 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1859 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1860 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1861 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1862 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1863 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1864 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1865 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1866 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1867 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1868 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1869 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1870 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1871 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1872 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001873 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001874 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1875 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1876 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1877 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1878 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1879 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1880 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1881 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1882 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1883 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1884 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1885 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1886 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1887 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1888 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1889 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1890 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1891 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1892 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1893 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1894 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1895 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1896 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1897 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1898 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1899 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1900 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1901 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1902 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1903 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1904 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1905 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1906 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1907 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1908 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1909 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1910 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1911 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001912]
1913
Marat Dukhan1b354632020-03-23 12:50:22 -07001914INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001915 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 "src/xnnpack/argmaxpool.h",
1917 "src/xnnpack/avgpool.h",
1918 "src/xnnpack/clamp.h",
1919 "src/xnnpack/common.h",
1920 "src/xnnpack/conv.h",
1921 "src/xnnpack/dwconv.h",
1922 "src/xnnpack/gavgpool.h",
1923 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001924 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001926 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001928 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001929 "src/xnnpack/lut.h",
1930 "src/xnnpack/math.h",
1931 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001932 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001933 "src/xnnpack/packx.h",
1934 "src/xnnpack/pad.h",
1935 "src/xnnpack/params.h",
1936 "src/xnnpack/pavgpool.h",
1937 "src/xnnpack/ppmm.h",
1938 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001939 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001940 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001941 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001942 "src/xnnpack/rmax.h",
1943 "src/xnnpack/scalar-utils.h",
1944 "src/xnnpack/spmm.h",
1945 "src/xnnpack/unpool.h",
1946 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001947 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001948 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001949 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001950 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001951 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001952 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001954]
1955
1956INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001957 "include/xnnpack.h",
1958 "src/xnnpack/allocator.h",
1959 "src/xnnpack/compute.h",
1960 "src/xnnpack/im2col.h",
1961 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001962 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001963 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001964 "src/xnnpack/operator.h",
1965 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001966 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001968 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001969 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001970]
1971
Marat Dukhan1b354632020-03-23 12:50:22 -07001972ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001973 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001974]
1975
Marat Dukhan1b354632020-03-23 12:50:22 -07001976MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001977 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978 "include/xnnpack.h",
1979]
1980
Marat Dukhan1b354632020-03-23 12:50:22 -07001981MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001982 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001983 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001984 "src/xnnpack/requantization.h",
1985 "include/xnnpack.h",
1986]
1987
1988OPERATOR_TEST_PARAMS_HDRS = [
1989 "src/xnnpack/params.h",
1990 "src/xnnpack/common.h",
1991]
1992
1993WEIGHTS_PACK_HDRS = [
1994 "src/xnnpack/pack.h",
1995 "src/xnnpack/operator.h",
1996 "src/xnnpack/compute.h",
1997]
1998
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001999LOGGING_COPTS = select({
2000 # No logging in optimized mode
2001 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2002 # Full logging in debug mode
2003 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2004 # Error-only logging in default (fastbuild) mode
2005 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2006})
2007
Marat Dukhan3b59de22020-06-03 20:15:19 -07002008LOGGING_SRCS = select({
2009 # No logging in optimized mode
2010 ":optimized_build": [],
2011 "//conditions:default": [
2012 "src/operator-strings.c",
2013 "src/subgraph-strings.c",
2014 ],
2015})
2016
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002017LOGGING_HDRS = [
2018 "src/xnnpack/log.h",
2019]
2020
Marat Dukhan08c4a432019-10-03 09:29:21 -07002021xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002022 name = "tables",
2023 srcs = TABLE_SRCS,
2024 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002025 gcc_copts = xnnpack_gcc_std_copts(),
2026 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002027)
2028
2029xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002030 name = "scalar_ukernels",
2031 srcs = SCALAR_UKERNELS,
2032 hdrs = INTERNAL_HDRS,
2033 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002034 gcc_copts = xnnpack_gcc_std_copts(),
2035 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002037 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038 "@FP16",
2039 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002040 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002041 ],
2042)
2043
2044xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002045 name = "scalar_ukernels_test_mode",
2046 srcs = SCALAR_UKERNELS,
2047 hdrs = INTERNAL_HDRS,
2048 aarch32_copts = ["-marm"],
2049 copts = [
2050 "-UNDEBUG",
2051 "-DXNN_TEST_MODE=1",
2052 ],
2053 gcc_copts = xnnpack_gcc_std_copts(),
2054 msvc_copts = xnnpack_msvc_std_copts(),
2055 deps = [
2056 ":tables",
2057 "@FP16",
2058 "@FXdiv",
2059 "@pthreadpool",
2060 ],
2061)
2062
2063xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002064 name = "wasm_ukernels",
2065 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002066 gcc_copts = xnnpack_gcc_std_copts(),
2067 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002068 wasm_srcs = WASM_UKERNELS,
Marat Dukhan290055c2020-06-09 12:24:29 -07002069 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002070 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002071 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002072 "@FP16",
2073 "@FXdiv",
2074 "@pthreadpool",
2075 ],
2076)
2077
2078xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002079 name = "wasm_ukernels_test_mode",
2080 hdrs = INTERNAL_HDRS,
2081 copts = [
2082 "-UNDEBUG",
2083 "-DXNN_TEST_MODE=1",
2084 ],
2085 gcc_copts = xnnpack_gcc_std_copts(),
2086 msvc_copts = xnnpack_msvc_std_copts(),
2087 wasm_srcs = WASM_UKERNELS,
Marat Dukhan290055c2020-06-09 12:24:29 -07002088 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002089 deps = [
2090 ":tables",
2091 "@FP16",
2092 "@FXdiv",
2093 "@pthreadpool",
2094 ],
2095)
2096
2097xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002098 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099 hdrs = INTERNAL_HDRS,
2100 aarch32_copts = [
2101 "-marm",
2102 "-mfpu=neon",
2103 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002104 gcc_copts = xnnpack_gcc_std_copts(),
2105 gcc_x86_copts = ["-msse2"],
2106 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002107 optimized_copts = [
2108 "-O3",
2109 "-ffast-math",
2110 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002111 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002113 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002114 "@FP16",
2115 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002116 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002117 ],
2118)
2119
2120xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002121 name = "psimd_fastmath_ukernels_test_mode",
2122 hdrs = INTERNAL_HDRS,
2123 aarch32_copts = [
2124 "-marm",
2125 "-mfpu=neon",
2126 ],
2127 copts = [
2128 "-UNDEBUG",
2129 "-DXNN_TEST_MODE=1",
2130 ],
2131 gcc_copts = xnnpack_gcc_std_copts(),
2132 gcc_x86_copts = ["-msse2"],
2133 msvc_copts = xnnpack_msvc_std_copts(),
2134 optimized_copts = [
2135 "-O3",
2136 "-ffast-math",
2137 ],
2138 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2139 deps = [
2140 ":tables",
2141 "@FP16",
2142 "@psimd",
2143 "@pthreadpool",
2144 ],
2145)
2146
2147xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002148 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002149 hdrs = INTERNAL_HDRS,
2150 aarch32_copts = [
2151 "-marm",
2152 "-mfpu=neon",
2153 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002154 gcc_copts = xnnpack_gcc_std_copts(),
2155 gcc_x86_copts = ["-msse2"],
2156 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002157 optimized_copts = [
2158 "-O3",
2159 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002160 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002161 deps = [
2162 ":tables",
2163 "@FP16",
2164 "@psimd",
2165 "@pthreadpool",
2166 ],
2167)
2168
2169xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002170 name = "psimd_accmath_ukernels_test_mode",
2171 hdrs = INTERNAL_HDRS,
2172 aarch32_copts = [
2173 "-marm",
2174 "-mfpu=neon",
2175 ],
2176 copts = [
2177 "-UNDEBUG",
2178 "-DXNN_TEST_MODE=1",
2179 ],
2180 gcc_copts = xnnpack_gcc_std_copts(),
2181 gcc_x86_copts = ["-msse2"],
2182 msvc_copts = xnnpack_msvc_std_copts(),
2183 optimized_copts = [
2184 "-O3",
2185 ],
2186 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2187 deps = [
2188 ":tables",
2189 "@FP16",
2190 "@psimd",
2191 "@pthreadpool",
2192 ],
2193)
2194
2195xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002196 name = "neon_ukernels",
2197 hdrs = INTERNAL_HDRS,
2198 aarch32_copts = [
2199 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002200 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002201 "-mfpu=neon",
2202 ],
2203 aarch32_srcs = NEON_UKERNELS,
2204 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002205 gcc_copts = xnnpack_gcc_std_copts(),
2206 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002207 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002208 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002209 "@FP16",
2210 "@pthreadpool",
2211 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002212)
2213
2214xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002215 name = "neon_ukernels_test_mode",
2216 hdrs = INTERNAL_HDRS,
2217 aarch32_copts = [
2218 "-marm",
2219 "-march=armv7-a",
2220 "-mfpu=neon",
2221 ],
2222 aarch32_srcs = NEON_UKERNELS,
2223 aarch64_srcs = NEON_UKERNELS,
2224 copts = [
2225 "-UNDEBUG",
2226 "-DXNN_TEST_MODE=1",
2227 ],
2228 gcc_copts = xnnpack_gcc_std_copts(),
2229 msvc_copts = xnnpack_msvc_std_copts(),
2230 deps = [
2231 ":tables",
2232 "@FP16",
2233 "@pthreadpool",
2234 ],
2235)
2236
2237xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 name = "neonfma_ukernels",
2239 hdrs = INTERNAL_HDRS,
2240 aarch32_copts = [
2241 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002242 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243 "-mfpu=neon-vfpv4",
2244 ],
2245 aarch32_srcs = NEONFMA_UKERNELS,
2246 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002247 apple_aarch32_copts = [
2248 "-mcpu=swift",
2249 "-mtune=generic",
2250 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002251 gcc_copts = xnnpack_gcc_std_copts(),
2252 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002253 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002254 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002255 "@FP16",
2256 "@pthreadpool",
2257 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258)
2259
2260xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002261 name = "neonfma_ukernels_test_mode",
2262 hdrs = INTERNAL_HDRS,
2263 aarch32_copts = [
2264 "-marm",
2265 "-march=armv7-a",
2266 "-mfpu=neon-vfpv4",
2267 ],
2268 aarch32_srcs = NEONFMA_UKERNELS,
2269 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002270 apple_aarch32_copts = [
2271 "-mcpu=swift",
2272 "-mtune=generic",
2273 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002274 copts = [
2275 "-UNDEBUG",
2276 "-DXNN_TEST_MODE=1",
2277 ],
2278 gcc_copts = xnnpack_gcc_std_copts(),
2279 msvc_copts = xnnpack_msvc_std_copts(),
2280 deps = [
2281 ":tables",
2282 "@FP16",
2283 "@pthreadpool",
2284 ],
2285)
2286
2287xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002288 name = "neonv8_ukernels",
2289 hdrs = INTERNAL_HDRS,
2290 aarch32_copts = [
2291 "-marm",
2292 "-march=armv8-a",
2293 "-mfpu=neon-fp-armv8",
2294 ],
2295 aarch32_srcs = NEONV8_UKERNELS,
2296 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002297 apple_aarch32_copts = [
2298 "-mcpu=cyclone",
2299 "-mtune=generic",
2300 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002301 gcc_copts = xnnpack_gcc_std_copts(),
2302 msvc_copts = xnnpack_msvc_std_copts(),
2303 deps = [
2304 ":tables",
2305 "@FP16",
2306 "@pthreadpool",
2307 ],
2308)
2309
2310xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002311 name = "neonv8_ukernels_test_mode",
2312 hdrs = INTERNAL_HDRS,
2313 aarch32_copts = [
2314 "-marm",
2315 "-march=armv8-a",
2316 "-mfpu=neon-fp-armv8",
2317 ],
2318 aarch32_srcs = NEONV8_UKERNELS,
2319 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002320 apple_aarch32_copts = [
2321 "-mcpu=cyclone",
2322 "-mtune=generic",
2323 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002324 copts = [
2325 "-UNDEBUG",
2326 "-DXNN_TEST_MODE=1",
2327 ],
2328 gcc_copts = xnnpack_gcc_std_copts(),
2329 msvc_copts = xnnpack_msvc_std_copts(),
2330 deps = [
2331 ":tables",
2332 "@FP16",
2333 "@pthreadpool",
2334 ],
2335)
2336
2337xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002338 name = "neonfp16arith_ukernels",
2339 hdrs = INTERNAL_HDRS,
2340 aarch64_copts = ["-march=armv8.2-a+fp16"],
2341 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002342 gcc_copts = xnnpack_gcc_std_copts(),
2343 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002344 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002345 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002346 "@FP16",
2347 "@pthreadpool",
2348 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349)
2350
2351xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002352 name = "neonfp16arith_ukernels_test_mode",
2353 hdrs = INTERNAL_HDRS,
2354 aarch64_copts = ["-march=armv8.2-a+fp16"],
2355 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2356 copts = [
2357 "-UNDEBUG",
2358 "-DXNN_TEST_MODE=1",
2359 ],
2360 gcc_copts = xnnpack_gcc_std_copts(),
2361 msvc_copts = xnnpack_msvc_std_copts(),
2362 deps = [
2363 ":tables",
2364 "@FP16",
2365 "@pthreadpool",
2366 ],
2367)
2368
2369xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002370 name = "sse2_ukernels",
2371 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002372 gcc_copts = xnnpack_gcc_std_copts(),
2373 gcc_x86_copts = ["-msse2"],
2374 msvc_copts = xnnpack_msvc_std_copts(),
2375 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002376 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002377 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002378 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002379 "@FP16",
2380 "@pthreadpool",
2381 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382)
2383
2384xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002385 name = "sse2_ukernels_test_mode",
2386 hdrs = INTERNAL_HDRS,
2387 copts = [
2388 "-UNDEBUG",
2389 "-DXNN_TEST_MODE=1",
2390 ],
2391 gcc_copts = xnnpack_gcc_std_copts(),
2392 gcc_x86_copts = ["-msse2"],
2393 msvc_copts = xnnpack_msvc_std_copts(),
2394 msvc_x86_32_copts = ["/arch:SSE2"],
2395 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2396 deps = [
2397 ":tables",
2398 "@FP16",
2399 "@pthreadpool",
2400 ],
2401)
2402
2403xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002404 name = "ssse3_ukernels",
2405 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002406 gcc_copts = xnnpack_gcc_std_copts(),
2407 gcc_x86_copts = ["-mssse3"],
2408 msvc_copts = xnnpack_msvc_std_copts(),
2409 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002410 x86_srcs = SSSE3_UKERNELS,
2411 deps = [
2412 ":tables",
2413 "@FP16",
2414 "@pthreadpool",
2415 ],
2416)
2417
2418xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002419 name = "ssse3_ukernels_test_mode",
2420 hdrs = INTERNAL_HDRS,
2421 copts = [
2422 "-UNDEBUG",
2423 "-DXNN_TEST_MODE=1",
2424 ],
2425 gcc_copts = xnnpack_gcc_std_copts(),
2426 gcc_x86_copts = ["-mssse3"],
2427 msvc_copts = xnnpack_msvc_std_copts(),
2428 msvc_x86_32_copts = ["/arch:SSE2"],
2429 x86_srcs = SSSE3_UKERNELS,
2430 deps = [
2431 ":tables",
2432 "@FP16",
2433 "@pthreadpool",
2434 ],
2435)
2436
2437xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002438 name = "sse41_ukernels",
2439 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002440 gcc_copts = xnnpack_gcc_std_copts(),
2441 gcc_x86_copts = ["-msse4.1"],
2442 msvc_copts = xnnpack_msvc_std_copts(),
2443 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002444 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002445 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002446 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002447 "@FP16",
2448 "@pthreadpool",
2449 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002450)
2451
2452xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002453 name = "sse41_ukernels_test_mode",
2454 hdrs = INTERNAL_HDRS,
2455 copts = [
2456 "-UNDEBUG",
2457 "-DXNN_TEST_MODE=1",
2458 ],
2459 gcc_copts = xnnpack_gcc_std_copts(),
2460 gcc_x86_copts = ["-msse4.1"],
2461 msvc_copts = xnnpack_msvc_std_copts(),
2462 msvc_x86_32_copts = ["/arch:SSE2"],
2463 x86_srcs = SSE41_UKERNELS,
2464 deps = [
2465 ":tables",
2466 "@FP16",
2467 "@pthreadpool",
2468 ],
2469)
2470
2471xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002472 name = "avx_ukernels",
2473 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002474 gcc_copts = xnnpack_gcc_std_copts(),
2475 gcc_x86_copts = ["-mavx"],
2476 msvc_copts = xnnpack_msvc_std_copts(),
2477 msvc_x86_32_copts = ["/arch:AVX"],
2478 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002479 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002480 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002481 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002482 "@FP16",
2483 "@pthreadpool",
2484 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002485)
2486
2487xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002488 name = "avx_ukernels_test_mode",
2489 hdrs = INTERNAL_HDRS,
2490 copts = [
2491 "-UNDEBUG",
2492 "-DXNN_TEST_MODE=1",
2493 ],
2494 gcc_copts = xnnpack_gcc_std_copts(),
2495 gcc_x86_copts = ["-mavx"],
2496 msvc_copts = xnnpack_msvc_std_copts(),
2497 msvc_x86_32_copts = ["/arch:AVX"],
2498 msvc_x86_64_copts = ["/arch:AVX"],
2499 x86_srcs = AVX_UKERNELS,
2500 deps = [
2501 ":tables",
2502 "@FP16",
2503 "@pthreadpool",
2504 ],
2505)
2506
2507xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002508 name = "fma3_ukernels",
2509 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002510 gcc_copts = xnnpack_gcc_std_copts(),
2511 gcc_x86_copts = ["-mfma"],
2512 msvc_copts = xnnpack_msvc_std_copts(),
2513 msvc_x86_32_copts = ["/arch:AVX"],
2514 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002515 x86_srcs = FMA3_UKERNELS,
2516 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002517 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002518 "@FP16",
2519 "@pthreadpool",
2520 ],
2521)
2522
2523xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002524 name = "fma3_ukernels_test_mode",
2525 hdrs = INTERNAL_HDRS,
2526 copts = [
2527 "-UNDEBUG",
2528 "-DXNN_TEST_MODE=1",
2529 ],
2530 gcc_copts = xnnpack_gcc_std_copts(),
2531 gcc_x86_copts = ["-mfma"],
2532 msvc_copts = xnnpack_msvc_std_copts(),
2533 msvc_x86_32_copts = ["/arch:AVX"],
2534 msvc_x86_64_copts = ["/arch:AVX"],
2535 x86_srcs = FMA3_UKERNELS,
2536 deps = [
2537 ":tables",
2538 "@FP16",
2539 "@pthreadpool",
2540 ],
2541)
2542
2543xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002544 name = "avx2_ukernels",
2545 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002546 gcc_copts = xnnpack_gcc_std_copts(),
2547 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002548 "-mfma",
2549 "-mavx2",
2550 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002551 msvc_copts = xnnpack_msvc_std_copts(),
2552 msvc_x86_32_copts = ["/arch:AVX2"],
2553 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002554 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002555 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002556 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002557 "@FP16",
2558 "@pthreadpool",
2559 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002560)
2561
2562xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002563 name = "avx2_ukernels_test_mode",
2564 hdrs = INTERNAL_HDRS,
2565 copts = [
2566 "-UNDEBUG",
2567 "-DXNN_TEST_MODE=1",
2568 ],
2569 gcc_copts = xnnpack_gcc_std_copts(),
2570 gcc_x86_copts = [
2571 "-mfma",
2572 "-mavx2",
2573 ],
2574 msvc_copts = xnnpack_msvc_std_copts(),
2575 msvc_x86_32_copts = ["/arch:AVX2"],
2576 msvc_x86_64_copts = ["/arch:AVX2"],
2577 x86_srcs = AVX2_UKERNELS,
2578 deps = [
2579 ":tables",
2580 "@FP16",
2581 "@pthreadpool",
2582 ],
2583)
2584
2585xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586 name = "avx512f_ukernels",
2587 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002588 gcc_copts = xnnpack_gcc_std_copts(),
2589 gcc_x86_copts = ["-mavx512f"],
2590 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2591 msvc_copts = xnnpack_msvc_std_copts(),
2592 msvc_x86_32_copts = ["/arch:AVX512"],
2593 msvc_x86_64_copts = ["/arch:AVX512"],
2594 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002596 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002597 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002598 "@FP16",
2599 "@pthreadpool",
2600 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601)
2602
2603xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002604 name = "avx512f_ukernels_test_mode",
2605 hdrs = INTERNAL_HDRS,
2606 copts = [
2607 "-UNDEBUG",
2608 "-DXNN_TEST_MODE=1",
2609 ],
2610 gcc_copts = xnnpack_gcc_std_copts(),
2611 gcc_x86_copts = ["-mavx512f"],
2612 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2613 msvc_copts = xnnpack_msvc_std_copts(),
2614 msvc_x86_32_copts = ["/arch:AVX512"],
2615 msvc_x86_64_copts = ["/arch:AVX512"],
2616 msys_copts = ["-fno-asynchronous-unwind-tables"],
2617 x86_srcs = AVX512F_UKERNELS,
2618 deps = [
2619 ":tables",
2620 "@FP16",
2621 "@pthreadpool",
2622 ],
2623)
2624
2625xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 name = "asm_ukernels",
2627 hdrs = ["src/xnnpack/assembly.h"],
2628 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002629 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630 aarch64_srcs = AARCH64_ASM_UKERNELS,
2631)
2632
Marat Dukhan3b59de22020-06-03 20:15:19 -07002633xnnpack_cc_library(
2634 name = "logging_utils",
2635 srcs = LOGGING_SRCS,
2636 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2637 copts = LOGGING_COPTS + [
2638 "-Isrc",
2639 "-Iinclude",
2640 ] + select({
2641 ":debug_build": [],
2642 "//conditions:default": xnnpack_min_size_copts(),
2643 }),
2644 gcc_copts = xnnpack_gcc_std_copts(),
2645 msvc_copts = xnnpack_msvc_std_copts(),
2646 visibility = xnnpack_visibility(),
2647 deps = [
2648 "@FP16",
2649 "@clog",
2650 "@pthreadpool",
2651 ],
2652)
2653
Marat Dukhan08c4a432019-10-03 09:29:21 -07002654xnnpack_aggregate_library(
2655 name = "ukernels",
2656 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 ":neon_ukernels",
2658 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002659 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660 ":asm_ukernels",
2661 ],
2662 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002663 ":neon_ukernels",
2664 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002665 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666 ":neonfp16arith_ukernels",
2667 ":asm_ukernels",
2668 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002669 generic_deps = [
2670 ":scalar_ukernels",
2671 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002672 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002673 ":psimd_fastmath_ukernels",
2674 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002676 wasm_deps = [
2677 ":wasm_ukernels",
2678 ],
2679 wasmsimd_deps = [
2680 ":wasm_ukernels",
2681 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002682 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002683 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002684 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002685 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002687 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002688 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002689 ":avx512f_ukernels",
2690 ],
2691)
2692
Marat Dukhan33fcf782020-05-24 14:27:15 -07002693xnnpack_aggregate_library(
2694 name = "ukernels_test_mode",
2695 aarch32_deps = [
2696 ":neon_ukernels_test_mode",
2697 ":neonfma_ukernels_test_mode",
2698 ":neonv8_ukernels_test_mode",
2699 ":asm_ukernels",
2700 ],
2701 aarch64_deps = [
2702 ":neon_ukernels_test_mode",
2703 ":neonfma_ukernels_test_mode",
2704 ":neonv8_ukernels_test_mode",
2705 ":neonfp16arith_ukernels_test_mode",
2706 ":asm_ukernels",
2707 ],
2708 generic_deps = [
2709 ":scalar_ukernels_test_mode",
2710 ],
2711 psimd_deps = [
2712 ":psimd_fastmath_ukernels_test_mode",
2713 ":psimd_accmath_ukernels_test_mode",
2714 ],
2715 wasm_deps = [
2716 ":wasm_ukernels_test_mode",
2717 ],
2718 wasmsimd_deps = [
2719 ":wasm_ukernels_test_mode",
2720 ],
2721 x86_deps = [
2722 ":sse2_ukernels_test_mode",
2723 ":ssse3_ukernels_test_mode",
2724 ":sse41_ukernels_test_mode",
2725 ":avx_ukernels_test_mode",
2726 ":fma3_ukernels_test_mode",
2727 ":avx2_ukernels_test_mode",
2728 ":avx512f_ukernels_test_mode",
2729 ],
2730)
2731
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732xnnpack_cc_library(
2733 name = "im2col",
2734 srcs = ["src/im2col.c"],
2735 hdrs = [
2736 "src/xnnpack/common.h",
2737 "src/xnnpack/im2col.h",
2738 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002739 gcc_copts = xnnpack_gcc_std_copts(),
2740 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741)
2742
2743xnnpack_cc_library(
2744 name = "indirection",
2745 srcs = ["src/indirection.c"],
2746 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002747 gcc_copts = xnnpack_gcc_std_copts(),
2748 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749 deps = [
2750 "@FP16",
2751 "@FXdiv",
2752 "@pthreadpool",
2753 ],
2754)
2755
2756xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002757 name = "indirection_test_mode",
2758 srcs = ["src/indirection.c"],
2759 hdrs = INTERNAL_HDRS,
2760 copts = [
2761 "-UNDEBUG",
2762 "-DXNN_TEST_MODE=1",
2763 ],
2764 gcc_copts = xnnpack_gcc_std_copts(),
2765 msvc_copts = xnnpack_msvc_std_copts(),
2766 deps = [
2767 "@FP16",
2768 "@FXdiv",
2769 "@pthreadpool",
2770 ],
2771)
2772
2773xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002774 name = "operator_run",
2775 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002776 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002777 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002778 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2779 "//conditions:default": [],
2780 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002781 gcc_copts = xnnpack_gcc_std_copts(),
2782 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002784 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002785 "@FP16",
2786 "@FXdiv",
2787 "@clog",
2788 "@pthreadpool",
2789 ],
2790)
2791
Chao Mei6ddfc602020-05-13 22:29:36 -07002792xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002793 name = "operator_run_test_mode",
2794 srcs = ["src/operator-run.c"],
2795 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2796 copts = LOGGING_COPTS + [
2797 "-UNDEBUG",
2798 "-DXNN_TEST_MODE=1",
2799 ] + select({
2800 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2801 "//conditions:default": [],
2802 }),
2803 gcc_copts = xnnpack_gcc_std_copts(),
2804 msvc_copts = xnnpack_msvc_std_copts(),
2805 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002806 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002807 "@FP16",
2808 "@FXdiv",
2809 "@clog",
2810 "@pthreadpool",
2811 ],
2812)
2813
2814xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07002815 name = "memory_planner",
2816 srcs = ["src/memory-planner.c"],
2817 hdrs = INTERNAL_HDRS,
2818 defines = select({
2819 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2820 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2821 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2822 }),
2823 gcc_copts = xnnpack_gcc_std_copts(),
2824 msvc_copts = xnnpack_msvc_std_copts(),
2825 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002826 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002827 "@pthreadpool",
2828 ],
2829)
2830
Marat Dukhan33fcf782020-05-24 14:27:15 -07002831xnnpack_cc_library(
2832 name = "memory_planner_test_mode",
2833 srcs = ["src/memory-planner.c"],
2834 hdrs = INTERNAL_HDRS,
2835 copts = [
2836 "-UNDEBUG",
2837 "-DXNN_TEST_MODE=1",
2838 ],
2839 defines = select({
2840 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2841 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2842 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2843 }),
2844 gcc_copts = xnnpack_gcc_std_copts(),
2845 msvc_copts = xnnpack_msvc_std_copts(),
2846 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002847 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002848 "@pthreadpool",
2849 ],
2850)
2851
Marat Dukhan08c4a432019-10-03 09:29:21 -07002852cc_library(
2853 name = "enable_assembly",
2854 defines = select({
2855 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2856 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002857 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002858 }),
2859)
2860
Marat Dukhancf056b22019-10-07 10:26:29 -07002861xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862 name = "operators",
2863 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002864 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002866 ],
2867 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002868 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002869 "-Isrc",
2870 "-Iinclude",
2871 ] + select({
2872 ":debug_build": [],
2873 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002874 }) + select({
2875 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2876 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002877 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002878 gcc_copts = xnnpack_gcc_std_copts(),
2879 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002882 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002883 "@FP16",
2884 "@FXdiv",
2885 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002886 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002887 ],
2888)
2889
Marat Dukhan10a38082020-04-17 03:58:35 -07002890xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002891 name = "operators_test_mode",
2892 srcs = OPERATOR_SRCS + [
2893 "src/memory.c",
2894 "src/operator-delete.c",
2895 ],
2896 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2897 copts = LOGGING_COPTS + [
2898 "-Isrc",
2899 "-Iinclude",
2900 "-UNDEBUG",
2901 "-DXNN_TEST_MODE=1",
2902 ] + select({
2903 ":debug_build": [],
2904 "//conditions:default": xnnpack_min_size_copts(),
2905 }) + select({
2906 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2907 "//conditions:default": [],
2908 }),
2909 gcc_copts = xnnpack_gcc_std_copts(),
2910 msvc_copts = xnnpack_msvc_std_copts(),
2911 deps = [
2912 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002913 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002914 "@FP16",
2915 "@FXdiv",
2916 "@clog",
2917 "@pthreadpool",
2918 ],
2919)
2920
2921xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002922 name = "XNNPACK",
2923 srcs = [
2924 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002925 "src/runtime.c",
2926 "src/subgraph.c",
2927 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07002928 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002929 hdrs = ["include/xnnpack.h"],
2930 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002931 "-Isrc",
2932 "-Iinclude",
2933 ] + select({
2934 ":debug_build": [],
2935 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002936 }) + select({
2937 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2938 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002939 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002940 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002941 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002942 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002943 visibility = xnnpack_visibility(),
2944 deps = [
2945 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002946 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002947 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002948 ":operator_run",
2949 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002950 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002951 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07002952 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002953 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002954 ] + select({
2955 ":emscripten": [],
2956 "//conditions:default": ["@cpuinfo"],
2957 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958)
2959
Marat Dukhan10a38082020-04-17 03:58:35 -07002960xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002961 name = "XNNPACK_test_mode",
2962 srcs = [
2963 "src/init.c",
2964 "src/runtime.c",
2965 "src/subgraph.c",
2966 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07002967 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002968 hdrs = ["include/xnnpack.h"],
2969 copts = LOGGING_COPTS + [
2970 "-Isrc",
2971 "-Iinclude",
2972 "-UNDEBUG",
2973 "-DXNN_TEST_MODE=1",
2974 ] + select({
2975 ":debug_build": [],
2976 "//conditions:default": xnnpack_min_size_copts(),
2977 }) + select({
2978 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2979 "//conditions:default": [],
2980 }),
2981 gcc_copts = xnnpack_gcc_std_copts(),
2982 includes = ["include"],
2983 msvc_copts = xnnpack_msvc_std_copts(),
2984 visibility = xnnpack_visibility(),
2985 deps = [
2986 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002987 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002988 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002989 ":operator_run_test_mode",
2990 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002991 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002992 "@clog",
2993 "@FP16",
2994 "@pthreadpool",
2995 ] + select({
2996 ":emscripten": [],
2997 "//conditions:default": ["@cpuinfo"],
2998 }),
2999)
3000
3001xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003002 name = "xnnpack_operators_nhwc_f32",
3003 srcs = [
3004 "src/init.c",
3005 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003006 hdrs = ["include/xnnpack.h"],
3007 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003008 "-Isrc",
3009 "-Iinclude",
3010 ] + select({
3011 ":debug_build": [],
3012 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003013 }) + select({
3014 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3015 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003016 }),
3017 defines = [
3018 "XNN_NO_Q8_OPERATORS",
3019 "XNN_NO_U8_OPERATORS",
3020 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003021 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003022 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003023 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003025 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 visibility = xnnpack_visibility(),
3027 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003028 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003029 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 ":operator_run",
3031 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003032 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003033 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003034 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003035 ] + select({
3036 ":emscripten": [],
3037 "//conditions:default": ["@cpuinfo"],
3038 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039)
3040
Marat Dukhancf056b22019-10-07 10:26:29 -07003041xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 name = "bench_utils",
3043 srcs = ["bench/utils.cc"],
3044 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003045 deps = [
3046 "@com_google_benchmark//:benchmark",
3047 "@cpuinfo",
3048 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049)
3050
Frank Barchard7e955972019-10-11 10:34:25 -07003051######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052
3053xnnpack_benchmark(
3054 name = "q8_gemm_bench",
3055 srcs = [
3056 "bench/gemm.h",
3057 "bench/q8-gemm.cc",
3058 "src/xnnpack/AlignedAllocator.h",
3059 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003060 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
3062)
3063
3064xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003065 name = "f16_igemm_bench",
3066 srcs = [
3067 "bench/f16-igemm.cc",
3068 "bench/conv.h",
3069 "bench/google/conv.h",
3070 "src/xnnpack/AlignedAllocator.h",
3071 ] + MICROKERNEL_BENCHMARK_HDRS,
3072 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3073)
3074
3075xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 name = "f16_gemm_bench",
3077 srcs = [
3078 "bench/f16-gemm.cc",
3079 "bench/gemm.h",
3080 "src/xnnpack/AlignedAllocator.h",
3081 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003082 deps = MICROKERNEL_BENCHMARK_DEPS,
3083)
3084
3085xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003086 name = "f16_spmm_bench",
3087 srcs = [
3088 "bench/f16-spmm.cc",
3089 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003090 "src/xnnpack/AlignedAllocator.h",
3091 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003092 deps = MICROKERNEL_BENCHMARK_DEPS,
3093)
3094
3095xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 name = "f32_igemm_bench",
3097 srcs = [
3098 "bench/f32-igemm.cc",
3099 "bench/conv.h",
3100 "src/xnnpack/AlignedAllocator.h",
3101 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07003102 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103)
3104
3105xnnpack_benchmark(
3106 name = "f32_conv_hwc_bench",
3107 srcs = [
3108 "bench/f32-conv-hwc.cc",
3109 "bench/dconv.h",
3110 "src/xnnpack/AlignedAllocator.h",
3111 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003112 deps = MICROKERNEL_BENCHMARK_DEPS,
3113)
3114
3115xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003116 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003117 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003118 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003119 "bench/dconv.h",
3120 "src/xnnpack/AlignedAllocator.h",
3121 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003122 deps = MICROKERNEL_BENCHMARK_DEPS,
3123)
3124
3125xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003126 name = "f16_dwconv_bench",
3127 srcs = [
3128 "bench/f16-dwconv.cc",
3129 "bench/dwconv.h",
3130 "bench/google/dwconv.h",
3131 "src/xnnpack/AlignedAllocator.h",
3132 ] + MICROKERNEL_BENCHMARK_HDRS,
3133 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3134)
3135
3136xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 name = "f32_dwconv_bench",
3138 srcs = [
3139 "bench/f32-dwconv.cc",
3140 "bench/dwconv.h",
3141 "src/xnnpack/AlignedAllocator.h",
3142 ] + MICROKERNEL_BENCHMARK_HDRS,
3143 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3144)
3145
3146xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003147 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003149 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 "bench/dwconv.h",
3151 "src/xnnpack/AlignedAllocator.h",
3152 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3153 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3154)
3155
3156xnnpack_benchmark(
3157 name = "f32_gemm_bench",
3158 srcs = [
3159 "bench/f32-gemm.cc",
3160 "bench/gemm.h",
3161 "src/xnnpack/AlignedAllocator.h",
3162 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003163 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003164 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165)
3166
3167xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003168 name = "f32_raddexpminusmax_bench",
3169 srcs = [
3170 "bench/f32-raddexpminusmax.cc",
3171 "src/xnnpack/AlignedAllocator.h",
3172 ] + MICROKERNEL_BENCHMARK_HDRS,
3173 deps = MICROKERNEL_BENCHMARK_DEPS,
3174)
3175
3176xnnpack_benchmark(
3177 name = "f32_raddextexp_bench",
3178 srcs = [
3179 "bench/f32-raddextexp.cc",
3180 "src/xnnpack/AlignedAllocator.h",
3181 ] + MICROKERNEL_BENCHMARK_HDRS,
3182 deps = MICROKERNEL_BENCHMARK_DEPS,
3183)
3184
3185xnnpack_benchmark(
3186 name = "f32_raddstoreexpminusmax_bench",
3187 srcs = [
3188 "bench/f32-raddstoreexpminusmax.cc",
3189 "src/xnnpack/AlignedAllocator.h",
3190 ] + MICROKERNEL_BENCHMARK_HDRS,
3191 deps = MICROKERNEL_BENCHMARK_DEPS,
3192)
3193
3194xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195 name = "f32_rmax_bench",
3196 srcs = [
3197 "bench/f32-rmax.cc",
3198 "src/xnnpack/AlignedAllocator.h",
3199 ] + MICROKERNEL_BENCHMARK_HDRS,
3200 deps = MICROKERNEL_BENCHMARK_DEPS,
3201)
3202
3203xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003204 name = "f32_sigmoid_bench",
3205 srcs = [
3206 "bench/f32-sigmoid.cc",
3207 "src/xnnpack/AlignedAllocator.h",
3208 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003209 deps = MICROKERNEL_BENCHMARK_DEPS,
3210)
3211
3212xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213 name = "f32_spmm_bench",
3214 srcs = [
3215 "bench/f32-spmm.cc",
3216 "bench/gemm.h",
3217 "src/xnnpack/AlignedAllocator.h",
3218 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 deps = MICROKERNEL_BENCHMARK_DEPS,
3220)
3221
3222xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003223 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003224 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003225 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003226 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003227 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003228 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003229)
3230
3231xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003232 name = "f32_vscaleexpminusmax_bench",
3233 srcs = [
3234 "bench/f32-vscaleexpminusmax.cc",
3235 "src/xnnpack/AlignedAllocator.h",
3236 ] + MICROKERNEL_BENCHMARK_HDRS,
3237 deps = MICROKERNEL_BENCHMARK_DEPS,
3238)
3239
3240xnnpack_benchmark(
3241 name = "f32_vscaleextexp_bench",
3242 srcs = [
3243 "bench/f32-vscaleextexp.cc",
3244 "src/xnnpack/AlignedAllocator.h",
3245 ] + MICROKERNEL_BENCHMARK_HDRS,
3246 deps = MICROKERNEL_BENCHMARK_DEPS,
3247)
3248
3249xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 name = "f32_im2col_gemm_bench",
3251 srcs = [
3252 "bench/f32-im2col-gemm.cc",
3253 "bench/conv.h",
3254 "src/xnnpack/AlignedAllocator.h",
3255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3256 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3257)
3258
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003259xnnpack_benchmark(
3260 name = "requantization_bench",
3261 srcs = [
3262 "bench/requantization.cc",
3263 "src/xnnpack/requantization-stubs.h",
3264 "src/xnnpack/AlignedAllocator.h",
3265 ] + MICROKERNEL_BENCHMARK_HDRS,
3266 deps = MICROKERNEL_BENCHMARK_DEPS,
3267)
3268
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003269xnnpack_benchmark(
3270 name = "rounding_bench",
3271 srcs = [
3272 "bench/rounding.cc",
3273 "src/xnnpack/math-stubs.h",
3274 "src/xnnpack/AlignedAllocator.h",
3275 ] + MICROKERNEL_BENCHMARK_HDRS,
3276 deps = MICROKERNEL_BENCHMARK_DEPS,
3277)
3278
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279########################### Benchmarks for operators ###########################
3280
3281xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003282 name = "average_pooling_bench",
3283 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003284 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003285 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003286 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287)
3288
3289xnnpack_benchmark(
3290 name = "channel_shuffle_bench",
3291 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003292 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293)
3294
3295xnnpack_benchmark(
3296 name = "convolution_bench",
3297 srcs = ["bench/convolution.cc"],
3298 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003299 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003300 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003301)
3302
3303xnnpack_benchmark(
3304 name = "deconvolution_bench",
3305 srcs = ["bench/deconvolution.cc"],
3306 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003307 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003308 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309)
3310
3311xnnpack_benchmark(
3312 name = "global_average_pooling_bench",
3313 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003314 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003315)
3316
3317xnnpack_benchmark(
3318 name = "max_pooling_bench",
3319 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003320 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321)
3322
3323xnnpack_benchmark(
3324 name = "sigmoid_bench",
3325 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003326 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003327 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003328 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003329)
3330
3331xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003332 name = "prelu_bench",
3333 srcs = ["bench/prelu.cc"],
3334 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003335 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003336 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003337)
3338
3339xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003340 name = "softmax_bench",
3341 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003342 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003343 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003344 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345)
3346
Marat Dukhanc068bb62019-10-04 13:24:39 -07003347############################# End-to-end benchmarks ############################
3348
3349cc_library(
3350 name = "mobilenet_v1",
3351 srcs = ["models/mobilenet-v1.cc"],
3352 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003353 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003354 linkstatic = True,
3355 deps = [
3356 ":XNNPACK",
3357 "@pthreadpool",
3358 ],
3359)
3360
3361cc_library(
3362 name = "mobilenet_v2",
3363 srcs = ["models/mobilenet-v2.cc"],
3364 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003365 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003366 linkstatic = True,
3367 deps = [
3368 ":XNNPACK",
3369 "@pthreadpool",
3370 ],
3371)
3372
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003373cc_library(
3374 name = "mobilenet_v3_large",
3375 srcs = ["models/mobilenet-v3-large.cc"],
3376 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003377 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003378 linkstatic = True,
3379 deps = [
3380 ":XNNPACK",
3381 "@pthreadpool",
3382 ],
3383)
3384
3385cc_library(
3386 name = "mobilenet_v3_small",
3387 srcs = ["models/mobilenet-v3-small.cc"],
3388 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003389 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003390 linkstatic = True,
3391 deps = [
3392 ":XNNPACK",
3393 "@pthreadpool",
3394 ],
3395)
3396
Marat Dukhanc068bb62019-10-04 13:24:39 -07003397xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003398 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003399 srcs = [
3400 "bench/f32-dwconv-e2e.cc",
3401 "bench/end2end.h",
3402 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003403 deps = MICROKERNEL_BENCHMARK_DEPS + [
3404 ":XNNPACK",
3405 ":mobilenet_v1",
3406 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003407 ":mobilenet_v3_large",
3408 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003409 ],
3410)
3411
3412xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003413 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003414 srcs = [
3415 "bench/f32-gemm-e2e.cc",
3416 "bench/end2end.h",
3417 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003418 deps = MICROKERNEL_BENCHMARK_DEPS + [
3419 ":XNNPACK",
3420 ":mobilenet_v1",
3421 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003422 ":mobilenet_v3_large",
3423 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003424 ],
3425)
3426
3427xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003428 name = "end2end_bench",
3429 srcs = ["bench/end2end.cc"],
3430 deps = [
3431 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003432 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003433 ":mobilenet_v1",
3434 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003435 ":mobilenet_v3_large",
3436 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003437 "@pthreadpool",
3438 ],
3439)
3440
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003441#################### Accuracy evaluation for math functions ####################
3442
3443xnnpack_benchmark(
3444 name = "f32_exp_eval",
3445 srcs = [
3446 "eval/f32-exp.cc",
3447 "src/xnnpack/AlignedAllocator.h",
3448 ] + ACCURACY_EVAL_HDRS,
3449 deps = ACCURACY_EVAL_DEPS,
3450)
3451
Marat Dukhan515c9772019-10-17 18:07:57 -07003452xnnpack_benchmark(
3453 name = "f32_expminus_eval",
3454 srcs = [
3455 "eval/f32-expminus.cc",
3456 "src/xnnpack/AlignedAllocator.h",
3457 ] + ACCURACY_EVAL_HDRS,
3458 deps = ACCURACY_EVAL_DEPS,
3459)
3460
Marat Dukhan98ba4412019-10-23 02:14:28 -07003461xnnpack_benchmark(
3462 name = "f32_extexp_eval",
3463 srcs = [
3464 "eval/f32-extexp.cc",
3465 "src/xnnpack/AlignedAllocator.h",
3466 ] + ACCURACY_EVAL_HDRS,
3467 deps = ACCURACY_EVAL_DEPS,
3468)
3469
Marat Dukhan8853b822020-05-07 12:19:01 -07003470xnnpack_unit_test(
3471 name = "f32_roundne_eval",
3472 srcs = [
3473 "eval/f32-roundne.cc",
3474 "src/xnnpack/AlignedAllocator.h",
3475 "src/xnnpack/math-stubs.h",
3476 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003477 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003478 deps = MICROKERNEL_TEST_DEPS,
3479)
3480
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003481xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003482 name = "f32_roundd_eval",
3483 srcs = [
3484 "eval/f32-roundd.cc",
3485 "src/xnnpack/AlignedAllocator.h",
3486 "src/xnnpack/math-stubs.h",
3487 ] + MICROKERNEL_TEST_HDRS,
3488 automatic = False,
3489 deps = MICROKERNEL_TEST_DEPS,
3490)
3491
3492xnnpack_unit_test(
3493 name = "f32_roundu_eval",
3494 srcs = [
3495 "eval/f32-roundu.cc",
3496 "src/xnnpack/AlignedAllocator.h",
3497 "src/xnnpack/math-stubs.h",
3498 ] + MICROKERNEL_TEST_HDRS,
3499 automatic = False,
3500 deps = MICROKERNEL_TEST_DEPS,
3501)
3502
3503xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003504 name = "f32_roundz_eval",
3505 srcs = [
3506 "eval/f32-roundz.cc",
3507 "src/xnnpack/AlignedAllocator.h",
3508 "src/xnnpack/math-stubs.h",
3509 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003510 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003511 deps = MICROKERNEL_TEST_DEPS,
3512)
3513
Marat Dukhan346a9e52019-11-15 09:06:30 -08003514xnnpack_benchmark(
3515 name = "f32_sigmoid_eval",
3516 srcs = [
3517 "eval/f32-sigmoid.cc",
3518 "src/xnnpack/AlignedAllocator.h",
3519 ] + ACCURACY_EVAL_HDRS,
3520 deps = ACCURACY_EVAL_DEPS,
3521)
3522
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523######################### Unit tests for micro-kernels #########################
3524
3525xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003526 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003528 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003529 "test/gemm-microkernel-tester.h",
3530 "src/xnnpack/AlignedAllocator.h",
3531 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3532 deps = MICROKERNEL_TEST_DEPS,
3533)
3534
3535xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003536 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003537 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003538 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003539 "test/spmm-microkernel-tester.h",
3540 "src/xnnpack/AlignedAllocator.h",
3541 ] + MICROKERNEL_TEST_HDRS,
3542 deps = MICROKERNEL_TEST_DEPS,
3543)
3544
3545xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 name = "f32_argmaxpool_test",
3547 srcs = [
3548 "test/f32-argmaxpool.cc",
3549 "test/argmaxpool-microkernel-tester.h",
3550 "src/xnnpack/AlignedAllocator.h",
3551 ] + MICROKERNEL_TEST_HDRS,
3552 deps = MICROKERNEL_TEST_DEPS,
3553)
3554
3555xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003556 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003558 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559 "test/avgpool-microkernel-tester.h",
3560 "src/xnnpack/AlignedAllocator.h",
3561 ] + MICROKERNEL_TEST_HDRS,
3562 deps = MICROKERNEL_TEST_DEPS,
3563)
3564
3565xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07003566 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003567 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07003568 "test/f32-ibilinear.cc",
3569 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003570 "src/xnnpack/AlignedAllocator.h",
3571 ] + MICROKERNEL_TEST_HDRS,
3572 deps = MICROKERNEL_TEST_DEPS,
3573)
3574
3575xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003576 name = "f16_clamp_test",
3577 srcs = [
3578 "test/f16-clamp.cc",
3579 "test/clamp-microkernel-tester.h",
3580 ] + MICROKERNEL_TEST_HDRS,
3581 deps = MICROKERNEL_TEST_DEPS,
3582)
3583
3584xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585 name = "f32_clamp_test",
3586 srcs = [
3587 "test/f32-clamp.cc",
3588 "test/clamp-microkernel-tester.h",
3589 ] + MICROKERNEL_TEST_HDRS,
3590 deps = MICROKERNEL_TEST_DEPS,
3591)
3592
3593xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003594 name = "f32_igemm_test",
3595 srcs = [
3596 "test/f32-igemm.cc",
3597 "test/gemm-microkernel-tester.h",
3598 "src/xnnpack/AlignedAllocator.h",
3599 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3600 deps = MICROKERNEL_TEST_DEPS,
3601)
3602
3603xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003604 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07003606 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 "test/gemm-microkernel-tester.h",
3608 "src/xnnpack/AlignedAllocator.h",
3609 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3610 deps = MICROKERNEL_TEST_DEPS,
3611)
3612
3613xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07003614 name = "f32_igemm_minmax_test",
3615 srcs = [
3616 "test/f32-igemm-minmax.cc",
3617 "test/gemm-microkernel-tester.h",
3618 "src/xnnpack/AlignedAllocator.h",
3619 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3620 deps = MICROKERNEL_TEST_DEPS,
3621)
3622
3623xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07003624 name = "f16_igemm_minmax_test",
3625 srcs = [
3626 "test/f16-igemm-minmax.cc",
3627 "test/gemm-microkernel-tester.h",
3628 "src/xnnpack/AlignedAllocator.h",
3629 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3630 deps = MICROKERNEL_TEST_DEPS,
3631)
3632
3633xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 name = "f32_conv_hwc_test",
3635 srcs = [
3636 "test/f32-conv-hwc.cc",
3637 "test/conv-hwc-microkernel-tester.h",
3638 "src/xnnpack/AlignedAllocator.h",
3639 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3640 deps = MICROKERNEL_TEST_DEPS,
3641)
3642
3643xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003644 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003646 "test/f32-conv-hwc2chw.cc",
3647 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 "src/xnnpack/AlignedAllocator.h",
3649 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3650 deps = MICROKERNEL_TEST_DEPS,
3651)
3652
3653xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003654 name = "f32_dwconv_test",
3655 srcs = [
3656 "test/f32-dwconv.cc",
3657 "test/dwconv-microkernel-tester.h",
3658 "src/xnnpack/AlignedAllocator.h",
3659 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3660 deps = MICROKERNEL_TEST_DEPS,
3661)
3662
3663xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07003664 name = "f16_dwconv_minmax_test",
3665 srcs = [
3666 "test/f16-dwconv-minmax.cc",
3667 "test/dwconv-microkernel-tester.h",
3668 "src/xnnpack/AlignedAllocator.h",
3669 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3670 deps = MICROKERNEL_TEST_DEPS,
3671)
3672
3673xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003674 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003675 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003676 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003677 "test/dwconv-microkernel-tester.h",
3678 "src/xnnpack/AlignedAllocator.h",
3679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3680 deps = MICROKERNEL_TEST_DEPS,
3681)
3682
3683xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003684 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003685 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003686 "test/f32-dwconv-chw.cc",
3687 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003688 "src/xnnpack/AlignedAllocator.h",
3689 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3690 deps = MICROKERNEL_TEST_DEPS,
3691)
3692
3693xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07003694 name = "f16_gavgpool_minmax_test",
3695 srcs = [
3696 "test/f16-gavgpool-minmax.cc",
3697 "test/gavgpool-microkernel-tester.h",
3698 "src/xnnpack/AlignedAllocator.h",
3699 ] + MICROKERNEL_TEST_HDRS,
3700 deps = MICROKERNEL_TEST_DEPS,
3701)
3702
3703xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003704 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003706 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 "test/gavgpool-microkernel-tester.h",
3708 "src/xnnpack/AlignedAllocator.h",
3709 ] + MICROKERNEL_TEST_HDRS,
3710 deps = MICROKERNEL_TEST_DEPS,
3711)
3712
3713xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003714 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003716 "test/f32-gavgpool-cw.cc",
3717 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 "src/xnnpack/AlignedAllocator.h",
3719 ] + MICROKERNEL_TEST_HDRS,
3720 deps = MICROKERNEL_TEST_DEPS,
3721)
3722
3723xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003724 name = "f32_gemm_test",
3725 srcs = [
3726 "test/f32-gemm.cc",
3727 "test/gemm-microkernel-tester.h",
3728 "src/xnnpack/AlignedAllocator.h",
3729 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3730 deps = MICROKERNEL_TEST_DEPS,
3731)
3732
3733xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003734 name = "f32_gemm_relu_test",
3735 srcs = [
3736 "test/f32-gemm-relu.cc",
3737 "test/gemm-microkernel-tester.h",
3738 "src/xnnpack/AlignedAllocator.h",
3739 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3740 deps = MICROKERNEL_TEST_DEPS,
3741)
3742
3743xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003744 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003746 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747 "test/gemm-microkernel-tester.h",
3748 "src/xnnpack/AlignedAllocator.h",
3749 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3750 deps = MICROKERNEL_TEST_DEPS,
3751)
3752
3753xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003754 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003756 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 "test/gemm-microkernel-tester.h",
3758 "src/xnnpack/AlignedAllocator.h",
3759 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3760 deps = MICROKERNEL_TEST_DEPS,
3761)
3762
3763xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003764 name = "f16_hswish_test",
3765 srcs = [
3766 "test/f16-hswish.cc",
3767 "test/hswish-microkernel-tester.h",
3768 ] + MICROKERNEL_TEST_HDRS,
3769 deps = MICROKERNEL_TEST_DEPS,
3770)
3771
3772xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 name = "f32_hswish_test",
3774 srcs = [
3775 "test/f32-hswish.cc",
3776 "test/hswish-microkernel-tester.h",
3777 ] + MICROKERNEL_TEST_HDRS,
3778 deps = MICROKERNEL_TEST_DEPS,
3779)
3780
3781xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003782 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003784 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003785 "test/maxpool-microkernel-tester.h",
3786 ] + MICROKERNEL_TEST_HDRS,
3787 deps = MICROKERNEL_TEST_DEPS,
3788)
3789
3790xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003791 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003792 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003793 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 "test/avgpool-microkernel-tester.h",
3795 "src/xnnpack/AlignedAllocator.h",
3796 ] + MICROKERNEL_TEST_HDRS,
3797 deps = MICROKERNEL_TEST_DEPS,
3798)
3799
3800xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003801 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003803 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804 "test/gemm-microkernel-tester.h",
3805 "src/xnnpack/AlignedAllocator.h",
3806 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3807 deps = MICROKERNEL_TEST_DEPS,
3808)
3809
3810xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003811 name = "f16_prelu_test",
3812 srcs = [
3813 "test/f16-prelu.cc",
3814 "test/prelu-microkernel-tester.h",
3815 "src/xnnpack/AlignedAllocator.h",
3816 ] + MICROKERNEL_TEST_HDRS,
3817 deps = MICROKERNEL_TEST_DEPS,
3818)
3819
3820xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003821 name = "f32_prelu_test",
3822 srcs = [
3823 "test/f32-prelu.cc",
3824 "test/prelu-microkernel-tester.h",
3825 "src/xnnpack/AlignedAllocator.h",
3826 ] + MICROKERNEL_TEST_HDRS,
3827 deps = MICROKERNEL_TEST_DEPS,
3828)
3829
3830xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003831 name = "f32_raddexpminusmax_test",
3832 srcs = [
3833 "test/f32-raddexpminusmax.cc",
3834 "test/raddexpminusmax-microkernel-tester.h",
3835 ] + MICROKERNEL_TEST_HDRS,
3836 deps = MICROKERNEL_TEST_DEPS,
3837)
3838
3839xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003840 name = "f32_raddextexp_test",
3841 srcs = [
3842 "test/f32-raddextexp.cc",
3843 "test/raddextexp-microkernel-tester.h",
3844 ] + MICROKERNEL_TEST_HDRS,
3845 deps = MICROKERNEL_TEST_DEPS,
3846)
3847
3848xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003849 name = "f32_raddstoreexpminusmax_test",
3850 srcs = [
3851 "test/f32-raddstoreexpminusmax.cc",
3852 "test/raddstoreexpminusmax-microkernel-tester.h",
3853 ] + MICROKERNEL_TEST_HDRS,
3854 deps = MICROKERNEL_TEST_DEPS,
3855)
3856
3857xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 name = "f32_rmax_test",
3859 srcs = [
3860 "test/f32-rmax.cc",
3861 "test/rmax-microkernel-tester.h",
3862 ] + MICROKERNEL_TEST_HDRS,
3863 deps = MICROKERNEL_TEST_DEPS,
3864)
3865
3866xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003867 name = "f32_sigmoid_test",
3868 srcs = [
3869 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003870 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003871 ] + MICROKERNEL_TEST_HDRS,
3872 deps = MICROKERNEL_TEST_DEPS,
3873)
3874
3875xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003876 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003877 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003878 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879 "test/spmm-microkernel-tester.h",
3880 "src/xnnpack/AlignedAllocator.h",
3881 ] + MICROKERNEL_TEST_HDRS,
3882 deps = MICROKERNEL_TEST_DEPS,
3883)
3884
3885xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003886 name = "f16_vadd_minmax_test",
3887 srcs = [
3888 "test/f16-vadd-minmax.cc",
3889 "test/vbinary-microkernel-tester.h",
3890 ] + MICROKERNEL_TEST_HDRS,
3891 deps = MICROKERNEL_TEST_DEPS,
3892)
3893
3894xnnpack_unit_test(
3895 name = "f16_vaddc_minmax_test",
3896 srcs = [
3897 "test/f16-vaddc-minmax.cc",
3898 "test/vbinaryc-microkernel-tester.h",
3899 ] + MICROKERNEL_TEST_HDRS,
3900 deps = MICROKERNEL_TEST_DEPS,
3901)
3902
3903xnnpack_unit_test(
3904 name = "f16_vdiv_minmax_test",
3905 srcs = [
3906 "test/f16-vdiv-minmax.cc",
3907 "test/vbinary-microkernel-tester.h",
3908 ] + MICROKERNEL_TEST_HDRS,
3909 deps = MICROKERNEL_TEST_DEPS,
3910)
3911
3912xnnpack_unit_test(
3913 name = "f16_vdivc_minmax_test",
3914 srcs = [
3915 "test/f16-vdivc-minmax.cc",
3916 "test/vbinaryc-microkernel-tester.h",
3917 ] + MICROKERNEL_TEST_HDRS,
3918 deps = MICROKERNEL_TEST_DEPS,
3919)
3920
3921xnnpack_unit_test(
3922 name = "f16_vrdivc_minmax_test",
3923 srcs = [
3924 "test/f16-vrdivc-minmax.cc",
3925 "test/vbinaryc-microkernel-tester.h",
3926 ] + MICROKERNEL_TEST_HDRS,
3927 deps = MICROKERNEL_TEST_DEPS,
3928)
3929
3930xnnpack_unit_test(
3931 name = "f16_vmax_test",
3932 srcs = [
3933 "test/f16-vmax.cc",
3934 "test/vbinary-microkernel-tester.h",
3935 ] + MICROKERNEL_TEST_HDRS,
3936 deps = MICROKERNEL_TEST_DEPS,
3937)
3938
3939xnnpack_unit_test(
3940 name = "f16_vmaxc_test",
3941 srcs = [
3942 "test/f16-vmaxc.cc",
3943 "test/vbinaryc-microkernel-tester.h",
3944 ] + MICROKERNEL_TEST_HDRS,
3945 deps = MICROKERNEL_TEST_DEPS,
3946)
3947
3948xnnpack_unit_test(
3949 name = "f16_vmin_test",
3950 srcs = [
3951 "test/f16-vmin.cc",
3952 "test/vbinary-microkernel-tester.h",
3953 ] + MICROKERNEL_TEST_HDRS,
3954 deps = MICROKERNEL_TEST_DEPS,
3955)
3956
3957xnnpack_unit_test(
3958 name = "f16_vminc_test",
3959 srcs = [
3960 "test/f16-vminc.cc",
3961 "test/vbinaryc-microkernel-tester.h",
3962 ] + MICROKERNEL_TEST_HDRS,
3963 deps = MICROKERNEL_TEST_DEPS,
3964)
3965
3966xnnpack_unit_test(
3967 name = "f16_vmul_minmax_test",
3968 srcs = [
3969 "test/f16-vmul-minmax.cc",
3970 "test/vbinary-microkernel-tester.h",
3971 ] + MICROKERNEL_TEST_HDRS,
3972 deps = MICROKERNEL_TEST_DEPS,
3973)
3974
3975xnnpack_unit_test(
3976 name = "f16_vmulc_minmax_test",
3977 srcs = [
3978 "test/f16-vmulc-minmax.cc",
3979 "test/vbinaryc-microkernel-tester.h",
3980 ] + MICROKERNEL_TEST_HDRS,
3981 deps = MICROKERNEL_TEST_DEPS,
3982)
3983
3984xnnpack_unit_test(
3985 name = "f16_vsub_minmax_test",
3986 srcs = [
3987 "test/f16-vsub-minmax.cc",
3988 "test/vbinary-microkernel-tester.h",
3989 ] + MICROKERNEL_TEST_HDRS,
3990 deps = MICROKERNEL_TEST_DEPS,
3991)
3992
3993xnnpack_unit_test(
3994 name = "f16_vsubc_minmax_test",
3995 srcs = [
3996 "test/f16-vsubc-minmax.cc",
3997 "test/vbinaryc-microkernel-tester.h",
3998 ] + MICROKERNEL_TEST_HDRS,
3999 deps = MICROKERNEL_TEST_DEPS,
4000)
4001
4002xnnpack_unit_test(
4003 name = "f16_vrsubc_minmax_test",
4004 srcs = [
4005 "test/f16-vrsubc-minmax.cc",
4006 "test/vbinaryc-microkernel-tester.h",
4007 ] + MICROKERNEL_TEST_HDRS,
4008 deps = MICROKERNEL_TEST_DEPS,
4009)
4010
4011xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004012 name = "f32_vabs_test",
4013 srcs = [
4014 "test/f32-vabs.cc",
4015 "test/vunary-microkernel-tester.h",
4016 ] + MICROKERNEL_TEST_HDRS,
4017 deps = MICROKERNEL_TEST_DEPS,
4018)
4019
4020xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004021 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004023 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004024 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004025 ] + MICROKERNEL_TEST_HDRS,
4026 deps = MICROKERNEL_TEST_DEPS,
4027)
4028
4029xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004030 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004031 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004032 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004033 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004034 ] + MICROKERNEL_TEST_HDRS,
4035 deps = MICROKERNEL_TEST_DEPS,
4036)
4037
4038xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004039 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004040 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004041 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004042 "test/vbinary-microkernel-tester.h",
4043 ] + MICROKERNEL_TEST_HDRS,
4044 deps = MICROKERNEL_TEST_DEPS,
4045)
4046
4047xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004048 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004049 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004050 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004051 "test/vbinaryc-microkernel-tester.h",
4052 ] + MICROKERNEL_TEST_HDRS,
4053 deps = MICROKERNEL_TEST_DEPS,
4054)
4055
4056xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004057 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004058 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004059 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004060 "test/vbinaryc-microkernel-tester.h",
4061 ] + MICROKERNEL_TEST_HDRS,
4062 deps = MICROKERNEL_TEST_DEPS,
4063)
4064
4065xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004066 name = "f32_vmax_test",
4067 srcs = [
4068 "test/f32-vmax.cc",
4069 "test/vbinary-microkernel-tester.h",
4070 ] + MICROKERNEL_TEST_HDRS,
4071 deps = MICROKERNEL_TEST_DEPS,
4072)
4073
4074xnnpack_unit_test(
4075 name = "f32_vmaxc_test",
4076 srcs = [
4077 "test/f32-vmaxc.cc",
4078 "test/vbinaryc-microkernel-tester.h",
4079 ] + MICROKERNEL_TEST_HDRS,
4080 deps = MICROKERNEL_TEST_DEPS,
4081)
4082
4083xnnpack_unit_test(
4084 name = "f32_vmin_test",
4085 srcs = [
4086 "test/f32-vmin.cc",
4087 "test/vbinary-microkernel-tester.h",
4088 ] + MICROKERNEL_TEST_HDRS,
4089 deps = MICROKERNEL_TEST_DEPS,
4090)
4091
4092xnnpack_unit_test(
4093 name = "f32_vminc_test",
4094 srcs = [
4095 "test/f32-vminc.cc",
4096 "test/vbinaryc-microkernel-tester.h",
4097 ] + MICROKERNEL_TEST_HDRS,
4098 deps = MICROKERNEL_TEST_DEPS,
4099)
4100
4101xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004102 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004103 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004104 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004105 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004106 ] + MICROKERNEL_TEST_HDRS,
4107 deps = MICROKERNEL_TEST_DEPS,
4108)
4109
4110xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004111 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004112 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004113 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004114 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004115 ] + MICROKERNEL_TEST_HDRS,
4116 deps = MICROKERNEL_TEST_DEPS,
4117)
4118
4119xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004120 name = "f16_vmulcaddc_minmax_test",
4121 srcs = [
4122 "test/f16-vmulcaddc-minmax.cc",
4123 "test/vmulcaddc-microkernel-tester.h",
4124 "src/xnnpack/AlignedAllocator.h",
4125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4126 deps = MICROKERNEL_TEST_DEPS,
4127)
4128
4129xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004130 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004132 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004133 "test/vmulcaddc-microkernel-tester.h",
4134 "src/xnnpack/AlignedAllocator.h",
4135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4136 deps = MICROKERNEL_TEST_DEPS,
4137)
4138
4139xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004140 name = "f32_vlrelu_test",
4141 srcs = [
4142 "test/f32-vlrelu.cc",
4143 "test/vunary-microkernel-tester.h",
4144 ] + MICROKERNEL_TEST_HDRS,
4145 deps = MICROKERNEL_TEST_DEPS,
4146)
4147
4148xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004149 name = "f32_vneg_test",
4150 srcs = [
4151 "test/f32-vneg.cc",
4152 "test/vunary-microkernel-tester.h",
4153 ] + MICROKERNEL_TEST_HDRS,
4154 deps = MICROKERNEL_TEST_DEPS,
4155)
4156
4157xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004158 name = "f32_vrndne_test",
4159 srcs = [
4160 "test/f32-vrndne.cc",
4161 "test/vunary-microkernel-tester.h",
4162 ] + MICROKERNEL_TEST_HDRS,
4163 deps = MICROKERNEL_TEST_DEPS,
4164)
4165
4166xnnpack_unit_test(
4167 name = "f32_vrndz_test",
4168 srcs = [
4169 "test/f32-vrndz.cc",
4170 "test/vunary-microkernel-tester.h",
4171 ] + MICROKERNEL_TEST_HDRS,
4172 deps = MICROKERNEL_TEST_DEPS,
4173)
4174
4175xnnpack_unit_test(
4176 name = "f32_vrndu_test",
4177 srcs = [
4178 "test/f32-vrndu.cc",
4179 "test/vunary-microkernel-tester.h",
4180 ] + MICROKERNEL_TEST_HDRS,
4181 deps = MICROKERNEL_TEST_DEPS,
4182)
4183
4184xnnpack_unit_test(
4185 name = "f32_vrndd_test",
4186 srcs = [
4187 "test/f32-vrndd.cc",
4188 "test/vunary-microkernel-tester.h",
4189 ] + MICROKERNEL_TEST_HDRS,
4190 deps = MICROKERNEL_TEST_DEPS,
4191)
4192
4193xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004194 name = "f32_vscale_test",
4195 srcs = [
4196 "test/f32-vscale.cc",
4197 "test/vscale-microkernel-tester.h",
4198 ] + MICROKERNEL_TEST_HDRS,
4199 deps = MICROKERNEL_TEST_DEPS,
4200)
4201
4202xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004203 name = "f32_vscaleexpminusmax_test",
4204 srcs = [
4205 "test/f32-vscaleexpminusmax.cc",
4206 "test/vscaleexpminusmax-microkernel-tester.h",
4207 ] + MICROKERNEL_TEST_HDRS,
4208 deps = MICROKERNEL_TEST_DEPS,
4209)
4210
4211xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004212 name = "f32_vscaleextexp_test",
4213 srcs = [
4214 "test/f32-vscaleextexp.cc",
4215 "test/vscaleextexp-microkernel-tester.h",
4216 ] + MICROKERNEL_TEST_HDRS,
4217 deps = MICROKERNEL_TEST_DEPS,
4218)
4219
4220xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004221 name = "f32_vsqr_test",
4222 srcs = [
4223 "test/f32-vsqr.cc",
4224 "test/vunary-microkernel-tester.h",
4225 ] + MICROKERNEL_TEST_HDRS,
4226 deps = MICROKERNEL_TEST_DEPS,
4227)
4228
4229xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004230 name = "f32_vsqrdiff_test",
4231 srcs = [
4232 "test/f32-vsqrdiff.cc",
4233 "test/vbinary-microkernel-tester.h",
4234 ] + MICROKERNEL_TEST_HDRS,
4235 deps = MICROKERNEL_TEST_DEPS,
4236)
4237
4238xnnpack_unit_test(
4239 name = "f32_vsqrdiffc_test",
4240 srcs = [
4241 "test/f32-vsqrdiffc.cc",
4242 "test/vbinaryc-microkernel-tester.h",
4243 ] + MICROKERNEL_TEST_HDRS,
4244 deps = MICROKERNEL_TEST_DEPS,
4245)
4246
4247xnnpack_unit_test(
4248 name = "f32_vrsqrdiffc_test",
4249 srcs = [
4250 "test/f32-vrsqrdiffc.cc",
4251 "test/vbinaryc-microkernel-tester.h",
4252 ] + MICROKERNEL_TEST_HDRS,
4253 deps = MICROKERNEL_TEST_DEPS,
4254)
4255
4256xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004257 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004258 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004259 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004260 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004261 ] + MICROKERNEL_TEST_HDRS,
4262 deps = MICROKERNEL_TEST_DEPS,
4263)
4264
4265xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004266 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004267 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004268 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004269 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004270 ] + MICROKERNEL_TEST_HDRS,
4271 deps = MICROKERNEL_TEST_DEPS,
4272)
4273
4274xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004275 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004276 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004277 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004278 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004279 ] + MICROKERNEL_TEST_HDRS,
4280 deps = MICROKERNEL_TEST_DEPS,
4281)
4282
4283xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004284 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004286 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287 "test/avgpool-microkernel-tester.h",
4288 "src/xnnpack/AlignedAllocator.h",
4289 ] + MICROKERNEL_TEST_HDRS,
4290 deps = MICROKERNEL_TEST_DEPS,
4291)
4292
4293xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004294 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004295 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004296 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297 "test/gemm-microkernel-tester.h",
4298 "src/xnnpack/AlignedAllocator.h",
4299 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4300 deps = MICROKERNEL_TEST_DEPS,
4301)
4302
4303xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004304 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004306 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 "test/dwconv-microkernel-tester.h",
4308 "src/xnnpack/AlignedAllocator.h",
4309 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4310 deps = MICROKERNEL_TEST_DEPS,
4311)
4312
4313xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004314 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004315 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004316 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317 "test/gavgpool-microkernel-tester.h",
4318 "src/xnnpack/AlignedAllocator.h",
4319 ] + MICROKERNEL_TEST_HDRS,
4320 deps = MICROKERNEL_TEST_DEPS,
4321)
4322
4323xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004324 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004326 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327 "test/gemm-microkernel-tester.h",
4328 "src/xnnpack/AlignedAllocator.h",
4329 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4330 deps = MICROKERNEL_TEST_DEPS,
4331)
4332
4333xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004334 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004336 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337 "test/vadd-microkernel-tester.h",
4338 ] + MICROKERNEL_TEST_HDRS,
4339 deps = MICROKERNEL_TEST_DEPS,
4340)
4341
4342xnnpack_unit_test(
4343 name = "u8_clamp_test",
4344 srcs = [
4345 "test/u8-clamp.cc",
4346 "test/clamp-microkernel-tester.h",
4347 ] + MICROKERNEL_TEST_HDRS,
4348 deps = MICROKERNEL_TEST_DEPS,
4349)
4350
4351xnnpack_unit_test(
4352 name = "u8_lut32norm_test",
4353 srcs = [
4354 "test/u8-lut32norm.cc",
4355 "test/lut-norm-microkernel-tester.h",
4356 ] + MICROKERNEL_TEST_HDRS,
4357 deps = MICROKERNEL_TEST_DEPS,
4358)
4359
4360xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004361 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004363 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 "test/maxpool-microkernel-tester.h",
4365 ] + MICROKERNEL_TEST_HDRS,
4366 deps = MICROKERNEL_TEST_DEPS,
4367)
4368
4369xnnpack_unit_test(
4370 name = "u8_rmax_test",
4371 srcs = [
4372 "test/u8-rmax.cc",
4373 "test/rmax-microkernel-tester.h",
4374 ] + MICROKERNEL_TEST_HDRS,
4375 deps = MICROKERNEL_TEST_DEPS,
4376)
4377
4378xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004379 name = "x32_fill_test",
4380 srcs = [
4381 "test/x32-fill.cc",
4382 "test/fill-microkernel-tester.h",
4383 ] + MICROKERNEL_TEST_HDRS,
4384 deps = MICROKERNEL_TEST_DEPS,
4385)
4386
4387xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 name = "x32_packx_test",
4389 srcs = [
4390 "test/x32-packx.cc",
4391 "test/pack-microkernel-tester.h",
4392 "src/xnnpack/AlignedAllocator.h",
4393 ] + MICROKERNEL_TEST_HDRS,
4394 deps = MICROKERNEL_TEST_DEPS,
4395)
4396
4397xnnpack_unit_test(
4398 name = "x32_pad_test",
4399 srcs = [
4400 "test/x32-pad.cc",
4401 "test/pad-microkernel-tester.h",
4402 ] + MICROKERNEL_TEST_HDRS,
4403 deps = MICROKERNEL_TEST_DEPS,
4404)
4405
4406xnnpack_unit_test(
4407 name = "x32_unpool_test",
4408 srcs = [
4409 "test/x32-unpool.cc",
4410 "test/unpool-microkernel-tester.h",
4411 ] + MICROKERNEL_TEST_HDRS,
4412 deps = MICROKERNEL_TEST_DEPS,
4413)
4414
4415xnnpack_unit_test(
4416 name = "x32_zip_test",
4417 srcs = [
4418 "test/x32-zip.cc",
4419 "test/zip-microkernel-tester.h",
4420 ] + MICROKERNEL_TEST_HDRS,
4421 deps = MICROKERNEL_TEST_DEPS,
4422)
4423
4424xnnpack_unit_test(
4425 name = "x8_lut_test",
4426 srcs = [
4427 "test/x8-lut.cc",
4428 "test/lut-microkernel-tester.h",
4429 ] + MICROKERNEL_TEST_HDRS,
4430 deps = MICROKERNEL_TEST_DEPS,
4431)
4432
4433xnnpack_unit_test(
4434 name = "x8_zip_test",
4435 srcs = [
4436 "test/x8-zip.cc",
4437 "test/zip-microkernel-tester.h",
4438 ] + MICROKERNEL_TEST_HDRS,
4439 deps = MICROKERNEL_TEST_DEPS,
4440)
4441
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004442xnnpack_unit_test(
4443 name = "requantization_test",
4444 srcs = [
4445 "src/xnnpack/requantization-stubs.h",
4446 "test/requantization.cc",
4447 "test/requantization-tester.h",
4448 ] + MICROKERNEL_TEST_HDRS,
4449 deps = MICROKERNEL_TEST_DEPS,
4450)
4451
Marat Dukhan20c3b922020-03-10 03:45:06 -07004452########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004453
4454xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07004455 name = "operator_size_test",
4456 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004457 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458)
4459
Marat Dukhan20c3b922020-03-10 03:45:06 -07004460xnnpack_binary(
4461 name = "subgraph_size_test",
4462 srcs = ["test/subgraph-size.c"],
4463 deps = [":XNNPACK"],
4464)
4465
4466########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467
4468xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004469 name = "abs_nc_test",
4470 srcs = [
4471 "test/abs-nc.cc",
4472 "test/abs-operator-tester.h",
4473 ],
4474 deps = OPERATOR_TEST_DEPS,
4475)
4476
4477xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004478 name = "add_nd_test",
4479 srcs = [
4480 "test/add-nd.cc",
4481 "test/binary-elementwise-operator-tester.h",
4482 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004483 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004484)
4485
4486xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004487 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004489 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490 "test/argmax-pooling-operator-tester.h",
4491 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004492 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493)
4494
4495xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004496 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004498 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 "test/average-pooling-operator-tester.h",
4500 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004501 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502)
4503
4504xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004505 name = "bankers_rounding_nc_test",
4506 srcs = [
4507 "test/bankers-rounding-nc.cc",
4508 "test/bankers-rounding-operator-tester.h",
4509 ],
4510 deps = OPERATOR_TEST_DEPS,
4511)
4512
4513xnnpack_unit_test(
4514 name = "ceiling_nc_test",
4515 srcs = [
4516 "test/ceiling-nc.cc",
4517 "test/ceiling-operator-tester.h",
4518 ],
4519 deps = OPERATOR_TEST_DEPS,
4520)
4521
4522xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004523 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004525 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 "test/channel-shuffle-operator-tester.h",
4527 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004528 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529)
4530
4531xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004532 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004533 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004534 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535 "test/clamp-operator-tester.h",
4536 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004537 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538)
4539
4540xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07004541 name = "constant_pad_nd_test",
4542 srcs = [
4543 "test/constant-pad-nd.cc",
4544 "test/constant-pad-operator-tester.h",
4545 ],
4546 deps = OPERATOR_TEST_DEPS,
4547)
4548
4549xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004550 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004552 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 "test/convolution-operator-tester.h",
4554 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004555 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556)
4557
4558xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004559 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004561 "test/convolution-nchw.cc",
4562 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004563 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004564 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004565)
4566
4567xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07004568 name = "copy_nc_test",
4569 srcs = [
4570 "test/copy-nc.cc",
4571 "test/copy-operator-tester.h",
4572 ],
4573 deps = OPERATOR_TEST_DEPS,
4574)
4575
4576xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004577 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004579 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580 "test/deconvolution-operator-tester.h",
4581 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004582 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004583)
4584
4585xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08004586 name = "divide_nd_test",
4587 srcs = [
4588 "test/binary-elementwise-operator-tester.h",
4589 "test/divide-nd.cc",
4590 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004591 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08004592)
4593
4594xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004595 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004596 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004597 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598 "test/fully-connected-operator-tester.h",
4599 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004600 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004601)
4602
4603xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004604 name = "floor_nc_test",
4605 srcs = [
4606 "test/floor-nc.cc",
4607 "test/floor-operator-tester.h",
4608 ],
4609 deps = OPERATOR_TEST_DEPS,
4610)
4611
4612xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004613 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004614 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004615 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004616 "test/global-average-pooling-operator-tester.h",
4617 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004618 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619)
4620
4621xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004622 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004624 "test/global-average-pooling-ncw.cc",
4625 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004627 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004628)
4629
4630xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004631 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004633 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004634 "test/hardswish-operator-tester.h",
4635 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004636 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004637)
4638
4639xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004640 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004641 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004642 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 "test/leaky-relu-operator-tester.h",
4644 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004645 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646)
4647
4648xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004649 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004650 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004651 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004652 "test/max-pooling-operator-tester.h",
4653 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004654 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004655)
4656
4657xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08004658 name = "maximum_nd_test",
4659 srcs = [
4660 "test/binary-elementwise-operator-tester.h",
4661 "test/maximum-nd.cc",
4662 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004663 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004664)
4665
4666xnnpack_unit_test(
4667 name = "minimum_nd_test",
4668 srcs = [
4669 "test/binary-elementwise-operator-tester.h",
4670 "test/minimum-nd.cc",
4671 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004672 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004673)
4674
4675xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004676 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004677 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004678 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004679 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004680 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004681 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08004682)
4683
4684xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004685 name = "negate_nc_test",
4686 srcs = [
4687 "test/negate-nc.cc",
4688 "test/negate-operator-tester.h",
4689 ],
4690 deps = OPERATOR_TEST_DEPS,
4691)
4692
4693xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004694 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004695 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004696 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004697 "test/prelu-operator-tester.h",
4698 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004699 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004700)
4701
4702xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004703 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08004704 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004705 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08004706 "test/resize-bilinear-operator-tester.h",
4707 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004708 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08004709)
4710
4711xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004712 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004713 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004714 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715 "test/sigmoid-operator-tester.h",
4716 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004717 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718)
4719
4720xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004721 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004723 "test/softmax-nc.cc",
4724 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004726 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004727)
4728
4729xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004730 name = "square_nc_test",
4731 srcs = [
4732 "test/square-nc.cc",
4733 "test/square-operator-tester.h",
4734 ],
4735 deps = OPERATOR_TEST_DEPS,
4736)
4737
4738xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07004739 name = "squared_difference_nd_test",
4740 srcs = [
4741 "test/binary-elementwise-operator-tester.h",
4742 "test/squared-difference-nd.cc",
4743 ],
4744 deps = OPERATOR_TEST_DEPS,
4745)
4746
4747xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004748 name = "subtract_nd_test",
4749 srcs = [
4750 "test/binary-elementwise-operator-tester.h",
4751 "test/subtract-nd.cc",
4752 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004753 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004754)
4755
4756xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004757 name = "truncation_nc_test",
4758 srcs = [
4759 "test/truncation-nc.cc",
4760 "test/truncation-operator-tester.h",
4761 ],
4762 deps = OPERATOR_TEST_DEPS,
4763)
4764
4765xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004766 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004767 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004768 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004769 "test/unpooling-operator-tester.h",
4770 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004771 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772)
4773
Chao Mei6ddfc602020-05-13 22:29:36 -07004774############################### Misc unit tests ###############################
4775
4776xnnpack_unit_test(
4777 name = "memory_planner_test",
4778 srcs = [
4779 "test/memory-planner-test.cc",
4780 ],
4781 deps = [
4782 ":XNNPACK",
4783 ":memory_planner",
4784 ],
4785)
4786
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787############################# Build configurations #############################
4788
Marat Dukhanb8642352019-10-30 15:43:02 -07004789# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07004790config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004791 name = "xnn_enable_assembly_explicit_true",
4792 define_values = {"xnn_enable_assembly": "true"},
4793)
4794
4795# Disables usage of assembly kernels.
4796config_setting(
4797 name = "xnn_enable_assembly_explicit_false",
4798 define_values = {"xnn_enable_assembly": "false"},
4799)
4800
Marat Dukhan05702cf2020-03-26 15:41:33 -07004801# Disables usage of HMP-aware optimizations.
4802config_setting(
4803 name = "xnn_enable_hmp_explicit_false",
4804 define_values = {"xnn_enable_hmp": "false"},
4805)
4806
Chao Mei6ddfc602020-05-13 22:29:36 -07004807# Enable usage of optimized memory allocation
4808config_setting(
4809 name = "xnn_enable_memopt_explicit_true",
4810 define_values = {"xnnpack_enable_memopt": "true"},
4811)
4812
4813# Disable usage of optimized memory allocation
4814config_setting(
4815 name = "xnn_enable_memopt_explicit_false",
4816 define_values = {"xnnpack_enable_memopt": "false"},
4817)
4818
Marat Dukhanb8642352019-10-30 15:43:02 -07004819# Builds with -c dbg
4820config_setting(
4821 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004822 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07004823 "compilation_mode": "dbg",
4824 },
4825)
4826
4827# Builds with -c opt
4828config_setting(
4829 name = "optimized_build",
4830 values = {
4831 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004832 },
4833)
4834
4835config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004836 name = "linux_k8",
4837 values = {"cpu": "k8"},
4838)
4839
4840config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004841 name = "linux_arm",
4842 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07004843)
4844
4845config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07004846 name = "linux_armhf",
4847 values = {"cpu": "armhf"},
4848)
4849
4850config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07004851 name = "linux_armv7a",
4852 values = {"cpu": "armv7a"},
4853)
4854
4855config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004856 name = "linux_aarch64",
4857 values = {"cpu": "aarch64"},
4858)
4859
4860config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 name = "android",
4862 values = {"crosstool_top": "//external:android/crosstool"},
4863)
4864
4865config_setting(
4866 name = "android_armv7",
4867 values = {
4868 "crosstool_top": "//external:android/crosstool",
4869 "cpu": "armeabi-v7a",
4870 },
4871)
4872
4873config_setting(
4874 name = "android_arm64",
4875 values = {
4876 "crosstool_top": "//external:android/crosstool",
4877 "cpu": "arm64-v8a",
4878 },
4879)
4880
4881config_setting(
4882 name = "android_x86",
4883 values = {
4884 "crosstool_top": "//external:android/crosstool",
4885 "cpu": "x86",
4886 },
4887)
4888
4889config_setting(
4890 name = "android_x86_64",
4891 values = {
4892 "crosstool_top": "//external:android/crosstool",
4893 "cpu": "x86_64",
4894 },
4895)
4896
4897config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004898 name = "windows_x86_64",
4899 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004900)
4901
4902config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004903 name = "windows_x86_64_clang",
4904 values = {
4905 "compiler": "clang-cl",
4906 "cpu": "x64_windows",
4907 },
4908)
4909
4910config_setting(
4911 name = "windows_x86_64_mingw",
4912 values = {
4913 "compiler": "mingw-gcc",
4914 "cpu": "x64_windows",
4915 },
4916)
4917
4918config_setting(
4919 name = "windows_x86_64_msys",
4920 values = {
4921 "compiler": "msys-gcc",
4922 "cpu": "x64_windows",
4923 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004924)
4925
4926config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07004927 name = "macos_x86_64",
4928 values = {
4929 "apple_platform_type": "macos",
4930 "cpu": "darwin",
4931 },
4932)
4933
4934config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004935 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004936 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937)
4938
4939config_setting(
4940 name = "emscripten_wasm",
4941 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004942 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004943 "cpu": "wasm",
4944 },
4945)
4946
4947config_setting(
4948 name = "emscripten_wasmsimd",
4949 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004950 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004951 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07004952 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004953 },
4954)
4955
4956config_setting(
4957 name = "emscripten_asmjs",
4958 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004959 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 "cpu": "asmjs",
4961 },
4962)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004963
4964config_setting(
4965 name = "ios_armv7",
4966 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004967 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004968 "cpu": "ios_armv7",
4969 },
4970)
4971
4972config_setting(
4973 name = "ios_arm64",
4974 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004975 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004976 "cpu": "ios_arm64",
4977 },
4978)
4979
4980config_setting(
4981 name = "ios_arm64e",
4982 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004983 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004984 "cpu": "ios_arm64e",
4985 },
4986)
4987
4988config_setting(
4989 name = "ios_x86",
4990 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004991 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004992 "cpu": "ios_i386",
4993 },
4994)
4995
4996config_setting(
4997 name = "ios_x86_64",
4998 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004999 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005000 "cpu": "ios_x86_64",
5001 },
5002)
5003
5004config_setting(
5005 name = "watchos_armv7k",
5006 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005007 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005008 "cpu": "watchos_armv7k",
5009 },
5010)
5011
5012config_setting(
5013 name = "watchos_arm64_32",
5014 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005015 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005016 "cpu": "watchos_arm64_32",
5017 },
5018)
5019
5020config_setting(
5021 name = "watchos_x86",
5022 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005023 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005024 "cpu": "watchos_i386",
5025 },
5026)
5027
5028config_setting(
5029 name = "watchos_x86_64",
5030 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005031 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005032 "cpu": "watchos_x86_64",
5033 },
5034)
5035
5036config_setting(
5037 name = "tvos_arm64",
5038 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005039 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005040 "cpu": "tvos_arm64",
5041 },
5042)
5043
5044config_setting(
5045 name = "tvos_x86_64",
5046 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005047 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005048 "cpu": "tvos_x86_64",
5049 },
5050)