blob: 4e36cbd910a1b97006bdfc79330c010130a7eac0 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700102 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700103 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700104 "src/subgraph/squared-difference.c",
105 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/unpooling-2d.c",
107]
108
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800109TABLE_SRCS = [
110 "src/tables/exp2-k-over-64.c",
111 "src/tables/exp2-k-over-2048.c",
112]
113
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800116 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800117 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700118 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
119 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700120 "src/f32-clamp/gen/scalar-x1.c",
121 "src/f32-clamp/gen/scalar-x2.c",
122 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800123 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700125 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
126 "src/f32-dwconv-chw/3x3p1-scalar.c",
127 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
128 "src/f32-dwconv-chw/5x5p2-scalar.c",
129 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700130 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x4-scalar.c",
132 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x9-scalar.c",
134 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
135 "src/f32-dwconv/gen/up1x25-scalar.c",
136 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x4-scalar.c",
138 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x9-scalar.c",
140 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
141 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700142 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
144 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700148 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700152 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700154 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700155 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
156 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
158 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700172 "src/f32-hswish/gen/hswish-scalar-x1.c",
173 "src/f32-hswish/gen/hswish-scalar-x2.c",
174 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700175 "src/f32-ibilinear/gen/scalar-c1.c",
176 "src/f32-ibilinear/gen/scalar-c2.c",
177 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700178 "src/f32-igemm/gen/1x4-scalar.c",
179 "src/f32-igemm/gen/2x4-scalar.c",
180 "src/f32-igemm/gen/4x2-scalar.c",
181 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700182 "src/f32-igemm/gen/1x4-relu-scalar.c",
183 "src/f32-igemm/gen/2x4-relu-scalar.c",
184 "src/f32-igemm/gen/4x2-relu-scalar.c",
185 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700186 "src/f32-igemm/gen/1x4-minmax-scalar.c",
187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
188 "src/f32-igemm/gen/4x2-minmax-scalar.c",
189 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700190 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
191 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
194 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
195 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800197 "src/f32-prelu/gen/scalar-2x1.c",
198 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700211 "src/f32-relu/gen/scalar-x1.c",
212 "src/f32-relu/gen/scalar-x2.c",
213 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700214 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800215 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
218 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
221 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700224 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
225 "src/f32-spmm/gen/1x1-minmax-scalar.c",
226 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar.c",
228 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar.c",
230 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar.c",
232 "src/f32-spmm/gen/8x2-minmax-scalar.c",
233 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700234 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
235 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700237 "src/f32-vbinary/gen/vadd-scalar-x1.c",
238 "src/f32-vbinary/gen/vadd-scalar-x2.c",
239 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700240 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700243 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
244 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700246 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
250 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700252 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700255 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
256 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800258 "src/f32-vbinary/gen/vmax-scalar-x1.c",
259 "src/f32-vbinary/gen/vmax-scalar-x2.c",
260 "src/f32-vbinary/gen/vmax-scalar-x4.c",
261 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
264 "src/f32-vbinary/gen/vmin-scalar-x1.c",
265 "src/f32-vbinary/gen/vmin-scalar-x2.c",
266 "src/f32-vbinary/gen/vmin-scalar-x4.c",
267 "src/f32-vbinary/gen/vminc-scalar-x1.c",
268 "src/f32-vbinary/gen/vminc-scalar-x2.c",
269 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700270 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
271 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "src/f32-vbinary/gen/vmul-scalar-x1.c",
274 "src/f32-vbinary/gen/vmul-scalar-x2.c",
275 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700279 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
280 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700282 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
286 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
292 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700294 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
295 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
297 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vsub-scalar-x1.c",
304 "src/f32-vbinary/gen/vsub-scalar-x2.c",
305 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700306 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700309 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
310 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700312 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
313 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700315 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
316 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700318 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
319 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
321 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
324 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
327 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700330 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
331 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700333 "src/f32-vunary/gen/vabs-scalar-x1.c",
334 "src/f32-vunary/gen/vabs-scalar-x2.c",
335 "src/f32-vunary/gen/vabs-scalar-x4.c",
336 "src/f32-vunary/gen/vneg-scalar-x1.c",
337 "src/f32-vunary/gen/vneg-scalar-x2.c",
338 "src/f32-vunary/gen/vneg-scalar-x4.c",
339 "src/f32-vunary/gen/vsqr-scalar-x1.c",
340 "src/f32-vunary/gen/vsqr-scalar-x2.c",
341 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800342 "src/math/expminus-scalar-lut2048-p1.c",
343 "src/math/expminus-scalar-lut64-p2.c",
344 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700345 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700346 "src/math/roundne-scalar-nearbyint.c",
347 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundd-scalar-addsub.c",
349 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700350 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700351 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700352 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700353 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700354 "src/math/roundz-scalar-addsub.c",
355 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700356 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800357 "src/math/sigmoid-scalar-lut2048-p1-div.c",
358 "src/math/sigmoid-scalar-lut64-p2-div.c",
359 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700360 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
361 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
362 "src/qu8-dwconv/up1x9-minmax-scalar.c",
363 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
364 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
365 "src/qu8-gemm/2x2-minmax-scalar.c",
366 "src/qu8-igemm/2x2-minmax-scalar.c",
367 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700368 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700369 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700370 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700371 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700372 "src/x32-fill/scalar-float.c",
373 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700374 "src/x32-packx/x2-scalar.c",
375 "src/x32-packx/x3-scalar.c",
376 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700377 "src/x32-pad/scalar-int.c",
378 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700379 "src/x32-unpool/scalar.c",
380 "src/x32-zip/x2-scalar.c",
381 "src/x32-zip/x3-scalar.c",
382 "src/x32-zip/x4-scalar.c",
383 "src/x32-zip/xm-scalar.c",
384 "src/x8-lut/scalar.c",
385 "src/x8-zip/x2-scalar.c",
386 "src/x8-zip/x3-scalar.c",
387 "src/x8-zip/x4-scalar.c",
388 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700389 "src/requantization/precise-scalar.c",
390 "src/requantization/fp32-scalar.c",
391 "src/requantization/q31-scalar.c",
392 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700393]
394
Marat Dukhan436ebe62019-12-04 15:10:12 -0800395WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700396 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
397 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700398 "src/f32-clamp/gen/wasm-x1.c",
399 "src/f32-clamp/gen/wasm-x2.c",
400 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700401 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
402 "src/f32-dwconv/gen/up1x4-wasm.c",
403 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
404 "src/f32-dwconv/gen/up1x9-wasm.c",
405 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
406 "src/f32-dwconv/gen/up1x25-wasm.c",
407 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
408 "src/f32-dwconv/gen/up2x4-wasm.c",
409 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
410 "src/f32-dwconv/gen/up2x9-wasm.c",
411 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
412 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700413 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
414 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
415 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
416 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700417 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
418 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700419 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
420 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
421 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
422 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700423 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
424 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700425 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
426 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700427 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
428 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
429 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700430 "src/f32-gemm/gen/1x4-wasm.c",
431 "src/f32-gemm/gen/2x4-wasm.c",
432 "src/f32-gemm/gen/4x2-wasm.c",
433 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700434 "src/f32-gemm/gen/1x4-relu-wasm.c",
435 "src/f32-gemm/gen/2x4-relu-wasm.c",
436 "src/f32-gemm/gen/4x2-relu-wasm.c",
437 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700438 "src/f32-gemm/gen/1x4-minmax-wasm.c",
439 "src/f32-gemm/gen/2x4-minmax-wasm.c",
440 "src/f32-gemm/gen/4x2-minmax-wasm.c",
441 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700442 "src/f32-hswish/gen/hswish-wasm-x1.c",
443 "src/f32-hswish/gen/hswish-wasm-x2.c",
444 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700445 "src/f32-igemm/gen/1x4-wasm.c",
446 "src/f32-igemm/gen/2x4-wasm.c",
447 "src/f32-igemm/gen/4x2-wasm.c",
448 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700449 "src/f32-igemm/gen/1x4-relu-wasm.c",
450 "src/f32-igemm/gen/2x4-relu-wasm.c",
451 "src/f32-igemm/gen/4x2-relu-wasm.c",
452 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700453 "src/f32-igemm/gen/1x4-minmax-wasm.c",
454 "src/f32-igemm/gen/2x4-minmax-wasm.c",
455 "src/f32-igemm/gen/4x2-minmax-wasm.c",
456 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700457 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
458 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
459 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700460 "src/f32-prelu/gen/wasm-2x1.c",
461 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700462 "src/f32-relu/gen/wasm-x1.c",
463 "src/f32-relu/gen/wasm-x2.c",
464 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700465 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
466 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
467 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
468 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
469 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
470 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
471 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
472 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
473 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
474 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800477 "src/f32-vbinary/gen/vmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
481 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
482 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
483 "src/f32-vbinary/gen/vmin-wasm-x1.c",
484 "src/f32-vbinary/gen/vmin-wasm-x2.c",
485 "src/f32-vbinary/gen/vmin-wasm-x4.c",
486 "src/f32-vbinary/gen/vminc-wasm-x1.c",
487 "src/f32-vbinary/gen/vminc-wasm-x2.c",
488 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700489 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
490 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
491 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
492 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
493 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
494 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
495 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
496 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
497 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
498 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700507 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
508 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
509 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
511 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
512 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800513]
514
Marat Dukhan290055c2020-06-09 12:24:29 -0700515WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700516 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
517 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
518 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700519 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
520 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
521 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
522 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700523 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
524 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
525 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
526 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700527 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
528 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
529 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
530 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
531 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
532 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700533 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
534 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
535 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
536 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
537 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
538 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
539 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
540 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
541 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
542 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
543 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
544 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
545 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
547 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
549 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
551 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
553 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
555 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700557 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
558 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
559 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700560 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700561 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
562 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
563 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
564 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700565 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
566 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
567 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
568 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
569 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
570 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
571 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
572 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
573 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
574 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
581 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
582 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
583 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
584 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
585 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
586 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
587 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
588 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
589 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700595 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
596 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700597 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
598 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
599 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
600 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
601 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
602 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
603 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
604 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
605 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
606 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
613 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
614 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
615 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
616 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
617 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
618 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
619 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
620 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
621 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700627 "src/f32-ibilinear/gen/wasmsimd-c4.c",
628 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700629 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
630 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
631 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700632 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700633 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
634 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
635 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
636 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700637 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
638 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
639 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
640 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
653 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
654 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
655 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700667 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
668 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700669 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
670 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
671 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
672 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700673 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
674 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
675 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700676 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
677 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700678 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700682 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
683 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
684 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
685 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700686 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
687 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
688 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
689 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
690 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
691 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
692 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
693 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
694 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
695 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
696 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
697 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700698 "src/f32-relu/gen/wasmsimd-x4.c",
699 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700700 "src/f32-rmax/wasmsimd-arm.c",
701 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700702 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
703 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
704 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
705 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
706 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
707 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700708 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
709 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
710 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
711 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700712 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
713 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700714 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
715 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
716 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
717 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700718 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700720 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
721 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
722 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
723 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700724 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
725 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700726 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
727 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
728 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
729 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700730 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
731 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700732 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
733 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
734 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
735 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700736 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
737 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
738 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
739 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700740 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
741 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
742 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
743 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700744 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
745 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
746 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
747 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700748 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
749 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
750 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
751 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700752 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
753 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700754 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
755 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
756 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
757 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700758 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
759 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700760 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
761 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
762 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
763 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700764 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
765 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700766 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
767 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
768 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
769 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700770 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
771 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700772 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
773 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700774 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
775 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700776 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
777 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
778 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
779 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700780 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
781 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700782 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
783 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
784 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
785 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700786 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
787 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700788 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
789 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
790 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
791 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700792 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
793 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
794 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700796 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
797 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
798 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
799 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
800 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
801 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
802 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
803 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
804 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
805 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
806 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
807 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
808 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
809 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700810 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
811 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700812 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
813 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
814 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
815 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
816 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
817 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700818 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700819 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700820 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700821 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700822 "src/x32-zip/x2-wasmsimd.c",
823 "src/x32-zip/x3-wasmsimd.c",
824 "src/x32-zip/x4-wasmsimd.c",
825 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700826 "src/math/roundne-wasmsimd-addsub.c",
827 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700828 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700829 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700830 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700831 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700832 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan9340b6c2020-07-16 22:59:48 -0700833 "src/math/sigmoid-wasmsimd-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700834 "src/requantization/fp32-wasmsimd.c",
835]
836
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800837PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800838 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800839 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800840 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700841 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
842 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700843 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700844 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700845 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700846 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700847 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700848 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700849 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
850 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700851 "src/f32-ibilinear/gen/psimd-c4.c",
852 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700853 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700854 "src/f32-spmm/gen/4x1-minmax-psimd.c",
855 "src/f32-spmm/gen/8x1-minmax-psimd.c",
856 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700857 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
858 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700859 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800860 "src/f32-rmax/psimd.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700861 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700862 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700863 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700864 "src/x32-unpool/psimd.c",
865 "src/x32-zip/x2-psimd.c",
866 "src/x32-zip/x3-psimd.c",
867 "src/x32-zip/x4-psimd.c",
868 "src/x32-zip/xm-psimd.c",
869]
870
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800871PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan249cd6a2020-07-17 14:05:30 -0700872 "src/requantization/precise-psimd.c",
873 "src/requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800874]
875
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876# ISA-specific micro-kernels
877NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700878 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
879 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700880 "src/f32-clamp/gen/neon-x4.c",
881 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700882 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
883 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700884 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
885 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700886 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
887 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700888 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
889 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700890 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
891 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
892 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
893 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700894 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
895 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
896 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
897 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700898 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
899 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
900 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
901 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700902 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700903 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
904 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700905 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
906 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
907 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
908 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
909 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
910 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
911 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
912 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
913 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
914 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
915 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
916 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
917 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
918 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
919 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
920 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
921 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
922 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
923 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
925 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
927 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
928 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
929 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
930 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
931 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
932 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
933 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
934 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
935 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700936 "src/f32-hswish/gen/hswish-neon-x4.c",
937 "src/f32-hswish/gen/hswish-neon-x8.c",
938 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700939 "src/f32-ibilinear/gen/neon-c4.c",
940 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700941 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
942 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
943 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
944 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
945 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
946 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
947 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
948 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
949 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
950 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
951 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
952 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
953 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
954 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
955 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
956 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700957 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
958 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
959 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-ppmm/gen/4x8-minmax-neon.c",
961 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800962 "src/f32-prelu/gen/neon-2x4.c",
963 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800964 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
965 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
966 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
967 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
968 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
969 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
970 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
971 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
972 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
973 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
974 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
975 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
976 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
977 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
978 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
979 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
980 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
981 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
987 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700988 "src/f32-relu/gen/neon-x4.c",
989 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700990 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800991 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800992 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
993 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
994 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
995 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
996 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
997 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
998 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
999 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1000 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1001 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1002 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1003 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1004 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1005 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1006 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1007 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1008 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1009 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001010 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1011 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1012 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1013 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001014 "src/f32-vbinary/gen/vmax-neon-x4.c",
1015 "src/f32-vbinary/gen/vmax-neon-x8.c",
1016 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1017 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1018 "src/f32-vbinary/gen/vmin-neon-x4.c",
1019 "src/f32-vbinary/gen/vmin-neon-x8.c",
1020 "src/f32-vbinary/gen/vminc-neon-x4.c",
1021 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001022 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1023 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1024 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1025 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1026 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1027 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001028 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1029 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1030 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1031 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001032 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1033 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1034 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1035 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001036 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1037 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001038 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1039 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001040 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1041 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1042 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1043 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1044 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1045 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1046 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1047 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001048 "src/f32-vunary/gen/vabs-neon-x4.c",
1049 "src/f32-vunary/gen/vabs-neon-x8.c",
1050 "src/f32-vunary/gen/vneg-neon-x4.c",
1051 "src/f32-vunary/gen/vneg-neon-x8.c",
1052 "src/f32-vunary/gen/vsqr-neon-x4.c",
1053 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001054 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1055 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1056 "src/qu8-dwconv/up8x9-minmax-neon.c",
1057 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1058 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1059 "src/qu8-gemm/4x8-minmax-neon.c",
1060 "src/qu8-gemm/8x8-minmax-neon.c",
1061 "src/qu8-igemm/4x8-minmax-neon.c",
1062 "src/qu8-igemm/8x8-minmax-neon.c",
1063 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001064 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001065 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001066 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001067 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001068 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001069 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001070 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001071 "src/x32-zip/x2-neon.c",
1072 "src/x32-zip/x3-neon.c",
1073 "src/x32-zip/x4-neon.c",
1074 "src/x32-zip/xm-neon.c",
1075 "src/x8-zip/x2-neon.c",
1076 "src/x8-zip/x3-neon.c",
1077 "src/x8-zip/x4-neon.c",
1078 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001079 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001080 "src/math/roundd-neon-addsub.c",
1081 "src/math/roundd-neon-cvt.c",
1082 "src/math/roundu-neon-addsub.c",
1083 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001084 "src/math/roundz-neon-addsub.c",
1085 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001086 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001087 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1088 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1089 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1090 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1091 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1092 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001093 "src/math/sqrt-neon-nr1rsqrts.c",
1094 "src/math/sqrt-neon-nr2rsqrts.c",
1095 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001096 "src/requantization/precise-neon.c",
1097 "src/requantization/fp32-neon.c",
1098 "src/requantization/q31-neon.c",
1099 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001100]
1101
1102NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001103 "src/f32-ibilinear/gen/neonfma-c4.c",
1104 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001105 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1106 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1107 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1108 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1109 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1110 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1111 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1112 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1113 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001114 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1115 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1116 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1117 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001118 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1119 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1120 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1121 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001122 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1123 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1124 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1125 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001126 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1127 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1128 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1129 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1130 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1131 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1132 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1133 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1134 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1135 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1136 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1137 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1138 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1139 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1140 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1141 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1142 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1143 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001144 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1145 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001146 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1147 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1148 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1149 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1150 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1151 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1152 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1153 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1154 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1155 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1156 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1157 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1158 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1159 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1160 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1161 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1162 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1163 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1169 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001170 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1171 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1172 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1173 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1174 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1175 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1176 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1177 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001224 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1225 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001226 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1227 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1228 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1229 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1230 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1231 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1232 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1233 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1234 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1235 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1236 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1237 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1238 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1239 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1240 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1241 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1242 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1243 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1244 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1245 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001246 "src/math/exp-neonfma-lut64-p2.c",
1247 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001248 "src/math/expminus-neonfma-lut2048-p1.c",
1249 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001250 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001251 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1252 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1253 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1254 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1255 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1256 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1257 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1258 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1259 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1260 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1261 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1262 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1263 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1264 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1265 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1266 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1267 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1268 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001269 "src/math/sqrt-neonfma-nr1fma.c",
1270 "src/math/sqrt-neonfma-nr2fma.c",
1271 "src/math/sqrt-neonfma-nr3fma.c",
1272 "src/math/sqrt-neonfma-nr2fma1adj.c",
1273 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001274]
1275
1276AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001283 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1284 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1285 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1286 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1287 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1288 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1289 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1290 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1291 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1292 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1293 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1294 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1295 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1296 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1297 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1298 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1299 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1300 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1301 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1302 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001303 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1304 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001305 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1306 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001307 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1308 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001309 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1310 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001311 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1312 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1313 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1314 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1315 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001316 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001334 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1335 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1336 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1337 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1338 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1339 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1340 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1341 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1342 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1343 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1344 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1345 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1346 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1347 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1348 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1349 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1350 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1351 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001352 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1353 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001354 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1355 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1356 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1357 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1358 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1359 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360]
1361
Marat Dukhan8853b822020-05-07 12:19:01 -07001362NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001363 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1364 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1365 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1366 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1367 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1368 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1369 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1370 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001371 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001372 "src/math/roundd-neonv8.c",
1373 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001374 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001375]
1376
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001378 "src/f16-clamp/gen/neonfp16arith-x16.c",
1379 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001380 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1381 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1382 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1383 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1384 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1385 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1386 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1387 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1388 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1389 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1390 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1391 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001392 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1393 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001394 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1395 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001396 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1397 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001398 "src/f16-relu/gen/neonfp16arith-x16.c",
1399 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001400 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1401 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1402 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1403 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1404 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1405 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1406 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1407 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1408 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1409 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1410 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1411 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1412 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1413 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1414 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1415 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1416 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1417 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1418 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1419 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1420 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1421 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1422 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1423 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1424 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1425 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1426 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1427 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001428 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1429 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001430 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1431 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1432 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1433 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1434 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1435 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1436 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1437 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001438 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1439 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1440 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1441 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001442 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1443 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1444 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1445 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1446 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1447 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1448 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1449 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1450 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1451 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1452 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1453 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001454 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1455 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1456 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1457 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1458 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1459 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1460 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1461 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462]
1463
1464SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001465 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1466 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001467 "src/f32-clamp/gen/sse-x4.c",
1468 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001469 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1470 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001471 "src/f32-dwconv-chw/3x3p1-sse.c",
1472 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001473 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1474 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1475 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1476 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1477 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1478 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1479 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1480 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1481 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1482 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1483 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1484 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1487 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1489 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1490 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1491 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1492 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1493 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1494 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1495 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1496 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1497 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1498 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1499 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1500 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001501 "src/f32-hswish/gen/hswish-sse-x4.c",
1502 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001503 "src/f32-ibilinear/gen/sse-c4.c",
1504 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001505 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1506 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1507 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1508 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1509 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1510 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1511 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001512 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1513 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1514 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001515 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001516 "src/f32-prelu/gen/sse-2x4.c",
1517 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001518 "src/f32-relu/gen/sse-x4.c",
1519 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001520 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001521 "src/f32-spmm/gen/4x1-minmax-sse.c",
1522 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001523 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001524 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1525 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1526 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1527 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1528 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1529 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1530 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1531 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001532 "src/f32-vbinary/gen/vmax-sse-x4.c",
1533 "src/f32-vbinary/gen/vmax-sse-x8.c",
1534 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1535 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1536 "src/f32-vbinary/gen/vmin-sse-x4.c",
1537 "src/f32-vbinary/gen/vmin-sse-x8.c",
1538 "src/f32-vbinary/gen/vminc-sse-x4.c",
1539 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001540 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1541 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1542 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1543 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1544 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1545 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1546 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1547 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001548 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1549 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1550 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1551 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001552 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1554 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1555 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001556 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1557 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001558 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1559 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001560 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1561 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001562 "src/f32-vunary/gen/vabs-sse-x4.c",
1563 "src/f32-vunary/gen/vabs-sse-x8.c",
1564 "src/f32-vunary/gen/vneg-sse-x4.c",
1565 "src/f32-vunary/gen/vneg-sse-x8.c",
1566 "src/f32-vunary/gen/vsqr-sse-x4.c",
1567 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001568 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001570 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001571 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001572 "src/math/roundd-sse-addsub.c",
1573 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001574 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001575 "src/math/sqrt-sse-hh1mac.c",
1576 "src/math/sqrt-sse-nr1mac.c",
1577 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001578]
1579
1580SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001581 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1582 "src/f32-argmaxpool/4x-sse2-c4.c",
1583 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001584 "src/f32-prelu/gen/sse2-2x4.c",
1585 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001586 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1587 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1588 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1589 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1590 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1591 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1592 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1593 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1595 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1596 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001598 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001599 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001600 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001601 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001602 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1603 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001604 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1605 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001606 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1607 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1608 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1609 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1610 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1611 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1612 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1613 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001614 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1615 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1616 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1617 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1618 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1619 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1620 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1621 "src/qu8-gemm/4x4c2-minmax-sse2.c",
1622 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001623 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001624 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001625 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001626 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001627 "src/x32-zip/x2-sse2.c",
1628 "src/x32-zip/x3-sse2.c",
1629 "src/x32-zip/x4-sse2.c",
1630 "src/x32-zip/xm-sse2.c",
1631 "src/x8-zip/x2-sse2.c",
1632 "src/x8-zip/x3-sse2.c",
1633 "src/x8-zip/x4-sse2.c",
1634 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001635 "src/math/exp-sse2-p5.c",
1636 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001637 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001638 "src/math/roundd-sse2-cvt.c",
1639 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001640 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001641 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001642 "src/requantization/precise-sse2.c",
1643 "src/requantization/fp32-sse2.c",
1644 "src/requantization/q31-sse2.c",
1645 "src/requantization/gemmlowp-sse2.c",
1646]
1647
1648SSSE3_UKERNELS = [
1649 "src/requantization/precise-ssse3.c",
1650 "src/requantization/q31-ssse3.c",
1651 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001652]
1653
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001654SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001655 "src/f32-prelu/gen/sse41-2x4.c",
1656 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001657 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001658 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001659 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001660 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001661 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1662 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001663 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1664 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001665 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1666 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1667 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1668 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1669 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1670 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1671 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1672 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001673 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001674 "src/math/roundd-sse41.c",
1675 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001676 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001677 "src/requantization/precise-sse4.c",
1678 "src/requantization/q31-sse4.c",
1679 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001680]
1681
Marat Dukhan08c4a432019-10-03 09:29:21 -07001682AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001683 "src/f32-clamp/gen/avx-x8.c",
1684 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001685 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1686 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1687 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1688 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1689 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1690 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1691 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1692 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1693 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1694 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1695 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1696 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1697 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1698 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1699 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1700 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1701 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1702 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1703 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1704 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1705 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1706 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1707 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1708 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1709 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1710 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1711 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1712 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1713 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1714 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001715 "src/f32-hswish/gen/hswish-avx-x8.c",
1716 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001717 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1718 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1719 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1720 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1721 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1722 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1723 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1724 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1725 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001726 "src/f32-prelu/gen/avx-2x8.c",
1727 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001728 "src/f32-relu/gen/avx-x16.c",
1729 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001731 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1732 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1733 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1734 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1735 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1736 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1737 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1738 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001739 "src/f32-vbinary/gen/vmax-avx-x8.c",
1740 "src/f32-vbinary/gen/vmax-avx-x16.c",
1741 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1742 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1743 "src/f32-vbinary/gen/vmin-avx-x8.c",
1744 "src/f32-vbinary/gen/vmin-avx-x16.c",
1745 "src/f32-vbinary/gen/vminc-avx-x8.c",
1746 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001747 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1748 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1749 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1750 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1751 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1752 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1753 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1754 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001755 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1756 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1757 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1758 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001759 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1760 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1761 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1762 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001763 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1764 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001765 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001766 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1767 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1768 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1769 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1770 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1771 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1772 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1773 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001774 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1775 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001776 "src/f32-vunary/gen/vabs-avx-x8.c",
1777 "src/f32-vunary/gen/vabs-avx-x16.c",
1778 "src/f32-vunary/gen/vneg-avx-x8.c",
1779 "src/f32-vunary/gen/vneg-avx-x16.c",
1780 "src/f32-vunary/gen/vsqr-avx-x8.c",
1781 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001782]
1783
Marat Dukhanfda12b82019-11-21 12:27:59 -08001784FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001785 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1786 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1787 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1788 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1789 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1790 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1791 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1792 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1793 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1794 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1795 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1796 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1797 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1798 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1799 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1800 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1801 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1802 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1803 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1804 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1805 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1806 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1807 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1808 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1809 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1824 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001825 "src/f32-hswish/gen/hswish-fma3-x8.c",
1826 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001827 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1828 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1829 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1830 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1831 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1832 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1833 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1834 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1835 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1836 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1837 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1838 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1839 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1840 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001841 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1842 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1843 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1844 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1845 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1846 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1847 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1848 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001849 "src/math/sqrt-fma3-nr1fma.c",
1850 "src/math/sqrt-fma3-nr2fma.c",
1851 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001852]
1853
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001854AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001855 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1856 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1857 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1858 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1859 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1860 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1861 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1862 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1863 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1864 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1865 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1866 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1867 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1868 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1869 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1870 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1871 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1872 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1873 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1874 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1875 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1876 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1877 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1878 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1879 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1880 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1881 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1882 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1883 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1884 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1885 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1886 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1887 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1888 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1889 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1890 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001891 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1892 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1893 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1894 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1895 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1896 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1897 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1898 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1899 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1900 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1901 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1902 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1903 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1920 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001921 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1924 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1925 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1926 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1927 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1928 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1929 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1930 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1931 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1932 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1933 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1934 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1935 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1936 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1937 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1938 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1939 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1940 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1941 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1942 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1943 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1944 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001945 "src/math/exp-avx2-p5.c",
1946 "src/math/exp-avx2-perm-p3.c",
1947 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001948 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001949 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001950 "src/math/sigmoid-avx2-rr2-p5-div.c",
1951 "src/math/sigmoid-avx2-rr1-p5-div.c",
1952 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1953 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1954 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1955 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001956]
1957
Marat Dukhan08c4a432019-10-03 09:29:21 -07001958AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001959 "src/f32-clamp/gen/avx512f-x16.c",
1960 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001961 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1962 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1963 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1964 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1965 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1966 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1967 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1968 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1969 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1970 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1971 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1972 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1973 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1974 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1975 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1976 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1977 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1978 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1979 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1980 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1981 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1982 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1983 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1984 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001985 "src/f32-hswish/gen/hswish-avx512f-x16.c",
1986 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001987 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1988 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1989 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1990 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1991 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1992 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001993 "src/f32-prelu/gen/avx512f-2x16.c",
1994 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001995 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1996 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1997 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1998 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1999 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2000 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2006 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2007 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2008 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2009 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2010 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2011 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2012 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2018 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2019 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2022 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2023 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2024 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2030 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002031 "src/f32-relu/gen/avx512f-x16.c",
2032 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002033 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002034 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2035 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2036 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2037 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2038 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2039 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2040 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2041 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002042 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2043 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2044 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2045 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2046 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2047 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2048 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2049 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002050 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2051 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2052 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2053 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2054 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2055 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2056 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2057 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002058 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2059 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2060 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2061 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002062 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2063 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2064 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2065 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002066 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2067 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002068 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2072 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2073 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2074 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2075 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2076 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2077 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2078 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2079 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2080 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2084 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2085 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2086 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2087 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2088 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2089 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2090 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2091 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2092 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002093 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2094 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2095 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2096 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2097 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2098 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2099 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2100 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002101 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2102 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2103 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2104 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2105 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2106 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2107 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2108 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002109 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2110 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2111 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2112 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2113 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2114 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002115 "src/math/exp-avx512f-p5-scalef.c",
2116 "src/math/exp-avx512f-p5.c",
2117 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002118 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002119 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002120 "src/math/sqrt-avx512f-nr1fma.c",
2121 "src/math/sqrt-avx512f-nr2fma.c",
2122 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002123]
2124
2125AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002126 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002127 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002128 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2129 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002130 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002131 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002132 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002133 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002134 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2135 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002136 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2137 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2138 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2139 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140]
2141
2142AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002143 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002144 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2145 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002146 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002147 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2148 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002149 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2150 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2151 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002152 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002153 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2154 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2155 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002156 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002157 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2158 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002159 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002160 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2161 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2162 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2163 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2164 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2165 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2166 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2167 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2168 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2169 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2170 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2171 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2172 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2173 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2174 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2175 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2176 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2177 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002178 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2179 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002180 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002181 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2182 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2183 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2184 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2185 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2186 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2187 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2188 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2189 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2190 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2191 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2192 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2193 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2194 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2195 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2196 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2197 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2198 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002199 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2200 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2201 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2202 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2203 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2204 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2205 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2206 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2207 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2208 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2209 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2210 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2211 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2212 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2213 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2214 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2215 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2216 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217]
2218
Marat Dukhan1b354632020-03-23 12:50:22 -07002219INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002220 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002221 "src/xnnpack/argmaxpool.h",
2222 "src/xnnpack/avgpool.h",
2223 "src/xnnpack/clamp.h",
2224 "src/xnnpack/common.h",
2225 "src/xnnpack/conv.h",
2226 "src/xnnpack/dwconv.h",
2227 "src/xnnpack/gavgpool.h",
2228 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002229 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002231 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002233 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002234 "src/xnnpack/lut.h",
2235 "src/xnnpack/math.h",
2236 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002237 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/xnnpack/packx.h",
2239 "src/xnnpack/pad.h",
2240 "src/xnnpack/params.h",
2241 "src/xnnpack/pavgpool.h",
2242 "src/xnnpack/ppmm.h",
2243 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002244 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002245 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002246 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002247 "src/xnnpack/rmax.h",
2248 "src/xnnpack/scalar-utils.h",
2249 "src/xnnpack/spmm.h",
2250 "src/xnnpack/unpool.h",
2251 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002252 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002253 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002254 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002255 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002256 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002257 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002259]
2260
2261INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002262 "include/xnnpack.h",
2263 "src/xnnpack/allocator.h",
2264 "src/xnnpack/compute.h",
2265 "src/xnnpack/im2col.h",
2266 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002267 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002268 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269 "src/xnnpack/operator.h",
2270 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002271 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002272 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002273 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002274 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002275]
2276
Marat Dukhan1b354632020-03-23 12:50:22 -07002277ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002278 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002279]
2280
Marat Dukhan1b354632020-03-23 12:50:22 -07002281MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002282 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002283 "include/xnnpack.h",
2284]
2285
Marat Dukhan1b354632020-03-23 12:50:22 -07002286MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002287 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002288 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "src/xnnpack/requantization.h",
2290 "include/xnnpack.h",
2291]
2292
2293OPERATOR_TEST_PARAMS_HDRS = [
2294 "src/xnnpack/params.h",
2295 "src/xnnpack/common.h",
2296]
2297
2298WEIGHTS_PACK_HDRS = [
2299 "src/xnnpack/pack.h",
2300 "src/xnnpack/operator.h",
2301 "src/xnnpack/compute.h",
2302]
2303
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002304LOGGING_COPTS = select({
2305 # No logging in optimized mode
2306 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2307 # Full logging in debug mode
2308 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2309 # Error-only logging in default (fastbuild) mode
2310 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2311})
2312
Marat Dukhan3b59de22020-06-03 20:15:19 -07002313LOGGING_SRCS = select({
2314 # No logging in optimized mode
2315 ":optimized_build": [],
2316 "//conditions:default": [
2317 "src/operator-strings.c",
2318 "src/subgraph-strings.c",
2319 ],
2320})
2321
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002322LOGGING_HDRS = [
2323 "src/xnnpack/log.h",
2324]
2325
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002327 name = "tables",
2328 srcs = TABLE_SRCS,
2329 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002330 gcc_copts = xnnpack_gcc_std_copts(),
2331 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002332)
2333
2334xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002335 name = "scalar_ukernels",
2336 srcs = SCALAR_UKERNELS,
2337 hdrs = INTERNAL_HDRS,
2338 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002339 gcc_copts = xnnpack_gcc_std_copts(),
2340 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002342 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343 "@FP16",
2344 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002345 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 ],
2347)
2348
2349xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002350 name = "scalar_ukernels_test_mode",
2351 srcs = SCALAR_UKERNELS,
2352 hdrs = INTERNAL_HDRS,
2353 aarch32_copts = ["-marm"],
2354 copts = [
2355 "-UNDEBUG",
2356 "-DXNN_TEST_MODE=1",
2357 ],
2358 gcc_copts = xnnpack_gcc_std_copts(),
2359 msvc_copts = xnnpack_msvc_std_copts(),
2360 deps = [
2361 ":tables",
2362 "@FP16",
2363 "@FXdiv",
2364 "@pthreadpool",
2365 ],
2366)
2367
2368xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002369 name = "wasm_ukernels",
2370 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002371 gcc_copts = xnnpack_gcc_std_copts(),
2372 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002373 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002374 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002375 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002376 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002377 "@FP16",
2378 "@FXdiv",
2379 "@pthreadpool",
2380 ],
2381)
2382
2383xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002384 name = "wasm_ukernels_test_mode",
2385 hdrs = INTERNAL_HDRS,
2386 copts = [
2387 "-UNDEBUG",
2388 "-DXNN_TEST_MODE=1",
2389 ],
2390 gcc_copts = xnnpack_gcc_std_copts(),
2391 msvc_copts = xnnpack_msvc_std_copts(),
2392 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002393 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002394 deps = [
2395 ":tables",
2396 "@FP16",
2397 "@FXdiv",
2398 "@pthreadpool",
2399 ],
2400)
2401
2402xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002403 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 hdrs = INTERNAL_HDRS,
2405 aarch32_copts = [
2406 "-marm",
2407 "-mfpu=neon",
2408 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002409 gcc_copts = xnnpack_gcc_std_copts(),
2410 gcc_x86_copts = ["-msse2"],
2411 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002412 optimized_copts = [
2413 "-O3",
2414 "-ffast-math",
2415 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002416 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002418 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002419 "@FP16",
2420 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002421 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422 ],
2423)
2424
2425xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002426 name = "psimd_fastmath_ukernels_test_mode",
2427 hdrs = INTERNAL_HDRS,
2428 aarch32_copts = [
2429 "-marm",
2430 "-mfpu=neon",
2431 ],
2432 copts = [
2433 "-UNDEBUG",
2434 "-DXNN_TEST_MODE=1",
2435 ],
2436 gcc_copts = xnnpack_gcc_std_copts(),
2437 gcc_x86_copts = ["-msse2"],
2438 msvc_copts = xnnpack_msvc_std_copts(),
2439 optimized_copts = [
2440 "-O3",
2441 "-ffast-math",
2442 ],
2443 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2444 deps = [
2445 ":tables",
2446 "@FP16",
2447 "@psimd",
2448 "@pthreadpool",
2449 ],
2450)
2451
2452xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002453 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002454 hdrs = INTERNAL_HDRS,
2455 aarch32_copts = [
2456 "-marm",
2457 "-mfpu=neon",
2458 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002459 gcc_copts = xnnpack_gcc_std_copts(),
2460 gcc_x86_copts = ["-msse2"],
2461 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002462 optimized_copts = [
2463 "-O3",
2464 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002465 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002466 deps = [
2467 ":tables",
2468 "@FP16",
2469 "@psimd",
2470 "@pthreadpool",
2471 ],
2472)
2473
2474xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002475 name = "psimd_accmath_ukernels_test_mode",
2476 hdrs = INTERNAL_HDRS,
2477 aarch32_copts = [
2478 "-marm",
2479 "-mfpu=neon",
2480 ],
2481 copts = [
2482 "-UNDEBUG",
2483 "-DXNN_TEST_MODE=1",
2484 ],
2485 gcc_copts = xnnpack_gcc_std_copts(),
2486 gcc_x86_copts = ["-msse2"],
2487 msvc_copts = xnnpack_msvc_std_copts(),
2488 optimized_copts = [
2489 "-O3",
2490 ],
2491 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2492 deps = [
2493 ":tables",
2494 "@FP16",
2495 "@psimd",
2496 "@pthreadpool",
2497 ],
2498)
2499
2500xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002501 name = "neon_ukernels",
2502 hdrs = INTERNAL_HDRS,
2503 aarch32_copts = [
2504 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002505 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002506 "-mfpu=neon",
2507 ],
2508 aarch32_srcs = NEON_UKERNELS,
2509 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002510 gcc_copts = xnnpack_gcc_std_copts(),
2511 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002512 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002513 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002514 "@FP16",
2515 "@pthreadpool",
2516 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002517)
2518
2519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002520 name = "neon_ukernels_test_mode",
2521 hdrs = INTERNAL_HDRS,
2522 aarch32_copts = [
2523 "-marm",
2524 "-march=armv7-a",
2525 "-mfpu=neon",
2526 ],
2527 aarch32_srcs = NEON_UKERNELS,
2528 aarch64_srcs = NEON_UKERNELS,
2529 copts = [
2530 "-UNDEBUG",
2531 "-DXNN_TEST_MODE=1",
2532 ],
2533 gcc_copts = xnnpack_gcc_std_copts(),
2534 msvc_copts = xnnpack_msvc_std_copts(),
2535 deps = [
2536 ":tables",
2537 "@FP16",
2538 "@pthreadpool",
2539 ],
2540)
2541
2542xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002543 name = "neonfma_ukernels",
2544 hdrs = INTERNAL_HDRS,
2545 aarch32_copts = [
2546 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002547 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002548 "-mfpu=neon-vfpv4",
2549 ],
2550 aarch32_srcs = NEONFMA_UKERNELS,
2551 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002552 apple_aarch32_copts = [
2553 "-mcpu=swift",
2554 "-mtune=generic",
2555 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002556 gcc_copts = xnnpack_gcc_std_copts(),
2557 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002558 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002559 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002560 "@FP16",
2561 "@pthreadpool",
2562 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002563)
2564
2565xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002566 name = "neonfma_ukernels_test_mode",
2567 hdrs = INTERNAL_HDRS,
2568 aarch32_copts = [
2569 "-marm",
2570 "-march=armv7-a",
2571 "-mfpu=neon-vfpv4",
2572 ],
2573 aarch32_srcs = NEONFMA_UKERNELS,
2574 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002575 apple_aarch32_copts = [
2576 "-mcpu=swift",
2577 "-mtune=generic",
2578 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002579 copts = [
2580 "-UNDEBUG",
2581 "-DXNN_TEST_MODE=1",
2582 ],
2583 gcc_copts = xnnpack_gcc_std_copts(),
2584 msvc_copts = xnnpack_msvc_std_copts(),
2585 deps = [
2586 ":tables",
2587 "@FP16",
2588 "@pthreadpool",
2589 ],
2590)
2591
2592xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002593 name = "neonv8_ukernels",
2594 hdrs = INTERNAL_HDRS,
2595 aarch32_copts = [
2596 "-marm",
2597 "-march=armv8-a",
2598 "-mfpu=neon-fp-armv8",
2599 ],
2600 aarch32_srcs = NEONV8_UKERNELS,
2601 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002602 apple_aarch32_copts = [
2603 "-mcpu=cyclone",
2604 "-mtune=generic",
2605 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002606 gcc_copts = xnnpack_gcc_std_copts(),
2607 msvc_copts = xnnpack_msvc_std_copts(),
2608 deps = [
2609 ":tables",
2610 "@FP16",
2611 "@pthreadpool",
2612 ],
2613)
2614
2615xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002616 name = "neonv8_ukernels_test_mode",
2617 hdrs = INTERNAL_HDRS,
2618 aarch32_copts = [
2619 "-marm",
2620 "-march=armv8-a",
2621 "-mfpu=neon-fp-armv8",
2622 ],
2623 aarch32_srcs = NEONV8_UKERNELS,
2624 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002625 apple_aarch32_copts = [
2626 "-mcpu=cyclone",
2627 "-mtune=generic",
2628 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002629 copts = [
2630 "-UNDEBUG",
2631 "-DXNN_TEST_MODE=1",
2632 ],
2633 gcc_copts = xnnpack_gcc_std_copts(),
2634 msvc_copts = xnnpack_msvc_std_copts(),
2635 deps = [
2636 ":tables",
2637 "@FP16",
2638 "@pthreadpool",
2639 ],
2640)
2641
2642xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002643 name = "neonfp16arith_ukernels",
2644 hdrs = INTERNAL_HDRS,
2645 aarch64_copts = ["-march=armv8.2-a+fp16"],
2646 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002647 gcc_copts = xnnpack_gcc_std_copts(),
2648 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002649 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002650 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002651 "@FP16",
2652 "@pthreadpool",
2653 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002654)
2655
2656xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002657 name = "neonfp16arith_ukernels_test_mode",
2658 hdrs = INTERNAL_HDRS,
2659 aarch64_copts = ["-march=armv8.2-a+fp16"],
2660 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2661 copts = [
2662 "-UNDEBUG",
2663 "-DXNN_TEST_MODE=1",
2664 ],
2665 gcc_copts = xnnpack_gcc_std_copts(),
2666 msvc_copts = xnnpack_msvc_std_copts(),
2667 deps = [
2668 ":tables",
2669 "@FP16",
2670 "@pthreadpool",
2671 ],
2672)
2673
2674xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 name = "sse2_ukernels",
2676 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002677 gcc_copts = xnnpack_gcc_std_copts(),
2678 gcc_x86_copts = ["-msse2"],
2679 msvc_copts = xnnpack_msvc_std_copts(),
2680 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002681 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002682 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002683 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002684 "@FP16",
2685 "@pthreadpool",
2686 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687)
2688
2689xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002690 name = "sse2_ukernels_test_mode",
2691 hdrs = INTERNAL_HDRS,
2692 copts = [
2693 "-UNDEBUG",
2694 "-DXNN_TEST_MODE=1",
2695 ],
2696 gcc_copts = xnnpack_gcc_std_copts(),
2697 gcc_x86_copts = ["-msse2"],
2698 msvc_copts = xnnpack_msvc_std_copts(),
2699 msvc_x86_32_copts = ["/arch:SSE2"],
2700 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2701 deps = [
2702 ":tables",
2703 "@FP16",
2704 "@pthreadpool",
2705 ],
2706)
2707
2708xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002709 name = "ssse3_ukernels",
2710 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002711 gcc_copts = xnnpack_gcc_std_copts(),
2712 gcc_x86_copts = ["-mssse3"],
2713 msvc_copts = xnnpack_msvc_std_copts(),
2714 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002715 x86_srcs = SSSE3_UKERNELS,
2716 deps = [
2717 ":tables",
2718 "@FP16",
2719 "@pthreadpool",
2720 ],
2721)
2722
2723xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002724 name = "ssse3_ukernels_test_mode",
2725 hdrs = INTERNAL_HDRS,
2726 copts = [
2727 "-UNDEBUG",
2728 "-DXNN_TEST_MODE=1",
2729 ],
2730 gcc_copts = xnnpack_gcc_std_copts(),
2731 gcc_x86_copts = ["-mssse3"],
2732 msvc_copts = xnnpack_msvc_std_copts(),
2733 msvc_x86_32_copts = ["/arch:SSE2"],
2734 x86_srcs = SSSE3_UKERNELS,
2735 deps = [
2736 ":tables",
2737 "@FP16",
2738 "@pthreadpool",
2739 ],
2740)
2741
2742xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002743 name = "sse41_ukernels",
2744 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002745 gcc_copts = xnnpack_gcc_std_copts(),
2746 gcc_x86_copts = ["-msse4.1"],
2747 msvc_copts = xnnpack_msvc_std_copts(),
2748 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002749 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002750 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002751 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002752 "@FP16",
2753 "@pthreadpool",
2754 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002755)
2756
2757xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002758 name = "sse41_ukernels_test_mode",
2759 hdrs = INTERNAL_HDRS,
2760 copts = [
2761 "-UNDEBUG",
2762 "-DXNN_TEST_MODE=1",
2763 ],
2764 gcc_copts = xnnpack_gcc_std_copts(),
2765 gcc_x86_copts = ["-msse4.1"],
2766 msvc_copts = xnnpack_msvc_std_copts(),
2767 msvc_x86_32_copts = ["/arch:SSE2"],
2768 x86_srcs = SSE41_UKERNELS,
2769 deps = [
2770 ":tables",
2771 "@FP16",
2772 "@pthreadpool",
2773 ],
2774)
2775
2776xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002777 name = "avx_ukernels",
2778 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002779 gcc_copts = xnnpack_gcc_std_copts(),
2780 gcc_x86_copts = ["-mavx"],
2781 msvc_copts = xnnpack_msvc_std_copts(),
2782 msvc_x86_32_copts = ["/arch:AVX"],
2783 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002784 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002785 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002786 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002787 "@FP16",
2788 "@pthreadpool",
2789 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790)
2791
2792xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002793 name = "avx_ukernels_test_mode",
2794 hdrs = INTERNAL_HDRS,
2795 copts = [
2796 "-UNDEBUG",
2797 "-DXNN_TEST_MODE=1",
2798 ],
2799 gcc_copts = xnnpack_gcc_std_copts(),
2800 gcc_x86_copts = ["-mavx"],
2801 msvc_copts = xnnpack_msvc_std_copts(),
2802 msvc_x86_32_copts = ["/arch:AVX"],
2803 msvc_x86_64_copts = ["/arch:AVX"],
2804 x86_srcs = AVX_UKERNELS,
2805 deps = [
2806 ":tables",
2807 "@FP16",
2808 "@pthreadpool",
2809 ],
2810)
2811
2812xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002813 name = "fma3_ukernels",
2814 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002815 gcc_copts = xnnpack_gcc_std_copts(),
2816 gcc_x86_copts = ["-mfma"],
2817 msvc_copts = xnnpack_msvc_std_copts(),
2818 msvc_x86_32_copts = ["/arch:AVX"],
2819 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002820 x86_srcs = FMA3_UKERNELS,
2821 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002822 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002823 "@FP16",
2824 "@pthreadpool",
2825 ],
2826)
2827
2828xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002829 name = "fma3_ukernels_test_mode",
2830 hdrs = INTERNAL_HDRS,
2831 copts = [
2832 "-UNDEBUG",
2833 "-DXNN_TEST_MODE=1",
2834 ],
2835 gcc_copts = xnnpack_gcc_std_copts(),
2836 gcc_x86_copts = ["-mfma"],
2837 msvc_copts = xnnpack_msvc_std_copts(),
2838 msvc_x86_32_copts = ["/arch:AVX"],
2839 msvc_x86_64_copts = ["/arch:AVX"],
2840 x86_srcs = FMA3_UKERNELS,
2841 deps = [
2842 ":tables",
2843 "@FP16",
2844 "@pthreadpool",
2845 ],
2846)
2847
2848xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002849 name = "avx2_ukernels",
2850 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002851 gcc_copts = xnnpack_gcc_std_copts(),
2852 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002853 "-mfma",
2854 "-mavx2",
2855 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002856 msvc_copts = xnnpack_msvc_std_copts(),
2857 msvc_x86_32_copts = ["/arch:AVX2"],
2858 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002859 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002860 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002861 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002862 "@FP16",
2863 "@pthreadpool",
2864 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002865)
2866
2867xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002868 name = "avx2_ukernels_test_mode",
2869 hdrs = INTERNAL_HDRS,
2870 copts = [
2871 "-UNDEBUG",
2872 "-DXNN_TEST_MODE=1",
2873 ],
2874 gcc_copts = xnnpack_gcc_std_copts(),
2875 gcc_x86_copts = [
2876 "-mfma",
2877 "-mavx2",
2878 ],
2879 msvc_copts = xnnpack_msvc_std_copts(),
2880 msvc_x86_32_copts = ["/arch:AVX2"],
2881 msvc_x86_64_copts = ["/arch:AVX2"],
2882 x86_srcs = AVX2_UKERNELS,
2883 deps = [
2884 ":tables",
2885 "@FP16",
2886 "@pthreadpool",
2887 ],
2888)
2889
2890xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 name = "avx512f_ukernels",
2892 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002893 gcc_copts = xnnpack_gcc_std_copts(),
2894 gcc_x86_copts = ["-mavx512f"],
2895 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 msvc_x86_32_copts = ["/arch:AVX512"],
2898 msvc_x86_64_copts = ["/arch:AVX512"],
2899 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002901 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002902 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002903 "@FP16",
2904 "@pthreadpool",
2905 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906)
2907
2908xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002909 name = "avx512f_ukernels_test_mode",
2910 hdrs = INTERNAL_HDRS,
2911 copts = [
2912 "-UNDEBUG",
2913 "-DXNN_TEST_MODE=1",
2914 ],
2915 gcc_copts = xnnpack_gcc_std_copts(),
2916 gcc_x86_copts = ["-mavx512f"],
2917 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2918 msvc_copts = xnnpack_msvc_std_copts(),
2919 msvc_x86_32_copts = ["/arch:AVX512"],
2920 msvc_x86_64_copts = ["/arch:AVX512"],
2921 msys_copts = ["-fno-asynchronous-unwind-tables"],
2922 x86_srcs = AVX512F_UKERNELS,
2923 deps = [
2924 ":tables",
2925 "@FP16",
2926 "@pthreadpool",
2927 ],
2928)
2929
2930xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 name = "asm_ukernels",
2932 hdrs = ["src/xnnpack/assembly.h"],
2933 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002934 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002935 aarch64_srcs = AARCH64_ASM_UKERNELS,
2936)
2937
Marat Dukhan3b59de22020-06-03 20:15:19 -07002938xnnpack_cc_library(
2939 name = "logging_utils",
2940 srcs = LOGGING_SRCS,
2941 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2942 copts = LOGGING_COPTS + [
2943 "-Isrc",
2944 "-Iinclude",
2945 ] + select({
2946 ":debug_build": [],
2947 "//conditions:default": xnnpack_min_size_copts(),
2948 }),
2949 gcc_copts = xnnpack_gcc_std_copts(),
2950 msvc_copts = xnnpack_msvc_std_copts(),
2951 visibility = xnnpack_visibility(),
2952 deps = [
2953 "@FP16",
2954 "@clog",
2955 "@pthreadpool",
2956 ],
2957)
2958
Marat Dukhan08c4a432019-10-03 09:29:21 -07002959xnnpack_aggregate_library(
2960 name = "ukernels",
2961 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 ":neon_ukernels",
2963 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002964 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002965 ":asm_ukernels",
2966 ],
2967 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 ":neon_ukernels",
2969 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002970 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 ":neonfp16arith_ukernels",
2972 ":asm_ukernels",
2973 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002974 generic_deps = [
2975 ":scalar_ukernels",
2976 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002977 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002978 ":psimd_fastmath_ukernels",
2979 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002980 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002981 wasm_deps = [
2982 ":wasm_ukernels",
2983 ],
2984 wasmsimd_deps = [
2985 ":wasm_ukernels",
2986 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002989 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002990 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002992 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002993 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 ":avx512f_ukernels",
2995 ],
2996)
2997
Marat Dukhan33fcf782020-05-24 14:27:15 -07002998xnnpack_aggregate_library(
2999 name = "ukernels_test_mode",
3000 aarch32_deps = [
3001 ":neon_ukernels_test_mode",
3002 ":neonfma_ukernels_test_mode",
3003 ":neonv8_ukernels_test_mode",
3004 ":asm_ukernels",
3005 ],
3006 aarch64_deps = [
3007 ":neon_ukernels_test_mode",
3008 ":neonfma_ukernels_test_mode",
3009 ":neonv8_ukernels_test_mode",
3010 ":neonfp16arith_ukernels_test_mode",
3011 ":asm_ukernels",
3012 ],
3013 generic_deps = [
3014 ":scalar_ukernels_test_mode",
3015 ],
3016 psimd_deps = [
3017 ":psimd_fastmath_ukernels_test_mode",
3018 ":psimd_accmath_ukernels_test_mode",
3019 ],
3020 wasm_deps = [
3021 ":wasm_ukernels_test_mode",
3022 ],
3023 wasmsimd_deps = [
3024 ":wasm_ukernels_test_mode",
3025 ],
3026 x86_deps = [
3027 ":sse2_ukernels_test_mode",
3028 ":ssse3_ukernels_test_mode",
3029 ":sse41_ukernels_test_mode",
3030 ":avx_ukernels_test_mode",
3031 ":fma3_ukernels_test_mode",
3032 ":avx2_ukernels_test_mode",
3033 ":avx512f_ukernels_test_mode",
3034 ],
3035)
3036
Marat Dukhan86df3842020-06-17 14:03:31 -07003037# This target must be compatible with Android NDK r17 and compile with Clang 6.
3038# It is similar to :ukernels target, but excludes microkernels which require
3039# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003040xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003041 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003042 aarch32_deps = [
3043 ":neon_ukernels",
3044 ":neonfma_ukernels",
3045 ":neonv8_ukernels",
3046 ":asm_ukernels",
3047 ],
3048 aarch64_deps = [
3049 ":neon_ukernels",
3050 ":neonfma_ukernels",
3051 ":neonv8_ukernels",
3052 ":asm_ukernels",
3053 ],
3054 generic_deps = [
3055 ":scalar_ukernels",
3056 ],
3057 psimd_deps = [
3058 ":psimd_fastmath_ukernels",
3059 ":psimd_accmath_ukernels",
3060 ],
3061 wasm_deps = [
3062 ":wasm_ukernels",
3063 ],
3064 wasmsimd_deps = [
3065 ":wasm_ukernels",
3066 ],
3067 x86_deps = [
3068 ":sse2_ukernels",
3069 ":ssse3_ukernels",
3070 ":sse41_ukernels",
3071 ":avx_ukernels",
3072 ":fma3_ukernels",
3073 ":avx2_ukernels",
3074 ":avx512f_ukernels",
3075 ],
3076)
3077
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078xnnpack_cc_library(
3079 name = "im2col",
3080 srcs = ["src/im2col.c"],
3081 hdrs = [
3082 "src/xnnpack/common.h",
3083 "src/xnnpack/im2col.h",
3084 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003085 gcc_copts = xnnpack_gcc_std_copts(),
3086 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087)
3088
3089xnnpack_cc_library(
3090 name = "indirection",
3091 srcs = ["src/indirection.c"],
3092 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003093 gcc_copts = xnnpack_gcc_std_copts(),
3094 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 deps = [
3096 "@FP16",
3097 "@FXdiv",
3098 "@pthreadpool",
3099 ],
3100)
3101
3102xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003103 name = "indirection_test_mode",
3104 srcs = ["src/indirection.c"],
3105 hdrs = INTERNAL_HDRS,
3106 copts = [
3107 "-UNDEBUG",
3108 "-DXNN_TEST_MODE=1",
3109 ],
3110 gcc_copts = xnnpack_gcc_std_copts(),
3111 msvc_copts = xnnpack_msvc_std_copts(),
3112 deps = [
3113 "@FP16",
3114 "@FXdiv",
3115 "@pthreadpool",
3116 ],
3117)
3118
3119xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003120 name = "packing",
3121 srcs = ["src/packing.c"],
3122 hdrs = INTERNAL_HDRS,
3123 gcc_copts = xnnpack_gcc_std_copts(),
3124 msvc_copts = xnnpack_msvc_std_copts(),
3125 deps = [
3126 "@FP16",
3127 "@FXdiv",
3128 "@pthreadpool",
3129 ],
3130)
3131
3132xnnpack_cc_library(
3133 name = "packing_test_mode",
3134 srcs = ["src/packing.c"],
3135 hdrs = INTERNAL_HDRS,
3136 copts = [
3137 "-UNDEBUG",
3138 "-DXNN_TEST_MODE=1",
3139 ],
3140 gcc_copts = xnnpack_gcc_std_copts(),
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 deps = [
3143 "@FP16",
3144 "@FXdiv",
3145 "@pthreadpool",
3146 ],
3147)
3148
3149xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 name = "operator_run",
3151 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003152 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003153 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003154 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3155 "//conditions:default": [],
3156 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003157 gcc_copts = xnnpack_gcc_std_copts(),
3158 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003160 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 "@FP16",
3162 "@FXdiv",
3163 "@clog",
3164 "@pthreadpool",
3165 ],
3166)
3167
Chao Mei6ddfc602020-05-13 22:29:36 -07003168xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003169 name = "operator_run_test_mode",
3170 srcs = ["src/operator-run.c"],
3171 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3172 copts = LOGGING_COPTS + [
3173 "-UNDEBUG",
3174 "-DXNN_TEST_MODE=1",
3175 ] + select({
3176 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3177 "//conditions:default": [],
3178 }),
3179 gcc_copts = xnnpack_gcc_std_copts(),
3180 msvc_copts = xnnpack_msvc_std_copts(),
3181 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003182 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003183 "@FP16",
3184 "@FXdiv",
3185 "@clog",
3186 "@pthreadpool",
3187 ],
3188)
3189
3190xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003191 name = "memory_planner",
3192 srcs = ["src/memory-planner.c"],
3193 hdrs = INTERNAL_HDRS,
3194 defines = select({
3195 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3196 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3197 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3198 }),
3199 gcc_copts = xnnpack_gcc_std_copts(),
3200 msvc_copts = xnnpack_msvc_std_copts(),
3201 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003202 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003203 "@pthreadpool",
3204 ],
3205)
3206
Marat Dukhan33fcf782020-05-24 14:27:15 -07003207xnnpack_cc_library(
3208 name = "memory_planner_test_mode",
3209 srcs = ["src/memory-planner.c"],
3210 hdrs = INTERNAL_HDRS,
3211 copts = [
3212 "-UNDEBUG",
3213 "-DXNN_TEST_MODE=1",
3214 ],
3215 defines = select({
3216 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3217 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3218 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3219 }),
3220 gcc_copts = xnnpack_gcc_std_copts(),
3221 msvc_copts = xnnpack_msvc_std_copts(),
3222 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003223 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003224 "@pthreadpool",
3225 ],
3226)
3227
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228cc_library(
3229 name = "enable_assembly",
3230 defines = select({
3231 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3232 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003233 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 }),
3235)
3236
Marat Dukhan9de90e02020-06-18 16:04:12 -07003237cc_library(
3238 name = "enable_sparse",
3239 defines = select({
3240 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3241 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003242 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003243 }),
3244)
3245
Marat Dukhancf056b22019-10-07 10:26:29 -07003246xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247 name = "operators",
3248 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003249 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003251 ],
3252 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003253 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 "-Isrc",
3255 "-Iinclude",
3256 ] + select({
3257 ":debug_build": [],
3258 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003259 }) + select({
3260 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3261 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003263 gcc_copts = xnnpack_gcc_std_copts(),
3264 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003267 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003268 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003269 "@FP16",
3270 "@FXdiv",
3271 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003273 ],
3274)
3275
Marat Dukhan10a38082020-04-17 03:58:35 -07003276xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003277 name = "operators_test_mode",
3278 srcs = OPERATOR_SRCS + [
3279 "src/memory.c",
3280 "src/operator-delete.c",
3281 ],
3282 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3283 copts = LOGGING_COPTS + [
3284 "-Isrc",
3285 "-Iinclude",
3286 "-UNDEBUG",
3287 "-DXNN_TEST_MODE=1",
3288 ] + select({
3289 ":debug_build": [],
3290 "//conditions:default": xnnpack_min_size_copts(),
3291 }) + select({
3292 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3293 "//conditions:default": [],
3294 }),
3295 gcc_copts = xnnpack_gcc_std_copts(),
3296 msvc_copts = xnnpack_msvc_std_copts(),
3297 deps = [
3298 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003299 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003300 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003301 "@FP16",
3302 "@FXdiv",
3303 "@clog",
3304 "@pthreadpool",
3305 ],
3306)
3307
3308xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003309 name = "XNNPACK",
3310 srcs = [
3311 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003312 "src/runtime.c",
3313 "src/subgraph.c",
3314 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003315 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003316 hdrs = ["include/xnnpack.h"],
3317 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003318 "-Isrc",
3319 "-Iinclude",
3320 ] + select({
3321 ":debug_build": [],
3322 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003323 }) + select({
3324 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3325 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003326 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003327 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003328 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003329 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003330 visibility = xnnpack_visibility(),
3331 deps = [
3332 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003333 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003334 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003335 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003336 ":operator_run",
3337 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003338 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003339 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003340 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003341 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003342 ] + select({
3343 ":emscripten": [],
3344 "//conditions:default": ["@cpuinfo"],
3345 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003346)
3347
Marat Dukhan10a38082020-04-17 03:58:35 -07003348xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003349 name = "XNNPACK_test_mode",
3350 srcs = [
3351 "src/init.c",
3352 "src/runtime.c",
3353 "src/subgraph.c",
3354 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003355 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003356 hdrs = ["include/xnnpack.h"],
3357 copts = LOGGING_COPTS + [
3358 "-Isrc",
3359 "-Iinclude",
3360 "-UNDEBUG",
3361 "-DXNN_TEST_MODE=1",
3362 ] + select({
3363 ":debug_build": [],
3364 "//conditions:default": xnnpack_min_size_copts(),
3365 }) + select({
3366 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3367 "//conditions:default": [],
3368 }),
3369 gcc_copts = xnnpack_gcc_std_copts(),
3370 includes = ["include"],
3371 msvc_copts = xnnpack_msvc_std_copts(),
3372 visibility = xnnpack_visibility(),
3373 deps = [
3374 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003375 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003376 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003377 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003378 ":operator_run_test_mode",
3379 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003380 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003381 "@clog",
3382 "@FP16",
3383 "@pthreadpool",
3384 ] + select({
3385 ":emscripten": [],
3386 "//conditions:default": ["@cpuinfo"],
3387 }),
3388)
3389
3390xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003391 name = "xnnpack_f32",
3392 srcs = [
3393 "src/init.c",
3394 "src/runtime.c",
3395 "src/subgraph.c",
3396 "src/tensor.c",
3397 ] + SUBGRAPH_SRCS,
3398 hdrs = ["include/xnnpack.h"],
3399 copts = LOGGING_COPTS + [
3400 "-Isrc",
3401 "-Iinclude",
3402 ] + select({
3403 ":debug_build": [],
3404 "//conditions:default": xnnpack_min_size_copts(),
3405 }) + select({
3406 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3407 "//conditions:default": [],
3408 }),
3409 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003410 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003411 "XNN_NO_U8_OPERATORS",
3412 "XNN_NO_X8_OPERATORS",
3413 "XNN_NO_F16_OPERATORS",
3414 "XNN_NO_X16_OPERATORS",
3415 ],
3416 gcc_copts = xnnpack_gcc_std_copts(),
3417 includes = ["include"],
3418 msvc_copts = xnnpack_msvc_std_copts(),
3419 visibility = xnnpack_visibility(),
3420 deps = [
3421 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003422 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003423 ":logging_utils",
3424 ":memory_planner",
3425 ":operator_run",
3426 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003427 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003428 "@clog",
3429 "@FP16",
3430 "@pthreadpool",
3431 ] + select({
3432 ":emscripten": [],
3433 "//conditions:default": ["@cpuinfo"],
3434 }),
3435)
3436
3437xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003438 name = "xnnpack_operators_nhwc_f32",
3439 srcs = [
3440 "src/init.c",
3441 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003442 hdrs = ["include/xnnpack.h"],
3443 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003444 "-Isrc",
3445 "-Iinclude",
3446 ] + select({
3447 ":debug_build": [],
3448 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003449 }) + select({
3450 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3451 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003452 }),
3453 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003454 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003455 "XNN_NO_U8_OPERATORS",
3456 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003457 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003458 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003459 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003460 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003461 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003462 visibility = xnnpack_visibility(),
3463 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003464 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003465 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466 ":operator_run",
3467 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003468 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003469 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003470 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003471 ] + select({
3472 ":emscripten": [],
3473 "//conditions:default": ["@cpuinfo"],
3474 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475)
3476
Marat Dukhancf056b22019-10-07 10:26:29 -07003477xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 name = "bench_utils",
3479 srcs = ["bench/utils.cc"],
3480 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003481 deps = [
3482 "@com_google_benchmark//:benchmark",
3483 "@cpuinfo",
3484 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485)
3486
Frank Barchard7e955972019-10-11 10:34:25 -07003487######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003488
3489xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003490 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491 srcs = [
3492 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003493 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003494 "src/xnnpack/AlignedAllocator.h",
3495 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003496 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003497 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003498)
3499
3500xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003501 name = "f16_igemm_bench",
3502 srcs = [
3503 "bench/f16-igemm.cc",
3504 "bench/conv.h",
3505 "bench/google/conv.h",
3506 "src/xnnpack/AlignedAllocator.h",
3507 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003508 deps = MICROKERNEL_BENCHMARK_DEPS + [
3509 ":indirection",
3510 ":packing",
3511 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003512)
3513
3514xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 name = "f16_gemm_bench",
3516 srcs = [
3517 "bench/f16-gemm.cc",
3518 "bench/gemm.h",
3519 "src/xnnpack/AlignedAllocator.h",
3520 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003521 deps = MICROKERNEL_BENCHMARK_DEPS + [
3522 ":packing",
3523 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524)
3525
3526xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003527 name = "f16_spmm_bench",
3528 srcs = [
3529 "bench/f16-spmm.cc",
3530 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003531 "src/xnnpack/AlignedAllocator.h",
3532 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003533 deps = MICROKERNEL_BENCHMARK_DEPS,
3534)
3535
3536xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003537 name = "f32_igemm_bench",
3538 srcs = [
3539 "bench/f32-igemm.cc",
3540 "bench/conv.h",
3541 "src/xnnpack/AlignedAllocator.h",
3542 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003543 deps = MICROKERNEL_BENCHMARK_DEPS + [
3544 ":indirection",
3545 ":packing",
3546 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003547)
3548
3549xnnpack_benchmark(
3550 name = "f32_conv_hwc_bench",
3551 srcs = [
3552 "bench/f32-conv-hwc.cc",
3553 "bench/dconv.h",
3554 "src/xnnpack/AlignedAllocator.h",
3555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003556 deps = MICROKERNEL_BENCHMARK_DEPS + [
3557 ":packing",
3558 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559)
3560
3561xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003562 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003563 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003564 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003565 "bench/dconv.h",
3566 "src/xnnpack/AlignedAllocator.h",
3567 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003568 deps = MICROKERNEL_BENCHMARK_DEPS + [
3569 ":packing",
3570 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003571)
3572
3573xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003574 name = "f16_dwconv_bench",
3575 srcs = [
3576 "bench/f16-dwconv.cc",
3577 "bench/dwconv.h",
3578 "bench/google/dwconv.h",
3579 "src/xnnpack/AlignedAllocator.h",
3580 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003581 deps = MICROKERNEL_BENCHMARK_DEPS + [
3582 ":indirection",
3583 ":packing",
3584 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003585)
3586
3587xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 name = "f32_dwconv_bench",
3589 srcs = [
3590 "bench/f32-dwconv.cc",
3591 "bench/dwconv.h",
3592 "src/xnnpack/AlignedAllocator.h",
3593 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003594 deps = MICROKERNEL_BENCHMARK_DEPS + [
3595 ":indirection",
3596 ":packing",
3597 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598)
3599
3600xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003601 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003603 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604 "bench/dwconv.h",
3605 "src/xnnpack/AlignedAllocator.h",
3606 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003607 deps = MICROKERNEL_BENCHMARK_DEPS + [
3608 ":indirection",
3609 ":packing",
3610 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611)
3612
3613xnnpack_benchmark(
3614 name = "f32_gemm_bench",
3615 srcs = [
3616 "bench/f32-gemm.cc",
3617 "bench/gemm.h",
3618 "src/xnnpack/AlignedAllocator.h",
3619 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003620 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003621 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003622)
3623
3624xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003625 name = "f32_hswish_bench",
3626 srcs = [
3627 "bench/f32-hswish.cc",
3628 "src/xnnpack/AlignedAllocator.h",
3629 ] + MICROKERNEL_BENCHMARK_HDRS,
3630 deps = MICROKERNEL_BENCHMARK_DEPS,
3631)
3632
3633xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003634 name = "f32_raddexpminusmax_bench",
3635 srcs = [
3636 "bench/f32-raddexpminusmax.cc",
3637 "src/xnnpack/AlignedAllocator.h",
3638 ] + MICROKERNEL_BENCHMARK_HDRS,
3639 deps = MICROKERNEL_BENCHMARK_DEPS,
3640)
3641
3642xnnpack_benchmark(
3643 name = "f32_raddextexp_bench",
3644 srcs = [
3645 "bench/f32-raddextexp.cc",
3646 "src/xnnpack/AlignedAllocator.h",
3647 ] + MICROKERNEL_BENCHMARK_HDRS,
3648 deps = MICROKERNEL_BENCHMARK_DEPS,
3649)
3650
3651xnnpack_benchmark(
3652 name = "f32_raddstoreexpminusmax_bench",
3653 srcs = [
3654 "bench/f32-raddstoreexpminusmax.cc",
3655 "src/xnnpack/AlignedAllocator.h",
3656 ] + MICROKERNEL_BENCHMARK_HDRS,
3657 deps = MICROKERNEL_BENCHMARK_DEPS,
3658)
3659
3660xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003661 name = "f32_rmax_bench",
3662 srcs = [
3663 "bench/f32-rmax.cc",
3664 "src/xnnpack/AlignedAllocator.h",
3665 ] + MICROKERNEL_BENCHMARK_HDRS,
3666 deps = MICROKERNEL_BENCHMARK_DEPS,
3667)
3668
3669xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003670 name = "f32_sigmoid_bench",
3671 srcs = [
3672 "bench/f32-sigmoid.cc",
3673 "src/xnnpack/AlignedAllocator.h",
3674 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003675 deps = MICROKERNEL_BENCHMARK_DEPS,
3676)
3677
3678xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003679 name = "f32_spmm_bench",
3680 srcs = [
3681 "bench/f32-spmm.cc",
3682 "bench/gemm.h",
3683 "src/xnnpack/AlignedAllocator.h",
3684 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003685 deps = MICROKERNEL_BENCHMARK_DEPS,
3686)
3687
3688xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003689 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003690 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003691 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003692 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003693 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003694 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003695)
3696
3697xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003698 name = "f32_vscaleexpminusmax_bench",
3699 srcs = [
3700 "bench/f32-vscaleexpminusmax.cc",
3701 "src/xnnpack/AlignedAllocator.h",
3702 ] + MICROKERNEL_BENCHMARK_HDRS,
3703 deps = MICROKERNEL_BENCHMARK_DEPS,
3704)
3705
3706xnnpack_benchmark(
3707 name = "f32_vscaleextexp_bench",
3708 srcs = [
3709 "bench/f32-vscaleextexp.cc",
3710 "src/xnnpack/AlignedAllocator.h",
3711 ] + MICROKERNEL_BENCHMARK_HDRS,
3712 deps = MICROKERNEL_BENCHMARK_DEPS,
3713)
3714
3715xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003716 name = "f32_vsqrt_bench",
3717 srcs = [
3718 "bench/f32-vsqrt.cc",
3719 "src/xnnpack/AlignedAllocator.h",
3720 ] + MICROKERNEL_BENCHMARK_HDRS,
3721 deps = MICROKERNEL_BENCHMARK_DEPS,
3722)
3723
3724xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725 name = "f32_im2col_gemm_bench",
3726 srcs = [
3727 "bench/f32-im2col-gemm.cc",
3728 "bench/conv.h",
3729 "src/xnnpack/AlignedAllocator.h",
3730 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003731 deps = MICROKERNEL_BENCHMARK_DEPS + [
3732 ":im2col",
3733 ":packing",
3734 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735)
3736
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003737xnnpack_benchmark(
3738 name = "requantization_bench",
3739 srcs = [
3740 "bench/requantization.cc",
3741 "src/xnnpack/requantization-stubs.h",
3742 "src/xnnpack/AlignedAllocator.h",
3743 ] + MICROKERNEL_BENCHMARK_HDRS,
3744 deps = MICROKERNEL_BENCHMARK_DEPS,
3745)
3746
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003747xnnpack_benchmark(
3748 name = "rounding_bench",
3749 srcs = [
3750 "bench/rounding.cc",
3751 "src/xnnpack/math-stubs.h",
3752 "src/xnnpack/AlignedAllocator.h",
3753 ] + MICROKERNEL_BENCHMARK_HDRS,
3754 deps = MICROKERNEL_BENCHMARK_DEPS,
3755)
3756
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757########################### Benchmarks for operators ###########################
3758
3759xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 name = "average_pooling_bench",
3761 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003762 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003763 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003764 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003765)
3766
3767xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003768 name = "bankers_rounding_bench",
3769 srcs = ["bench/bankers-rounding.cc"],
3770 copts = xnnpack_optional_tflite_copts(),
3771 tags = ["nowin32"],
3772 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3773)
3774
3775xnnpack_benchmark(
3776 name = "ceiling_bench",
3777 srcs = ["bench/ceiling.cc"],
3778 copts = xnnpack_optional_tflite_copts(),
3779 tags = ["nowin32"],
3780 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3781)
3782
3783xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003784 name = "channel_shuffle_bench",
3785 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003786 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003787)
3788
3789xnnpack_benchmark(
3790 name = "convolution_bench",
3791 srcs = ["bench/convolution.cc"],
3792 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003793 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003794 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795)
3796
3797xnnpack_benchmark(
3798 name = "deconvolution_bench",
3799 srcs = ["bench/deconvolution.cc"],
3800 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003801 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003802 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003803)
3804
3805xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003806 name = "floor_bench",
3807 srcs = ["bench/floor.cc"],
3808 copts = xnnpack_optional_tflite_copts(),
3809 tags = ["nowin32"],
3810 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3811)
3812
3813xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 name = "global_average_pooling_bench",
3815 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003816 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003817)
3818
3819xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003820 name = "hardswish_bench",
3821 srcs = ["bench/hardswish.cc"],
3822 copts = xnnpack_optional_tflite_copts(),
3823 tags = ["nowin32"],
3824 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3825)
3826
3827xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828 name = "max_pooling_bench",
3829 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003830 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003831)
3832
3833xnnpack_benchmark(
3834 name = "sigmoid_bench",
3835 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003836 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003837 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003838 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003839)
3840
3841xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003842 name = "prelu_bench",
3843 srcs = ["bench/prelu.cc"],
3844 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003845 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003846 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003847)
3848
3849xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003850 name = "softmax_bench",
3851 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003852 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003853 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003854 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855)
3856
Marat Dukhan87727142020-06-24 15:24:10 -07003857xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003858 name = "square_root_bench",
3859 srcs = ["bench/square-root.cc"],
3860 copts = xnnpack_optional_tflite_copts(),
3861 tags = ["nowin32"],
3862 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3863)
3864
3865xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003866 name = "truncation_bench",
3867 srcs = ["bench/truncation.cc"],
3868 deps = OPERATOR_BENCHMARK_DEPS,
3869)
3870
Marat Dukhanc068bb62019-10-04 13:24:39 -07003871############################# End-to-end benchmarks ############################
3872
3873cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003874 name = "fp32_mobilenet_v1",
3875 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003876 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003877 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003878 linkstatic = True,
3879 deps = [
3880 ":XNNPACK",
3881 "@pthreadpool",
3882 ],
3883)
3884
3885cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003886 name = "fp16_mobilenet_v1",
3887 srcs = ["models/fp16-mobilenet-v1.cc"],
3888 hdrs = ["models/models.h"],
3889 copts = xnnpack_std_cxxopts(),
3890 linkstatic = True,
3891 deps = [
3892 ":XNNPACK",
3893 "@FP16",
3894 "@pthreadpool",
3895 ],
3896)
3897
3898cc_library(
3899 name = "fp32_mobilenet_v2",
3900 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003901 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003902 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003903 linkstatic = True,
3904 deps = [
3905 ":XNNPACK",
3906 "@pthreadpool",
3907 ],
3908)
3909
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003910cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003911 name = "fp16_mobilenet_v2",
3912 srcs = ["models/fp16-mobilenet-v2.cc"],
3913 hdrs = ["models/models.h"],
3914 copts = xnnpack_std_cxxopts(),
3915 linkstatic = True,
3916 deps = [
3917 ":XNNPACK",
3918 "@FP16",
3919 "@pthreadpool",
3920 ],
3921)
3922
3923cc_library(
3924 name = "fp32_mobilenet_v3_large",
3925 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003926 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003927 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003928 linkstatic = True,
3929 deps = [
3930 ":XNNPACK",
3931 "@pthreadpool",
3932 ],
3933)
3934
3935cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003936 name = "fp32_mobilenet_v3_small",
3937 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003938 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003939 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003940 linkstatic = True,
3941 deps = [
3942 ":XNNPACK",
3943 "@pthreadpool",
3944 ],
3945)
3946
Marat Dukhanc068bb62019-10-04 13:24:39 -07003947xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003948 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003949 srcs = [
3950 "bench/f32-dwconv-e2e.cc",
3951 "bench/end2end.h",
3952 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003953 deps = MICROKERNEL_BENCHMARK_DEPS + [
3954 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003955 ":fp32_mobilenet_v1",
3956 ":fp32_mobilenet_v2",
3957 ":fp32_mobilenet_v3_large",
3958 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003959 ],
3960)
3961
3962xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003963 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003964 srcs = [
3965 "bench/f32-gemm-e2e.cc",
3966 "bench/end2end.h",
3967 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003968 deps = MICROKERNEL_BENCHMARK_DEPS + [
3969 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003970 ":fp32_mobilenet_v1",
3971 ":fp32_mobilenet_v2",
3972 ":fp32_mobilenet_v3_large",
3973 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003974 ],
3975)
3976
3977xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003978 name = "end2end_bench",
3979 srcs = ["bench/end2end.cc"],
3980 deps = [
3981 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003982 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003983 ":fp16_mobilenet_v1",
3984 ":fp16_mobilenet_v2",
3985 ":fp32_mobilenet_v1",
3986 ":fp32_mobilenet_v2",
3987 ":fp32_mobilenet_v3_large",
3988 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003989 "@pthreadpool",
3990 ],
3991)
3992
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003993#################### Accuracy evaluation for math functions ####################
3994
3995xnnpack_benchmark(
3996 name = "f32_exp_eval",
3997 srcs = [
3998 "eval/f32-exp.cc",
3999 "src/xnnpack/AlignedAllocator.h",
4000 ] + ACCURACY_EVAL_HDRS,
4001 deps = ACCURACY_EVAL_DEPS,
4002)
4003
Marat Dukhan515c9772019-10-17 18:07:57 -07004004xnnpack_benchmark(
4005 name = "f32_expminus_eval",
4006 srcs = [
4007 "eval/f32-expminus.cc",
4008 "src/xnnpack/AlignedAllocator.h",
4009 ] + ACCURACY_EVAL_HDRS,
4010 deps = ACCURACY_EVAL_DEPS,
4011)
4012
Marat Dukhan98ba4412019-10-23 02:14:28 -07004013xnnpack_benchmark(
4014 name = "f32_extexp_eval",
4015 srcs = [
4016 "eval/f32-extexp.cc",
4017 "src/xnnpack/AlignedAllocator.h",
4018 ] + ACCURACY_EVAL_HDRS,
4019 deps = ACCURACY_EVAL_DEPS,
4020)
4021
Marat Dukhan8853b822020-05-07 12:19:01 -07004022xnnpack_unit_test(
4023 name = "f32_roundne_eval",
4024 srcs = [
4025 "eval/f32-roundne.cc",
4026 "src/xnnpack/AlignedAllocator.h",
4027 "src/xnnpack/math-stubs.h",
4028 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004029 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004030 deps = MICROKERNEL_TEST_DEPS,
4031)
4032
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004033xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004034 name = "f32_roundd_eval",
4035 srcs = [
4036 "eval/f32-roundd.cc",
4037 "src/xnnpack/AlignedAllocator.h",
4038 "src/xnnpack/math-stubs.h",
4039 ] + MICROKERNEL_TEST_HDRS,
4040 automatic = False,
4041 deps = MICROKERNEL_TEST_DEPS,
4042)
4043
4044xnnpack_unit_test(
4045 name = "f32_roundu_eval",
4046 srcs = [
4047 "eval/f32-roundu.cc",
4048 "src/xnnpack/AlignedAllocator.h",
4049 "src/xnnpack/math-stubs.h",
4050 ] + MICROKERNEL_TEST_HDRS,
4051 automatic = False,
4052 deps = MICROKERNEL_TEST_DEPS,
4053)
4054
4055xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004056 name = "f32_roundz_eval",
4057 srcs = [
4058 "eval/f32-roundz.cc",
4059 "src/xnnpack/AlignedAllocator.h",
4060 "src/xnnpack/math-stubs.h",
4061 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004062 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004063 deps = MICROKERNEL_TEST_DEPS,
4064)
4065
Marat Dukhan346a9e52019-11-15 09:06:30 -08004066xnnpack_benchmark(
4067 name = "f32_sigmoid_eval",
4068 srcs = [
4069 "eval/f32-sigmoid.cc",
4070 "src/xnnpack/AlignedAllocator.h",
4071 ] + ACCURACY_EVAL_HDRS,
4072 deps = ACCURACY_EVAL_DEPS,
4073)
4074
Marat Dukhan84000762020-06-29 18:38:43 -07004075xnnpack_benchmark(
4076 name = "f32_sqrt_eval",
4077 srcs = [
4078 "eval/f32-sqrt.cc",
4079 "src/xnnpack/AlignedAllocator.h",
4080 ] + ACCURACY_EVAL_HDRS,
4081 deps = ACCURACY_EVAL_DEPS + [
4082 ":bench_utils",
4083 ],
4084)
4085
Marat Dukhan08c4a432019-10-03 09:29:21 -07004086######################### Unit tests for micro-kernels #########################
4087
4088xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004089 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004091 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004092 "test/gemm-microkernel-tester.h",
4093 "src/xnnpack/AlignedAllocator.h",
4094 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004095 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096)
4097
4098xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004099 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004100 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004101 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004102 "test/spmm-microkernel-tester.h",
4103 "src/xnnpack/AlignedAllocator.h",
4104 ] + MICROKERNEL_TEST_HDRS,
4105 deps = MICROKERNEL_TEST_DEPS,
4106)
4107
4108xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004109 name = "f32_argmaxpool_test",
4110 srcs = [
4111 "test/f32-argmaxpool.cc",
4112 "test/argmaxpool-microkernel-tester.h",
4113 "src/xnnpack/AlignedAllocator.h",
4114 ] + MICROKERNEL_TEST_HDRS,
4115 deps = MICROKERNEL_TEST_DEPS,
4116)
4117
4118xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004119 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004120 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004121 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122 "test/avgpool-microkernel-tester.h",
4123 "src/xnnpack/AlignedAllocator.h",
4124 ] + MICROKERNEL_TEST_HDRS,
4125 deps = MICROKERNEL_TEST_DEPS,
4126)
4127
4128xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004129 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004130 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004131 "test/f32-ibilinear.cc",
4132 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004133 "src/xnnpack/AlignedAllocator.h",
4134 ] + MICROKERNEL_TEST_HDRS,
4135 deps = MICROKERNEL_TEST_DEPS,
4136)
4137
4138xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004139 name = "f16_clamp_test",
4140 srcs = [
4141 "test/f16-clamp.cc",
4142 "test/clamp-microkernel-tester.h",
4143 ] + MICROKERNEL_TEST_HDRS,
4144 deps = MICROKERNEL_TEST_DEPS,
4145)
4146
4147xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004148 name = "f32_clamp_test",
4149 srcs = [
4150 "test/f32-clamp.cc",
4151 "test/clamp-microkernel-tester.h",
4152 ] + MICROKERNEL_TEST_HDRS,
4153 deps = MICROKERNEL_TEST_DEPS,
4154)
4155
4156xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004157 name = "f32_igemm_test",
4158 srcs = [
4159 "test/f32-igemm.cc",
4160 "test/gemm-microkernel-tester.h",
4161 "src/xnnpack/AlignedAllocator.h",
4162 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004163 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004164)
4165
4166xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004167 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004168 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004169 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004170 "test/gemm-microkernel-tester.h",
4171 "src/xnnpack/AlignedAllocator.h",
4172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004173 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004174)
4175
4176xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004177 name = "f32_igemm_minmax_test",
4178 srcs = [
4179 "test/f32-igemm-minmax.cc",
4180 "test/gemm-microkernel-tester.h",
4181 "src/xnnpack/AlignedAllocator.h",
4182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004183 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004184)
4185
4186xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004187 name = "f16_igemm_minmax_test",
4188 srcs = [
4189 "test/f16-igemm-minmax.cc",
4190 "test/gemm-microkernel-tester.h",
4191 "src/xnnpack/AlignedAllocator.h",
4192 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004193 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004194)
4195
4196xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004197 name = "f32_conv_hwc_test",
4198 srcs = [
4199 "test/f32-conv-hwc.cc",
4200 "test/conv-hwc-microkernel-tester.h",
4201 "src/xnnpack/AlignedAllocator.h",
4202 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004203 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004204)
4205
4206xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004207 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004208 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004209 "test/f32-conv-hwc2chw.cc",
4210 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211 "src/xnnpack/AlignedAllocator.h",
4212 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004213 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214)
4215
4216xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004217 name = "f32_dwconv_test",
4218 srcs = [
4219 "test/f32-dwconv.cc",
4220 "test/dwconv-microkernel-tester.h",
4221 "src/xnnpack/AlignedAllocator.h",
4222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004223 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004224)
4225
4226xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004227 name = "f16_dwconv_minmax_test",
4228 srcs = [
4229 "test/f16-dwconv-minmax.cc",
4230 "test/dwconv-microkernel-tester.h",
4231 "src/xnnpack/AlignedAllocator.h",
4232 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004233 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004234)
4235
4236xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004237 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004238 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004239 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240 "test/dwconv-microkernel-tester.h",
4241 "src/xnnpack/AlignedAllocator.h",
4242 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004243 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004244)
4245
4246xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004247 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004249 "test/f32-dwconv-chw.cc",
4250 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 "src/xnnpack/AlignedAllocator.h",
4252 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004253 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004254)
4255
4256xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004257 name = "f16_gavgpool_minmax_test",
4258 srcs = [
4259 "test/f16-gavgpool-minmax.cc",
4260 "test/gavgpool-microkernel-tester.h",
4261 "src/xnnpack/AlignedAllocator.h",
4262 ] + MICROKERNEL_TEST_HDRS,
4263 deps = MICROKERNEL_TEST_DEPS,
4264)
4265
4266xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004267 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004269 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004270 "test/gavgpool-microkernel-tester.h",
4271 "src/xnnpack/AlignedAllocator.h",
4272 ] + MICROKERNEL_TEST_HDRS,
4273 deps = MICROKERNEL_TEST_DEPS,
4274)
4275
4276xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004277 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004279 "test/f32-gavgpool-cw.cc",
4280 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281 "src/xnnpack/AlignedAllocator.h",
4282 ] + MICROKERNEL_TEST_HDRS,
4283 deps = MICROKERNEL_TEST_DEPS,
4284)
4285
4286xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004287 name = "f32_gemm_test",
4288 srcs = [
4289 "test/f32-gemm.cc",
4290 "test/gemm-microkernel-tester.h",
4291 "src/xnnpack/AlignedAllocator.h",
4292 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004293 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004294)
4295
4296xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004297 name = "f32_gemm_relu_test",
4298 srcs = [
4299 "test/f32-gemm-relu.cc",
4300 "test/gemm-microkernel-tester.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004303 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004304)
4305
4306xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004307 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004309 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 "test/gemm-microkernel-tester.h",
4311 "src/xnnpack/AlignedAllocator.h",
4312 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004313 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314)
4315
4316xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004317 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004319 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004320 "test/gemm-microkernel-tester.h",
4321 "src/xnnpack/AlignedAllocator.h",
4322 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004323 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324)
4325
4326xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004327 name = "f16_hswish_test",
4328 srcs = [
4329 "test/f16-hswish.cc",
4330 "test/hswish-microkernel-tester.h",
4331 ] + MICROKERNEL_TEST_HDRS,
4332 deps = MICROKERNEL_TEST_DEPS,
4333)
4334
4335xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004336 name = "f32_hswish_test",
4337 srcs = [
4338 "test/f32-hswish.cc",
4339 "test/hswish-microkernel-tester.h",
4340 ] + MICROKERNEL_TEST_HDRS,
4341 deps = MICROKERNEL_TEST_DEPS,
4342)
4343
4344xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004345 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004347 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 "test/maxpool-microkernel-tester.h",
4349 ] + MICROKERNEL_TEST_HDRS,
4350 deps = MICROKERNEL_TEST_DEPS,
4351)
4352
4353xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004354 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004356 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357 "test/avgpool-microkernel-tester.h",
4358 "src/xnnpack/AlignedAllocator.h",
4359 ] + MICROKERNEL_TEST_HDRS,
4360 deps = MICROKERNEL_TEST_DEPS,
4361)
4362
4363xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004364 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004366 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 "test/gemm-microkernel-tester.h",
4368 "src/xnnpack/AlignedAllocator.h",
4369 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004370 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371)
4372
4373xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004374 name = "f16_prelu_test",
4375 srcs = [
4376 "test/f16-prelu.cc",
4377 "test/prelu-microkernel-tester.h",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + MICROKERNEL_TEST_HDRS,
4380 deps = MICROKERNEL_TEST_DEPS,
4381)
4382
4383xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_prelu_test",
4385 srcs = [
4386 "test/f32-prelu.cc",
4387 "test/prelu-microkernel-tester.h",
4388 "src/xnnpack/AlignedAllocator.h",
4389 ] + MICROKERNEL_TEST_HDRS,
4390 deps = MICROKERNEL_TEST_DEPS,
4391)
4392
4393xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004394 name = "f32_raddexpminusmax_test",
4395 srcs = [
4396 "test/f32-raddexpminusmax.cc",
4397 "test/raddexpminusmax-microkernel-tester.h",
4398 ] + MICROKERNEL_TEST_HDRS,
4399 deps = MICROKERNEL_TEST_DEPS,
4400)
4401
4402xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004403 name = "f32_raddextexp_test",
4404 srcs = [
4405 "test/f32-raddextexp.cc",
4406 "test/raddextexp-microkernel-tester.h",
4407 ] + MICROKERNEL_TEST_HDRS,
4408 deps = MICROKERNEL_TEST_DEPS,
4409)
4410
4411xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004412 name = "f32_raddstoreexpminusmax_test",
4413 srcs = [
4414 "test/f32-raddstoreexpminusmax.cc",
4415 "test/raddstoreexpminusmax-microkernel-tester.h",
4416 ] + MICROKERNEL_TEST_HDRS,
4417 deps = MICROKERNEL_TEST_DEPS,
4418)
4419
4420xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004421 name = "f32_relu_test",
4422 srcs = [
4423 "test/f32-relu.cc",
4424 "test/vunary-microkernel-tester.h",
4425 ] + MICROKERNEL_TEST_HDRS,
4426 deps = MICROKERNEL_TEST_DEPS,
4427)
4428
4429xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430 name = "f32_rmax_test",
4431 srcs = [
4432 "test/f32-rmax.cc",
4433 "test/rmax-microkernel-tester.h",
4434 ] + MICROKERNEL_TEST_HDRS,
4435 deps = MICROKERNEL_TEST_DEPS,
4436)
4437
4438xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004439 name = "f32_sigmoid_test",
4440 srcs = [
4441 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004442 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004443 ] + MICROKERNEL_TEST_HDRS,
4444 deps = MICROKERNEL_TEST_DEPS,
4445)
4446
4447xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004448 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004449 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004450 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 "test/spmm-microkernel-tester.h",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + MICROKERNEL_TEST_HDRS,
4454 deps = MICROKERNEL_TEST_DEPS,
4455)
4456
4457xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004458 name = "f16_vadd_minmax_test",
4459 srcs = [
4460 "test/f16-vadd-minmax.cc",
4461 "test/vbinary-microkernel-tester.h",
4462 ] + MICROKERNEL_TEST_HDRS,
4463 deps = MICROKERNEL_TEST_DEPS,
4464)
4465
4466xnnpack_unit_test(
4467 name = "f16_vaddc_minmax_test",
4468 srcs = [
4469 "test/f16-vaddc-minmax.cc",
4470 "test/vbinaryc-microkernel-tester.h",
4471 ] + MICROKERNEL_TEST_HDRS,
4472 deps = MICROKERNEL_TEST_DEPS,
4473)
4474
4475xnnpack_unit_test(
4476 name = "f16_vdiv_minmax_test",
4477 srcs = [
4478 "test/f16-vdiv-minmax.cc",
4479 "test/vbinary-microkernel-tester.h",
4480 ] + MICROKERNEL_TEST_HDRS,
4481 deps = MICROKERNEL_TEST_DEPS,
4482)
4483
4484xnnpack_unit_test(
4485 name = "f16_vdivc_minmax_test",
4486 srcs = [
4487 "test/f16-vdivc-minmax.cc",
4488 "test/vbinaryc-microkernel-tester.h",
4489 ] + MICROKERNEL_TEST_HDRS,
4490 deps = MICROKERNEL_TEST_DEPS,
4491)
4492
4493xnnpack_unit_test(
4494 name = "f16_vrdivc_minmax_test",
4495 srcs = [
4496 "test/f16-vrdivc-minmax.cc",
4497 "test/vbinaryc-microkernel-tester.h",
4498 ] + MICROKERNEL_TEST_HDRS,
4499 deps = MICROKERNEL_TEST_DEPS,
4500)
4501
4502xnnpack_unit_test(
4503 name = "f16_vmax_test",
4504 srcs = [
4505 "test/f16-vmax.cc",
4506 "test/vbinary-microkernel-tester.h",
4507 ] + MICROKERNEL_TEST_HDRS,
4508 deps = MICROKERNEL_TEST_DEPS,
4509)
4510
4511xnnpack_unit_test(
4512 name = "f16_vmaxc_test",
4513 srcs = [
4514 "test/f16-vmaxc.cc",
4515 "test/vbinaryc-microkernel-tester.h",
4516 ] + MICROKERNEL_TEST_HDRS,
4517 deps = MICROKERNEL_TEST_DEPS,
4518)
4519
4520xnnpack_unit_test(
4521 name = "f16_vmin_test",
4522 srcs = [
4523 "test/f16-vmin.cc",
4524 "test/vbinary-microkernel-tester.h",
4525 ] + MICROKERNEL_TEST_HDRS,
4526 deps = MICROKERNEL_TEST_DEPS,
4527)
4528
4529xnnpack_unit_test(
4530 name = "f16_vminc_test",
4531 srcs = [
4532 "test/f16-vminc.cc",
4533 "test/vbinaryc-microkernel-tester.h",
4534 ] + MICROKERNEL_TEST_HDRS,
4535 deps = MICROKERNEL_TEST_DEPS,
4536)
4537
4538xnnpack_unit_test(
4539 name = "f16_vmul_minmax_test",
4540 srcs = [
4541 "test/f16-vmul-minmax.cc",
4542 "test/vbinary-microkernel-tester.h",
4543 ] + MICROKERNEL_TEST_HDRS,
4544 deps = MICROKERNEL_TEST_DEPS,
4545)
4546
4547xnnpack_unit_test(
4548 name = "f16_vmulc_minmax_test",
4549 srcs = [
4550 "test/f16-vmulc-minmax.cc",
4551 "test/vbinaryc-microkernel-tester.h",
4552 ] + MICROKERNEL_TEST_HDRS,
4553 deps = MICROKERNEL_TEST_DEPS,
4554)
4555
4556xnnpack_unit_test(
4557 name = "f16_vsub_minmax_test",
4558 srcs = [
4559 "test/f16-vsub-minmax.cc",
4560 "test/vbinary-microkernel-tester.h",
4561 ] + MICROKERNEL_TEST_HDRS,
4562 deps = MICROKERNEL_TEST_DEPS,
4563)
4564
4565xnnpack_unit_test(
4566 name = "f16_vsubc_minmax_test",
4567 srcs = [
4568 "test/f16-vsubc-minmax.cc",
4569 "test/vbinaryc-microkernel-tester.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
4575 name = "f16_vrsubc_minmax_test",
4576 srcs = [
4577 "test/f16-vrsubc-minmax.cc",
4578 "test/vbinaryc-microkernel-tester.h",
4579 ] + MICROKERNEL_TEST_HDRS,
4580 deps = MICROKERNEL_TEST_DEPS,
4581)
4582
4583xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004584 name = "f32_vabs_test",
4585 srcs = [
4586 "test/f32-vabs.cc",
4587 "test/vunary-microkernel-tester.h",
4588 ] + MICROKERNEL_TEST_HDRS,
4589 deps = MICROKERNEL_TEST_DEPS,
4590)
4591
4592xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004593 name = "f32_vadd_test",
4594 srcs = [
4595 "test/f32-vadd.cc",
4596 "test/vbinary-microkernel-tester.h",
4597 ] + MICROKERNEL_TEST_HDRS,
4598 deps = MICROKERNEL_TEST_DEPS,
4599)
4600
4601xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004602 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004603 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004604 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004605 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004606 ] + MICROKERNEL_TEST_HDRS,
4607 deps = MICROKERNEL_TEST_DEPS,
4608)
4609
4610xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004611 name = "f32_vaddc_test",
4612 srcs = [
4613 "test/f32-vaddc.cc",
4614 "test/vbinaryc-microkernel-tester.h",
4615 ] + MICROKERNEL_TEST_HDRS,
4616 deps = MICROKERNEL_TEST_DEPS,
4617)
4618
4619xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004620 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004621 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004622 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004623 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 ] + MICROKERNEL_TEST_HDRS,
4625 deps = MICROKERNEL_TEST_DEPS,
4626)
4627
4628xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004629 name = "f32_vdiv_test",
4630 srcs = [
4631 "test/f32-vdiv.cc",
4632 "test/vbinary-microkernel-tester.h",
4633 ] + MICROKERNEL_TEST_HDRS,
4634 deps = MICROKERNEL_TEST_DEPS,
4635)
4636
4637xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004638 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004639 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004640 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004641 "test/vbinary-microkernel-tester.h",
4642 ] + MICROKERNEL_TEST_HDRS,
4643 deps = MICROKERNEL_TEST_DEPS,
4644)
4645
4646xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004647 name = "f32_vdivc_test",
4648 srcs = [
4649 "test/f32-vdivc.cc",
4650 "test/vbinaryc-microkernel-tester.h",
4651 ] + MICROKERNEL_TEST_HDRS,
4652 deps = MICROKERNEL_TEST_DEPS,
4653)
4654
4655xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004656 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004657 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004658 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004659 "test/vbinaryc-microkernel-tester.h",
4660 ] + MICROKERNEL_TEST_HDRS,
4661 deps = MICROKERNEL_TEST_DEPS,
4662)
4663
4664xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004665 name = "f32_vrdivc_test",
4666 srcs = [
4667 "test/f32-vrdivc.cc",
4668 "test/vbinaryc-microkernel-tester.h",
4669 ] + MICROKERNEL_TEST_HDRS,
4670 deps = MICROKERNEL_TEST_DEPS,
4671)
4672
4673xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004674 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004675 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004676 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004677 "test/vbinaryc-microkernel-tester.h",
4678 ] + MICROKERNEL_TEST_HDRS,
4679 deps = MICROKERNEL_TEST_DEPS,
4680)
4681
4682xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004683 name = "f32_vmax_test",
4684 srcs = [
4685 "test/f32-vmax.cc",
4686 "test/vbinary-microkernel-tester.h",
4687 ] + MICROKERNEL_TEST_HDRS,
4688 deps = MICROKERNEL_TEST_DEPS,
4689)
4690
4691xnnpack_unit_test(
4692 name = "f32_vmaxc_test",
4693 srcs = [
4694 "test/f32-vmaxc.cc",
4695 "test/vbinaryc-microkernel-tester.h",
4696 ] + MICROKERNEL_TEST_HDRS,
4697 deps = MICROKERNEL_TEST_DEPS,
4698)
4699
4700xnnpack_unit_test(
4701 name = "f32_vmin_test",
4702 srcs = [
4703 "test/f32-vmin.cc",
4704 "test/vbinary-microkernel-tester.h",
4705 ] + MICROKERNEL_TEST_HDRS,
4706 deps = MICROKERNEL_TEST_DEPS,
4707)
4708
4709xnnpack_unit_test(
4710 name = "f32_vminc_test",
4711 srcs = [
4712 "test/f32-vminc.cc",
4713 "test/vbinaryc-microkernel-tester.h",
4714 ] + MICROKERNEL_TEST_HDRS,
4715 deps = MICROKERNEL_TEST_DEPS,
4716)
4717
4718xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004719 name = "f32_vmul_test",
4720 srcs = [
4721 "test/f32-vmul.cc",
4722 "test/vbinary-microkernel-tester.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004728 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004729 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004730 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004731 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004737 name = "f32_vmulc_test",
4738 srcs = [
4739 "test/f32-vmulc.cc",
4740 "test/vbinaryc-microkernel-tester.h",
4741 ] + MICROKERNEL_TEST_HDRS,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004746 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004747 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004748 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004749 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004750 ] + MICROKERNEL_TEST_HDRS,
4751 deps = MICROKERNEL_TEST_DEPS,
4752)
4753
4754xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004755 name = "f16_vmulcaddc_minmax_test",
4756 srcs = [
4757 "test/f16-vmulcaddc-minmax.cc",
4758 "test/vmulcaddc-microkernel-tester.h",
4759 "src/xnnpack/AlignedAllocator.h",
4760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004761 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004762)
4763
4764xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004765 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004766 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004767 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004768 "test/vmulcaddc-microkernel-tester.h",
4769 "src/xnnpack/AlignedAllocator.h",
4770 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004771 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772)
4773
4774xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004775 name = "f32_vlrelu_test",
4776 srcs = [
4777 "test/f32-vlrelu.cc",
4778 "test/vunary-microkernel-tester.h",
4779 ] + MICROKERNEL_TEST_HDRS,
4780 deps = MICROKERNEL_TEST_DEPS,
4781)
4782
4783xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004784 name = "f32_vneg_test",
4785 srcs = [
4786 "test/f32-vneg.cc",
4787 "test/vunary-microkernel-tester.h",
4788 ] + MICROKERNEL_TEST_HDRS,
4789 deps = MICROKERNEL_TEST_DEPS,
4790)
4791
4792xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004793 name = "f32_vrndne_test",
4794 srcs = [
4795 "test/f32-vrndne.cc",
4796 "test/vunary-microkernel-tester.h",
4797 ] + MICROKERNEL_TEST_HDRS,
4798 deps = MICROKERNEL_TEST_DEPS,
4799)
4800
4801xnnpack_unit_test(
4802 name = "f32_vrndz_test",
4803 srcs = [
4804 "test/f32-vrndz.cc",
4805 "test/vunary-microkernel-tester.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
4811 name = "f32_vrndu_test",
4812 srcs = [
4813 "test/f32-vrndu.cc",
4814 "test/vunary-microkernel-tester.h",
4815 ] + MICROKERNEL_TEST_HDRS,
4816 deps = MICROKERNEL_TEST_DEPS,
4817)
4818
4819xnnpack_unit_test(
4820 name = "f32_vrndd_test",
4821 srcs = [
4822 "test/f32-vrndd.cc",
4823 "test/vunary-microkernel-tester.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004829 name = "f32_vscale_test",
4830 srcs = [
4831 "test/f32-vscale.cc",
4832 "test/vscale-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004838 name = "f32_vscaleexpminusmax_test",
4839 srcs = [
4840 "test/f32-vscaleexpminusmax.cc",
4841 "test/vscaleexpminusmax-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004847 name = "f32_vscaleextexp_test",
4848 srcs = [
4849 "test/f32-vscaleextexp.cc",
4850 "test/vscaleextexp-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
4855xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004856 name = "f32_vsqr_test",
4857 srcs = [
4858 "test/f32-vsqr.cc",
4859 "test/vunary-microkernel-tester.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004865 name = "f32_vsqrdiff_test",
4866 srcs = [
4867 "test/f32-vsqrdiff.cc",
4868 "test/vbinary-microkernel-tester.h",
4869 ] + MICROKERNEL_TEST_HDRS,
4870 deps = MICROKERNEL_TEST_DEPS,
4871)
4872
4873xnnpack_unit_test(
4874 name = "f32_vsqrdiffc_test",
4875 srcs = [
4876 "test/f32-vsqrdiffc.cc",
4877 "test/vbinaryc-microkernel-tester.h",
4878 ] + MICROKERNEL_TEST_HDRS,
4879 deps = MICROKERNEL_TEST_DEPS,
4880)
4881
4882xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004883 name = "f32_vsqrt_test",
4884 srcs = [
4885 "test/f32-vsqrt.cc",
4886 "test/vunary-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004892 name = "f32_vsub_test",
4893 srcs = [
4894 "test/f32-vsub.cc",
4895 "test/vbinary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004901 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004902 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004903 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004904 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004910 name = "f32_vsubc_test",
4911 srcs = [
4912 "test/f32-vsubc.cc",
4913 "test/vbinaryc-microkernel-tester.h",
4914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004919 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004920 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004921 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004922 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004928 name = "f32_vrsubc_test",
4929 srcs = [
4930 "test/f32-vrsubc.cc",
4931 "test/vbinaryc-microkernel-tester.h",
4932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004937 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004938 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004939 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004940 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004946 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004947 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004948 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004949 "test/avgpool-microkernel-tester.h",
4950 "src/xnnpack/AlignedAllocator.h",
4951 ] + MICROKERNEL_TEST_HDRS,
4952 deps = MICROKERNEL_TEST_DEPS,
4953)
4954
4955xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004956 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004957 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004958 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004959 "test/gemm-microkernel-tester.h",
4960 "src/xnnpack/AlignedAllocator.h",
4961 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004962 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004963)
4964
4965xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004966 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004967 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004968 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969 "test/dwconv-microkernel-tester.h",
4970 "src/xnnpack/AlignedAllocator.h",
4971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004972 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973)
4974
4975xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004976 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004977 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004978 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979 "test/gavgpool-microkernel-tester.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + MICROKERNEL_TEST_HDRS,
4982 deps = MICROKERNEL_TEST_DEPS,
4983)
4984
4985xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004986 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004988 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 "test/gemm-microkernel-tester.h",
4990 "src/xnnpack/AlignedAllocator.h",
4991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004992 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
4995xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004996 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004998 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 "test/vadd-microkernel-tester.h",
5000 ] + MICROKERNEL_TEST_HDRS,
5001 deps = MICROKERNEL_TEST_DEPS,
5002)
5003
5004xnnpack_unit_test(
5005 name = "u8_clamp_test",
5006 srcs = [
5007 "test/u8-clamp.cc",
5008 "test/clamp-microkernel-tester.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
5014 name = "u8_lut32norm_test",
5015 srcs = [
5016 "test/u8-lut32norm.cc",
5017 "test/lut-norm-microkernel-tester.h",
5018 ] + MICROKERNEL_TEST_HDRS,
5019 deps = MICROKERNEL_TEST_DEPS,
5020)
5021
5022xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005023 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005024 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005025 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026 "test/maxpool-microkernel-tester.h",
5027 ] + MICROKERNEL_TEST_HDRS,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
5032 name = "u8_rmax_test",
5033 srcs = [
5034 "test/u8-rmax.cc",
5035 "test/rmax-microkernel-tester.h",
5036 ] + MICROKERNEL_TEST_HDRS,
5037 deps = MICROKERNEL_TEST_DEPS,
5038)
5039
5040xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005041 name = "x32_fill_test",
5042 srcs = [
5043 "test/x32-fill.cc",
5044 "test/fill-microkernel-tester.h",
5045 ] + MICROKERNEL_TEST_HDRS,
5046 deps = MICROKERNEL_TEST_DEPS,
5047)
5048
5049xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050 name = "x32_packx_test",
5051 srcs = [
5052 "test/x32-packx.cc",
5053 "test/pack-microkernel-tester.h",
5054 "src/xnnpack/AlignedAllocator.h",
5055 ] + MICROKERNEL_TEST_HDRS,
5056 deps = MICROKERNEL_TEST_DEPS,
5057)
5058
5059xnnpack_unit_test(
5060 name = "x32_pad_test",
5061 srcs = [
5062 "test/x32-pad.cc",
5063 "test/pad-microkernel-tester.h",
5064 ] + MICROKERNEL_TEST_HDRS,
5065 deps = MICROKERNEL_TEST_DEPS,
5066)
5067
5068xnnpack_unit_test(
5069 name = "x32_unpool_test",
5070 srcs = [
5071 "test/x32-unpool.cc",
5072 "test/unpool-microkernel-tester.h",
5073 ] + MICROKERNEL_TEST_HDRS,
5074 deps = MICROKERNEL_TEST_DEPS,
5075)
5076
5077xnnpack_unit_test(
5078 name = "x32_zip_test",
5079 srcs = [
5080 "test/x32-zip.cc",
5081 "test/zip-microkernel-tester.h",
5082 ] + MICROKERNEL_TEST_HDRS,
5083 deps = MICROKERNEL_TEST_DEPS,
5084)
5085
5086xnnpack_unit_test(
5087 name = "x8_lut_test",
5088 srcs = [
5089 "test/x8-lut.cc",
5090 "test/lut-microkernel-tester.h",
5091 ] + MICROKERNEL_TEST_HDRS,
5092 deps = MICROKERNEL_TEST_DEPS,
5093)
5094
5095xnnpack_unit_test(
5096 name = "x8_zip_test",
5097 srcs = [
5098 "test/x8-zip.cc",
5099 "test/zip-microkernel-tester.h",
5100 ] + MICROKERNEL_TEST_HDRS,
5101 deps = MICROKERNEL_TEST_DEPS,
5102)
5103
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005104xnnpack_unit_test(
5105 name = "requantization_test",
5106 srcs = [
5107 "src/xnnpack/requantization-stubs.h",
5108 "test/requantization.cc",
5109 "test/requantization-tester.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
Marat Dukhan20c3b922020-03-10 03:45:06 -07005114########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005115
5116xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005117 name = "operator_size_test",
5118 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005119 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120)
5121
Marat Dukhan20c3b922020-03-10 03:45:06 -07005122xnnpack_binary(
5123 name = "subgraph_size_test",
5124 srcs = ["test/subgraph-size.c"],
5125 deps = [":XNNPACK"],
5126)
5127
5128########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129
5130xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005131 name = "abs_nc_test",
5132 srcs = [
5133 "test/abs-nc.cc",
5134 "test/abs-operator-tester.h",
5135 ],
5136 deps = OPERATOR_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005140 name = "add_nd_test",
5141 srcs = [
5142 "test/add-nd.cc",
5143 "test/binary-elementwise-operator-tester.h",
5144 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005145 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005146)
5147
5148xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005149 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005150 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005151 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005152 "test/argmax-pooling-operator-tester.h",
5153 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005154 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155)
5156
5157xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005158 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005160 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161 "test/average-pooling-operator-tester.h",
5162 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005163 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164)
5165
5166xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005167 name = "bankers_rounding_nc_test",
5168 srcs = [
5169 "test/bankers-rounding-nc.cc",
5170 "test/bankers-rounding-operator-tester.h",
5171 ],
5172 deps = OPERATOR_TEST_DEPS,
5173)
5174
5175xnnpack_unit_test(
5176 name = "ceiling_nc_test",
5177 srcs = [
5178 "test/ceiling-nc.cc",
5179 "test/ceiling-operator-tester.h",
5180 ],
5181 deps = OPERATOR_TEST_DEPS,
5182)
5183
5184xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005185 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005186 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005187 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005188 "test/channel-shuffle-operator-tester.h",
5189 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005190 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005191)
5192
5193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005194 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005196 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 "test/clamp-operator-tester.h",
5198 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200)
5201
5202xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005203 name = "constant_pad_nd_test",
5204 srcs = [
5205 "test/constant-pad-nd.cc",
5206 "test/constant-pad-operator-tester.h",
5207 ],
5208 deps = OPERATOR_TEST_DEPS,
5209)
5210
5211xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005212 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005214 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215 "test/convolution-operator-tester.h",
5216 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005217 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005218)
5219
5220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005221 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005223 "test/convolution-nchw.cc",
5224 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227)
5228
5229xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005230 name = "copy_nc_test",
5231 srcs = [
5232 "test/copy-nc.cc",
5233 "test/copy-operator-tester.h",
5234 ],
5235 deps = OPERATOR_TEST_DEPS,
5236)
5237
5238xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005239 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005241 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242 "test/deconvolution-operator-tester.h",
5243 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245)
5246
5247xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005248 name = "divide_nd_test",
5249 srcs = [
5250 "test/binary-elementwise-operator-tester.h",
5251 "test/divide-nd.cc",
5252 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005254)
5255
5256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005257 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005259 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 "test/fully-connected-operator-tester.h",
5261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263)
5264
5265xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005266 name = "floor_nc_test",
5267 srcs = [
5268 "test/floor-nc.cc",
5269 "test/floor-operator-tester.h",
5270 ],
5271 deps = OPERATOR_TEST_DEPS,
5272)
5273
5274xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005275 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005277 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005279 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005280 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281)
5282
5283xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005284 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005286 "test/global-average-pooling-ncw.cc",
5287 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290)
5291
5292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005293 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005295 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "test/hardswish-operator-tester.h",
5297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005302 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005304 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 "test/leaky-relu-operator-tester.h",
5306 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308)
5309
5310xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005311 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005313 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314 "test/max-pooling-operator-tester.h",
5315 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005316 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317)
5318
5319xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005320 name = "maximum_nd_test",
5321 srcs = [
5322 "test/binary-elementwise-operator-tester.h",
5323 "test/maximum-nd.cc",
5324 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005326)
5327
5328xnnpack_unit_test(
5329 name = "minimum_nd_test",
5330 srcs = [
5331 "test/binary-elementwise-operator-tester.h",
5332 "test/minimum-nd.cc",
5333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005335)
5336
5337xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005338 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005339 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005340 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005341 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005343 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005344)
5345
5346xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005347 name = "negate_nc_test",
5348 srcs = [
5349 "test/negate-nc.cc",
5350 "test/negate-operator-tester.h",
5351 ],
5352 deps = OPERATOR_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005356 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005358 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 "test/prelu-operator-tester.h",
5360 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005362)
5363
5364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005365 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005367 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005368 "test/resize-bilinear-operator-tester.h",
5369 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005371)
5372
5373xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005374 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005376 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005377 "test/sigmoid-operator-tester.h",
5378 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380)
5381
5382xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005383 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005385 "test/softmax-nc.cc",
5386 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005387 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005388 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389)
5390
5391xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005392 name = "square_nc_test",
5393 srcs = [
5394 "test/square-nc.cc",
5395 "test/square-operator-tester.h",
5396 ],
5397 deps = OPERATOR_TEST_DEPS,
5398)
5399
5400xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005401 name = "square_root_nc_test",
5402 srcs = [
5403 "test/square-root-nc.cc",
5404 "test/square-root-operator-tester.h",
5405 ],
5406 deps = OPERATOR_TEST_DEPS,
5407)
5408
5409xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005410 name = "squared_difference_nd_test",
5411 srcs = [
5412 "test/binary-elementwise-operator-tester.h",
5413 "test/squared-difference-nd.cc",
5414 ],
5415 deps = OPERATOR_TEST_DEPS,
5416)
5417
5418xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005419 name = "subtract_nd_test",
5420 srcs = [
5421 "test/binary-elementwise-operator-tester.h",
5422 "test/subtract-nd.cc",
5423 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005425)
5426
5427xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005428 name = "truncation_nc_test",
5429 srcs = [
5430 "test/truncation-nc.cc",
5431 "test/truncation-operator-tester.h",
5432 ],
5433 deps = OPERATOR_TEST_DEPS,
5434)
5435
5436xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005437 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005439 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005440 "test/unpooling-operator-tester.h",
5441 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005442 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443)
5444
Chao Mei6ddfc602020-05-13 22:29:36 -07005445############################### Misc unit tests ###############################
5446
5447xnnpack_unit_test(
5448 name = "memory_planner_test",
5449 srcs = [
5450 "test/memory-planner-test.cc",
5451 ],
5452 deps = [
5453 ":XNNPACK",
5454 ":memory_planner",
5455 ],
5456)
5457
Marat Dukhan08c4a432019-10-03 09:29:21 -07005458############################# Build configurations #############################
5459
Marat Dukhanb8642352019-10-30 15:43:02 -07005460# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005462 name = "xnn_enable_assembly_explicit_true",
5463 define_values = {"xnn_enable_assembly": "true"},
5464)
5465
5466# Disables usage of assembly kernels.
5467config_setting(
5468 name = "xnn_enable_assembly_explicit_false",
5469 define_values = {"xnn_enable_assembly": "false"},
5470)
5471
Marat Dukhan9de90e02020-06-18 16:04:12 -07005472# Enables usage of sparse inference.
5473config_setting(
5474 name = "xnn_enable_sparse_explicit_true",
5475 define_values = {"xnn_enable_sparse": "true"},
5476)
5477
5478# Disables usage of sparse inference.
5479config_setting(
5480 name = "xnn_enable_sparse_explicit_false",
5481 define_values = {"xnn_enable_sparse": "false"},
5482)
5483
Marat Dukhan05702cf2020-03-26 15:41:33 -07005484# Disables usage of HMP-aware optimizations.
5485config_setting(
5486 name = "xnn_enable_hmp_explicit_false",
5487 define_values = {"xnn_enable_hmp": "false"},
5488)
5489
Chao Mei6ddfc602020-05-13 22:29:36 -07005490# Enable usage of optimized memory allocation
5491config_setting(
5492 name = "xnn_enable_memopt_explicit_true",
5493 define_values = {"xnnpack_enable_memopt": "true"},
5494)
5495
5496# Disable usage of optimized memory allocation
5497config_setting(
5498 name = "xnn_enable_memopt_explicit_false",
5499 define_values = {"xnnpack_enable_memopt": "false"},
5500)
5501
Marat Dukhanb8642352019-10-30 15:43:02 -07005502# Builds with -c dbg
5503config_setting(
5504 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005505 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005506 "compilation_mode": "dbg",
5507 },
5508)
5509
5510# Builds with -c opt
5511config_setting(
5512 name = "optimized_build",
5513 values = {
5514 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005515 },
5516)
5517
5518config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005519 name = "linux_k8",
5520 values = {"cpu": "k8"},
5521)
5522
5523config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005524 name = "linux_arm",
5525 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005526)
5527
5528config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005529 name = "linux_armeabi",
5530 values = {"cpu": "armeabi"},
5531)
5532
5533config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005534 name = "linux_armhf",
5535 values = {"cpu": "armhf"},
5536)
5537
5538config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005539 name = "linux_armv7a",
5540 values = {"cpu": "armv7a"},
5541)
5542
5543config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005544 name = "linux_aarch64",
5545 values = {"cpu": "aarch64"},
5546)
5547
5548config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005549 name = "android",
5550 values = {"crosstool_top": "//external:android/crosstool"},
5551)
5552
5553config_setting(
5554 name = "android_armv7",
5555 values = {
5556 "crosstool_top": "//external:android/crosstool",
5557 "cpu": "armeabi-v7a",
5558 },
5559)
5560
5561config_setting(
5562 name = "android_arm64",
5563 values = {
5564 "crosstool_top": "//external:android/crosstool",
5565 "cpu": "arm64-v8a",
5566 },
5567)
5568
5569config_setting(
5570 name = "android_x86",
5571 values = {
5572 "crosstool_top": "//external:android/crosstool",
5573 "cpu": "x86",
5574 },
5575)
5576
5577config_setting(
5578 name = "android_x86_64",
5579 values = {
5580 "crosstool_top": "//external:android/crosstool",
5581 "cpu": "x86_64",
5582 },
5583)
5584
5585config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005586 name = "windows_x86_64",
5587 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005588)
5589
5590config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005591 name = "windows_x86_64_clang",
5592 values = {
5593 "compiler": "clang-cl",
5594 "cpu": "x64_windows",
5595 },
5596)
5597
5598config_setting(
5599 name = "windows_x86_64_mingw",
5600 values = {
5601 "compiler": "mingw-gcc",
5602 "cpu": "x64_windows",
5603 },
5604)
5605
5606config_setting(
5607 name = "windows_x86_64_msys",
5608 values = {
5609 "compiler": "msys-gcc",
5610 "cpu": "x64_windows",
5611 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005612)
5613
5614config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005615 name = "macos_x86_64",
5616 values = {
5617 "apple_platform_type": "macos",
5618 "cpu": "darwin",
5619 },
5620)
5621
5622config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005623 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005624 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005625)
5626
5627config_setting(
5628 name = "emscripten_wasm",
5629 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005630 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005631 "cpu": "wasm",
5632 },
5633)
5634
5635config_setting(
5636 name = "emscripten_wasmsimd",
5637 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005638 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005640 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641 },
5642)
5643
5644config_setting(
5645 name = "emscripten_asmjs",
5646 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005647 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005648 "cpu": "asmjs",
5649 },
5650)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005651
5652config_setting(
5653 name = "ios_armv7",
5654 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005655 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005656 "cpu": "ios_armv7",
5657 },
5658)
5659
5660config_setting(
5661 name = "ios_arm64",
5662 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005663 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005664 "cpu": "ios_arm64",
5665 },
5666)
5667
5668config_setting(
5669 name = "ios_arm64e",
5670 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005671 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005672 "cpu": "ios_arm64e",
5673 },
5674)
5675
5676config_setting(
5677 name = "ios_x86",
5678 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005679 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005680 "cpu": "ios_i386",
5681 },
5682)
5683
5684config_setting(
5685 name = "ios_x86_64",
5686 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005687 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005688 "cpu": "ios_x86_64",
5689 },
5690)
5691
5692config_setting(
5693 name = "watchos_armv7k",
5694 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005695 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005696 "cpu": "watchos_armv7k",
5697 },
5698)
5699
5700config_setting(
5701 name = "watchos_arm64_32",
5702 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005703 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005704 "cpu": "watchos_arm64_32",
5705 },
5706)
5707
5708config_setting(
5709 name = "watchos_x86",
5710 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005711 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005712 "cpu": "watchos_i386",
5713 },
5714)
5715
5716config_setting(
5717 name = "watchos_x86_64",
5718 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005719 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005720 "cpu": "watchos_x86_64",
5721 },
5722)
5723
5724config_setting(
5725 name = "tvos_arm64",
5726 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005727 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005728 "cpu": "tvos_arm64",
5729 },
5730)
5731
5732config_setting(
5733 name = "tvos_x86_64",
5734 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005735 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005736 "cpu": "tvos_x86_64",
5737 },
5738)