blob: 276d3e09eded559f2ff28f5e1ad28f2715eca48f [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800219 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
225 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700228 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/1x1-minmax-scalar.c",
230 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar.c",
232 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar.c",
234 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar.c",
236 "src/f32-spmm/gen/8x2-minmax-scalar.c",
237 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700238 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700241 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vadd-scalar-x1.c",
245 "src/f32-vbinary/gen/vadd-scalar-x2.c",
246 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700250 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
251 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
252 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700253 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700259 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
260 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
261 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700262 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
263 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
264 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
272 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
273 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800274 "src/f32-vbinary/gen/vmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vmax-scalar-x4.c",
277 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
278 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
279 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
280 "src/f32-vbinary/gen/vmin-scalar-x1.c",
281 "src/f32-vbinary/gen/vmin-scalar-x2.c",
282 "src/f32-vbinary/gen/vmin-scalar-x4.c",
283 "src/f32-vbinary/gen/vminc-scalar-x1.c",
284 "src/f32-vbinary/gen/vminc-scalar-x2.c",
285 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700289 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
290 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
291 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vmul-scalar-x1.c",
293 "src/f32-vbinary/gen/vmul-scalar-x2.c",
294 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700298 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
299 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
300 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700301 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
302 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
303 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
311 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
312 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700313 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
314 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
315 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700316 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
317 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
318 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700319 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
320 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
321 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700322 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
323 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
324 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
325 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
326 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
327 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700334 "src/f32-vbinary/gen/vsub-scalar-x1.c",
335 "src/f32-vbinary/gen/vsub-scalar-x2.c",
336 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700337 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
338 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
339 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700340 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
341 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
342 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700343 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
344 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
345 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700346 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
347 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
348 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700349 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
350 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
351 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700352 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
353 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
354 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
355 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
356 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
357 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
358 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
359 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
360 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700361 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
362 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
363 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700364 "src/f32-vunary/gen/vabs-scalar-x1.c",
365 "src/f32-vunary/gen/vabs-scalar-x2.c",
366 "src/f32-vunary/gen/vabs-scalar-x4.c",
367 "src/f32-vunary/gen/vneg-scalar-x1.c",
368 "src/f32-vunary/gen/vneg-scalar-x2.c",
369 "src/f32-vunary/gen/vneg-scalar-x4.c",
370 "src/f32-vunary/gen/vsqr-scalar-x1.c",
371 "src/f32-vunary/gen/vsqr-scalar-x2.c",
372 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800373 "src/math/expminus-scalar-lut2048-p1.c",
374 "src/math/expminus-scalar-lut64-p2.c",
375 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700376 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700377 "src/math/roundne-scalar-nearbyint.c",
378 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700379 "src/math/roundd-scalar-addsub.c",
380 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700381 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700382 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700383 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700384 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700385 "src/math/roundz-scalar-addsub.c",
386 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700387 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800388 "src/math/sigmoid-scalar-lut2048-p1-div.c",
389 "src/math/sigmoid-scalar-lut64-p2-div.c",
390 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700391 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
392 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700393 "src/qs8-requantization/fp32-scalar-lrintf.c",
394 "src/qs8-requantization/fp32-scalar-magic.c",
395 "src/qs8-requantization/precise-scalar-signed64.c",
396 "src/qs8-requantization/precise-scalar-unsigned32.c",
397 "src/qs8-requantization/precise-scalar-unsigned64.c",
398 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700399 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
400 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
401 "src/qu8-dwconv/up1x9-minmax-scalar.c",
402 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
403 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700404 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
405 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700406 "src/qu8-gemm/2x2-minmax-scalar.c",
407 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700408 "src/qu8-requantization/fp32-scalar-lrintf.c",
409 "src/qu8-requantization/fp32-scalar-magic.c",
410 "src/qu8-requantization/precise-scalar-signed64.c",
411 "src/qu8-requantization/precise-scalar-unsigned32.c",
412 "src/qu8-requantization/precise-scalar-unsigned64.c",
413 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700414 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700415 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700416 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700417 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700419 "src/x32-fill/scalar-float.c",
420 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x2-scalar.c",
422 "src/x32-packx/x3-scalar.c",
423 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700424 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700425 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700426 "src/x32-unpool/scalar.c",
427 "src/x32-zip/x2-scalar.c",
428 "src/x32-zip/x3-scalar.c",
429 "src/x32-zip/x4-scalar.c",
430 "src/x32-zip/xm-scalar.c",
431 "src/x8-lut/scalar.c",
432 "src/x8-zip/x2-scalar.c",
433 "src/x8-zip/x3-scalar.c",
434 "src/x8-zip/x4-scalar.c",
435 "src/x8-zip/xm-scalar.c",
436]
437
Marat Dukhan436ebe62019-12-04 15:10:12 -0800438WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700439 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
440 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700441 "src/f32-clamp/gen/wasm-x1.c",
442 "src/f32-clamp/gen/wasm-x2.c",
443 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700444 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
445 "src/f32-dwconv/gen/up1x4-wasm.c",
446 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
447 "src/f32-dwconv/gen/up1x9-wasm.c",
448 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
449 "src/f32-dwconv/gen/up1x25-wasm.c",
450 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
451 "src/f32-dwconv/gen/up2x4-wasm.c",
452 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
453 "src/f32-dwconv/gen/up2x9-wasm.c",
454 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
455 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700456 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
457 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
458 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
459 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700460 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
461 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
463 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
464 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
465 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700466 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
467 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
469 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700470 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
471 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
472 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700473 "src/f32-gemm/gen/1x4-wasm.c",
474 "src/f32-gemm/gen/2x4-wasm.c",
475 "src/f32-gemm/gen/4x2-wasm.c",
476 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700477 "src/f32-gemm/gen/1x4-relu-wasm.c",
478 "src/f32-gemm/gen/2x4-relu-wasm.c",
479 "src/f32-gemm/gen/4x2-relu-wasm.c",
480 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700481 "src/f32-gemm/gen/1x4-minmax-wasm.c",
482 "src/f32-gemm/gen/2x4-minmax-wasm.c",
483 "src/f32-gemm/gen/4x2-minmax-wasm.c",
484 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700485 "src/f32-hswish/gen/hswish-wasm-x1.c",
486 "src/f32-hswish/gen/hswish-wasm-x2.c",
487 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700488 "src/f32-igemm/gen/1x4-wasm.c",
489 "src/f32-igemm/gen/2x4-wasm.c",
490 "src/f32-igemm/gen/4x2-wasm.c",
491 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700492 "src/f32-igemm/gen/1x4-relu-wasm.c",
493 "src/f32-igemm/gen/2x4-relu-wasm.c",
494 "src/f32-igemm/gen/4x2-relu-wasm.c",
495 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700496 "src/f32-igemm/gen/1x4-minmax-wasm.c",
497 "src/f32-igemm/gen/2x4-minmax-wasm.c",
498 "src/f32-igemm/gen/4x2-minmax-wasm.c",
499 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700500 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
501 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
502 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700503 "src/f32-prelu/gen/wasm-2x1.c",
504 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700505 "src/f32-relu/gen/wasm-x1.c",
506 "src/f32-relu/gen/wasm-x2.c",
507 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700508 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
509 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
510 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
511 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
512 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
513 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
514 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
515 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
516 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
517 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
518 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
519 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700520 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
521 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
522 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
523 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
524 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
525 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
526 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
527 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
528 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
529 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
530 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
531 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800532 "src/f32-vbinary/gen/vmax-wasm-x1.c",
533 "src/f32-vbinary/gen/vmax-wasm-x2.c",
534 "src/f32-vbinary/gen/vmax-wasm-x4.c",
535 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
536 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
537 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
538 "src/f32-vbinary/gen/vmin-wasm-x1.c",
539 "src/f32-vbinary/gen/vmin-wasm-x2.c",
540 "src/f32-vbinary/gen/vmin-wasm-x4.c",
541 "src/f32-vbinary/gen/vminc-wasm-x1.c",
542 "src/f32-vbinary/gen/vminc-wasm-x2.c",
543 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700544 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
545 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
546 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
547 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
548 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
549 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
550 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
553 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
554 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
555 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
556 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
557 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
558 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
559 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
560 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
561 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
565 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
566 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
567 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
568 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
569 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
570 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
571 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
572 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
573 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
574 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
577 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
578 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
579 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700580 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
581 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
582 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700583 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
584 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
585 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800586]
587
Marat Dukhan290055c2020-06-09 12:24:29 -0700588WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700589 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
590 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
591 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700592 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
593 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
594 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
595 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700596 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
597 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
598 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
599 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700600 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
601 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
602 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
603 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
604 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
605 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700606 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
607 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
608 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
609 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
610 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
611 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
612 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
613 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
614 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
615 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
616 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
617 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
618 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
619 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
620 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
621 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
622 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
623 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
624 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
625 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
626 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
627 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
628 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
629 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700630 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
631 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
632 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700633 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700634 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
635 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
636 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
637 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700638 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
639 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
640 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
641 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
642 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
643 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
644 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
645 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
646 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
647 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
648 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
649 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
650 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
651 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
652 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
653 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
654 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
655 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
656 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
657 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
658 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
659 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
660 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
661 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
662 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
663 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
664 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
665 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
666 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
667 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700668 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
669 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700670 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
671 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
672 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
673 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
674 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
675 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
676 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
677 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
678 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
679 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
680 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
681 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
682 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
683 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
684 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
685 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
686 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
687 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
688 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
689 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
690 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
691 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
692 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
693 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
694 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
695 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
696 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
697 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
698 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
699 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700700 "src/f32-ibilinear/gen/wasmsimd-c4.c",
701 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700702 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
703 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
704 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700705 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700706 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
707 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
708 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
709 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700710 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
711 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
712 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
713 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
714 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
715 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
716 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
717 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
718 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
719 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
720 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
721 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
722 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
723 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
724 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
725 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
726 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
727 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
728 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
729 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
730 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
731 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
732 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
733 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
734 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
735 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
736 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
737 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
738 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
739 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700740 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
741 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700742 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
743 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
744 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
745 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700746 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
747 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700748 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
749 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
750 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700751 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
752 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700753 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
754 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
755 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
756 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700757 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
758 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700759 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
760 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
761 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700762 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
763 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700764 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
765 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
766 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
767 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
768 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
769 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
770 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700771 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
772 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700773 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
774 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
775 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
776 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700777 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
778 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
779 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
780 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
781 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
782 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
783 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
784 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
785 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
786 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
787 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
788 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700789 "src/f32-relu/gen/wasmsimd-x4.c",
790 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700791 "src/f32-rmax/wasmsimd-arm.c",
792 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700793 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
794 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
795 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
796 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
797 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
798 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700799 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
800 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
801 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
802 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
803 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
804 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
806 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
807 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
808 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700809 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
810 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700811 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
812 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700813 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
814 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
815 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
816 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700817 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
818 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700821 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
822 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
823 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
824 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700825 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
826 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700827 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
828 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700829 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
830 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
831 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
832 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700833 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
834 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700835 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
836 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700837 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
838 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
839 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
840 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700841 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
842 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
843 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
844 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700845 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
846 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
847 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
848 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700849 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
850 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
851 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
852 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700853 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
854 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
855 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
856 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700857 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
858 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700859 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
860 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700861 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
862 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
863 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
864 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700865 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
866 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700867 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
868 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700869 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
870 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
871 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700873 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
874 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700875 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
876 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700877 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
878 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
879 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
880 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700881 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
882 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700883 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
884 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700885 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
886 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700887 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
888 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700889 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
890 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
891 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
892 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700893 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
894 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700895 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
896 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700897 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
898 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
899 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
900 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700901 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700903 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
904 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700905 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
906 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
907 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
908 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700909 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
910 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
911 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
912 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700913 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
914 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
915 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
916 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
917 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
918 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
919 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
920 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
921 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
922 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
923 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
924 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
925 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
926 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700927 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
928 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700929 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
930 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
931 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
932 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
933 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
934 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700935 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
936 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
937 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700938 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
939 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
940 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700941 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
942 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
943 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700944 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
945 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
946 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
947 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
948 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
949 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
950 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
951 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
952 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
953 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
954 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
955 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
956 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
957 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
958 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700959 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700960 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700961 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700962 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700963 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700964 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700965 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700966 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700967 "src/x32-zip/x2-wasmsimd.c",
968 "src/x32-zip/x3-wasmsimd.c",
969 "src/x32-zip/x4-wasmsimd.c",
970 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700971 "src/math/roundne-wasmsimd-addsub.c",
972 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700973 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700974 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700975 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700976 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700977 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700978 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
979 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700980]
981
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800982PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700983 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700984 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700985 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700986 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700987 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700988 "src/f32-spmm/gen/4x1-minmax-psimd.c",
989 "src/f32-spmm/gen/8x1-minmax-psimd.c",
990 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700991]
992
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800993PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700994 "src/qs8-requantization/precise-psimd.c",
995 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700996 "src/qu8-requantization/precise-psimd.c",
997 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800998]
999
Marat Dukhan08c4a432019-10-03 09:29:21 -07001000# ISA-specific micro-kernels
1001NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001002 "src/f32-argmaxpool/4x-neon-c4.c",
1003 "src/f32-argmaxpool/9p8x-neon-c4.c",
1004 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001005 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1006 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001007 "src/f32-clamp/gen/neon-x4.c",
1008 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001009 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1010 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001011 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1012 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001013 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1014 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001015 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1016 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001017 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1018 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1019 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1020 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001021 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1022 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1023 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1024 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001025 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1026 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1027 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1028 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001029 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001030 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1031 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001032 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1033 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1034 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1035 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1036 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1038 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1039 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1040 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1041 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1042 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1044 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1045 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1046 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1047 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1048 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1049 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1050 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1051 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1052 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1053 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1054 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1055 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1056 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1057 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1058 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1059 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1060 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1061 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1062 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001063 "src/f32-hswish/gen/hswish-neon-x4.c",
1064 "src/f32-hswish/gen/hswish-neon-x8.c",
1065 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001066 "src/f32-ibilinear/gen/neon-c4.c",
1067 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001068 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1069 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1070 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1071 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1072 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1073 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1074 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1075 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1076 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1077 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1078 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1079 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1080 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1081 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1082 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1083 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001084 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1085 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1086 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001087 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1088 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001089 "src/f32-prelu/gen/neon-1x4.c",
1090 "src/f32-prelu/gen/neon-1x8.c",
1091 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001092 "src/f32-prelu/gen/neon-2x4.c",
1093 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001094 "src/f32-prelu/gen/neon-2x16.c",
1095 "src/f32-prelu/gen/neon-4x4.c",
1096 "src/f32-prelu/gen/neon-4x8.c",
1097 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001098 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1099 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1100 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1101 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1103 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1104 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1105 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1106 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1107 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1108 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1109 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1110 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1111 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1112 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1113 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1114 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1115 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1116 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1117 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1118 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1119 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1120 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1121 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001122 "src/f32-relu/gen/neon-x4.c",
1123 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001125 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001126 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1127 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1128 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1129 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1130 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1131 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1132 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1133 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1134 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1135 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1136 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1137 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1138 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1139 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1140 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1141 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1142 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1143 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001144 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1145 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1146 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1147 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001148 "src/f32-vbinary/gen/vmax-neon-x4.c",
1149 "src/f32-vbinary/gen/vmax-neon-x8.c",
1150 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1151 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1152 "src/f32-vbinary/gen/vmin-neon-x4.c",
1153 "src/f32-vbinary/gen/vmin-neon-x8.c",
1154 "src/f32-vbinary/gen/vminc-neon-x4.c",
1155 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001156 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1157 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1158 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1159 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1160 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1161 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001162 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1163 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1164 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1165 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001166 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1167 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1168 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1169 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001170 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1171 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001172 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1173 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001174 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1175 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1176 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1177 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1178 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1179 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1180 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1181 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001182 "src/f32-vunary/gen/vabs-neon-x4.c",
1183 "src/f32-vunary/gen/vabs-neon-x8.c",
1184 "src/f32-vunary/gen/vneg-neon-x4.c",
1185 "src/f32-vunary/gen/vneg-neon-x8.c",
1186 "src/f32-vunary/gen/vsqr-neon-x4.c",
1187 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001188 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1189 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1190 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1191 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001192 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1193 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1194 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1195 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1196 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1197 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1198 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1199 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001200 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1201 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1202 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1203 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1204 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1205 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1206 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1207 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001208 "src/qs8-requantization/precise-neon.c",
1209 "src/qs8-requantization/fp32-neon.c",
1210 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001211 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1212 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1213 "src/qu8-dwconv/up8x9-minmax-neon.c",
1214 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1215 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1216 "src/qu8-gemm/4x8-minmax-neon.c",
1217 "src/qu8-gemm/8x8-minmax-neon.c",
1218 "src/qu8-igemm/4x8-minmax-neon.c",
1219 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001220 "src/qu8-requantization/precise-neon.c",
1221 "src/qu8-requantization/fp32-neon.c",
1222 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001223 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001224 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001225 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001226 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001227 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001228 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001229 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001230 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001231 "src/x32-zip/x2-neon.c",
1232 "src/x32-zip/x3-neon.c",
1233 "src/x32-zip/x4-neon.c",
1234 "src/x32-zip/xm-neon.c",
1235 "src/x8-zip/x2-neon.c",
1236 "src/x8-zip/x3-neon.c",
1237 "src/x8-zip/x4-neon.c",
1238 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001239 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001240 "src/math/roundd-neon-addsub.c",
1241 "src/math/roundd-neon-cvt.c",
1242 "src/math/roundu-neon-addsub.c",
1243 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001244 "src/math/roundz-neon-addsub.c",
1245 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001246 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001247 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1248 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1249 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1250 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1251 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1252 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001253 "src/math/sqrt-neon-nr1rsqrts.c",
1254 "src/math/sqrt-neon-nr2rsqrts.c",
1255 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001256]
1257
1258NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001259 "src/f32-ibilinear/gen/neonfma-c4.c",
1260 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001261 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1262 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1263 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1264 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1265 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1266 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1267 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1268 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1269 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001270 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1271 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1272 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1273 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001274 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1275 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1276 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1277 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001278 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1279 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1280 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1281 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001282 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1283 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1284 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1285 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1286 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1287 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1288 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1289 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1290 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1291 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1292 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1293 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1294 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1295 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1296 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1297 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1298 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1299 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001300 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1301 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001302 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1303 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1304 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1305 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1306 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1307 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1308 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1309 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1310 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1311 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1312 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1313 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1314 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1315 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1316 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1317 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1318 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1319 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1320 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1321 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1322 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1323 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1324 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1325 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001326 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001380 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1381 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001382 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1383 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1384 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1385 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1386 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1387 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1388 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1389 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1390 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1391 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1392 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1393 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1394 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1395 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1396 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1397 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1398 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1399 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1400 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1401 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001402 "src/math/exp-neonfma-lut64-p2.c",
1403 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001404 "src/math/expminus-neonfma-lut2048-p1.c",
1405 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001406 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001407 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1408 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1409 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1410 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1411 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1412 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1413 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1414 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1415 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1416 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1417 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1418 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1419 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1420 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1421 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1422 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1423 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1424 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001425 "src/math/sqrt-neonfma-nr1fma.c",
1426 "src/math/sqrt-neonfma-nr2fma.c",
1427 "src/math/sqrt-neonfma-nr3fma.c",
1428 "src/math/sqrt-neonfma-nr2fma1adj.c",
1429 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001430]
1431
1432AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001433 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1434 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1435 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1436 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1437 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1438 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001439 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1440 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1441 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1442 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1443 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1444 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1445 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1446 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1447 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1448 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1449 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1450 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1451 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1452 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1453 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1454 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1455 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1456 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1457 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1458 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001459 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1460 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001461 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1462 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001463 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1464 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001465 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1466 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001467 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1468 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1469 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1470 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1471 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001472 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001490 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1491 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1492 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1493 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1494 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1495 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1496 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1497 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1498 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1499 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1500 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1501 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1502 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1503 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1504 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1505 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1506 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1507 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001508 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1509 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001510 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1511 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1512 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1513 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1514 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1515 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001516]
1517
Marat Dukhan8853b822020-05-07 12:19:01 -07001518NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001519 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1520 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1521 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1522 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1523 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1524 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1525 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1526 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001527 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001528 "src/math/roundd-neonv8.c",
1529 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001530 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001531]
1532
Marat Dukhan08c4a432019-10-03 09:29:21 -07001533AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001534 "src/f16-clamp/gen/neonfp16arith-x16.c",
1535 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001536 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1537 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1538 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1539 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1540 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1541 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1542 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1543 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1544 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1545 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1546 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1547 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001548 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1549 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001550 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1551 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001552 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001553 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001554 "src/f16-relu/gen/neonfp16arith-x16.c",
1555 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001556 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1557 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1558 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1559 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1560 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1561 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1562 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1563 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1564 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1565 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1566 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1567 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1568 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1569 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1570 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1571 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1572 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1573 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1574 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1575 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1576 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1577 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1578 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1579 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1580 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1581 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1582 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1583 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001584 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1585 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001586 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1587 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1588 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1589 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1590 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1591 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1592 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1593 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001594 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1595 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1596 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1597 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001598 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1599 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1600 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1601 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1602 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1603 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1604 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1605 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1606 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1607 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1608 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1609 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001610 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1611 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1612 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1613 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1614 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1615 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1616 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1617 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618]
1619
1620SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001621 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1622 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001623 "src/f32-clamp/gen/sse-x4.c",
1624 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001625 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1626 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001627 "src/f32-dwconv-chw/3x3p1-sse.c",
1628 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001629 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1630 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1631 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1632 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1633 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1634 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1635 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1636 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1637 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1638 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1639 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1640 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001641 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001642 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1643 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001644 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1645 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1646 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1647 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1648 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1649 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1650 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1651 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1652 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1653 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1654 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1655 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1656 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001657 "src/f32-hswish/gen/hswish-sse-x4.c",
1658 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001659 "src/f32-ibilinear/gen/sse-c4.c",
1660 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001661 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1662 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1663 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1664 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1665 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1666 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1667 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001668 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1669 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1670 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001671 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001672 "src/f32-prelu/gen/sse-2x4.c",
1673 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001674 "src/f32-relu/gen/sse-x4.c",
1675 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001677 "src/f32-spmm/gen/4x1-minmax-sse.c",
1678 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001679 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001680 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1681 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1682 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1683 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1684 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1685 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1686 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1687 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001688 "src/f32-vbinary/gen/vmax-sse-x4.c",
1689 "src/f32-vbinary/gen/vmax-sse-x8.c",
1690 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1691 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1692 "src/f32-vbinary/gen/vmin-sse-x4.c",
1693 "src/f32-vbinary/gen/vmin-sse-x8.c",
1694 "src/f32-vbinary/gen/vminc-sse-x4.c",
1695 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001696 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1697 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1698 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1699 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1700 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1701 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1702 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1703 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001704 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1705 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1706 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1707 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001708 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1709 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1710 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1711 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001712 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1713 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001714 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1715 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001716 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1717 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001718 "src/f32-vunary/gen/vabs-sse-x4.c",
1719 "src/f32-vunary/gen/vabs-sse-x8.c",
1720 "src/f32-vunary/gen/vneg-sse-x4.c",
1721 "src/f32-vunary/gen/vneg-sse-x8.c",
1722 "src/f32-vunary/gen/vsqr-sse-x4.c",
1723 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001724 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001725 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001726 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001727 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001728 "src/math/roundd-sse-addsub.c",
1729 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001730 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001731 "src/math/sqrt-sse-hh1mac.c",
1732 "src/math/sqrt-sse-nr1mac.c",
1733 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734]
1735
1736SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001737 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1738 "src/f32-argmaxpool/4x-sse2-c4.c",
1739 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001740 "src/f32-prelu/gen/sse2-2x4.c",
1741 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001742 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1743 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1744 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1745 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1746 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1747 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1748 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1749 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1750 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1751 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1752 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1753 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001754 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001755 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001756 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001757 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001758 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1759 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001760 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1761 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001762 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1763 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1764 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1765 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1766 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1767 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1768 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1769 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001770 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1771 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1772 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001773 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1774 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1775 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001776 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1777 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1778 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001779 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1780 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001781 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1782 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001783 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1784 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001785 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1786 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001787 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001788 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1789 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001790 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001791 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1792 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1793 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001794 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1795 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001796 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1797 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001798 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1799 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001800 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001801 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1802 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001803 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001804 "src/qs8-requantization/precise-sse2.c",
1805 "src/qs8-requantization/fp32-sse2.c",
1806 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001807 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1808 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1809 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1810 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001811 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1812 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1813 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1814 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1815 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1816 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1817 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1818 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001819 "src/qu8-requantization/precise-sse2.c",
1820 "src/qu8-requantization/fp32-sse2.c",
1821 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001822 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001823 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001824 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001826 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827 "src/x32-zip/x2-sse2.c",
1828 "src/x32-zip/x3-sse2.c",
1829 "src/x32-zip/x4-sse2.c",
1830 "src/x32-zip/xm-sse2.c",
1831 "src/x8-zip/x2-sse2.c",
1832 "src/x8-zip/x3-sse2.c",
1833 "src/x8-zip/x4-sse2.c",
1834 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001835 "src/math/exp-sse2-p5.c",
1836 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001837 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001838 "src/math/roundd-sse2-cvt.c",
1839 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001840 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001841 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001842]
1843
1844SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001845 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1846 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1847 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001848 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1849 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1850 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001851 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1852 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1853 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001854 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1855 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001856 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1857 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001858 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1859 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001860 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1861 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001862 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001863 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1864 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001865 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001866 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1867 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1868 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001869 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1870 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001871 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1872 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001873 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1874 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001875 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001876 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1877 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001878 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001879 "src/qs8-requantization/precise-ssse3.c",
1880 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001881 "src/qu8-requantization/precise-ssse3.c",
1882 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001883]
1884
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001885SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001886 "src/f32-prelu/gen/sse41-2x4.c",
1887 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001888 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001889 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001890 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001891 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001892 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1893 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001894 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1895 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001896 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1897 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1898 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1899 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1900 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1901 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1902 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1903 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001904 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1905 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1906 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001907 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1908 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1909 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001910 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1911 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1912 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001913 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1914 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001915 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1916 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001917 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1918 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001919 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1920 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001921 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001922 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1923 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001924 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001925 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1926 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1927 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001928 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1929 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001930 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1931 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001932 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1933 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001934 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001935 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1936 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001937 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001938 "src/qs8-requantization/fp32-sse4.c",
1939 "src/qs8-requantization/precise-sse4.c",
1940 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001941 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1942 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1943 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1944 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001945 "src/qu8-requantization/precise-sse4.c",
1946 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001947 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001948 "src/math/roundd-sse41.c",
1949 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001950 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001951]
1952
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001954 "src/f32-clamp/gen/avx-x8.c",
1955 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001956 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1957 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1958 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1959 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1960 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1961 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1962 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1963 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1964 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1965 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1966 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1967 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1968 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1969 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1970 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1971 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1972 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1973 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1974 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1975 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1976 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1977 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1978 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1979 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1980 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1981 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1982 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1983 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1984 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1985 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001986 "src/f32-hswish/gen/hswish-avx-x8.c",
1987 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001988 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1989 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1990 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1991 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1992 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1993 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1994 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1995 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1996 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001997 "src/f32-prelu/gen/avx-2x8.c",
1998 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001999 "src/f32-relu/gen/avx-x16.c",
2000 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002002 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2003 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2004 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2005 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2006 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2007 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2008 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2009 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002010 "src/f32-vbinary/gen/vmax-avx-x8.c",
2011 "src/f32-vbinary/gen/vmax-avx-x16.c",
2012 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2013 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2014 "src/f32-vbinary/gen/vmin-avx-x8.c",
2015 "src/f32-vbinary/gen/vmin-avx-x16.c",
2016 "src/f32-vbinary/gen/vminc-avx-x8.c",
2017 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002018 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2019 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2020 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2021 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2022 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2023 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2024 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2025 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002026 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2027 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2028 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2029 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002030 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2031 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2032 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2033 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002034 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2035 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002036 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002037 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2038 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2039 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2040 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2041 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2042 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2043 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2044 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002045 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2046 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002047 "src/f32-vunary/gen/vabs-avx-x8.c",
2048 "src/f32-vunary/gen/vabs-avx-x16.c",
2049 "src/f32-vunary/gen/vneg-avx-x8.c",
2050 "src/f32-vunary/gen/vneg-avx-x16.c",
2051 "src/f32-vunary/gen/vsqr-avx-x8.c",
2052 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002053]
2054
Marat Dukhan1566fee2020-08-02 21:55:41 -07002055XOP_UKERNELS = [
2056 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2057 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2058 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2059 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002060 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2061 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002062 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2063 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002064 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002065 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2066 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002067 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002068 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2069 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2070 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002071 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2072 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2073 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2074 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2075 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2076 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002077 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002078 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2079 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002080 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002081]
2082
Marat Dukhanfda12b82019-11-21 12:27:59 -08002083FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002084 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2085 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2086 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2087 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2088 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2089 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2090 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2091 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2092 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2093 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2094 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2095 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2096 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2097 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2098 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2099 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2100 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2101 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2102 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2103 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2104 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2105 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2106 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2107 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2108 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2109 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2110 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2111 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2112 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2113 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2114 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2115 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2116 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2117 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2118 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2119 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2120 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2121 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2122 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2123 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002124 "src/f32-hswish/gen/hswish-fma3-x8.c",
2125 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002126 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2127 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2128 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2129 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2130 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2131 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2132 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2133 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2134 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2135 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2136 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2137 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2138 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2139 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002140 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2141 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2142 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2143 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2144 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2145 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2146 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2147 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002148 "src/math/sqrt-fma3-nr1fma.c",
2149 "src/math/sqrt-fma3-nr2fma.c",
2150 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002151]
2152
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002153AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002154 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2155 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2156 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2157 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2158 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2159 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2160 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2161 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2162 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2163 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2164 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2165 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2166 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2167 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2168 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2169 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2170 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2171 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2172 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2173 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2174 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2175 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2176 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2177 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2178 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2179 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2180 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2181 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2182 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2183 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2184 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2185 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2186 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2187 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2188 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2189 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002190 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2191 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2192 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2193 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2194 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2195 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2196 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2197 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2198 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2199 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2200 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2201 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2202 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2203 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2204 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2205 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2206 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2207 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2208 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2209 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2210 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2211 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2212 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2213 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2214 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2215 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2216 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2217 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2218 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2219 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002220 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2221 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2222 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2223 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2224 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2225 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2226 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2227 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2228 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2229 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2230 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2231 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2232 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2233 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2234 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2235 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2236 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2237 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2238 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2239 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2240 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2241 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2242 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2243 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002244 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2245 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002246 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2247 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2248 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2249 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002250 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2251 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2252 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002253 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2254 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2255 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002256 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2257 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2258 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002259 "src/math/exp-avx2-p5.c",
2260 "src/math/exp-avx2-perm-p3.c",
2261 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002262 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002263 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002264 "src/math/sigmoid-avx2-rr2-p5-div.c",
2265 "src/math/sigmoid-avx2-rr1-p5-div.c",
2266 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2267 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2268 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2269 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002270]
2271
Marat Dukhan08c4a432019-10-03 09:29:21 -07002272AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002273 "src/f32-clamp/gen/avx512f-x16.c",
2274 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002275 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2276 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2277 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2278 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2279 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2280 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2281 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2282 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2283 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2284 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2285 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2286 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2287 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2288 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2289 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2290 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2291 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2292 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2293 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2294 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2295 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2296 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2297 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2298 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002299 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2300 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002301 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2302 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2303 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2304 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2305 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2306 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002307 "src/f32-prelu/gen/avx512f-2x16.c",
2308 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002309 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2310 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2311 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2312 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2313 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2314 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2315 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2316 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2317 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2318 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2319 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2320 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2321 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2322 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2323 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2324 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2325 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2326 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2327 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2328 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2329 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2330 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2331 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2332 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2333 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2334 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2335 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2336 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2337 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2338 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2339 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2340 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2341 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2342 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2343 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2344 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002345 "src/f32-relu/gen/avx512f-x16.c",
2346 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002347 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002348 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2349 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2350 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2351 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2352 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2353 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2354 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2355 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002356 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2357 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2358 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2359 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2360 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2361 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2362 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2363 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002364 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2365 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2366 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2367 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2368 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2369 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2370 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2371 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002372 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2373 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2374 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2375 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002376 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2377 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2378 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2379 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002380 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2381 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002382 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002383 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2384 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2388 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2389 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2390 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2391 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2392 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2393 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2394 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2395 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2396 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2400 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2401 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2402 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2403 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2404 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2405 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2406 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002407 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2408 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2409 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2410 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2411 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2412 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2413 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2414 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002415 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2416 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2417 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2418 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2419 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2420 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2421 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2422 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002423 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2424 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2425 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2426 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2427 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2428 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002429 "src/math/exp-avx512f-p5-scalef.c",
2430 "src/math/exp-avx512f-p5.c",
2431 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002432 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002433 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002434 "src/math/sqrt-avx512f-nr1fma.c",
2435 "src/math/sqrt-avx512f-nr2fma.c",
2436 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002437]
2438
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002439AVX512SKX_UKERNELS = [
2440 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2441 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2442 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2443 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2444 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2445 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2446 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2447 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2448]
2449
Marat Dukhan08c4a432019-10-03 09:29:21 -07002450AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002451 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002452 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002453 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2454 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002455 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002456 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002457 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002458 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002459 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2460 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002461 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2462 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2463 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2464 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465]
2466
2467AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002468 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002469 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2470 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002471 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002472 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2473 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002474 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2475 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2476 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002477 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002478 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2479 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2480 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002481 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002482 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2483 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002484 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002485 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2486 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2487 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2488 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2489 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2490 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2491 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2492 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2493 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2494 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2495 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2496 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2497 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2498 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2499 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2500 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2501 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2502 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002503 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2504 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002505 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002506 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2507 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2508 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2510 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2511 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2512 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2513 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2517 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2518 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2519 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2520 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2526 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2527 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2528 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2529 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2530 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2531 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2532 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2533 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2534 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2535 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2536 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2537 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2538 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2539 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2540 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2541 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542]
2543
Marat Dukhan1b354632020-03-23 12:50:22 -07002544INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002545 "src/xnnpack/argmaxpool.h",
2546 "src/xnnpack/avgpool.h",
2547 "src/xnnpack/clamp.h",
2548 "src/xnnpack/common.h",
2549 "src/xnnpack/conv.h",
2550 "src/xnnpack/dwconv.h",
2551 "src/xnnpack/gavgpool.h",
2552 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002553 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002555 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002556 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002557 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002558 "src/xnnpack/lut.h",
2559 "src/xnnpack/math.h",
2560 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002561 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 "src/xnnpack/packx.h",
2563 "src/xnnpack/pad.h",
2564 "src/xnnpack/params.h",
2565 "src/xnnpack/pavgpool.h",
2566 "src/xnnpack/ppmm.h",
2567 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002568 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002569 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002570 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002571 "src/xnnpack/rmax.h",
2572 "src/xnnpack/scalar-utils.h",
2573 "src/xnnpack/spmm.h",
2574 "src/xnnpack/unpool.h",
2575 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002576 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002577 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002578 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002579 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002580 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002581 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002582 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002583]
2584
2585INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586 "include/xnnpack.h",
2587 "src/xnnpack/allocator.h",
2588 "src/xnnpack/compute.h",
2589 "src/xnnpack/im2col.h",
2590 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002591 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002592 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593 "src/xnnpack/operator.h",
2594 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002595 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002597 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002598 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002599]
2600
Marat Dukhan1b354632020-03-23 12:50:22 -07002601ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002602 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603]
2604
Marat Dukhan1b354632020-03-23 12:50:22 -07002605MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002606 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 "include/xnnpack.h",
2608]
2609
Marat Dukhan1b354632020-03-23 12:50:22 -07002610MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002612 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 "src/xnnpack/requantization.h",
2614 "include/xnnpack.h",
2615]
2616
2617OPERATOR_TEST_PARAMS_HDRS = [
2618 "src/xnnpack/params.h",
2619 "src/xnnpack/common.h",
2620]
2621
2622WEIGHTS_PACK_HDRS = [
2623 "src/xnnpack/pack.h",
2624 "src/xnnpack/operator.h",
2625 "src/xnnpack/compute.h",
2626]
2627
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002628LOGGING_COPTS = select({
2629 # No logging in optimized mode
2630 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2631 # Full logging in debug mode
2632 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2633 # Error-only logging in default (fastbuild) mode
2634 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2635})
2636
Marat Dukhan3b59de22020-06-03 20:15:19 -07002637LOGGING_SRCS = select({
2638 # No logging in optimized mode
2639 ":optimized_build": [],
2640 "//conditions:default": [
2641 "src/operator-strings.c",
2642 "src/subgraph-strings.c",
2643 ],
2644})
2645
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002646LOGGING_HDRS = [
2647 "src/xnnpack/log.h",
2648]
2649
Marat Dukhan08c4a432019-10-03 09:29:21 -07002650xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002651 name = "tables",
2652 srcs = TABLE_SRCS,
2653 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002654 gcc_copts = xnnpack_gcc_std_copts(),
2655 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002656)
2657
2658xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 name = "scalar_ukernels",
2660 srcs = SCALAR_UKERNELS,
2661 hdrs = INTERNAL_HDRS,
2662 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002663 gcc_copts = xnnpack_gcc_std_copts(),
2664 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002666 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002667 "@FP16",
2668 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002669 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002670 ],
2671)
2672
2673xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002674 name = "scalar_ukernels_test_mode",
2675 srcs = SCALAR_UKERNELS,
2676 hdrs = INTERNAL_HDRS,
2677 aarch32_copts = ["-marm"],
2678 copts = [
2679 "-UNDEBUG",
2680 "-DXNN_TEST_MODE=1",
2681 ],
2682 gcc_copts = xnnpack_gcc_std_copts(),
2683 msvc_copts = xnnpack_msvc_std_copts(),
2684 deps = [
2685 ":tables",
2686 "@FP16",
2687 "@FXdiv",
2688 "@pthreadpool",
2689 ],
2690)
2691
2692xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002693 name = "wasm_ukernels",
2694 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002695 gcc_copts = xnnpack_gcc_std_copts(),
2696 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002697 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002698 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002699 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002700 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002701 "@FP16",
2702 "@FXdiv",
2703 "@pthreadpool",
2704 ],
2705)
2706
2707xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002708 name = "wasm_ukernels_test_mode",
2709 hdrs = INTERNAL_HDRS,
2710 copts = [
2711 "-UNDEBUG",
2712 "-DXNN_TEST_MODE=1",
2713 ],
2714 gcc_copts = xnnpack_gcc_std_copts(),
2715 msvc_copts = xnnpack_msvc_std_copts(),
2716 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002717 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002718 deps = [
2719 ":tables",
2720 "@FP16",
2721 "@FXdiv",
2722 "@pthreadpool",
2723 ],
2724)
2725
2726xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002727 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002728 hdrs = INTERNAL_HDRS,
2729 aarch32_copts = [
2730 "-marm",
2731 "-mfpu=neon",
2732 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002733 gcc_copts = xnnpack_gcc_std_copts(),
2734 gcc_x86_copts = ["-msse2"],
2735 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002736 optimized_copts = [
2737 "-O3",
2738 "-ffast-math",
2739 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002740 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002742 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 "@FP16",
2744 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002745 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746 ],
2747)
2748
2749xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002750 name = "psimd_fastmath_ukernels_test_mode",
2751 hdrs = INTERNAL_HDRS,
2752 aarch32_copts = [
2753 "-marm",
2754 "-mfpu=neon",
2755 ],
2756 copts = [
2757 "-UNDEBUG",
2758 "-DXNN_TEST_MODE=1",
2759 ],
2760 gcc_copts = xnnpack_gcc_std_copts(),
2761 gcc_x86_copts = ["-msse2"],
2762 msvc_copts = xnnpack_msvc_std_copts(),
2763 optimized_copts = [
2764 "-O3",
2765 "-ffast-math",
2766 ],
2767 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2768 deps = [
2769 ":tables",
2770 "@FP16",
2771 "@psimd",
2772 "@pthreadpool",
2773 ],
2774)
2775
2776xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002777 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002778 hdrs = INTERNAL_HDRS,
2779 aarch32_copts = [
2780 "-marm",
2781 "-mfpu=neon",
2782 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002783 gcc_copts = xnnpack_gcc_std_copts(),
2784 gcc_x86_copts = ["-msse2"],
2785 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002786 optimized_copts = [
2787 "-O3",
2788 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002789 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002790 deps = [
2791 ":tables",
2792 "@FP16",
2793 "@psimd",
2794 "@pthreadpool",
2795 ],
2796)
2797
2798xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002799 name = "psimd_accmath_ukernels_test_mode",
2800 hdrs = INTERNAL_HDRS,
2801 aarch32_copts = [
2802 "-marm",
2803 "-mfpu=neon",
2804 ],
2805 copts = [
2806 "-UNDEBUG",
2807 "-DXNN_TEST_MODE=1",
2808 ],
2809 gcc_copts = xnnpack_gcc_std_copts(),
2810 gcc_x86_copts = ["-msse2"],
2811 msvc_copts = xnnpack_msvc_std_copts(),
2812 optimized_copts = [
2813 "-O3",
2814 ],
2815 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2816 deps = [
2817 ":tables",
2818 "@FP16",
2819 "@psimd",
2820 "@pthreadpool",
2821 ],
2822)
2823
2824xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002825 name = "neon_ukernels",
2826 hdrs = INTERNAL_HDRS,
2827 aarch32_copts = [
2828 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002829 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002830 "-mfpu=neon",
2831 ],
2832 aarch32_srcs = NEON_UKERNELS,
2833 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002834 gcc_copts = xnnpack_gcc_std_copts(),
2835 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002836 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002837 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002838 "@FP16",
2839 "@pthreadpool",
2840 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002841)
2842
2843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002844 name = "neon_ukernels_test_mode",
2845 hdrs = INTERNAL_HDRS,
2846 aarch32_copts = [
2847 "-marm",
2848 "-march=armv7-a",
2849 "-mfpu=neon",
2850 ],
2851 aarch32_srcs = NEON_UKERNELS,
2852 aarch64_srcs = NEON_UKERNELS,
2853 copts = [
2854 "-UNDEBUG",
2855 "-DXNN_TEST_MODE=1",
2856 ],
2857 gcc_copts = xnnpack_gcc_std_copts(),
2858 msvc_copts = xnnpack_msvc_std_copts(),
2859 deps = [
2860 ":tables",
2861 "@FP16",
2862 "@pthreadpool",
2863 ],
2864)
2865
2866xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 name = "neonfma_ukernels",
2868 hdrs = INTERNAL_HDRS,
2869 aarch32_copts = [
2870 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002871 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 "-mfpu=neon-vfpv4",
2873 ],
2874 aarch32_srcs = NEONFMA_UKERNELS,
2875 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002876 apple_aarch32_copts = [
2877 "-mcpu=swift",
2878 "-mtune=generic",
2879 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002880 gcc_copts = xnnpack_gcc_std_copts(),
2881 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002882 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002883 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002884 "@FP16",
2885 "@pthreadpool",
2886 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887)
2888
2889xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002890 name = "neonfma_ukernels_test_mode",
2891 hdrs = INTERNAL_HDRS,
2892 aarch32_copts = [
2893 "-marm",
2894 "-march=armv7-a",
2895 "-mfpu=neon-vfpv4",
2896 ],
2897 aarch32_srcs = NEONFMA_UKERNELS,
2898 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002899 apple_aarch32_copts = [
2900 "-mcpu=swift",
2901 "-mtune=generic",
2902 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002903 copts = [
2904 "-UNDEBUG",
2905 "-DXNN_TEST_MODE=1",
2906 ],
2907 gcc_copts = xnnpack_gcc_std_copts(),
2908 msvc_copts = xnnpack_msvc_std_copts(),
2909 deps = [
2910 ":tables",
2911 "@FP16",
2912 "@pthreadpool",
2913 ],
2914)
2915
2916xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002917 name = "neonv8_ukernels",
2918 hdrs = INTERNAL_HDRS,
2919 aarch32_copts = [
2920 "-marm",
2921 "-march=armv8-a",
2922 "-mfpu=neon-fp-armv8",
2923 ],
2924 aarch32_srcs = NEONV8_UKERNELS,
2925 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002926 apple_aarch32_copts = [
2927 "-mcpu=cyclone",
2928 "-mtune=generic",
2929 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002930 gcc_copts = xnnpack_gcc_std_copts(),
2931 msvc_copts = xnnpack_msvc_std_copts(),
2932 deps = [
2933 ":tables",
2934 "@FP16",
2935 "@pthreadpool",
2936 ],
2937)
2938
2939xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002940 name = "neonv8_ukernels_test_mode",
2941 hdrs = INTERNAL_HDRS,
2942 aarch32_copts = [
2943 "-marm",
2944 "-march=armv8-a",
2945 "-mfpu=neon-fp-armv8",
2946 ],
2947 aarch32_srcs = NEONV8_UKERNELS,
2948 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002949 apple_aarch32_copts = [
2950 "-mcpu=cyclone",
2951 "-mtune=generic",
2952 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002953 copts = [
2954 "-UNDEBUG",
2955 "-DXNN_TEST_MODE=1",
2956 ],
2957 gcc_copts = xnnpack_gcc_std_copts(),
2958 msvc_copts = xnnpack_msvc_std_copts(),
2959 deps = [
2960 ":tables",
2961 "@FP16",
2962 "@pthreadpool",
2963 ],
2964)
2965
2966xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 name = "neonfp16arith_ukernels",
2968 hdrs = INTERNAL_HDRS,
2969 aarch64_copts = ["-march=armv8.2-a+fp16"],
2970 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002971 gcc_copts = xnnpack_gcc_std_copts(),
2972 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002973 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002974 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002975 "@FP16",
2976 "@pthreadpool",
2977 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002978)
2979
2980xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002981 name = "neonfp16arith_ukernels_test_mode",
2982 hdrs = INTERNAL_HDRS,
2983 aarch64_copts = ["-march=armv8.2-a+fp16"],
2984 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2985 copts = [
2986 "-UNDEBUG",
2987 "-DXNN_TEST_MODE=1",
2988 ],
2989 gcc_copts = xnnpack_gcc_std_copts(),
2990 msvc_copts = xnnpack_msvc_std_copts(),
2991 deps = [
2992 ":tables",
2993 "@FP16",
2994 "@pthreadpool",
2995 ],
2996)
2997
2998xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 name = "sse2_ukernels",
3000 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003001 gcc_copts = xnnpack_gcc_std_copts(),
3002 gcc_x86_copts = ["-msse2"],
3003 msvc_copts = xnnpack_msvc_std_copts(),
3004 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003006 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003007 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003008 "@FP16",
3009 "@pthreadpool",
3010 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011)
3012
3013xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003014 name = "sse2_ukernels_test_mode",
3015 hdrs = INTERNAL_HDRS,
3016 copts = [
3017 "-UNDEBUG",
3018 "-DXNN_TEST_MODE=1",
3019 ],
3020 gcc_copts = xnnpack_gcc_std_copts(),
3021 gcc_x86_copts = ["-msse2"],
3022 msvc_copts = xnnpack_msvc_std_copts(),
3023 msvc_x86_32_copts = ["/arch:SSE2"],
3024 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3025 deps = [
3026 ":tables",
3027 "@FP16",
3028 "@pthreadpool",
3029 ],
3030)
3031
3032xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003033 name = "ssse3_ukernels",
3034 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003035 gcc_copts = xnnpack_gcc_std_copts(),
3036 gcc_x86_copts = ["-mssse3"],
3037 msvc_copts = xnnpack_msvc_std_copts(),
3038 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003039 x86_srcs = SSSE3_UKERNELS,
3040 deps = [
3041 ":tables",
3042 "@FP16",
3043 "@pthreadpool",
3044 ],
3045)
3046
3047xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003048 name = "ssse3_ukernels_test_mode",
3049 hdrs = INTERNAL_HDRS,
3050 copts = [
3051 "-UNDEBUG",
3052 "-DXNN_TEST_MODE=1",
3053 ],
3054 gcc_copts = xnnpack_gcc_std_copts(),
3055 gcc_x86_copts = ["-mssse3"],
3056 msvc_copts = xnnpack_msvc_std_copts(),
3057 msvc_x86_32_copts = ["/arch:SSE2"],
3058 x86_srcs = SSSE3_UKERNELS,
3059 deps = [
3060 ":tables",
3061 "@FP16",
3062 "@pthreadpool",
3063 ],
3064)
3065
3066xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003067 name = "sse41_ukernels",
3068 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003069 gcc_copts = xnnpack_gcc_std_copts(),
3070 gcc_x86_copts = ["-msse4.1"],
3071 msvc_copts = xnnpack_msvc_std_copts(),
3072 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003073 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003074 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003075 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003076 "@FP16",
3077 "@pthreadpool",
3078 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003079)
3080
3081xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003082 name = "sse41_ukernels_test_mode",
3083 hdrs = INTERNAL_HDRS,
3084 copts = [
3085 "-UNDEBUG",
3086 "-DXNN_TEST_MODE=1",
3087 ],
3088 gcc_copts = xnnpack_gcc_std_copts(),
3089 gcc_x86_copts = ["-msse4.1"],
3090 msvc_copts = xnnpack_msvc_std_copts(),
3091 msvc_x86_32_copts = ["/arch:SSE2"],
3092 x86_srcs = SSE41_UKERNELS,
3093 deps = [
3094 ":tables",
3095 "@FP16",
3096 "@pthreadpool",
3097 ],
3098)
3099
3100xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101 name = "avx_ukernels",
3102 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003103 gcc_copts = xnnpack_gcc_std_copts(),
3104 gcc_x86_copts = ["-mavx"],
3105 msvc_copts = xnnpack_msvc_std_copts(),
3106 msvc_x86_32_copts = ["/arch:AVX"],
3107 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003109 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003110 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003111 "@FP16",
3112 "@pthreadpool",
3113 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114)
3115
3116xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003117 name = "avx_ukernels_test_mode",
3118 hdrs = INTERNAL_HDRS,
3119 copts = [
3120 "-UNDEBUG",
3121 "-DXNN_TEST_MODE=1",
3122 ],
3123 gcc_copts = xnnpack_gcc_std_copts(),
3124 gcc_x86_copts = ["-mavx"],
3125 msvc_copts = xnnpack_msvc_std_copts(),
3126 msvc_x86_32_copts = ["/arch:AVX"],
3127 msvc_x86_64_copts = ["/arch:AVX"],
3128 x86_srcs = AVX_UKERNELS,
3129 deps = [
3130 ":tables",
3131 "@FP16",
3132 "@pthreadpool",
3133 ],
3134)
3135
3136xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003137 name = "xop_ukernels",
3138 hdrs = INTERNAL_HDRS,
3139 gcc_copts = xnnpack_gcc_std_copts(),
3140 gcc_x86_copts = ["-mxop"],
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 msvc_x86_32_copts = ["/arch:AVX"],
3143 msvc_x86_64_copts = ["/arch:AVX"],
3144 x86_srcs = XOP_UKERNELS,
3145 deps = [
3146 ":tables",
3147 "@FP16",
3148 "@pthreadpool",
3149 ],
3150)
3151
3152xnnpack_cc_library(
3153 name = "xop_ukernels_test_mode",
3154 hdrs = INTERNAL_HDRS,
3155 copts = [
3156 "-UNDEBUG",
3157 "-DXNN_TEST_MODE=1",
3158 ],
3159 gcc_copts = xnnpack_gcc_std_copts(),
3160 gcc_x86_copts = ["-mxop"],
3161 msvc_copts = xnnpack_msvc_std_copts(),
3162 msvc_x86_32_copts = ["/arch:AVX"],
3163 msvc_x86_64_copts = ["/arch:AVX"],
3164 x86_srcs = XOP_UKERNELS,
3165 deps = [
3166 ":tables",
3167 "@FP16",
3168 "@pthreadpool",
3169 ],
3170)
3171
3172xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003173 name = "fma3_ukernels",
3174 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003175 gcc_copts = xnnpack_gcc_std_copts(),
3176 gcc_x86_copts = ["-mfma"],
3177 msvc_copts = xnnpack_msvc_std_copts(),
3178 msvc_x86_32_copts = ["/arch:AVX"],
3179 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003180 x86_srcs = FMA3_UKERNELS,
3181 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003182 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003183 "@FP16",
3184 "@pthreadpool",
3185 ],
3186)
3187
3188xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003189 name = "fma3_ukernels_test_mode",
3190 hdrs = INTERNAL_HDRS,
3191 copts = [
3192 "-UNDEBUG",
3193 "-DXNN_TEST_MODE=1",
3194 ],
3195 gcc_copts = xnnpack_gcc_std_copts(),
3196 gcc_x86_copts = ["-mfma"],
3197 msvc_copts = xnnpack_msvc_std_copts(),
3198 msvc_x86_32_copts = ["/arch:AVX"],
3199 msvc_x86_64_copts = ["/arch:AVX"],
3200 x86_srcs = FMA3_UKERNELS,
3201 deps = [
3202 ":tables",
3203 "@FP16",
3204 "@pthreadpool",
3205 ],
3206)
3207
3208xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003209 name = "avx2_ukernels",
3210 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003211 gcc_copts = xnnpack_gcc_std_copts(),
3212 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003213 "-mfma",
3214 "-mavx2",
3215 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003216 msvc_copts = xnnpack_msvc_std_copts(),
3217 msvc_x86_32_copts = ["/arch:AVX2"],
3218 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003219 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003220 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003221 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003222 "@FP16",
3223 "@pthreadpool",
3224 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003225)
3226
3227xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003228 name = "avx2_ukernels_test_mode",
3229 hdrs = INTERNAL_HDRS,
3230 copts = [
3231 "-UNDEBUG",
3232 "-DXNN_TEST_MODE=1",
3233 ],
3234 gcc_copts = xnnpack_gcc_std_copts(),
3235 gcc_x86_copts = [
3236 "-mfma",
3237 "-mavx2",
3238 ],
3239 msvc_copts = xnnpack_msvc_std_copts(),
3240 msvc_x86_32_copts = ["/arch:AVX2"],
3241 msvc_x86_64_copts = ["/arch:AVX2"],
3242 x86_srcs = AVX2_UKERNELS,
3243 deps = [
3244 ":tables",
3245 "@FP16",
3246 "@pthreadpool",
3247 ],
3248)
3249
3250xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003251 name = "avx512f_ukernels",
3252 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003253 gcc_copts = xnnpack_gcc_std_copts(),
3254 gcc_x86_copts = ["-mavx512f"],
3255 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3256 msvc_copts = xnnpack_msvc_std_copts(),
3257 msvc_x86_32_copts = ["/arch:AVX512"],
3258 msvc_x86_64_copts = ["/arch:AVX512"],
3259 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003261 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003262 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003263 "@FP16",
3264 "@pthreadpool",
3265 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266)
3267
3268xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003269 name = "avx512f_ukernels_test_mode",
3270 hdrs = INTERNAL_HDRS,
3271 copts = [
3272 "-UNDEBUG",
3273 "-DXNN_TEST_MODE=1",
3274 ],
3275 gcc_copts = xnnpack_gcc_std_copts(),
3276 gcc_x86_copts = ["-mavx512f"],
3277 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3278 msvc_copts = xnnpack_msvc_std_copts(),
3279 msvc_x86_32_copts = ["/arch:AVX512"],
3280 msvc_x86_64_copts = ["/arch:AVX512"],
3281 msys_copts = ["-fno-asynchronous-unwind-tables"],
3282 x86_srcs = AVX512F_UKERNELS,
3283 deps = [
3284 ":tables",
3285 "@FP16",
3286 "@pthreadpool",
3287 ],
3288)
3289
3290xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003291 name = "avx512skx_ukernels",
3292 hdrs = INTERNAL_HDRS,
3293 gcc_copts = xnnpack_gcc_std_copts(),
3294 gcc_x86_copts = [
3295 "-mavx512f",
3296 "-mavx512cd",
3297 "-mavx512bw",
3298 "-mavx512dq",
3299 "-mavx512vl",
3300 ],
3301 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3302 msvc_copts = xnnpack_msvc_std_copts(),
3303 msvc_x86_32_copts = ["/arch:AVX512"],
3304 msvc_x86_64_copts = ["/arch:AVX512"],
3305 msys_copts = ["-fno-asynchronous-unwind-tables"],
3306 x86_srcs = AVX512SKX_UKERNELS,
3307 deps = [
3308 ":tables",
3309 "@FP16",
3310 "@pthreadpool",
3311 ],
3312)
3313
3314xnnpack_cc_library(
3315 name = "avx512skx_ukernels_test_mode",
3316 hdrs = INTERNAL_HDRS,
3317 copts = [
3318 "-UNDEBUG",
3319 "-DXNN_TEST_MODE=1",
3320 ],
3321 gcc_copts = xnnpack_gcc_std_copts(),
3322 gcc_x86_copts = [
3323 "-mavx512f",
3324 "-mavx512cd",
3325 "-mavx512bw",
3326 "-mavx512dq",
3327 "-mavx512vl",
3328 ],
3329 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3330 msvc_copts = xnnpack_msvc_std_copts(),
3331 msvc_x86_32_copts = ["/arch:AVX512"],
3332 msvc_x86_64_copts = ["/arch:AVX512"],
3333 msys_copts = ["-fno-asynchronous-unwind-tables"],
3334 x86_srcs = AVX512SKX_UKERNELS,
3335 deps = [
3336 ":tables",
3337 "@FP16",
3338 "@pthreadpool",
3339 ],
3340)
3341
3342xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003343 name = "asm_ukernels",
3344 hdrs = ["src/xnnpack/assembly.h"],
3345 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003346 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 aarch64_srcs = AARCH64_ASM_UKERNELS,
3348)
3349
Marat Dukhan3b59de22020-06-03 20:15:19 -07003350xnnpack_cc_library(
3351 name = "logging_utils",
3352 srcs = LOGGING_SRCS,
3353 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3354 copts = LOGGING_COPTS + [
3355 "-Isrc",
3356 "-Iinclude",
3357 ] + select({
3358 ":debug_build": [],
3359 "//conditions:default": xnnpack_min_size_copts(),
3360 }),
3361 gcc_copts = xnnpack_gcc_std_copts(),
3362 msvc_copts = xnnpack_msvc_std_copts(),
3363 visibility = xnnpack_visibility(),
3364 deps = [
3365 "@FP16",
3366 "@clog",
3367 "@pthreadpool",
3368 ],
3369)
3370
Marat Dukhan08c4a432019-10-03 09:29:21 -07003371xnnpack_aggregate_library(
3372 name = "ukernels",
3373 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003374 ":neon_ukernels",
3375 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003376 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 ":asm_ukernels",
3378 ],
3379 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 ":neon_ukernels",
3381 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003382 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 ":neonfp16arith_ukernels",
3384 ":asm_ukernels",
3385 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003386 generic_deps = [
3387 ":scalar_ukernels",
3388 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003389 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003390 ":psimd_fastmath_ukernels",
3391 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003393 wasm_deps = [
3394 ":wasm_ukernels",
3395 ],
3396 wasmsimd_deps = [
3397 ":wasm_ukernels",
3398 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003400 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003401 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003402 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003403 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003404 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003405 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003406 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003407 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003408 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 ],
3410)
3411
Marat Dukhan33fcf782020-05-24 14:27:15 -07003412xnnpack_aggregate_library(
3413 name = "ukernels_test_mode",
3414 aarch32_deps = [
3415 ":neon_ukernels_test_mode",
3416 ":neonfma_ukernels_test_mode",
3417 ":neonv8_ukernels_test_mode",
3418 ":asm_ukernels",
3419 ],
3420 aarch64_deps = [
3421 ":neon_ukernels_test_mode",
3422 ":neonfma_ukernels_test_mode",
3423 ":neonv8_ukernels_test_mode",
3424 ":neonfp16arith_ukernels_test_mode",
3425 ":asm_ukernels",
3426 ],
3427 generic_deps = [
3428 ":scalar_ukernels_test_mode",
3429 ],
3430 psimd_deps = [
3431 ":psimd_fastmath_ukernels_test_mode",
3432 ":psimd_accmath_ukernels_test_mode",
3433 ],
3434 wasm_deps = [
3435 ":wasm_ukernels_test_mode",
3436 ],
3437 wasmsimd_deps = [
3438 ":wasm_ukernels_test_mode",
3439 ],
3440 x86_deps = [
3441 ":sse2_ukernels_test_mode",
3442 ":ssse3_ukernels_test_mode",
3443 ":sse41_ukernels_test_mode",
3444 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003445 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003446 ":fma3_ukernels_test_mode",
3447 ":avx2_ukernels_test_mode",
3448 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003449 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003450 ],
3451)
3452
Marat Dukhan86df3842020-06-17 14:03:31 -07003453# This target must be compatible with Android NDK r17 and compile with Clang 6.
3454# It is similar to :ukernels target, but excludes microkernels which require
3455# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003456xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003457 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003458 aarch32_deps = [
3459 ":neon_ukernels",
3460 ":neonfma_ukernels",
3461 ":neonv8_ukernels",
3462 ":asm_ukernels",
3463 ],
3464 aarch64_deps = [
3465 ":neon_ukernels",
3466 ":neonfma_ukernels",
3467 ":neonv8_ukernels",
3468 ":asm_ukernels",
3469 ],
3470 generic_deps = [
3471 ":scalar_ukernels",
3472 ],
3473 psimd_deps = [
3474 ":psimd_fastmath_ukernels",
3475 ":psimd_accmath_ukernels",
3476 ],
3477 wasm_deps = [
3478 ":wasm_ukernels",
3479 ],
3480 wasmsimd_deps = [
3481 ":wasm_ukernels",
3482 ],
3483 x86_deps = [
3484 ":sse2_ukernels",
3485 ":ssse3_ukernels",
3486 ":sse41_ukernels",
3487 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003488 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003489 ":fma3_ukernels",
3490 ":avx2_ukernels",
3491 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003492 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003493 ],
3494)
3495
Marat Dukhan08c4a432019-10-03 09:29:21 -07003496xnnpack_cc_library(
3497 name = "im2col",
3498 srcs = ["src/im2col.c"],
3499 hdrs = [
3500 "src/xnnpack/common.h",
3501 "src/xnnpack/im2col.h",
3502 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003503 gcc_copts = xnnpack_gcc_std_copts(),
3504 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505)
3506
3507xnnpack_cc_library(
3508 name = "indirection",
3509 srcs = ["src/indirection.c"],
3510 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003511 gcc_copts = xnnpack_gcc_std_copts(),
3512 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 deps = [
3514 "@FP16",
3515 "@FXdiv",
3516 "@pthreadpool",
3517 ],
3518)
3519
3520xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003521 name = "indirection_test_mode",
3522 srcs = ["src/indirection.c"],
3523 hdrs = INTERNAL_HDRS,
3524 copts = [
3525 "-UNDEBUG",
3526 "-DXNN_TEST_MODE=1",
3527 ],
3528 gcc_copts = xnnpack_gcc_std_copts(),
3529 msvc_copts = xnnpack_msvc_std_copts(),
3530 deps = [
3531 "@FP16",
3532 "@FXdiv",
3533 "@pthreadpool",
3534 ],
3535)
3536
3537xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003538 name = "packing",
3539 srcs = ["src/packing.c"],
3540 hdrs = INTERNAL_HDRS,
3541 gcc_copts = xnnpack_gcc_std_copts(),
3542 msvc_copts = xnnpack_msvc_std_copts(),
3543 deps = [
3544 "@FP16",
3545 "@FXdiv",
3546 "@pthreadpool",
3547 ],
3548)
3549
3550xnnpack_cc_library(
3551 name = "packing_test_mode",
3552 srcs = ["src/packing.c"],
3553 hdrs = INTERNAL_HDRS,
3554 copts = [
3555 "-UNDEBUG",
3556 "-DXNN_TEST_MODE=1",
3557 ],
3558 gcc_copts = xnnpack_gcc_std_copts(),
3559 msvc_copts = xnnpack_msvc_std_copts(),
3560 deps = [
3561 "@FP16",
3562 "@FXdiv",
3563 "@pthreadpool",
3564 ],
3565)
3566
3567xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 name = "operator_run",
3569 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003570 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003571 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003572 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3573 "//conditions:default": [],
3574 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003575 gcc_copts = xnnpack_gcc_std_copts(),
3576 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003578 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 "@FP16",
3580 "@FXdiv",
3581 "@clog",
3582 "@pthreadpool",
3583 ],
3584)
3585
Chao Mei6ddfc602020-05-13 22:29:36 -07003586xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003587 name = "operator_run_test_mode",
3588 srcs = ["src/operator-run.c"],
3589 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3590 copts = LOGGING_COPTS + [
3591 "-UNDEBUG",
3592 "-DXNN_TEST_MODE=1",
3593 ] + select({
3594 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3595 "//conditions:default": [],
3596 }),
3597 gcc_copts = xnnpack_gcc_std_copts(),
3598 msvc_copts = xnnpack_msvc_std_copts(),
3599 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003600 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003601 "@FP16",
3602 "@FXdiv",
3603 "@clog",
3604 "@pthreadpool",
3605 ],
3606)
3607
3608xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003609 name = "memory_planner",
3610 srcs = ["src/memory-planner.c"],
3611 hdrs = INTERNAL_HDRS,
3612 defines = select({
3613 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3614 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3615 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3616 }),
3617 gcc_copts = xnnpack_gcc_std_copts(),
3618 msvc_copts = xnnpack_msvc_std_copts(),
3619 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003620 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003621 "@pthreadpool",
3622 ],
3623)
3624
Marat Dukhan33fcf782020-05-24 14:27:15 -07003625xnnpack_cc_library(
3626 name = "memory_planner_test_mode",
3627 srcs = ["src/memory-planner.c"],
3628 hdrs = INTERNAL_HDRS,
3629 copts = [
3630 "-UNDEBUG",
3631 "-DXNN_TEST_MODE=1",
3632 ],
3633 defines = select({
3634 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3635 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3636 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3637 }),
3638 gcc_copts = xnnpack_gcc_std_copts(),
3639 msvc_copts = xnnpack_msvc_std_copts(),
3640 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003641 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003642 "@pthreadpool",
3643 ],
3644)
3645
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646cc_library(
3647 name = "enable_assembly",
3648 defines = select({
3649 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3650 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003651 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 }),
3653)
3654
Marat Dukhan9de90e02020-06-18 16:04:12 -07003655cc_library(
3656 name = "enable_sparse",
3657 defines = select({
3658 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3659 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003660 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003661 }),
3662)
3663
Marat Dukhancf056b22019-10-07 10:26:29 -07003664xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665 name = "operators",
3666 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003667 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003668 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003669 ],
3670 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003671 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672 "-Isrc",
3673 "-Iinclude",
3674 ] + select({
3675 ":debug_build": [],
3676 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003677 }) + select({
3678 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3679 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003680 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003681 gcc_copts = xnnpack_gcc_std_copts(),
3682 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003685 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003686 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003687 "@FP16",
3688 "@FXdiv",
3689 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003691 ],
3692)
3693
Marat Dukhan10a38082020-04-17 03:58:35 -07003694xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003695 name = "operators_test_mode",
3696 srcs = OPERATOR_SRCS + [
3697 "src/memory.c",
3698 "src/operator-delete.c",
3699 ],
3700 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3701 copts = LOGGING_COPTS + [
3702 "-Isrc",
3703 "-Iinclude",
3704 "-UNDEBUG",
3705 "-DXNN_TEST_MODE=1",
3706 ] + select({
3707 ":debug_build": [],
3708 "//conditions:default": xnnpack_min_size_copts(),
3709 }) + select({
3710 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3711 "//conditions:default": [],
3712 }),
3713 gcc_copts = xnnpack_gcc_std_copts(),
3714 msvc_copts = xnnpack_msvc_std_copts(),
3715 deps = [
3716 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003717 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003718 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003719 "@FP16",
3720 "@FXdiv",
3721 "@clog",
3722 "@pthreadpool",
3723 ],
3724)
3725
3726xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003727 name = "XNNPACK",
3728 srcs = [
3729 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003730 "src/runtime.c",
3731 "src/subgraph.c",
3732 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003733 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003734 hdrs = ["include/xnnpack.h"],
3735 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003736 "-Isrc",
3737 "-Iinclude",
3738 ] + select({
3739 ":debug_build": [],
3740 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003741 }) + select({
3742 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3743 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003744 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003745 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003746 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003747 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003748 visibility = xnnpack_visibility(),
3749 deps = [
3750 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003751 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003752 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003753 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003754 ":operator_run",
3755 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003756 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003757 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003758 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003759 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003760 ] + select({
3761 ":emscripten": [],
3762 "//conditions:default": ["@cpuinfo"],
3763 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764)
3765
Marat Dukhan10a38082020-04-17 03:58:35 -07003766xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003767 name = "XNNPACK_test_mode",
3768 srcs = [
3769 "src/init.c",
3770 "src/runtime.c",
3771 "src/subgraph.c",
3772 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003773 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003774 hdrs = ["include/xnnpack.h"],
3775 copts = LOGGING_COPTS + [
3776 "-Isrc",
3777 "-Iinclude",
3778 "-UNDEBUG",
3779 "-DXNN_TEST_MODE=1",
3780 ] + select({
3781 ":debug_build": [],
3782 "//conditions:default": xnnpack_min_size_copts(),
3783 }) + select({
3784 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3785 "//conditions:default": [],
3786 }),
3787 gcc_copts = xnnpack_gcc_std_copts(),
3788 includes = ["include"],
3789 msvc_copts = xnnpack_msvc_std_copts(),
3790 visibility = xnnpack_visibility(),
3791 deps = [
3792 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003793 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003794 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003795 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003796 ":operator_run_test_mode",
3797 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003798 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003799 "@clog",
3800 "@FP16",
3801 "@pthreadpool",
3802 ] + select({
3803 ":emscripten": [],
3804 "//conditions:default": ["@cpuinfo"],
3805 }),
3806)
3807
3808xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003809 name = "xnnpack_f32",
3810 srcs = [
3811 "src/init.c",
3812 "src/runtime.c",
3813 "src/subgraph.c",
3814 "src/tensor.c",
3815 ] + SUBGRAPH_SRCS,
3816 hdrs = ["include/xnnpack.h"],
3817 copts = LOGGING_COPTS + [
3818 "-Isrc",
3819 "-Iinclude",
3820 ] + select({
3821 ":debug_build": [],
3822 "//conditions:default": xnnpack_min_size_copts(),
3823 }) + select({
3824 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3825 "//conditions:default": [],
3826 }),
3827 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003828 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003829 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003830 "XNN_NO_U8_OPERATORS",
3831 "XNN_NO_X8_OPERATORS",
3832 "XNN_NO_F16_OPERATORS",
3833 "XNN_NO_X16_OPERATORS",
3834 ],
3835 gcc_copts = xnnpack_gcc_std_copts(),
3836 includes = ["include"],
3837 msvc_copts = xnnpack_msvc_std_copts(),
3838 visibility = xnnpack_visibility(),
3839 deps = [
3840 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003841 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003842 ":logging_utils",
3843 ":memory_planner",
3844 ":operator_run",
3845 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003846 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003847 "@clog",
3848 "@FP16",
3849 "@pthreadpool",
3850 ] + select({
3851 ":emscripten": [],
3852 "//conditions:default": ["@cpuinfo"],
3853 }),
3854)
3855
3856xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003857 name = "xnnpack_operators_nhwc_f32",
3858 srcs = [
3859 "src/init.c",
3860 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003861 hdrs = ["include/xnnpack.h"],
3862 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003863 "-Isrc",
3864 "-Iinclude",
3865 ] + select({
3866 ":debug_build": [],
3867 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003868 }) + select({
3869 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3870 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003871 }),
3872 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003873 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003874 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003875 "XNN_NO_U8_OPERATORS",
3876 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003877 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003878 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003879 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003880 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003881 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882 visibility = xnnpack_visibility(),
3883 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003884 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003885 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003886 ":operator_run",
3887 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003888 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003889 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003891 ] + select({
3892 ":emscripten": [],
3893 "//conditions:default": ["@cpuinfo"],
3894 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003895)
3896
Marat Dukhancf056b22019-10-07 10:26:29 -07003897xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898 name = "bench_utils",
3899 srcs = ["bench/utils.cc"],
3900 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003901 deps = [
3902 "@com_google_benchmark//:benchmark",
3903 "@cpuinfo",
3904 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905)
3906
Frank Barchard7e955972019-10-11 10:34:25 -07003907######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908
3909xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003910 name = "qs8_gemm_bench",
3911 srcs = [
3912 "bench/gemm.h",
3913 "bench/qs8-gemm.cc",
3914 "src/xnnpack/AlignedAllocator.h",
3915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3916 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3917)
3918
3919xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003920 name = "qs8_requantization_bench",
3921 srcs = [
3922 "bench/qs8-requantization.cc",
3923 "src/xnnpack/requantization-stubs.h",
3924 "src/xnnpack/AlignedAllocator.h",
3925 ] + MICROKERNEL_BENCHMARK_HDRS,
3926 deps = MICROKERNEL_BENCHMARK_DEPS,
3927)
3928
3929xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003930 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931 srcs = [
3932 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003933 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934 "src/xnnpack/AlignedAllocator.h",
3935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003936 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003937 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003938)
3939
3940xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003941 name = "qu8_requantization_bench",
3942 srcs = [
3943 "bench/qu8-requantization.cc",
3944 "src/xnnpack/requantization-stubs.h",
3945 "src/xnnpack/AlignedAllocator.h",
3946 ] + MICROKERNEL_BENCHMARK_HDRS,
3947 deps = MICROKERNEL_BENCHMARK_DEPS,
3948)
3949
3950xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003951 name = "f16_igemm_bench",
3952 srcs = [
3953 "bench/f16-igemm.cc",
3954 "bench/conv.h",
3955 "bench/google/conv.h",
3956 "src/xnnpack/AlignedAllocator.h",
3957 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003958 deps = MICROKERNEL_BENCHMARK_DEPS + [
3959 ":indirection",
3960 ":packing",
3961 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003962)
3963
3964xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003965 name = "f16_gemm_bench",
3966 srcs = [
3967 "bench/f16-gemm.cc",
3968 "bench/gemm.h",
3969 "src/xnnpack/AlignedAllocator.h",
3970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003971 deps = MICROKERNEL_BENCHMARK_DEPS + [
3972 ":packing",
3973 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974)
3975
3976xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003977 name = "f16_spmm_bench",
3978 srcs = [
3979 "bench/f16-spmm.cc",
3980 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003981 "src/xnnpack/AlignedAllocator.h",
3982 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003983 deps = MICROKERNEL_BENCHMARK_DEPS,
3984)
3985
3986xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987 name = "f32_igemm_bench",
3988 srcs = [
3989 "bench/f32-igemm.cc",
3990 "bench/conv.h",
3991 "src/xnnpack/AlignedAllocator.h",
3992 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003993 deps = MICROKERNEL_BENCHMARK_DEPS + [
3994 ":indirection",
3995 ":packing",
3996 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997)
3998
3999xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004000 name = "f16_relu_bench",
4001 srcs = [
4002 "bench/f16-relu.cc",
4003 "src/xnnpack/AlignedAllocator.h",
4004 ] + MICROKERNEL_BENCHMARK_HDRS,
4005 deps = MICROKERNEL_BENCHMARK_DEPS,
4006)
4007
4008xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009 name = "f32_conv_hwc_bench",
4010 srcs = [
4011 "bench/f32-conv-hwc.cc",
4012 "bench/dconv.h",
4013 "src/xnnpack/AlignedAllocator.h",
4014 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004015 deps = MICROKERNEL_BENCHMARK_DEPS + [
4016 ":packing",
4017 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004018)
4019
4020xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004021 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004022 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004023 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004024 "bench/dconv.h",
4025 "src/xnnpack/AlignedAllocator.h",
4026 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004027 deps = MICROKERNEL_BENCHMARK_DEPS + [
4028 ":packing",
4029 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004030)
4031
4032xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004033 name = "f16_dwconv_bench",
4034 srcs = [
4035 "bench/f16-dwconv.cc",
4036 "bench/dwconv.h",
4037 "bench/google/dwconv.h",
4038 "src/xnnpack/AlignedAllocator.h",
4039 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004040 deps = MICROKERNEL_BENCHMARK_DEPS + [
4041 ":indirection",
4042 ":packing",
4043 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004044)
4045
4046xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004047 name = "f32_dwconv_bench",
4048 srcs = [
4049 "bench/f32-dwconv.cc",
4050 "bench/dwconv.h",
4051 "src/xnnpack/AlignedAllocator.h",
4052 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004053 deps = MICROKERNEL_BENCHMARK_DEPS + [
4054 ":indirection",
4055 ":packing",
4056 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057)
4058
4059xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004060 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004061 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004062 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004063 "bench/dwconv.h",
4064 "src/xnnpack/AlignedAllocator.h",
4065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004066 deps = MICROKERNEL_BENCHMARK_DEPS + [
4067 ":indirection",
4068 ":packing",
4069 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004070)
4071
4072xnnpack_benchmark(
4073 name = "f32_gemm_bench",
4074 srcs = [
4075 "bench/f32-gemm.cc",
4076 "bench/gemm.h",
4077 "src/xnnpack/AlignedAllocator.h",
4078 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004079 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004080 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081)
4082
4083xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004084 name = "f32_hswish_bench",
4085 srcs = [
4086 "bench/f32-hswish.cc",
4087 "src/xnnpack/AlignedAllocator.h",
4088 ] + MICROKERNEL_BENCHMARK_HDRS,
4089 deps = MICROKERNEL_BENCHMARK_DEPS,
4090)
4091
4092xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004093 name = "f32_raddexpminusmax_bench",
4094 srcs = [
4095 "bench/f32-raddexpminusmax.cc",
4096 "src/xnnpack/AlignedAllocator.h",
4097 ] + MICROKERNEL_BENCHMARK_HDRS,
4098 deps = MICROKERNEL_BENCHMARK_DEPS,
4099)
4100
4101xnnpack_benchmark(
4102 name = "f32_raddextexp_bench",
4103 srcs = [
4104 "bench/f32-raddextexp.cc",
4105 "src/xnnpack/AlignedAllocator.h",
4106 ] + MICROKERNEL_BENCHMARK_HDRS,
4107 deps = MICROKERNEL_BENCHMARK_DEPS,
4108)
4109
4110xnnpack_benchmark(
4111 name = "f32_raddstoreexpminusmax_bench",
4112 srcs = [
4113 "bench/f32-raddstoreexpminusmax.cc",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + MICROKERNEL_BENCHMARK_HDRS,
4116 deps = MICROKERNEL_BENCHMARK_DEPS,
4117)
4118
4119xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004120 name = "f32_relu_bench",
4121 srcs = [
4122 "bench/f32-relu.cc",
4123 "src/xnnpack/AlignedAllocator.h",
4124 ] + MICROKERNEL_BENCHMARK_HDRS,
4125 deps = MICROKERNEL_BENCHMARK_DEPS,
4126)
4127
4128xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 name = "f32_rmax_bench",
4130 srcs = [
4131 "bench/f32-rmax.cc",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + MICROKERNEL_BENCHMARK_HDRS,
4134 deps = MICROKERNEL_BENCHMARK_DEPS,
4135)
4136
4137xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004138 name = "f32_sigmoid_bench",
4139 srcs = [
4140 "bench/f32-sigmoid.cc",
4141 "src/xnnpack/AlignedAllocator.h",
4142 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004143 deps = MICROKERNEL_BENCHMARK_DEPS,
4144)
4145
4146xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 name = "f32_spmm_bench",
4148 srcs = [
4149 "bench/f32-spmm.cc",
4150 "bench/gemm.h",
4151 "src/xnnpack/AlignedAllocator.h",
4152 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 deps = MICROKERNEL_BENCHMARK_DEPS,
4154)
4155
4156xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004157 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004158 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004159 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004160 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004161 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004162 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004163)
4164
4165xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004166 name = "f32_vscaleexpminusmax_bench",
4167 srcs = [
4168 "bench/f32-vscaleexpminusmax.cc",
4169 "src/xnnpack/AlignedAllocator.h",
4170 ] + MICROKERNEL_BENCHMARK_HDRS,
4171 deps = MICROKERNEL_BENCHMARK_DEPS,
4172)
4173
4174xnnpack_benchmark(
4175 name = "f32_vscaleextexp_bench",
4176 srcs = [
4177 "bench/f32-vscaleextexp.cc",
4178 "src/xnnpack/AlignedAllocator.h",
4179 ] + MICROKERNEL_BENCHMARK_HDRS,
4180 deps = MICROKERNEL_BENCHMARK_DEPS,
4181)
4182
4183xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004184 name = "f32_vsqrt_bench",
4185 srcs = [
4186 "bench/f32-vsqrt.cc",
4187 "src/xnnpack/AlignedAllocator.h",
4188 ] + MICROKERNEL_BENCHMARK_HDRS,
4189 deps = MICROKERNEL_BENCHMARK_DEPS,
4190)
4191
4192xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 name = "f32_im2col_gemm_bench",
4194 srcs = [
4195 "bench/f32-im2col-gemm.cc",
4196 "bench/conv.h",
4197 "src/xnnpack/AlignedAllocator.h",
4198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004199 deps = MICROKERNEL_BENCHMARK_DEPS + [
4200 ":im2col",
4201 ":packing",
4202 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203)
4204
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004205xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004206 name = "rounding_bench",
4207 srcs = [
4208 "bench/rounding.cc",
4209 "src/xnnpack/math-stubs.h",
4210 "src/xnnpack/AlignedAllocator.h",
4211 ] + MICROKERNEL_BENCHMARK_HDRS,
4212 deps = MICROKERNEL_BENCHMARK_DEPS,
4213)
4214
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215########################### Benchmarks for operators ###########################
4216
4217xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 name = "average_pooling_bench",
4219 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004220 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004221 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004222 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004223)
4224
4225xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004226 name = "bankers_rounding_bench",
4227 srcs = ["bench/bankers-rounding.cc"],
4228 copts = xnnpack_optional_tflite_copts(),
4229 tags = ["nowin32"],
4230 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4231)
4232
4233xnnpack_benchmark(
4234 name = "ceiling_bench",
4235 srcs = ["bench/ceiling.cc"],
4236 copts = xnnpack_optional_tflite_copts(),
4237 tags = ["nowin32"],
4238 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4239)
4240
4241xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 name = "channel_shuffle_bench",
4243 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004244 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245)
4246
4247xnnpack_benchmark(
4248 name = "convolution_bench",
4249 srcs = ["bench/convolution.cc"],
4250 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004251 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004252 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253)
4254
4255xnnpack_benchmark(
4256 name = "deconvolution_bench",
4257 srcs = ["bench/deconvolution.cc"],
4258 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004259 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004260 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004261)
4262
4263xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004264 name = "floor_bench",
4265 srcs = ["bench/floor.cc"],
4266 copts = xnnpack_optional_tflite_copts(),
4267 tags = ["nowin32"],
4268 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4269)
4270
4271xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272 name = "global_average_pooling_bench",
4273 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004274 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275)
4276
4277xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004278 name = "hardswish_bench",
4279 srcs = ["bench/hardswish.cc"],
4280 copts = xnnpack_optional_tflite_copts(),
4281 tags = ["nowin32"],
4282 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4283)
4284
4285xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004286 name = "max_pooling_bench",
4287 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004288 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_benchmark(
4292 name = "sigmoid_bench",
4293 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004294 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004295 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004296 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297)
4298
4299xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004300 name = "prelu_bench",
4301 srcs = ["bench/prelu.cc"],
4302 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004303 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004304 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004305)
4306
4307xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004308 name = "softmax_bench",
4309 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004310 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004311 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004312 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313)
4314
Marat Dukhan87727142020-06-24 15:24:10 -07004315xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004316 name = "square_root_bench",
4317 srcs = ["bench/square-root.cc"],
4318 copts = xnnpack_optional_tflite_copts(),
4319 tags = ["nowin32"],
4320 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4321)
4322
4323xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004324 name = "truncation_bench",
4325 srcs = ["bench/truncation.cc"],
4326 deps = OPERATOR_BENCHMARK_DEPS,
4327)
4328
Marat Dukhanc068bb62019-10-04 13:24:39 -07004329############################# End-to-end benchmarks ############################
4330
4331cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004332 name = "fp32_mobilenet_v1",
4333 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004334 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004335 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004336 linkstatic = True,
4337 deps = [
4338 ":XNNPACK",
4339 "@pthreadpool",
4340 ],
4341)
4342
4343cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004344 name = "fp16_mobilenet_v1",
4345 srcs = ["models/fp16-mobilenet-v1.cc"],
4346 hdrs = ["models/models.h"],
4347 copts = xnnpack_std_cxxopts(),
4348 linkstatic = True,
4349 deps = [
4350 ":XNNPACK",
4351 "@FP16",
4352 "@pthreadpool",
4353 ],
4354)
4355
4356cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004357 name = "qs8_mobilenet_v1",
4358 srcs = ["models/qs8-mobilenet-v1.cc"],
4359 hdrs = ["models/models.h"],
4360 copts = xnnpack_std_cxxopts(),
4361 linkstatic = True,
4362 deps = [
4363 ":XNNPACK",
4364 "@pthreadpool",
4365 ],
4366)
4367
4368cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004369 name = "fp32_mobilenet_v2",
4370 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004371 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004372 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004373 linkstatic = True,
4374 deps = [
4375 ":XNNPACK",
4376 "@pthreadpool",
4377 ],
4378)
4379
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004380cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004381 name = "fp16_mobilenet_v2",
4382 srcs = ["models/fp16-mobilenet-v2.cc"],
4383 hdrs = ["models/models.h"],
4384 copts = xnnpack_std_cxxopts(),
4385 linkstatic = True,
4386 deps = [
4387 ":XNNPACK",
4388 "@FP16",
4389 "@pthreadpool",
4390 ],
4391)
4392
4393cc_library(
4394 name = "fp32_mobilenet_v3_large",
4395 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004396 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004397 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004398 linkstatic = True,
4399 deps = [
4400 ":XNNPACK",
4401 "@pthreadpool",
4402 ],
4403)
4404
4405cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004406 name = "fp32_mobilenet_v3_small",
4407 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004408 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004409 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004410 linkstatic = True,
4411 deps = [
4412 ":XNNPACK",
4413 "@pthreadpool",
4414 ],
4415)
4416
Marat Dukhanc068bb62019-10-04 13:24:39 -07004417xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004418 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004419 srcs = [
4420 "bench/f32-dwconv-e2e.cc",
4421 "bench/end2end.h",
4422 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004423 deps = MICROKERNEL_BENCHMARK_DEPS + [
4424 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004425 ":fp32_mobilenet_v1",
4426 ":fp32_mobilenet_v2",
4427 ":fp32_mobilenet_v3_large",
4428 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004429 ],
4430)
4431
4432xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004433 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004434 srcs = [
4435 "bench/f32-gemm-e2e.cc",
4436 "bench/end2end.h",
4437 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004438 deps = MICROKERNEL_BENCHMARK_DEPS + [
4439 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004440 ":fp32_mobilenet_v1",
4441 ":fp32_mobilenet_v2",
4442 ":fp32_mobilenet_v3_large",
4443 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004444 ],
4445)
4446
4447xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004448 name = "end2end_bench",
4449 srcs = ["bench/end2end.cc"],
4450 deps = [
4451 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004452 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004453 ":fp16_mobilenet_v1",
4454 ":fp16_mobilenet_v2",
4455 ":fp32_mobilenet_v1",
4456 ":fp32_mobilenet_v2",
4457 ":fp32_mobilenet_v3_large",
4458 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004459 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004460 "@pthreadpool",
4461 ],
4462)
4463
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004464#################### Accuracy evaluation for math functions ####################
4465
4466xnnpack_benchmark(
4467 name = "f32_exp_eval",
4468 srcs = [
4469 "eval/f32-exp.cc",
4470 "src/xnnpack/AlignedAllocator.h",
4471 ] + ACCURACY_EVAL_HDRS,
4472 deps = ACCURACY_EVAL_DEPS,
4473)
4474
Marat Dukhan515c9772019-10-17 18:07:57 -07004475xnnpack_benchmark(
4476 name = "f32_expminus_eval",
4477 srcs = [
4478 "eval/f32-expminus.cc",
4479 "src/xnnpack/AlignedAllocator.h",
4480 ] + ACCURACY_EVAL_HDRS,
4481 deps = ACCURACY_EVAL_DEPS,
4482)
4483
Marat Dukhan98ba4412019-10-23 02:14:28 -07004484xnnpack_benchmark(
4485 name = "f32_extexp_eval",
4486 srcs = [
4487 "eval/f32-extexp.cc",
4488 "src/xnnpack/AlignedAllocator.h",
4489 ] + ACCURACY_EVAL_HDRS,
4490 deps = ACCURACY_EVAL_DEPS,
4491)
4492
Marat Dukhan8853b822020-05-07 12:19:01 -07004493xnnpack_unit_test(
4494 name = "f32_roundne_eval",
4495 srcs = [
4496 "eval/f32-roundne.cc",
4497 "src/xnnpack/AlignedAllocator.h",
4498 "src/xnnpack/math-stubs.h",
4499 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004500 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004501 deps = MICROKERNEL_TEST_DEPS,
4502)
4503
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004504xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004505 name = "f32_roundd_eval",
4506 srcs = [
4507 "eval/f32-roundd.cc",
4508 "src/xnnpack/AlignedAllocator.h",
4509 "src/xnnpack/math-stubs.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 automatic = False,
4512 deps = MICROKERNEL_TEST_DEPS,
4513)
4514
4515xnnpack_unit_test(
4516 name = "f32_roundu_eval",
4517 srcs = [
4518 "eval/f32-roundu.cc",
4519 "src/xnnpack/AlignedAllocator.h",
4520 "src/xnnpack/math-stubs.h",
4521 ] + MICROKERNEL_TEST_HDRS,
4522 automatic = False,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004527 name = "f32_roundz_eval",
4528 srcs = [
4529 "eval/f32-roundz.cc",
4530 "src/xnnpack/AlignedAllocator.h",
4531 "src/xnnpack/math-stubs.h",
4532 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004533 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004534 deps = MICROKERNEL_TEST_DEPS,
4535)
4536
Marat Dukhan346a9e52019-11-15 09:06:30 -08004537xnnpack_benchmark(
4538 name = "f32_sigmoid_eval",
4539 srcs = [
4540 "eval/f32-sigmoid.cc",
4541 "src/xnnpack/AlignedAllocator.h",
4542 ] + ACCURACY_EVAL_HDRS,
4543 deps = ACCURACY_EVAL_DEPS,
4544)
4545
Marat Dukhan84000762020-06-29 18:38:43 -07004546xnnpack_benchmark(
4547 name = "f32_sqrt_eval",
4548 srcs = [
4549 "eval/f32-sqrt.cc",
4550 "src/xnnpack/AlignedAllocator.h",
4551 ] + ACCURACY_EVAL_HDRS,
4552 deps = ACCURACY_EVAL_DEPS + [
4553 ":bench_utils",
4554 ],
4555)
4556
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557######################### Unit tests for micro-kernels #########################
4558
4559xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004560 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004561 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004562 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004563 "test/gemm-microkernel-tester.h",
4564 "src/xnnpack/AlignedAllocator.h",
4565 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004566 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567)
4568
4569xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004570 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004571 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004572 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004573 "test/spmm-microkernel-tester.h",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_TEST_HDRS,
4576 deps = MICROKERNEL_TEST_DEPS,
4577)
4578
4579xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580 name = "f32_argmaxpool_test",
4581 srcs = [
4582 "test/f32-argmaxpool.cc",
4583 "test/argmaxpool-microkernel-tester.h",
4584 "src/xnnpack/AlignedAllocator.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004590 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004592 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 "test/avgpool-microkernel-tester.h",
4594 "src/xnnpack/AlignedAllocator.h",
4595 ] + MICROKERNEL_TEST_HDRS,
4596 deps = MICROKERNEL_TEST_DEPS,
4597)
4598
4599xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004600 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004601 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004602 "test/f32-ibilinear.cc",
4603 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004604 "src/xnnpack/AlignedAllocator.h",
4605 ] + MICROKERNEL_TEST_HDRS,
4606 deps = MICROKERNEL_TEST_DEPS,
4607)
4608
4609xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004610 name = "f16_clamp_test",
4611 srcs = [
4612 "test/f16-clamp.cc",
4613 "test/clamp-microkernel-tester.h",
4614 ] + MICROKERNEL_TEST_HDRS,
4615 deps = MICROKERNEL_TEST_DEPS,
4616)
4617
4618xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619 name = "f32_clamp_test",
4620 srcs = [
4621 "test/f32-clamp.cc",
4622 "test/clamp-microkernel-tester.h",
4623 ] + MICROKERNEL_TEST_HDRS,
4624 deps = MICROKERNEL_TEST_DEPS,
4625)
4626
4627xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004628 name = "f32_igemm_test",
4629 srcs = [
4630 "test/f32-igemm.cc",
4631 "test/gemm-microkernel-tester.h",
4632 "src/xnnpack/AlignedAllocator.h",
4633 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004634 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004635)
4636
4637xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004638 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004639 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004640 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004641 "test/gemm-microkernel-tester.h",
4642 "src/xnnpack/AlignedAllocator.h",
4643 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004644 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645)
4646
4647xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004648 name = "f32_igemm_minmax_test",
4649 srcs = [
4650 "test/f32-igemm-minmax.cc",
4651 "test/gemm-microkernel-tester.h",
4652 "src/xnnpack/AlignedAllocator.h",
4653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004654 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004655)
4656
4657xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004658 name = "f16_igemm_minmax_test",
4659 srcs = [
4660 "test/f16-igemm-minmax.cc",
4661 "test/gemm-microkernel-tester.h",
4662 "src/xnnpack/AlignedAllocator.h",
4663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004664 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004665)
4666
4667xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004668 name = "f32_conv_hwc_test",
4669 srcs = [
4670 "test/f32-conv-hwc.cc",
4671 "test/conv-hwc-microkernel-tester.h",
4672 "src/xnnpack/AlignedAllocator.h",
4673 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004674 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
4677xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004678 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004680 "test/f32-conv-hwc2chw.cc",
4681 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682 "src/xnnpack/AlignedAllocator.h",
4683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004684 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004685)
4686
4687xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004688 name = "f32_dwconv_test",
4689 srcs = [
4690 "test/f32-dwconv.cc",
4691 "test/dwconv-microkernel-tester.h",
4692 "src/xnnpack/AlignedAllocator.h",
4693 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004694 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004695)
4696
4697xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004698 name = "f16_dwconv_minmax_test",
4699 srcs = [
4700 "test/f16-dwconv-minmax.cc",
4701 "test/dwconv-microkernel-tester.h",
4702 "src/xnnpack/AlignedAllocator.h",
4703 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004704 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004705)
4706
4707xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004708 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004709 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004710 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711 "test/dwconv-microkernel-tester.h",
4712 "src/xnnpack/AlignedAllocator.h",
4713 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004714 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004715)
4716
4717xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004718 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004720 "test/f32-dwconv-chw.cc",
4721 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722 "src/xnnpack/AlignedAllocator.h",
4723 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004724 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725)
4726
4727xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004728 name = "f16_gavgpool_minmax_test",
4729 srcs = [
4730 "test/f16-gavgpool-minmax.cc",
4731 "test/gavgpool-microkernel-tester.h",
4732 "src/xnnpack/AlignedAllocator.h",
4733 ] + MICROKERNEL_TEST_HDRS,
4734 deps = MICROKERNEL_TEST_DEPS,
4735)
4736
4737xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004738 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004740 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004741 "test/gavgpool-microkernel-tester.h",
4742 "src/xnnpack/AlignedAllocator.h",
4743 ] + MICROKERNEL_TEST_HDRS,
4744 deps = MICROKERNEL_TEST_DEPS,
4745)
4746
4747xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004748 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004749 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004750 "test/f32-gavgpool-cw.cc",
4751 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004752 "src/xnnpack/AlignedAllocator.h",
4753 ] + MICROKERNEL_TEST_HDRS,
4754 deps = MICROKERNEL_TEST_DEPS,
4755)
4756
4757xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004758 name = "f32_gemm_test",
4759 srcs = [
4760 "test/f32-gemm.cc",
4761 "test/gemm-microkernel-tester.h",
4762 "src/xnnpack/AlignedAllocator.h",
4763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004765)
4766
4767xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004768 name = "f32_gemm_relu_test",
4769 srcs = [
4770 "test/f32-gemm-relu.cc",
4771 "test/gemm-microkernel-tester.h",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004774 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004775)
4776
4777xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004778 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004779 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004780 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781 "test/gemm-microkernel-tester.h",
4782 "src/xnnpack/AlignedAllocator.h",
4783 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004784 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785)
4786
4787xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004788 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004790 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004791 "test/gemm-microkernel-tester.h",
4792 "src/xnnpack/AlignedAllocator.h",
4793 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004794 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004795)
4796
4797xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004798 name = "f16_hswish_test",
4799 srcs = [
4800 "test/f16-hswish.cc",
4801 "test/hswish-microkernel-tester.h",
4802 ] + MICROKERNEL_TEST_HDRS,
4803 deps = MICROKERNEL_TEST_DEPS,
4804)
4805
4806xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807 name = "f32_hswish_test",
4808 srcs = [
4809 "test/f32-hswish.cc",
4810 "test/hswish-microkernel-tester.h",
4811 ] + MICROKERNEL_TEST_HDRS,
4812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
4815xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004816 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004817 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004818 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004819 "test/maxpool-microkernel-tester.h",
4820 ] + MICROKERNEL_TEST_HDRS,
4821 deps = MICROKERNEL_TEST_DEPS,
4822)
4823
4824xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004825 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004826 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004827 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004828 "test/avgpool-microkernel-tester.h",
4829 "src/xnnpack/AlignedAllocator.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004835 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004837 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004838 "test/gemm-microkernel-tester.h",
4839 "src/xnnpack/AlignedAllocator.h",
4840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004841 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004842)
4843
4844xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004845 name = "f16_prelu_test",
4846 srcs = [
4847 "test/f16-prelu.cc",
4848 "test/prelu-microkernel-tester.h",
4849 "src/xnnpack/AlignedAllocator.h",
4850 ] + MICROKERNEL_TEST_HDRS,
4851 deps = MICROKERNEL_TEST_DEPS,
4852)
4853
4854xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855 name = "f32_prelu_test",
4856 srcs = [
4857 "test/f32-prelu.cc",
4858 "test/prelu-microkernel-tester.h",
4859 "src/xnnpack/AlignedAllocator.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004865 name = "f32_raddexpminusmax_test",
4866 srcs = [
4867 "test/f32-raddexpminusmax.cc",
4868 "test/raddexpminusmax-microkernel-tester.h",
4869 ] + MICROKERNEL_TEST_HDRS,
4870 deps = MICROKERNEL_TEST_DEPS,
4871)
4872
4873xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004874 name = "f32_raddextexp_test",
4875 srcs = [
4876 "test/f32-raddextexp.cc",
4877 "test/raddextexp-microkernel-tester.h",
4878 ] + MICROKERNEL_TEST_HDRS,
4879 deps = MICROKERNEL_TEST_DEPS,
4880)
4881
4882xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004883 name = "f32_raddstoreexpminusmax_test",
4884 srcs = [
4885 "test/f32-raddstoreexpminusmax.cc",
4886 "test/raddstoreexpminusmax-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004892 name = "f32_relu_test",
4893 srcs = [
4894 "test/f32-relu.cc",
4895 "test/vunary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 name = "f32_rmax_test",
4902 srcs = [
4903 "test/f32-rmax.cc",
4904 "test/rmax-microkernel-tester.h",
4905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004910 name = "f32_sigmoid_test",
4911 srcs = [
4912 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004913 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004919 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004920 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004921 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004922 "test/spmm-microkernel-tester.h",
4923 "src/xnnpack/AlignedAllocator.h",
4924 ] + MICROKERNEL_TEST_HDRS,
4925 deps = MICROKERNEL_TEST_DEPS,
4926)
4927
4928xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004929 name = "f16_vadd_minmax_test",
4930 srcs = [
4931 "test/f16-vadd-minmax.cc",
4932 "test/vbinary-microkernel-tester.h",
4933 ] + MICROKERNEL_TEST_HDRS,
4934 deps = MICROKERNEL_TEST_DEPS,
4935)
4936
4937xnnpack_unit_test(
4938 name = "f16_vaddc_minmax_test",
4939 srcs = [
4940 "test/f16-vaddc-minmax.cc",
4941 "test/vbinaryc-microkernel-tester.h",
4942 ] + MICROKERNEL_TEST_HDRS,
4943 deps = MICROKERNEL_TEST_DEPS,
4944)
4945
4946xnnpack_unit_test(
4947 name = "f16_vdiv_minmax_test",
4948 srcs = [
4949 "test/f16-vdiv-minmax.cc",
4950 "test/vbinary-microkernel-tester.h",
4951 ] + MICROKERNEL_TEST_HDRS,
4952 deps = MICROKERNEL_TEST_DEPS,
4953)
4954
4955xnnpack_unit_test(
4956 name = "f16_vdivc_minmax_test",
4957 srcs = [
4958 "test/f16-vdivc-minmax.cc",
4959 "test/vbinaryc-microkernel-tester.h",
4960 ] + MICROKERNEL_TEST_HDRS,
4961 deps = MICROKERNEL_TEST_DEPS,
4962)
4963
4964xnnpack_unit_test(
4965 name = "f16_vrdivc_minmax_test",
4966 srcs = [
4967 "test/f16-vrdivc-minmax.cc",
4968 "test/vbinaryc-microkernel-tester.h",
4969 ] + MICROKERNEL_TEST_HDRS,
4970 deps = MICROKERNEL_TEST_DEPS,
4971)
4972
4973xnnpack_unit_test(
4974 name = "f16_vmax_test",
4975 srcs = [
4976 "test/f16-vmax.cc",
4977 "test/vbinary-microkernel-tester.h",
4978 ] + MICROKERNEL_TEST_HDRS,
4979 deps = MICROKERNEL_TEST_DEPS,
4980)
4981
4982xnnpack_unit_test(
4983 name = "f16_vmaxc_test",
4984 srcs = [
4985 "test/f16-vmaxc.cc",
4986 "test/vbinaryc-microkernel-tester.h",
4987 ] + MICROKERNEL_TEST_HDRS,
4988 deps = MICROKERNEL_TEST_DEPS,
4989)
4990
4991xnnpack_unit_test(
4992 name = "f16_vmin_test",
4993 srcs = [
4994 "test/f16-vmin.cc",
4995 "test/vbinary-microkernel-tester.h",
4996 ] + MICROKERNEL_TEST_HDRS,
4997 deps = MICROKERNEL_TEST_DEPS,
4998)
4999
5000xnnpack_unit_test(
5001 name = "f16_vminc_test",
5002 srcs = [
5003 "test/f16-vminc.cc",
5004 "test/vbinaryc-microkernel-tester.h",
5005 ] + MICROKERNEL_TEST_HDRS,
5006 deps = MICROKERNEL_TEST_DEPS,
5007)
5008
5009xnnpack_unit_test(
5010 name = "f16_vmul_minmax_test",
5011 srcs = [
5012 "test/f16-vmul-minmax.cc",
5013 "test/vbinary-microkernel-tester.h",
5014 ] + MICROKERNEL_TEST_HDRS,
5015 deps = MICROKERNEL_TEST_DEPS,
5016)
5017
5018xnnpack_unit_test(
5019 name = "f16_vmulc_minmax_test",
5020 srcs = [
5021 "test/f16-vmulc-minmax.cc",
5022 "test/vbinaryc-microkernel-tester.h",
5023 ] + MICROKERNEL_TEST_HDRS,
5024 deps = MICROKERNEL_TEST_DEPS,
5025)
5026
5027xnnpack_unit_test(
5028 name = "f16_vsub_minmax_test",
5029 srcs = [
5030 "test/f16-vsub-minmax.cc",
5031 "test/vbinary-microkernel-tester.h",
5032 ] + MICROKERNEL_TEST_HDRS,
5033 deps = MICROKERNEL_TEST_DEPS,
5034)
5035
5036xnnpack_unit_test(
5037 name = "f16_vsubc_minmax_test",
5038 srcs = [
5039 "test/f16-vsubc-minmax.cc",
5040 "test/vbinaryc-microkernel-tester.h",
5041 ] + MICROKERNEL_TEST_HDRS,
5042 deps = MICROKERNEL_TEST_DEPS,
5043)
5044
5045xnnpack_unit_test(
5046 name = "f16_vrsubc_minmax_test",
5047 srcs = [
5048 "test/f16-vrsubc-minmax.cc",
5049 "test/vbinaryc-microkernel-tester.h",
5050 ] + MICROKERNEL_TEST_HDRS,
5051 deps = MICROKERNEL_TEST_DEPS,
5052)
5053
5054xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005055 name = "f32_vabs_test",
5056 srcs = [
5057 "test/f32-vabs.cc",
5058 "test/vunary-microkernel-tester.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005064 name = "f32_vadd_test",
5065 srcs = [
5066 "test/f32-vadd.cc",
5067 "test/vbinary-microkernel-tester.h",
5068 ] + MICROKERNEL_TEST_HDRS,
5069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
5072xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005073 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005075 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005076 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005082 name = "f32_vadd_relu_test",
5083 srcs = [
5084 "test/f32-vadd-relu.cc",
5085 "test/vbinary-microkernel-tester.h",
5086 ] + MICROKERNEL_TEST_HDRS,
5087 deps = MICROKERNEL_TEST_DEPS,
5088)
5089
5090xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005091 name = "f32_vaddc_test",
5092 srcs = [
5093 "test/f32-vaddc.cc",
5094 "test/vbinaryc-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005100 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005101 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005102 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005103 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005109 name = "f32_vaddc_relu_test",
5110 srcs = [
5111 "test/f32-vaddc-relu.cc",
5112 "test/vbinaryc-microkernel-tester.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005118 name = "f32_vdiv_test",
5119 srcs = [
5120 "test/f32-vdiv.cc",
5121 "test/vbinary-microkernel-tester.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 deps = MICROKERNEL_TEST_DEPS,
5124)
5125
5126xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005127 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005128 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005129 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005130 "test/vbinary-microkernel-tester.h",
5131 ] + MICROKERNEL_TEST_HDRS,
5132 deps = MICROKERNEL_TEST_DEPS,
5133)
5134
5135xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005136 name = "f32_vdiv_relu_test",
5137 srcs = [
5138 "test/f32-vdiv-relu.cc",
5139 "test/vbinary-microkernel-tester.h",
5140 ] + MICROKERNEL_TEST_HDRS,
5141 deps = MICROKERNEL_TEST_DEPS,
5142)
5143
5144xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005145 name = "f32_vdivc_test",
5146 srcs = [
5147 "test/f32-vdivc.cc",
5148 "test/vbinaryc-microkernel-tester.h",
5149 ] + MICROKERNEL_TEST_HDRS,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005154 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005155 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005156 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005157 "test/vbinaryc-microkernel-tester.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005163 name = "f32_vdivc_relu_test",
5164 srcs = [
5165 "test/f32-vdivc-relu.cc",
5166 "test/vbinaryc-microkernel-tester.h",
5167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005172 name = "f32_vrdivc_test",
5173 srcs = [
5174 "test/f32-vrdivc.cc",
5175 "test/vbinaryc-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005181 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005182 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005183 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005184 "test/vbinaryc-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005190 name = "f32_vrdivc_relu_test",
5191 srcs = [
5192 "test/f32-vrdivc-relu.cc",
5193 "test/vbinaryc-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005199 name = "f32_vmax_test",
5200 srcs = [
5201 "test/f32-vmax.cc",
5202 "test/vbinary-microkernel-tester.h",
5203 ] + MICROKERNEL_TEST_HDRS,
5204 deps = MICROKERNEL_TEST_DEPS,
5205)
5206
5207xnnpack_unit_test(
5208 name = "f32_vmaxc_test",
5209 srcs = [
5210 "test/f32-vmaxc.cc",
5211 "test/vbinaryc-microkernel-tester.h",
5212 ] + MICROKERNEL_TEST_HDRS,
5213 deps = MICROKERNEL_TEST_DEPS,
5214)
5215
5216xnnpack_unit_test(
5217 name = "f32_vmin_test",
5218 srcs = [
5219 "test/f32-vmin.cc",
5220 "test/vbinary-microkernel-tester.h",
5221 ] + MICROKERNEL_TEST_HDRS,
5222 deps = MICROKERNEL_TEST_DEPS,
5223)
5224
5225xnnpack_unit_test(
5226 name = "f32_vminc_test",
5227 srcs = [
5228 "test/f32-vminc.cc",
5229 "test/vbinaryc-microkernel-tester.h",
5230 ] + MICROKERNEL_TEST_HDRS,
5231 deps = MICROKERNEL_TEST_DEPS,
5232)
5233
5234xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005235 name = "f32_vmul_test",
5236 srcs = [
5237 "test/f32-vmul.cc",
5238 "test/vbinary-microkernel-tester.h",
5239 ] + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS,
5241)
5242
5243xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005244 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005246 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005247 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005253 name = "f32_vmul_relu_test",
5254 srcs = [
5255 "test/f32-vmul-relu.cc",
5256 "test/vbinary-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005262 name = "f32_vmulc_test",
5263 srcs = [
5264 "test/f32-vmulc.cc",
5265 "test/vbinaryc-microkernel-tester.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005271 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005272 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005273 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005274 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 ] + MICROKERNEL_TEST_HDRS,
5276 deps = MICROKERNEL_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005280 name = "f32_vmulc_relu_test",
5281 srcs = [
5282 "test/f32-vmulc-relu.cc",
5283 "test/vbinaryc-microkernel-tester.h",
5284 ] + MICROKERNEL_TEST_HDRS,
5285 deps = MICROKERNEL_TEST_DEPS,
5286)
5287
5288xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005289 name = "f16_vmulcaddc_minmax_test",
5290 srcs = [
5291 "test/f16-vmulcaddc-minmax.cc",
5292 "test/vmulcaddc-microkernel-tester.h",
5293 "src/xnnpack/AlignedAllocator.h",
5294 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005295 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005296)
5297
5298xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005299 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005301 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 "test/vmulcaddc-microkernel-tester.h",
5303 "src/xnnpack/AlignedAllocator.h",
5304 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005305 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306)
5307
5308xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005309 name = "f32_vlrelu_test",
5310 srcs = [
5311 "test/f32-vlrelu.cc",
5312 "test/vunary-microkernel-tester.h",
5313 ] + MICROKERNEL_TEST_HDRS,
5314 deps = MICROKERNEL_TEST_DEPS,
5315)
5316
5317xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005318 name = "f32_vneg_test",
5319 srcs = [
5320 "test/f32-vneg.cc",
5321 "test/vunary-microkernel-tester.h",
5322 ] + MICROKERNEL_TEST_HDRS,
5323 deps = MICROKERNEL_TEST_DEPS,
5324)
5325
5326xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005327 name = "f32_vrndne_test",
5328 srcs = [
5329 "test/f32-vrndne.cc",
5330 "test/vunary-microkernel-tester.h",
5331 ] + MICROKERNEL_TEST_HDRS,
5332 deps = MICROKERNEL_TEST_DEPS,
5333)
5334
5335xnnpack_unit_test(
5336 name = "f32_vrndz_test",
5337 srcs = [
5338 "test/f32-vrndz.cc",
5339 "test/vunary-microkernel-tester.h",
5340 ] + MICROKERNEL_TEST_HDRS,
5341 deps = MICROKERNEL_TEST_DEPS,
5342)
5343
5344xnnpack_unit_test(
5345 name = "f32_vrndu_test",
5346 srcs = [
5347 "test/f32-vrndu.cc",
5348 "test/vunary-microkernel-tester.h",
5349 ] + MICROKERNEL_TEST_HDRS,
5350 deps = MICROKERNEL_TEST_DEPS,
5351)
5352
5353xnnpack_unit_test(
5354 name = "f32_vrndd_test",
5355 srcs = [
5356 "test/f32-vrndd.cc",
5357 "test/vunary-microkernel-tester.h",
5358 ] + MICROKERNEL_TEST_HDRS,
5359 deps = MICROKERNEL_TEST_DEPS,
5360)
5361
5362xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005363 name = "f32_vscale_test",
5364 srcs = [
5365 "test/f32-vscale.cc",
5366 "test/vscale-microkernel-tester.h",
5367 ] + MICROKERNEL_TEST_HDRS,
5368 deps = MICROKERNEL_TEST_DEPS,
5369)
5370
5371xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005372 name = "f32_vscaleexpminusmax_test",
5373 srcs = [
5374 "test/f32-vscaleexpminusmax.cc",
5375 "test/vscaleexpminusmax-microkernel-tester.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005381 name = "f32_vscaleextexp_test",
5382 srcs = [
5383 "test/f32-vscaleextexp.cc",
5384 "test/vscaleextexp-microkernel-tester.h",
5385 ] + MICROKERNEL_TEST_HDRS,
5386 deps = MICROKERNEL_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005390 name = "f32_vsqr_test",
5391 srcs = [
5392 "test/f32-vsqr.cc",
5393 "test/vunary-microkernel-tester.h",
5394 ] + MICROKERNEL_TEST_HDRS,
5395 deps = MICROKERNEL_TEST_DEPS,
5396)
5397
5398xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005399 name = "f32_vsqrdiff_test",
5400 srcs = [
5401 "test/f32-vsqrdiff.cc",
5402 "test/vbinary-microkernel-tester.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 deps = MICROKERNEL_TEST_DEPS,
5405)
5406
5407xnnpack_unit_test(
5408 name = "f32_vsqrdiffc_test",
5409 srcs = [
5410 "test/f32-vsqrdiffc.cc",
5411 "test/vbinaryc-microkernel-tester.h",
5412 ] + MICROKERNEL_TEST_HDRS,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005417 name = "f32_vsqrt_test",
5418 srcs = [
5419 "test/f32-vsqrt.cc",
5420 "test/vunary-microkernel-tester.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005426 name = "f32_vsub_test",
5427 srcs = [
5428 "test/f32-vsub.cc",
5429 "test/vbinary-microkernel-tester.h",
5430 ] + MICROKERNEL_TEST_HDRS,
5431 deps = MICROKERNEL_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005435 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005436 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005437 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005438 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005444 name = "f32_vsub_relu_test",
5445 srcs = [
5446 "test/f32-vsub-relu.cc",
5447 "test/vbinary-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005453 name = "f32_vsubc_test",
5454 srcs = [
5455 "test/f32-vsubc.cc",
5456 "test/vbinaryc-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005462 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005463 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005464 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005465 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005471 name = "f32_vsubc_relu_test",
5472 srcs = [
5473 "test/f32-vsubc-relu.cc",
5474 "test/vbinaryc-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005480 name = "f32_vrsubc_test",
5481 srcs = [
5482 "test/f32-vrsubc.cc",
5483 "test/vbinaryc-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005489 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005490 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005491 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005492 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005498 name = "f32_vrsubc_relu_test",
5499 srcs = [
5500 "test/f32-vrsubc-relu.cc",
5501 "test/vbinaryc-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005507 name = "qs8_dwconv_minmax_test",
5508 srcs = [
5509 "test/qs8-dwconv-minmax.cc",
5510 "test/dwconv-microkernel-tester.h",
5511 "src/xnnpack/AlignedAllocator.h",
5512 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5513 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5514)
5515
5516xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005517 name = "qs8_gavgpool_minmax_test",
5518 srcs = [
5519 "test/qs8-gavgpool-minmax.cc",
5520 "test/gavgpool-microkernel-tester.h",
5521 "src/xnnpack/AlignedAllocator.h",
5522 ] + MICROKERNEL_TEST_HDRS,
5523 deps = MICROKERNEL_TEST_DEPS,
5524)
5525
5526xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005527 name = "qs8_gemm_minmax_test",
5528 srcs = [
5529 "test/qs8-gemm-minmax.cc",
5530 "test/gemm-microkernel-tester.h",
5531 "src/xnnpack/AlignedAllocator.h",
5532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5534)
5535
5536xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005537 name = "qs8_igemm_minmax_test",
5538 srcs = [
5539 "test/qs8-igemm-minmax.cc",
5540 "test/gemm-microkernel-tester.h",
5541 "src/xnnpack/AlignedAllocator.h",
5542 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5543 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5544)
5545
5546xnnpack_unit_test(
5547 name = "qs8_requantization_test",
5548 srcs = [
5549 "src/xnnpack/requantization-stubs.h",
5550 "test/qs8-requantization.cc",
5551 "test/requantization-tester.h",
5552 ] + MICROKERNEL_TEST_HDRS,
5553 deps = MICROKERNEL_TEST_DEPS,
5554)
5555
5556xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005557 name = "qs8_vadd_minmax_test",
5558 srcs = [
5559 "test/qs8-vadd-minmax.cc",
5560 "test/vadd-microkernel-tester.h",
5561 ] + MICROKERNEL_TEST_HDRS,
5562 deps = MICROKERNEL_TEST_DEPS,
5563)
5564
5565xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005566 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005567 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005568 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005569 "test/avgpool-microkernel-tester.h",
5570 "src/xnnpack/AlignedAllocator.h",
5571 ] + MICROKERNEL_TEST_HDRS,
5572 deps = MICROKERNEL_TEST_DEPS,
5573)
5574
5575xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005576 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005577 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005578 "test/qu8-dwconv-minmax.cc",
5579 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005580 "src/xnnpack/AlignedAllocator.h",
5581 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005582 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005583)
5584
5585xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005586 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005587 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005588 "test/qu8-igemm-minmax.cc",
5589 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005590 "src/xnnpack/AlignedAllocator.h",
5591 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005592 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005593)
5594
5595xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005596 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005598 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005599 "test/gavgpool-microkernel-tester.h",
5600 "src/xnnpack/AlignedAllocator.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005606 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005607 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005608 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 "test/gemm-microkernel-tester.h",
5610 "src/xnnpack/AlignedAllocator.h",
5611 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005612 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005613)
5614
5615xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005616 name = "qu8_requantization_test",
5617 srcs = [
5618 "src/xnnpack/requantization-stubs.h",
5619 "test/qu8-requantization.cc",
5620 "test/requantization-tester.h",
5621 ] + MICROKERNEL_TEST_HDRS,
5622 deps = MICROKERNEL_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005626 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005628 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005629 "test/vadd-microkernel-tester.h",
5630 ] + MICROKERNEL_TEST_HDRS,
5631 deps = MICROKERNEL_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
5635 name = "u8_clamp_test",
5636 srcs = [
5637 "test/u8-clamp.cc",
5638 "test/clamp-microkernel-tester.h",
5639 ] + MICROKERNEL_TEST_HDRS,
5640 deps = MICROKERNEL_TEST_DEPS,
5641)
5642
5643xnnpack_unit_test(
5644 name = "u8_lut32norm_test",
5645 srcs = [
5646 "test/u8-lut32norm.cc",
5647 "test/lut-norm-microkernel-tester.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005653 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005655 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005656 "test/maxpool-microkernel-tester.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 deps = MICROKERNEL_TEST_DEPS,
5659)
5660
5661xnnpack_unit_test(
5662 name = "u8_rmax_test",
5663 srcs = [
5664 "test/u8-rmax.cc",
5665 "test/rmax-microkernel-tester.h",
5666 ] + MICROKERNEL_TEST_HDRS,
5667 deps = MICROKERNEL_TEST_DEPS,
5668)
5669
5670xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005671 name = "x32_fill_test",
5672 srcs = [
5673 "test/x32-fill.cc",
5674 "test/fill-microkernel-tester.h",
5675 ] + MICROKERNEL_TEST_HDRS,
5676 deps = MICROKERNEL_TEST_DEPS,
5677)
5678
5679xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005680 name = "x32_packx_test",
5681 srcs = [
5682 "test/x32-packx.cc",
5683 "test/pack-microkernel-tester.h",
5684 "src/xnnpack/AlignedAllocator.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 deps = MICROKERNEL_TEST_DEPS,
5687)
5688
5689xnnpack_unit_test(
5690 name = "x32_pad_test",
5691 srcs = [
5692 "test/x32-pad.cc",
5693 "test/pad-microkernel-tester.h",
5694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
5699 name = "x32_unpool_test",
5700 srcs = [
5701 "test/x32-unpool.cc",
5702 "test/unpool-microkernel-tester.h",
5703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
5708 name = "x32_zip_test",
5709 srcs = [
5710 "test/x32-zip.cc",
5711 "test/zip-microkernel-tester.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
5717 name = "x8_lut_test",
5718 srcs = [
5719 "test/x8-lut.cc",
5720 "test/lut-microkernel-tester.h",
5721 ] + MICROKERNEL_TEST_HDRS,
5722 deps = MICROKERNEL_TEST_DEPS,
5723)
5724
5725xnnpack_unit_test(
5726 name = "x8_zip_test",
5727 srcs = [
5728 "test/x8-zip.cc",
5729 "test/zip-microkernel-tester.h",
5730 ] + MICROKERNEL_TEST_HDRS,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
Marat Dukhan20c3b922020-03-10 03:45:06 -07005734########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005735
5736xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005737 name = "operator_size_test",
5738 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005739 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005740)
5741
Marat Dukhan20c3b922020-03-10 03:45:06 -07005742xnnpack_binary(
5743 name = "subgraph_size_test",
5744 srcs = ["test/subgraph-size.c"],
5745 deps = [":XNNPACK"],
5746)
5747
5748########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005749
5750xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005751 name = "abs_nc_test",
5752 srcs = [
5753 "test/abs-nc.cc",
5754 "test/abs-operator-tester.h",
5755 ],
5756 deps = OPERATOR_TEST_DEPS,
5757)
5758
5759xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005760 name = "add_nd_test",
5761 srcs = [
5762 "test/add-nd.cc",
5763 "test/binary-elementwise-operator-tester.h",
5764 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005765 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005766)
5767
5768xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005769 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005770 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005771 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005772 "test/argmax-pooling-operator-tester.h",
5773 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005774 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005775)
5776
5777xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005778 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005779 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005780 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005781 "test/average-pooling-operator-tester.h",
5782 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005783 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005784)
5785
5786xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005787 name = "bankers_rounding_nc_test",
5788 srcs = [
5789 "test/bankers-rounding-nc.cc",
5790 "test/bankers-rounding-operator-tester.h",
5791 ],
5792 deps = OPERATOR_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
5796 name = "ceiling_nc_test",
5797 srcs = [
5798 "test/ceiling-nc.cc",
5799 "test/ceiling-operator-tester.h",
5800 ],
5801 deps = OPERATOR_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005805 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005807 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005808 "test/channel-shuffle-operator-tester.h",
5809 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005810 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005811)
5812
5813xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005814 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005816 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005817 "test/clamp-operator-tester.h",
5818 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005819 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005820)
5821
5822xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005823 name = "constant_pad_nd_test",
5824 srcs = [
5825 "test/constant-pad-nd.cc",
5826 "test/constant-pad-operator-tester.h",
5827 ],
5828 deps = OPERATOR_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005832 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005833 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005834 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005835 "test/convolution-operator-tester.h",
5836 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005837 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005838)
5839
5840xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005841 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005842 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005843 "test/convolution-nchw.cc",
5844 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005845 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005846 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005847)
5848
5849xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005850 name = "copy_nc_test",
5851 srcs = [
5852 "test/copy-nc.cc",
5853 "test/copy-operator-tester.h",
5854 ],
5855 deps = OPERATOR_TEST_DEPS,
5856)
5857
5858xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005859 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005860 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005861 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005862 "test/deconvolution-operator-tester.h",
5863 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005864 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005865)
5866
5867xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005868 name = "divide_nd_test",
5869 srcs = [
5870 "test/binary-elementwise-operator-tester.h",
5871 "test/divide-nd.cc",
5872 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005873 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005874)
5875
5876xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005877 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005878 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005879 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005880 "test/fully-connected-operator-tester.h",
5881 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005882 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005883)
5884
5885xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005886 name = "floor_nc_test",
5887 srcs = [
5888 "test/floor-nc.cc",
5889 "test/floor-operator-tester.h",
5890 ],
5891 deps = OPERATOR_TEST_DEPS,
5892)
5893
5894xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005895 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005897 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005899 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005900 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005901)
5902
5903xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005904 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005905 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005906 "test/global-average-pooling-ncw.cc",
5907 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005909 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910)
5911
5912xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005913 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005914 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005915 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005916 "test/hardswish-operator-tester.h",
5917 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005918 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005919)
5920
5921xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005922 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005923 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005924 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005925 "test/leaky-relu-operator-tester.h",
5926 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005927 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005928)
5929
5930xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005931 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005932 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005933 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934 "test/max-pooling-operator-tester.h",
5935 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005936 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005937)
5938
5939xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005940 name = "maximum_nd_test",
5941 srcs = [
5942 "test/binary-elementwise-operator-tester.h",
5943 "test/maximum-nd.cc",
5944 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005945 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005946)
5947
5948xnnpack_unit_test(
5949 name = "minimum_nd_test",
5950 srcs = [
5951 "test/binary-elementwise-operator-tester.h",
5952 "test/minimum-nd.cc",
5953 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005954 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005955)
5956
5957xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005958 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005959 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005960 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005961 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005962 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005963 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005964)
5965
5966xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005967 name = "negate_nc_test",
5968 srcs = [
5969 "test/negate-nc.cc",
5970 "test/negate-operator-tester.h",
5971 ],
5972 deps = OPERATOR_TEST_DEPS,
5973)
5974
5975xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005976 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005977 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005978 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005979 "test/prelu-operator-tester.h",
5980 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005981 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982)
5983
5984xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005985 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005986 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005987 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005988 "test/resize-bilinear-operator-tester.h",
5989 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005990 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005991)
5992
5993xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005994 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005995 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005996 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005997 "test/sigmoid-operator-tester.h",
5998 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005999 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000)
6001
6002xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006003 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006004 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006005 "test/softmax-nc.cc",
6006 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006007 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006008 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009)
6010
6011xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006012 name = "square_nc_test",
6013 srcs = [
6014 "test/square-nc.cc",
6015 "test/square-operator-tester.h",
6016 ],
6017 deps = OPERATOR_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006021 name = "square_root_nc_test",
6022 srcs = [
6023 "test/square-root-nc.cc",
6024 "test/square-root-operator-tester.h",
6025 ],
6026 deps = OPERATOR_TEST_DEPS,
6027)
6028
6029xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006030 name = "squared_difference_nd_test",
6031 srcs = [
6032 "test/binary-elementwise-operator-tester.h",
6033 "test/squared-difference-nd.cc",
6034 ],
6035 deps = OPERATOR_TEST_DEPS,
6036)
6037
6038xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006039 name = "subtract_nd_test",
6040 srcs = [
6041 "test/binary-elementwise-operator-tester.h",
6042 "test/subtract-nd.cc",
6043 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006044 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006045)
6046
6047xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006048 name = "truncation_nc_test",
6049 srcs = [
6050 "test/truncation-nc.cc",
6051 "test/truncation-operator-tester.h",
6052 ],
6053 deps = OPERATOR_TEST_DEPS,
6054)
6055
6056xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006057 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006058 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006059 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006060 "test/unpooling-operator-tester.h",
6061 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006062 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063)
6064
Chao Mei6ddfc602020-05-13 22:29:36 -07006065############################### Misc unit tests ###############################
6066
6067xnnpack_unit_test(
6068 name = "memory_planner_test",
6069 srcs = [
6070 "test/memory-planner-test.cc",
6071 ],
6072 deps = [
6073 ":XNNPACK",
6074 ":memory_planner",
6075 ],
6076)
6077
Marat Dukhan08c4a432019-10-03 09:29:21 -07006078############################# Build configurations #############################
6079
Marat Dukhanb8642352019-10-30 15:43:02 -07006080# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006081config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006082 name = "xnn_enable_assembly_explicit_true",
6083 define_values = {"xnn_enable_assembly": "true"},
6084)
6085
6086# Disables usage of assembly kernels.
6087config_setting(
6088 name = "xnn_enable_assembly_explicit_false",
6089 define_values = {"xnn_enable_assembly": "false"},
6090)
6091
Marat Dukhan9de90e02020-06-18 16:04:12 -07006092# Enables usage of sparse inference.
6093config_setting(
6094 name = "xnn_enable_sparse_explicit_true",
6095 define_values = {"xnn_enable_sparse": "true"},
6096)
6097
6098# Disables usage of sparse inference.
6099config_setting(
6100 name = "xnn_enable_sparse_explicit_false",
6101 define_values = {"xnn_enable_sparse": "false"},
6102)
6103
Marat Dukhan05702cf2020-03-26 15:41:33 -07006104# Disables usage of HMP-aware optimizations.
6105config_setting(
6106 name = "xnn_enable_hmp_explicit_false",
6107 define_values = {"xnn_enable_hmp": "false"},
6108)
6109
Chao Mei6ddfc602020-05-13 22:29:36 -07006110# Enable usage of optimized memory allocation
6111config_setting(
6112 name = "xnn_enable_memopt_explicit_true",
6113 define_values = {"xnnpack_enable_memopt": "true"},
6114)
6115
6116# Disable usage of optimized memory allocation
6117config_setting(
6118 name = "xnn_enable_memopt_explicit_false",
6119 define_values = {"xnnpack_enable_memopt": "false"},
6120)
6121
Marat Dukhanb8642352019-10-30 15:43:02 -07006122# Builds with -c dbg
6123config_setting(
6124 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006125 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006126 "compilation_mode": "dbg",
6127 },
6128)
6129
6130# Builds with -c opt
6131config_setting(
6132 name = "optimized_build",
6133 values = {
6134 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135 },
6136)
6137
6138config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006139 name = "linux_k8",
6140 values = {"cpu": "k8"},
6141)
6142
6143config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006144 name = "linux_arm",
6145 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006146)
6147
6148config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006149 name = "linux_armeabi",
6150 values = {"cpu": "armeabi"},
6151)
6152
6153config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006154 name = "linux_armhf",
6155 values = {"cpu": "armhf"},
6156)
6157
6158config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006159 name = "linux_armv7a",
6160 values = {"cpu": "armv7a"},
6161)
6162
6163config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006164 name = "linux_aarch64",
6165 values = {"cpu": "aarch64"},
6166)
6167
6168config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006169 name = "android",
6170 values = {"crosstool_top": "//external:android/crosstool"},
6171)
6172
6173config_setting(
6174 name = "android_armv7",
6175 values = {
6176 "crosstool_top": "//external:android/crosstool",
6177 "cpu": "armeabi-v7a",
6178 },
6179)
6180
6181config_setting(
6182 name = "android_arm64",
6183 values = {
6184 "crosstool_top": "//external:android/crosstool",
6185 "cpu": "arm64-v8a",
6186 },
6187)
6188
6189config_setting(
6190 name = "android_x86",
6191 values = {
6192 "crosstool_top": "//external:android/crosstool",
6193 "cpu": "x86",
6194 },
6195)
6196
6197config_setting(
6198 name = "android_x86_64",
6199 values = {
6200 "crosstool_top": "//external:android/crosstool",
6201 "cpu": "x86_64",
6202 },
6203)
6204
6205config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006206 name = "windows_x86_64",
6207 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006208)
6209
6210config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006211 name = "windows_x86_64_clang",
6212 values = {
6213 "compiler": "clang-cl",
6214 "cpu": "x64_windows",
6215 },
6216)
6217
6218config_setting(
6219 name = "windows_x86_64_mingw",
6220 values = {
6221 "compiler": "mingw-gcc",
6222 "cpu": "x64_windows",
6223 },
6224)
6225
6226config_setting(
6227 name = "windows_x86_64_msys",
6228 values = {
6229 "compiler": "msys-gcc",
6230 "cpu": "x64_windows",
6231 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006232)
6233
6234config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006235 name = "macos_x86_64",
6236 values = {
6237 "apple_platform_type": "macos",
6238 "cpu": "darwin",
6239 },
6240)
6241
6242config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006244 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245)
6246
6247config_setting(
6248 name = "emscripten_wasm",
6249 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006250 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006251 "cpu": "wasm",
6252 },
6253)
6254
6255config_setting(
6256 name = "emscripten_wasmsimd",
6257 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006258 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006260 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 },
6262)
6263
6264config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006265 name = "ios_armv7",
6266 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006267 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006268 "cpu": "ios_armv7",
6269 },
6270)
6271
6272config_setting(
6273 name = "ios_arm64",
6274 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006275 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006276 "cpu": "ios_arm64",
6277 },
6278)
6279
6280config_setting(
6281 name = "ios_arm64e",
6282 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006283 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006284 "cpu": "ios_arm64e",
6285 },
6286)
6287
6288config_setting(
6289 name = "ios_x86",
6290 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006291 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006292 "cpu": "ios_i386",
6293 },
6294)
6295
6296config_setting(
6297 name = "ios_x86_64",
6298 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006299 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006300 "cpu": "ios_x86_64",
6301 },
6302)
6303
6304config_setting(
6305 name = "watchos_armv7k",
6306 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006307 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006308 "cpu": "watchos_armv7k",
6309 },
6310)
6311
6312config_setting(
6313 name = "watchos_arm64_32",
6314 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006315 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006316 "cpu": "watchos_arm64_32",
6317 },
6318)
6319
6320config_setting(
6321 name = "watchos_x86",
6322 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006323 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006324 "cpu": "watchos_i386",
6325 },
6326)
6327
6328config_setting(
6329 name = "watchos_x86_64",
6330 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006331 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006332 "cpu": "watchos_x86_64",
6333 },
6334)
6335
6336config_setting(
6337 name = "tvos_arm64",
6338 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006339 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006340 "cpu": "tvos_arm64",
6341 },
6342)
6343
6344config_setting(
6345 name = "tvos_x86_64",
6346 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006347 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006348 "cpu": "tvos_x86_64",
6349 },
6350)