blob: 4a984f17e885ba644509af61eac20046929bdc63 [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 Dukhan5fab4092020-06-10 01:28:28 -070087 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070088 "src/subgraph/fully-connected.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",
Frank Barchard04336c12020-10-22 16:48:55 -0700100 "src/subgraph/square-root.c",
101 "src/subgraph/square.c",
102 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700103 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700104 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700105 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700106 "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",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700113 "src/tables/exp2minus-k-over-64.c",
114 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115]
116
Marat Dukhan08c4a432019-10-03 09:29:21 -0700117SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800119 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700121 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
122 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700123 "src/f32-clamp/gen/scalar-x1.c",
124 "src/f32-clamp/gen/scalar-x2.c",
125 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700126 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700127 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700128 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700129 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700130 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700133 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700134 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
135 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
143 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
144 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
147 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
148 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -0700153 "src/f32-dwconv2d-chw/3x3p1-scalar-1x1-acc3.c",
154 "src/f32-dwconv2d-chw/3x3s2p1-scalar-1x1-acc3.c",
155 "src/f32-dwconv2d-chw/5x5p2-scalar-1x1-acc5.c",
156 "src/f32-dwconv2d-chw/5x5s2p2-scalar-1x1-acc5.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700157 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700158 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
159 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700160 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
161 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
162 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700163 "src/f32-gemm/gen/1x4-minmax-scalar.c",
164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/1x4-scalar.c",
166 "src/f32-gemm/gen/2x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-relu-scalar.c",
168 "src/f32-gemm/gen/2x4-scalar.c",
169 "src/f32-gemm/gen/4x2-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-relu-scalar.c",
171 "src/f32-gemm/gen/4x2-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
173 "src/f32-gemm/gen/4x4-relu-scalar.c",
174 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700175 "src/f32-hswish/gen/hswish-scalar-x1.c",
176 "src/f32-hswish/gen/hswish-scalar-x2.c",
177 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700178 "src/f32-ibilinear-chw/gen/scalar-p1.c",
179 "src/f32-ibilinear-chw/gen/scalar-p2.c",
180 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700181 "src/f32-ibilinear/gen/scalar-c1.c",
182 "src/f32-ibilinear/gen/scalar-c2.c",
183 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700184 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700185 "src/f32-igemm/gen/1x4-relu-scalar.c",
186 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700188 "src/f32-igemm/gen/2x4-relu-scalar.c",
189 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700190 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700191 "src/f32-igemm/gen/4x2-relu-scalar.c",
192 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700194 "src/f32-igemm/gen/4x4-relu-scalar.c",
195 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700196 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
197 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
198 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700199 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
200 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
201 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
202 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800203 "src/f32-prelu/gen/scalar-2x1.c",
204 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700210 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
215 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700216 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700217 "src/f32-relu/gen/scalar-x1.c",
218 "src/f32-relu/gen/scalar-x2.c",
219 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700220 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700221 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800222 "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",
Frank Barchard04336c12020-10-22 16:48:55 -0700225 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800228 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
229 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
230 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700231 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/1x1-minmax-scalar.c",
233 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/2x1-minmax-scalar.c",
235 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/4x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
238 "src/f32-spmm/gen/8x1-minmax-scalar.c",
239 "src/f32-spmm/gen/8x2-minmax-scalar.c",
240 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700244 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700245 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
246 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
247 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700248 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "src/f32-vbinary/gen/vadd-scalar-x1.c",
250 "src/f32-vbinary/gen/vadd-scalar-x2.c",
251 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700252 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700256 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700257 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
258 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
259 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700260 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700261 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
262 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
263 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700264 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700268 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700269 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
270 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
271 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700272 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
274 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
275 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700276 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700280 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700281 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
282 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
283 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700284 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
286 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
287 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700288 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800289 "src/f32-vbinary/gen/vmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700292 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800293 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
294 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
295 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700296 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800297 "src/f32-vbinary/gen/vmin-scalar-x1.c",
298 "src/f32-vbinary/gen/vmin-scalar-x2.c",
299 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700300 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800301 "src/f32-vbinary/gen/vminc-scalar-x1.c",
302 "src/f32-vbinary/gen/vminc-scalar-x2.c",
303 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700304 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700305 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
306 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
307 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700308 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700309 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
310 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
311 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700312 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700313 "src/f32-vbinary/gen/vmul-scalar-x1.c",
314 "src/f32-vbinary/gen/vmul-scalar-x2.c",
315 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700316 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
319 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700320 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700321 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
322 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
323 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700324 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700325 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
326 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
327 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700328 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
331 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700332 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
335 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700336 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700337 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
338 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
339 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700340 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
343 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700344 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
347 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700348 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700349 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
350 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
351 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700352 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700353 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
354 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
355 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700356 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
359 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700360 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700361 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
362 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
363 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700364 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700365 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
366 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
367 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700368 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700369 "src/f32-vbinary/gen/vsub-scalar-x1.c",
370 "src/f32-vbinary/gen/vsub-scalar-x2.c",
371 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700372 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
375 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700376 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700377 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
378 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
379 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700380 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700381 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
382 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
383 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700384 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700385 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
386 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
387 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700388 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
389 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
390 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700391 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
392 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
393 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700394 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
395 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
396 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
397 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
398 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
399 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
400 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
401 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
402 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700403 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
404 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
405 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700406 "src/f32-vunary/gen/vabs-scalar-x1.c",
407 "src/f32-vunary/gen/vabs-scalar-x2.c",
408 "src/f32-vunary/gen/vabs-scalar-x4.c",
409 "src/f32-vunary/gen/vneg-scalar-x1.c",
410 "src/f32-vunary/gen/vneg-scalar-x2.c",
411 "src/f32-vunary/gen/vneg-scalar-x4.c",
412 "src/f32-vunary/gen/vsqr-scalar-x1.c",
413 "src/f32-vunary/gen/vsqr-scalar-x2.c",
414 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800415 "src/math/expminus-scalar-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700416 "src/math/expminus-scalar-lut2048-p1.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800417 "src/math/expminus-scalar-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700418 "src/math/roundd-scalar-addsub.c",
419 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700421 "src/math/roundne-scalar-addsub.c",
422 "src/math/roundne-scalar-nearbyint.c",
423 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700424 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700425 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700426 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700427 "src/math/roundz-scalar-addsub.c",
428 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700429 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700430 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700431 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700432 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700433 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
434 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700435 "src/qs8-requantization/fp32-scalar-lrintf.c",
436 "src/qs8-requantization/fp32-scalar-magic.c",
437 "src/qs8-requantization/precise-scalar-signed64.c",
438 "src/qs8-requantization/precise-scalar-unsigned32.c",
439 "src/qs8-requantization/precise-scalar-unsigned64.c",
440 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700441 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
442 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
443 "src/qu8-dwconv/up1x9-minmax-scalar.c",
444 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
445 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700446 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700447 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
448 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700449 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700450 "src/qu8-requantization/fp32-scalar-lrintf.c",
451 "src/qu8-requantization/fp32-scalar-magic.c",
452 "src/qu8-requantization/precise-scalar-signed64.c",
453 "src/qu8-requantization/precise-scalar-unsigned32.c",
454 "src/qu8-requantization/precise-scalar-unsigned64.c",
455 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700456 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700457 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700459 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700461 "src/x8-lut/scalar.c",
462 "src/x8-zip/x2-scalar.c",
463 "src/x8-zip/x3-scalar.c",
464 "src/x8-zip/x4-scalar.c",
465 "src/x8-zip/xm-scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700466 "src/x32-fill/scalar-float.c",
467 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700468 "src/x32-packx/x2-scalar.c",
469 "src/x32-packx/x3-scalar.c",
470 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700471 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700472 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473 "src/x32-unpool/scalar.c",
474 "src/x32-zip/x2-scalar.c",
475 "src/x32-zip/x3-scalar.c",
476 "src/x32-zip/x4-scalar.c",
477 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan436ebe62019-12-04 15:10:12 -0800480WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700481 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
482 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700483 "src/f32-clamp/gen/wasm-x1.c",
484 "src/f32-clamp/gen/wasm-x2.c",
485 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700486 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
487 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700488 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
489 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700490 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
491 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
493 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700494 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
495 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700496 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
497 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700498 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
499 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700500 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
501 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700502 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
503 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700504 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
505 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700506 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
507 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700508 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
509 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
511 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700512 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
513 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
514 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
515 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700516 "src/f32-gemm/gen/1x4-relu-wasm.c",
517 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700519 "src/f32-gemm/gen/2x4-relu-wasm.c",
520 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700521 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700522 "src/f32-gemm/gen/4x2-relu-wasm.c",
523 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700524 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700525 "src/f32-gemm/gen/4x4-relu-wasm.c",
526 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700527 "src/f32-hswish/gen/hswish-wasm-x1.c",
528 "src/f32-hswish/gen/hswish-wasm-x2.c",
529 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700530 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-igemm/gen/1x4-relu-wasm.c",
532 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700534 "src/f32-igemm/gen/2x4-relu-wasm.c",
535 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700536 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700537 "src/f32-igemm/gen/4x2-relu-wasm.c",
538 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700539 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700540 "src/f32-igemm/gen/4x4-relu-wasm.c",
541 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700542 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
543 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
544 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700545 "src/f32-prelu/gen/wasm-2x1.c",
546 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700547 "src/f32-relu/gen/wasm-x1.c",
548 "src/f32-relu/gen/wasm-x2.c",
549 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700550 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700551 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
552 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
553 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700554 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700555 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
556 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
557 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700558 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
560 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
561 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
562 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700563 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
564 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
565 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
568 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
569 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
570 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700571 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
573 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700574 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
576 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
577 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
578 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700579 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
580 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
581 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700582 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800583 "src/f32-vbinary/gen/vmax-wasm-x1.c",
584 "src/f32-vbinary/gen/vmax-wasm-x2.c",
585 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700586 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800587 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
588 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
589 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700590 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800591 "src/f32-vbinary/gen/vmin-wasm-x1.c",
592 "src/f32-vbinary/gen/vmin-wasm-x2.c",
593 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700594 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800595 "src/f32-vbinary/gen/vminc-wasm-x1.c",
596 "src/f32-vbinary/gen/vminc-wasm-x2.c",
597 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700598 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700599 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
600 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
601 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700602 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700603 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
604 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
605 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700606 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700607 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
608 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
609 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
610 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700611 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
612 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
613 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700614 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700615 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
616 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
617 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
618 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700619 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
620 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
621 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700622 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700623 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
624 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
625 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
626 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700627 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
628 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
629 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700631 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
632 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
633 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
634 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700635 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
637 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700638 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700639 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
640 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
641 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
642 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700643 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
644 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
645 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700646 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700647 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
648 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
649 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
651 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
652 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800653]
654
Marat Dukhan290055c2020-06-09 12:24:29 -0700655WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700656 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
657 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
658 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700659 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
660 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
661 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
662 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700663 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
664 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
665 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
666 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700667 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700669 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700671 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700675 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700676 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
677 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
678 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
679 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
681 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
682 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700683 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700684 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700685 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700686 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
687 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700688 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700689 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700691 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
692 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
693 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
694 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
695 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
696 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
697 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
698 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700699 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
700 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
701 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
702 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700703 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
705 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
706 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
707 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
708 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
709 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
711 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
712 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
713 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
714 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
715 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
716 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
717 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
718 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
719 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
720 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
722 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
723 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
724 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
725 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
726 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
728 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
729 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
730 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
731 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
732 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
733 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
734 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
735 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
736 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
737 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
738 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
739 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
740 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
741 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
743 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
744 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
745 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
746 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
747 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
748 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
749 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
750 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
751 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
752 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
753 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
754 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
756 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
757 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
758 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
760 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
761 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
762 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
763 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
764 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
765 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
766 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
767 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
768 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
769 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
770 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
771 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
772 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700773 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
774 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
775 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700776 "src/f32-ibilinear/gen/wasmsimd-c4.c",
777 "src/f32-ibilinear/gen/wasmsimd-c8.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700778 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
779 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700780 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
782 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
784 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
785 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
786 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
788 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
789 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
793 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
794 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
795 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
797 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
798 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
799 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
801 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
802 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
803 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
804 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
805 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
806 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
807 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
808 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
809 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
810 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
811 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
812 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
813 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
814 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
815 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
816 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
817 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
818 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
819 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700820 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
821 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700822 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
823 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
824 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
825 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700826 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
827 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700828 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
830 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700831 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
832 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700833 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
834 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
835 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
836 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
839 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700840 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
841 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700842 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
843 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
844 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
845 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700851 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700852 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
853 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700854 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700855 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
856 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700857 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700858 "src/f32-relu/gen/wasmsimd-x4.c",
859 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700860 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700861 "src/f32-rmax/wasmsimd-arm.c",
862 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
868 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700869 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
870 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
871 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
872 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
873 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
874 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
875 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll2.c",
876 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll4.c",
877 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
878 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll2.c",
879 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll4.c",
880 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
881 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll2.c",
882 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll4.c",
883 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
884 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll2.c",
885 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll4.c",
886 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
887 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll2.c",
888 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll4.c",
889 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
890 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll2.c",
891 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll4.c",
892 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700893 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
894 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700895 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700896 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
897 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700898 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700899 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
900 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700901 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700902 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
903 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700904 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700905 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
906 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700907 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700908 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
909 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700910 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700911 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
912 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700913 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700914 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
915 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700916 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700917 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
918 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700919 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700920 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
921 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700922 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700923 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
924 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700925 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700926 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
927 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700928 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700929 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
930 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700931 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700932 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
933 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700934 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700935 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
936 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700937 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700938 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
939 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700940 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700941 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
942 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700943 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700944 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
945 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700946 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700947 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
948 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700949 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700950 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
951 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700952 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700953 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
954 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700955 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700956 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
957 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700958 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700959 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
960 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700961 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700962 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
963 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700964 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700965 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
966 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700967 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700968 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
969 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700970 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700971 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
972 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700973 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700974 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
975 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700976 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700977 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
978 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700979 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700980 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
981 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700982 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700983 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
984 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700985 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700986 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
987 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700988 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700989 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
990 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700991 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700992 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
993 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700994 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700995 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
996 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700997 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700998 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
999 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001000 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001001 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1002 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001003 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001004 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1005 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001006 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001007 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1008 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001009 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001010 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1011 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001012 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001013 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1014 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001015 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001016 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1017 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001018 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001019 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1020 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001021 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001022 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1023 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001024 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001025 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1026 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001027 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001028 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1029 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001030 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001031 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1032 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001033 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001034 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1035 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001036 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001037 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1038 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001039 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001040 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1041 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001042 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001043 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1044 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1045 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1046 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001047 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001048 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001049 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001050 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001051 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1052 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1053 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1054 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001055 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1056 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1057 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1058 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1059 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1060 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1061 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1062 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1063 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1064 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001065 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1066 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001067 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1068 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1069 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1070 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1071 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1072 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001073 "src/math/roundd-wasmsimd-addsub.c",
1074 "src/math/roundd-wasmsimd-cvt.c",
1075 "src/math/roundne-wasmsimd-addsub.c",
1076 "src/math/roundu-wasmsimd-addsub.c",
1077 "src/math/roundu-wasmsimd-cvt.c",
1078 "src/math/roundz-wasmsimd-addsub.c",
1079 "src/math/roundz-wasmsimd-cvt.c",
1080 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1081 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001082 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1083 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1084 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001085 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1086 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1087 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001088 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1089 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1090 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001091 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001092 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001093 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001094 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1095 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001096 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001097 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1098 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001099 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1100 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001101 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001102 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001103 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001104 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001105 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001106 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001107 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001108 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1109 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1110 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1111 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1112 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1113 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1114 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1115 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001116 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001117 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001118 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001119 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001120 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001121 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001122 "src/x32-zip/x2-wasmsimd.c",
1123 "src/x32-zip/x3-wasmsimd.c",
1124 "src/x32-zip/x4-wasmsimd.c",
1125 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001126]
1127
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001128PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001129 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001130 "src/f32-dwconv2d-chw/3x3p1-psimd-1x4-acc3.c",
1131 "src/f32-dwconv2d-chw/3x3s2p1-psimd-1x4-acc3.c",
1132 "src/f32-dwconv2d-chw/5x5p2-psimd-3x4.c",
1133 "src/f32-dwconv2d-chw/5x5s2p2-psimd-1x4-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001134]
1135
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001136PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001137 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001138 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001139 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001141]
1142
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143# ISA-specific micro-kernels
1144NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001145 "src/f32-argmaxpool/4x-neon-c4.c",
1146 "src/f32-argmaxpool/9p8x-neon-c4.c",
1147 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001148 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1149 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001150 "src/f32-clamp/gen/neon-x4.c",
1151 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001152 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001153 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001154 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001155 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001156 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001157 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001158 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001159 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001160 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001161 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001162 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001163 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001164 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001165 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1166 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1167 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1168 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1169 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001170 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001171 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001172 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001173 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1174 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001176 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1177 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001178 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001179 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1181 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1182 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001184 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1185 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001186 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1187 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001188 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1189 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001190 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1191 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1192 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1193 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1194 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1195 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1196 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1197 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1198 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1199 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1200 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1201 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1202 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1203 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1204 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1205 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001206 "src/f32-hswish/gen/hswish-neon-x4.c",
1207 "src/f32-hswish/gen/hswish-neon-x8.c",
1208 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001209 "src/f32-ibilinear/gen/neon-c4.c",
1210 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001211 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001212 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001213 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001214 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1215 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001216 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001217 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1218 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1219 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1220 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001221 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1222 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001223 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1224 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001225 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1226 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001227 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1228 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1229 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001230 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1231 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001232 "src/f32-prelu/gen/neon-1x4.c",
1233 "src/f32-prelu/gen/neon-1x8.c",
1234 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001235 "src/f32-prelu/gen/neon-2x4.c",
1236 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001237 "src/f32-prelu/gen/neon-2x16.c",
1238 "src/f32-prelu/gen/neon-4x4.c",
1239 "src/f32-prelu/gen/neon-4x8.c",
1240 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001243 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001244 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1245 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001246 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001247 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1248 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001249 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001250 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1251 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001252 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1253 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1254 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1255 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1256 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1257 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1258 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1259 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1260 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1261 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1262 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1263 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1264 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001265 "src/f32-relu/gen/neon-x4.c",
1266 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001267 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001268 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001269 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1270 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1271 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1272 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1273 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1274 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1275 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1276 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1277 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1278 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1279 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1280 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001281 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1282 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1283 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1284 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1285 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1286 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001291 "src/f32-vbinary/gen/vmax-neon-x4.c",
1292 "src/f32-vbinary/gen/vmax-neon-x8.c",
1293 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1294 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1295 "src/f32-vbinary/gen/vmin-neon-x4.c",
1296 "src/f32-vbinary/gen/vmin-neon-x8.c",
1297 "src/f32-vbinary/gen/vminc-neon-x4.c",
1298 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001299 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1300 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1301 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1302 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1303 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1304 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001305 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1306 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1307 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1308 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001309 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1310 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1311 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1312 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001313 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1314 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001315 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1316 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001317 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1318 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001319 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1320 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1321 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1322 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1323 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1324 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001325 "src/f32-vunary/gen/vabs-neon-x4.c",
1326 "src/f32-vunary/gen/vabs-neon-x8.c",
1327 "src/f32-vunary/gen/vneg-neon-x4.c",
1328 "src/f32-vunary/gen/vneg-neon-x8.c",
1329 "src/f32-vunary/gen/vsqr-neon-x4.c",
1330 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/math/roundd-neon-addsub.c",
1332 "src/math/roundd-neon-cvt.c",
1333 "src/math/roundne-neon-addsub.c",
1334 "src/math/roundu-neon-addsub.c",
1335 "src/math/roundu-neon-cvt.c",
1336 "src/math/roundz-neon-addsub.c",
1337 "src/math/roundz-neon-cvt.c",
1338 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1339 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1340 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1341 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1342 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1343 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1344 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1345 "src/math/sqrt-neon-nr1rsqrts.c",
1346 "src/math/sqrt-neon-nr2rsqrts.c",
1347 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001348 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1349 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1350 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1351 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001352 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1353 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1354 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1355 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001356 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1357 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1358 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1359 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001360 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001361 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001362 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001363 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1364 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001365 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001366 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001367 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001368 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001369 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001370 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001371 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1372 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1373 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1374 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1375 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1376 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1377 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1378 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001379 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1380 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1381 "src/qu8-dwconv/up8x9-minmax-neon.c",
1382 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1383 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1384 "src/qu8-gemm/4x8-minmax-neon.c",
1385 "src/qu8-gemm/8x8-minmax-neon.c",
1386 "src/qu8-igemm/4x8-minmax-neon.c",
1387 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001388 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001389 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001390 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001391 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001392 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001393 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001394 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001395 "src/x8-zip/x2-neon.c",
1396 "src/x8-zip/x3-neon.c",
1397 "src/x8-zip/x4-neon.c",
1398 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001399 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001400 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001401 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001402 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403 "src/x32-zip/x2-neon.c",
1404 "src/x32-zip/x3-neon.c",
1405 "src/x32-zip/x4-neon.c",
1406 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001407]
1408
1409NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001410 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1411 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1412 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1413 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1414 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1415 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1416 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1417 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1418 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1419 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1420 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1421 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1422 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1424 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1425 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1428 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1429 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1430 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1431 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1432 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1433 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1434 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1435 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1436 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1437 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1438 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1439 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001440 "src/f32-ibilinear/gen/neonfma-c4.c",
1441 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001442 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001444 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001445 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1446 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001447 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1448 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001449 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1450 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001451 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1452 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001455 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001456 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1457 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001459 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1460 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001461 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001462 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1463 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001464 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1465 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1466 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1467 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1468 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1469 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1470 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1471 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1472 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1473 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1474 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1475 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1476 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1510 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1511 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1512 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001513 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1514 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1515 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1516 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1517 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1518 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1519 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1520 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1521 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1522 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1523 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1524 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1525 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1526 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1527 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1528 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1529 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1530 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001531 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1532 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001533 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1534 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1535 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1536 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1537 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1538 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1539 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1540 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1541 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1542 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1543 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1544 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1545 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1546 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1547 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1548 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1549 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1550 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1551 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1552 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001553 "src/math/exp-neonfma-lut64-p2.c",
1554 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001555 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001556 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001557 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001558 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1559 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1560 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1562 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1563 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001564 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1565 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1566 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001567 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1568 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1569 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001570 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1571 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1572 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001573 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1574 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1575 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001576 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001577 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001578 "src/math/sqrt-neonfma-nr2fma.c",
1579 "src/math/sqrt-neonfma-nr2fma1adj.c",
1580 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581]
1582
1583AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001584 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001585 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001586 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001587 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001589 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001590 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001591 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001592 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001593 "src/f32-dwconv2d-chw/3x3p1-neonfma-3x4.c",
1594 "src/f32-dwconv2d-chw/3x3s2p1-neonfma-1x4-acc3.c",
1595 "src/f32-dwconv2d-chw/5x5p2-neonfma-3x4.c",
1596 "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001597 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1598 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1599 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1600 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1601 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1602 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1603 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1604 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1605 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1606 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1607 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1608 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1609 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1610 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1611 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1612 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1613 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1614 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1615 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1616 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001629 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1633 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001635 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1636 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1637 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1638 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1639 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1640 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1641 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1642 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1643 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1644 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001645 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1646 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1647 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1648 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1649 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1650 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1651 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1652 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1653 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1654 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1655 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1656 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1657 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1658 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001659 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1660 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001661 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001662 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001663 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001664 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001665 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001666 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001667]
1668
Marat Dukhan8853b822020-05-07 12:19:01 -07001669NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001670 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1671 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1673 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1674 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1675 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1676 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1677 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001678 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001680 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001681 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001682]
1683
Marat Dukhan08c4a432019-10-03 09:29:21 -07001684AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001685 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001686 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001687 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1688 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1689 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1690 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001691 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1692 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1693 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1694 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1695 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1696 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1697 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1698 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001699 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1700 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001701 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1702 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1703 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1704 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1705 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1706 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1707 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1708 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1709 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1710 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1711 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1712 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1713 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1714 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1715 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001717 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001718 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1719 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1720 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1721 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1722 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1723 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001727 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001728 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001729 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001730 "src/f16-relu/gen/neonfp16arith-x16.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001731 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001732 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001733 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001734 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001735 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001736 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001737 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001738 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1739 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1740 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1741 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1742 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1743 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1744 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1745 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1746 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1747 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1748 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1749 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1750 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1751 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1752 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1753 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1754 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1755 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1756 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1757 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1758 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1759 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1760 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1761 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1762 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1763 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1764 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1765 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1766 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1767 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1768 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769]
1770
Benoit Jacoba9644732020-08-13 12:48:55 -07001771NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001772 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001773 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001774 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001775 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001776 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001777 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001778 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001779 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001780 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001781 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001782 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001783 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001784 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001785 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001786 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001787 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001788 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001789 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001790]
1791
Marat Dukhan08c4a432019-10-03 09:29:21 -07001792SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001793 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1794 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001795 "src/f32-clamp/gen/sse-x4.c",
1796 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001797 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1798 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001799 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1800 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1801 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1802 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001803 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1804 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001805 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1806 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1807 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1808 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001809 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1810 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001811 "src/f32-dwconv2d-chw/3x3p1-sse-1x4-acc3.c",
1812 "src/f32-dwconv2d-chw/3x3s2p1-sse-1x4-acc3.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001813 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001814 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1815 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001816 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1817 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1818 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1819 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1820 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1821 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001822 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1823 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1824 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1825 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1826 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1827 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1828 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001829 "src/f32-hswish/gen/hswish-sse-x4.c",
1830 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001831 "src/f32-ibilinear/gen/sse-c4.c",
1832 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001833 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1834 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1835 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1836 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1837 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1838 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1839 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001840 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1841 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1842 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001843 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001844 "src/f32-prelu/gen/sse-2x4.c",
1845 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001846 "src/f32-relu/gen/sse-x4.c",
1847 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001849 "src/f32-spmm/gen/4x1-minmax-sse.c",
1850 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001851 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001852 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1854 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1855 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1856 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1857 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1858 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1859 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001860 "src/f32-vbinary/gen/vmax-sse-x4.c",
1861 "src/f32-vbinary/gen/vmax-sse-x8.c",
1862 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1863 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1864 "src/f32-vbinary/gen/vmin-sse-x4.c",
1865 "src/f32-vbinary/gen/vmin-sse-x8.c",
1866 "src/f32-vbinary/gen/vminc-sse-x4.c",
1867 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001868 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1869 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1870 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1871 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1872 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1873 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1874 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1875 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001876 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1877 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1878 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1879 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001880 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1881 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1882 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1883 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001884 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1885 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001886 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1887 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001888 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1889 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001890 "src/f32-vunary/gen/vabs-sse-x4.c",
1891 "src/f32-vunary/gen/vabs-sse-x8.c",
1892 "src/f32-vunary/gen/vneg-sse-x4.c",
1893 "src/f32-vunary/gen/vneg-sse-x8.c",
1894 "src/f32-vunary/gen/vsqr-sse-x4.c",
1895 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001896 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001897 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001898 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001899 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001900 "src/math/sqrt-sse-hh1mac.c",
1901 "src/math/sqrt-sse-nr1mac.c",
1902 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001903 "src/x32-fill/sse.c",
1904 "src/x32-packx/x4-sse.c",
1905 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906]
1907
1908SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001909 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001910 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08001911 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001912 "src/f32-prelu/gen/sse2-2x4.c",
1913 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001914 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001915 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001916 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001917 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001919 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001920 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1921 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001922 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001923 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1924 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001926 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1927 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1928 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1929 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1930 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1931 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001932 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001933 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001934 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001935 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001936 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1937 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001938 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1939 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001940 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1941 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001942 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1943 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1944 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1945 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1946 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1947 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1948 "src/math/exp-sse2-lut64-p2.c",
1949 "src/math/exp-sse2-p5.c",
1950 "src/math/expminus-sse2-p5.c",
1951 "src/math/roundd-sse2-cvt.c",
1952 "src/math/roundne-sse2-cvt.c",
1953 "src/math/roundu-sse2-cvt.c",
1954 "src/math/roundz-sse2-cvt.c",
1955 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1956 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1957 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1958 "src/math/sigmoid-sse2-rr2-p5-div.c",
1959 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1960 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001961 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1962 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1963 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001964 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1965 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1966 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001967 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1968 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1969 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001970 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001971 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001972 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001973 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001974 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001975 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001976 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
1977 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001978 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001979 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
1980 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001981 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001982 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
1983 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
1984 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001985 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001986 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001987 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001988 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001989 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001990 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001991 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001992 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001993 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
1994 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001995 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001996 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001997 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001998 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1999 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2000 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2001 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002002 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2003 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2004 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2005 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002006 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2007 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002008 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2009 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2010 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2011 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2012 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002013 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002014 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002015 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002016 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002017 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002018 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002019 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002021 "src/x8-zip/x2-sse2.c",
2022 "src/x8-zip/x3-sse2.c",
2023 "src/x8-zip/x4-sse2.c",
2024 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002025 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002026 "src/x32-zip/x2-sse2.c",
2027 "src/x32-zip/x3-sse2.c",
2028 "src/x32-zip/x4-sse2.c",
2029 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002030]
2031
2032SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002033 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2034 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2035 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002036 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2037 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2038 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002039 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2040 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2041 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002042 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002043 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002044 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002045 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002046 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002047 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2049 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002050 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002051 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2052 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002053 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002054 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2055 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2056 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002057 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002058 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002059 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002060 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002061 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002062 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002063 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002064 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002065 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2066 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002067 "src/qs8-requantization/precise-ssse3.c",
2068 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002069 "src/qu8-requantization/precise-ssse3.c",
2070 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002071]
2072
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002073SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002074 "src/f32-prelu/gen/sse41-2x4.c",
2075 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002076 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2077 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2078 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2079 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2080 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2081 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002082 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002083 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002084 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002085 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002086 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2087 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002088 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2089 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002090 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2091 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002092 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2093 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2094 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2095 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2096 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2097 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2098 "src/math/roundd-sse41.c",
2099 "src/math/roundne-sse41.c",
2100 "src/math/roundu-sse41.c",
2101 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002102 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2103 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2104 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002105 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2106 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2107 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002108 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2109 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2110 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002111 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002112 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002113 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002114 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002115 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002116 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002117 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2118 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002119 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002120 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2121 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002122 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002123 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2124 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2125 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002126 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002127 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002128 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002129 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002130 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002131 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002133 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002134 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2135 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002136 "src/qs8-requantization/fp32-sse4.c",
2137 "src/qs8-requantization/precise-sse4.c",
2138 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002139 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2140 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2141 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2142 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002143 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2144 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2145 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2146 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002147 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2148 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2149 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2150 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002151 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2152 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2153 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2154 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002155 "src/qu8-requantization/precise-sse4.c",
2156 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002157]
2158
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002160 "src/f32-clamp/gen/avx-x8.c",
2161 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002162 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2163 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002164 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2165 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002166 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2167 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002168 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2169 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2170 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2171 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2172 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2173 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002174 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002175 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2176 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002178 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002180 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002181 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2182 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2183 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2184 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2185 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2186 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2187 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2188 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2189 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2190 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2191 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002192 "src/f32-hswish/gen/hswish-avx-x8.c",
2193 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002194 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002195 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2196 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002197 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002198 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002199 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002200 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2202 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002203 "src/f32-prelu/gen/avx-2x8.c",
2204 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002205 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002206 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002207 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002208 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2209 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2210 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2211 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2212 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2213 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2214 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2215 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2216 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2217 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2218 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2219 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2220 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2221 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2222 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2223 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2224 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2225 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2226 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2227 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002228 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2229 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2230 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2231 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2232 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2233 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2234 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2235 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002236 "src/f32-vbinary/gen/vmax-avx-x8.c",
2237 "src/f32-vbinary/gen/vmax-avx-x16.c",
2238 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2239 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2240 "src/f32-vbinary/gen/vmin-avx-x8.c",
2241 "src/f32-vbinary/gen/vmin-avx-x16.c",
2242 "src/f32-vbinary/gen/vminc-avx-x8.c",
2243 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002244 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2245 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2246 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2247 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2248 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2249 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2250 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2251 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002252 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2253 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2254 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2255 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002256 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2257 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2258 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2259 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002260 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2261 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002262 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2263 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002264 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2265 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2266 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2267 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2268 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2269 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2270 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002271 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2272 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002273 "src/f32-vunary/gen/vabs-avx-x8.c",
2274 "src/f32-vunary/gen/vabs-avx-x16.c",
2275 "src/f32-vunary/gen/vneg-avx-x8.c",
2276 "src/f32-vunary/gen/vneg-avx-x16.c",
2277 "src/f32-vunary/gen/vsqr-avx-x8.c",
2278 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002279 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002280 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002281 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002282 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2283 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002284]
2285
Marat Dukhan1566fee2020-08-02 21:55:41 -07002286XOP_UKERNELS = [
2287 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002288 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002289 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002290 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002291 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002292 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002293 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2294 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002295 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002296 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2297 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002298 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002299 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2300 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2301 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002302 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002303 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002304 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002305 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002307 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002308 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002309 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002310 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2311 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002312 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2313 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2314 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2315 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2316 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2317 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2318 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2319 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002320]
2321
Marat Dukhanfda12b82019-11-21 12:27:59 -08002322FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002323 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2324 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002325 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2326 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002327 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2328 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002329 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2330 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2331 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2332 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2333 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2334 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002335 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002336 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2337 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2338 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2339 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002340 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002341 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2342 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002343 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002344 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2345 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002346 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2347 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2348 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002349 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2350 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2351 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2352 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2353 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2354 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2355 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2356 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2357 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2358 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2359 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2360 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2361 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2362 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002363 "src/f32-hswish/gen/hswish-fma3-x8.c",
2364 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002366 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2367 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2368 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2369 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002371 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2372 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002373 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002374 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2375 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002376 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2377 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2378 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002379 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2380 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2381 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2382 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2383 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2384 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2385 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2386 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002387 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002388 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002389 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002390]
2391
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002392AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002393 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2394 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002395 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002396 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002397 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002398 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2399 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002400 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002401 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2402 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2403 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002404 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002405 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2406 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002407 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002408 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002409 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002410 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2411 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002412 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002413 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2414 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2415 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002416 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002417 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2418 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002419 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002420 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002422 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2423 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002424 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002425 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2426 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2427 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002429 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2430 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2431 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2432 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2433 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2434 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2435 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2436 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2437 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2438 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2439 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2440 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2441 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2442 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2443 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2444 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2445 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2446 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2447 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2448 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2449 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2450 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2451 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2452 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2453 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2454 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2455 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2456 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2457 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2458 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002459 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2460 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2461 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2462 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2463 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2464 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2465 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2466 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2467 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2468 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2469 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2470 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2471 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2472 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2473 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2474 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2475 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2476 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2477 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2478 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2479 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2480 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2481 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2482 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002483 "src/math/exp-avx2-lut8-p3-perm.c",
2484 "src/math/exp-avx2-lut8-p4-perm.c",
2485 "src/math/exp-avx2-p5.c",
2486 "src/math/expminus-avx2-p5.c",
2487 "src/math/extexp-avx2-p5.c",
2488 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2489 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2490 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2491 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2492 "src/math/sigmoid-avx2-rr1-p5-div.c",
2493 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2494 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2495 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2496 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2497 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2498 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2499 "src/math/sigmoid-avx2-rr2-p5-div.c",
2500 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2501 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002502 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002504 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2505 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002506 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002507 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002508 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002509 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002510 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002511 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002512 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002513 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002514 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2515 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2516 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002517 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2518 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2519 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2520 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2521 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2522 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2523 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2524 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002525]
2526
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002528 "src/f32-clamp/gen/avx512f-x16.c",
2529 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002530 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2531 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002532 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2533 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002534 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2535 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002536 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2537 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2538 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2539 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2540 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2541 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002542 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2543 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2544 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2545 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2546 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2547 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2549 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2550 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2551 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2552 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2553 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002554 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2555 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002556 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2557 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2558 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2559 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2560 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2561 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002562 "src/f32-prelu/gen/avx512f-2x16.c",
2563 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002564 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2565 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002566 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002567 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002568 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002569 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2570 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002571 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002572 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2573 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2574 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002575 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002576 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2577 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002578 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002579 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002580 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002581 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2582 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002583 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002584 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2585 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2586 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002588 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2589 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002590 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002592 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002593 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2594 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002596 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2597 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2598 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002599 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002600 "src/f32-relu/gen/avx512f-x16.c",
2601 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002602 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002603 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2604 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2605 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2606 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2607 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2608 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2609 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2610 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2611 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2612 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2613 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2614 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2615 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2616 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2617 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2618 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002619 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2620 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2621 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2622 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2623 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2624 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2625 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2626 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2627 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2628 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2629 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2630 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2631 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2632 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2633 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2634 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002635 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2636 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2637 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2638 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2639 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2640 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2641 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2642 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2643 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2644 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2645 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2646 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2647 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2648 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2649 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2650 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002651 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2652 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2653 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2654 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2655 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2656 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2657 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2658 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002659 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2660 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2661 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2662 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2663 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2664 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2665 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2666 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002667 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2668 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2669 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2670 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2671 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2672 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2673 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2674 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002675 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2676 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2677 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2678 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002679 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2680 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2681 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2682 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002683 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2684 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002685 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2686 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2687 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2688 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2689 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2690 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2691 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2692 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002693 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002694 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2695 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2696 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2697 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2698 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2699 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2700 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2701 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2702 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2703 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2704 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2705 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2706 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2707 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2708 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2709 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2710 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2711 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2712 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2713 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2714 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2715 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2716 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2717 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002718 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2719 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2720 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2721 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2722 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2723 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2724 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2725 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002726 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2727 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2728 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2729 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2730 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2731 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002732 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2733 "src/math/exp-avx512f-lut16-p3-perm.c",
2734 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2735 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002736 "src/math/exp-avx512f-p5-scalef.c",
2737 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002738 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002739 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002740 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002741 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002742 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002743 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002744 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002745 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002746 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002747 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2748 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2749 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2750 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2751 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2752 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2753 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2754 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2755 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2756 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002757 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002758 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002759 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2760 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2761 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2762 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002763 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002764 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002765 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002766]
2767
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002768AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002769 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2770 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002771 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2772 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2773 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2774 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2775 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2776 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2777 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2778 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2779]
2780
Frank Barchardbcedc082020-08-17 18:00:51 -07002781WASM32_ASM_UKERNELS = [
2782 "src/f32-relu/wasm_shr_x1.S",
2783 "src/f32-relu/wasm_shr_x2.S",
2784 "src/f32-relu/wasm_shr_x4.S",
2785]
2786
Marat Dukhan08c4a432019-10-03 09:29:21 -07002787AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002788 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002789 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002790 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2791 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002792 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002793 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002794 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002795 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002796 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2797 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002798 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2799 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2800 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2801 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802]
2803
2804AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002805 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002807 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002808 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002809 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002810 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002811 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2813 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2814 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2815 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2816 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2817 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2818 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002819 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2820 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002821 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2822 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2823 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002824 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2825 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002826 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2827 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2828 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2829 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002830 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002831 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2832 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002833 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2834 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2835 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2836 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002837 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002838 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002839 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002840 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002841 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2842 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2843 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2844 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2845 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2846 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2847 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2848 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2849 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2850 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2851 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2852 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2853 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2854 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2855 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2856 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2857 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2858 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2859 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2860 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2861 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2862 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002863 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002864 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002865 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2866 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002867 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2868 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2869 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2870 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2871 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2872 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002873 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2874 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2875 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2876 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002877 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2878 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002879 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002880 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002881 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882]
2883
Marat Dukhan1b354632020-03-23 12:50:22 -07002884INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 "src/xnnpack/argmaxpool.h",
2886 "src/xnnpack/avgpool.h",
2887 "src/xnnpack/clamp.h",
2888 "src/xnnpack/common.h",
2889 "src/xnnpack/conv.h",
2890 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002891 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 "src/xnnpack/gavgpool.h",
2893 "src/xnnpack/gemm.h",
2894 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002895 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002896 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002897 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 "src/xnnpack/lut.h",
2899 "src/xnnpack/math.h",
2900 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002901 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002902 "src/xnnpack/packx.h",
2903 "src/xnnpack/pad.h",
2904 "src/xnnpack/params.h",
2905 "src/xnnpack/pavgpool.h",
2906 "src/xnnpack/ppmm.h",
2907 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002908 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002909 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002910 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 "src/xnnpack/rmax.h",
2912 "src/xnnpack/scalar-utils.h",
2913 "src/xnnpack/spmm.h",
2914 "src/xnnpack/unpool.h",
2915 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002916 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002918 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002919 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002920 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002921 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002923]
2924
2925INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 "include/xnnpack.h",
2927 "src/xnnpack/allocator.h",
2928 "src/xnnpack/compute.h",
2929 "src/xnnpack/im2col.h",
2930 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002931 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002932 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933 "src/xnnpack/operator.h",
2934 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002935 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002937 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002938 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002939]
2940
Marat Dukhan1b354632020-03-23 12:50:22 -07002941ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002942 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943]
2944
Marat Dukhan1b354632020-03-23 12:50:22 -07002945MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002946 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002947 "include/xnnpack.h",
2948]
2949
Marat Dukhan1b354632020-03-23 12:50:22 -07002950MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002952 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002953 "src/xnnpack/requantization.h",
2954 "include/xnnpack.h",
2955]
2956
2957OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002959 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960]
2961
2962WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002963 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002964 "src/xnnpack/operator.h",
2965 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966]
2967
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002968LOGGING_COPTS = select({
2969 # No logging in optimized mode
2970 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2971 # Full logging in debug mode
2972 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2973 # Error-only logging in default (fastbuild) mode
2974 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2975})
2976
Marat Dukhan3b59de22020-06-03 20:15:19 -07002977LOGGING_SRCS = select({
2978 # No logging in optimized mode
2979 ":optimized_build": [],
2980 "//conditions:default": [
2981 "src/operator-strings.c",
2982 "src/subgraph-strings.c",
2983 ],
2984})
2985
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002986LOGGING_HDRS = [
2987 "src/xnnpack/log.h",
2988]
2989
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002991 name = "tables",
2992 srcs = TABLE_SRCS,
2993 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002994 gcc_copts = xnnpack_gcc_std_copts(),
2995 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002996)
2997
2998xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 name = "scalar_ukernels",
3000 srcs = SCALAR_UKERNELS,
3001 hdrs = INTERNAL_HDRS,
3002 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003003 gcc_copts = xnnpack_gcc_std_copts(),
3004 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003006 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 "@FP16",
3008 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003009 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 ],
3011)
3012
3013xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003014 name = "scalar_ukernels_test_mode",
3015 srcs = SCALAR_UKERNELS,
3016 hdrs = INTERNAL_HDRS,
3017 aarch32_copts = ["-marm"],
3018 copts = [
3019 "-UNDEBUG",
3020 "-DXNN_TEST_MODE=1",
3021 ],
3022 gcc_copts = xnnpack_gcc_std_copts(),
3023 msvc_copts = xnnpack_msvc_std_copts(),
3024 deps = [
3025 ":tables",
3026 "@FP16",
3027 "@FXdiv",
3028 "@pthreadpool",
3029 ],
3030)
3031
3032xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003033 name = "wasm_ukernels",
3034 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003035 gcc_copts = xnnpack_gcc_std_copts(),
3036 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003037 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003038 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003039 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003040 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003041 "@FP16",
3042 "@FXdiv",
3043 "@pthreadpool",
3044 ],
3045)
3046
3047xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003048 name = "wasm_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 msvc_copts = xnnpack_msvc_std_copts(),
3056 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003057 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003058 deps = [
3059 ":tables",
3060 "@FP16",
3061 "@FXdiv",
3062 "@pthreadpool",
3063 ],
3064)
3065
3066xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003067 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 hdrs = INTERNAL_HDRS,
3069 aarch32_copts = [
3070 "-marm",
3071 "-mfpu=neon",
3072 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003073 gcc_copts = xnnpack_gcc_std_copts(),
3074 gcc_x86_copts = ["-msse2"],
3075 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 optimized_copts = [
3077 "-O3",
3078 "-ffast-math",
3079 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003080 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003082 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 "@FP16",
3084 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003085 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 ],
3087)
3088
3089xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003090 name = "psimd_fastmath_ukernels_test_mode",
3091 hdrs = INTERNAL_HDRS,
3092 aarch32_copts = [
3093 "-marm",
3094 "-mfpu=neon",
3095 ],
3096 copts = [
3097 "-UNDEBUG",
3098 "-DXNN_TEST_MODE=1",
3099 ],
3100 gcc_copts = xnnpack_gcc_std_copts(),
3101 gcc_x86_copts = ["-msse2"],
3102 msvc_copts = xnnpack_msvc_std_copts(),
3103 optimized_copts = [
3104 "-O3",
3105 "-ffast-math",
3106 ],
3107 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3108 deps = [
3109 ":tables",
3110 "@FP16",
3111 "@psimd",
3112 "@pthreadpool",
3113 ],
3114)
3115
3116xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003117 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003118 hdrs = INTERNAL_HDRS,
3119 aarch32_copts = [
3120 "-marm",
3121 "-mfpu=neon",
3122 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003123 gcc_copts = xnnpack_gcc_std_copts(),
3124 gcc_x86_copts = ["-msse2"],
3125 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003126 optimized_copts = [
3127 "-O3",
3128 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003129 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003130 deps = [
3131 ":tables",
3132 "@FP16",
3133 "@psimd",
3134 "@pthreadpool",
3135 ],
3136)
3137
3138xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003139 name = "psimd_accmath_ukernels_test_mode",
3140 hdrs = INTERNAL_HDRS,
3141 aarch32_copts = [
3142 "-marm",
3143 "-mfpu=neon",
3144 ],
3145 copts = [
3146 "-UNDEBUG",
3147 "-DXNN_TEST_MODE=1",
3148 ],
3149 gcc_copts = xnnpack_gcc_std_copts(),
3150 gcc_x86_copts = ["-msse2"],
3151 msvc_copts = xnnpack_msvc_std_copts(),
3152 optimized_copts = [
3153 "-O3",
3154 ],
3155 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3156 deps = [
3157 ":tables",
3158 "@FP16",
3159 "@psimd",
3160 "@pthreadpool",
3161 ],
3162)
3163
3164xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165 name = "neon_ukernels",
3166 hdrs = INTERNAL_HDRS,
3167 aarch32_copts = [
3168 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003169 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 "-mfpu=neon",
3171 ],
3172 aarch32_srcs = NEON_UKERNELS,
3173 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003174 gcc_copts = xnnpack_gcc_std_copts(),
3175 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003176 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003177 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003178 "@FP16",
3179 "@pthreadpool",
3180 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181)
3182
3183xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003184 name = "neon_ukernels_test_mode",
3185 hdrs = INTERNAL_HDRS,
3186 aarch32_copts = [
3187 "-marm",
3188 "-march=armv7-a",
3189 "-mfpu=neon",
3190 ],
3191 aarch32_srcs = NEON_UKERNELS,
3192 aarch64_srcs = NEON_UKERNELS,
3193 copts = [
3194 "-UNDEBUG",
3195 "-DXNN_TEST_MODE=1",
3196 ],
3197 gcc_copts = xnnpack_gcc_std_copts(),
3198 msvc_copts = xnnpack_msvc_std_copts(),
3199 deps = [
3200 ":tables",
3201 "@FP16",
3202 "@pthreadpool",
3203 ],
3204)
3205
3206xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 name = "neonfma_ukernels",
3208 hdrs = INTERNAL_HDRS,
3209 aarch32_copts = [
3210 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003211 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 "-mfpu=neon-vfpv4",
3213 ],
3214 aarch32_srcs = NEONFMA_UKERNELS,
3215 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003216 apple_aarch32_copts = [
3217 "-mcpu=swift",
3218 "-mtune=generic",
3219 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003220 gcc_copts = xnnpack_gcc_std_copts(),
3221 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003222 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003223 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003224 "@FP16",
3225 "@pthreadpool",
3226 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227)
3228
3229xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003230 name = "neonfma_ukernels_test_mode",
3231 hdrs = INTERNAL_HDRS,
3232 aarch32_copts = [
3233 "-marm",
3234 "-march=armv7-a",
3235 "-mfpu=neon-vfpv4",
3236 ],
3237 aarch32_srcs = NEONFMA_UKERNELS,
3238 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003239 apple_aarch32_copts = [
3240 "-mcpu=swift",
3241 "-mtune=generic",
3242 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003243 copts = [
3244 "-UNDEBUG",
3245 "-DXNN_TEST_MODE=1",
3246 ],
3247 gcc_copts = xnnpack_gcc_std_copts(),
3248 msvc_copts = xnnpack_msvc_std_copts(),
3249 deps = [
3250 ":tables",
3251 "@FP16",
3252 "@pthreadpool",
3253 ],
3254)
3255
3256xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003257 name = "neonv8_ukernels",
3258 hdrs = INTERNAL_HDRS,
3259 aarch32_copts = [
3260 "-marm",
3261 "-march=armv8-a",
3262 "-mfpu=neon-fp-armv8",
3263 ],
3264 aarch32_srcs = NEONV8_UKERNELS,
3265 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003266 apple_aarch32_copts = [
3267 "-mcpu=cyclone",
3268 "-mtune=generic",
3269 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003270 gcc_copts = xnnpack_gcc_std_copts(),
3271 msvc_copts = xnnpack_msvc_std_copts(),
3272 deps = [
3273 ":tables",
3274 "@FP16",
3275 "@pthreadpool",
3276 ],
3277)
3278
3279xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003280 name = "neonv8_ukernels_test_mode",
3281 hdrs = INTERNAL_HDRS,
3282 aarch32_copts = [
3283 "-marm",
3284 "-march=armv8-a",
3285 "-mfpu=neon-fp-armv8",
3286 ],
3287 aarch32_srcs = NEONV8_UKERNELS,
3288 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003289 apple_aarch32_copts = [
3290 "-mcpu=cyclone",
3291 "-mtune=generic",
3292 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003293 copts = [
3294 "-UNDEBUG",
3295 "-DXNN_TEST_MODE=1",
3296 ],
3297 gcc_copts = xnnpack_gcc_std_copts(),
3298 msvc_copts = xnnpack_msvc_std_copts(),
3299 deps = [
3300 ":tables",
3301 "@FP16",
3302 "@pthreadpool",
3303 ],
3304)
3305
3306xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003307 name = "neonfp16arith_ukernels",
3308 hdrs = INTERNAL_HDRS,
3309 aarch64_copts = ["-march=armv8.2-a+fp16"],
3310 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003311 gcc_copts = xnnpack_gcc_std_copts(),
3312 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003313 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003314 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003315 "@FP16",
3316 "@pthreadpool",
3317 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003318)
3319
3320xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003321 name = "neonfp16arith_ukernels_test_mode",
3322 hdrs = INTERNAL_HDRS,
3323 aarch64_copts = ["-march=armv8.2-a+fp16"],
3324 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3325 copts = [
3326 "-UNDEBUG",
3327 "-DXNN_TEST_MODE=1",
3328 ],
3329 gcc_copts = xnnpack_gcc_std_copts(),
3330 msvc_copts = xnnpack_msvc_std_copts(),
3331 deps = [
3332 ":tables",
3333 "@FP16",
3334 "@pthreadpool",
3335 ],
3336)
3337
3338xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003339 name = "neondot_ukernels",
3340 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003341 aarch32_copts = [
3342 "-marm",
3343 "-march=armv8.2-a+dotprod",
3344 "-mfpu=neon-fp-armv8",
3345 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003346 aarch32_srcs = NEONDOT_UKERNELS,
3347 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3348 aarch64_srcs = NEONDOT_UKERNELS,
3349 gcc_copts = xnnpack_gcc_std_copts(),
3350 msvc_copts = xnnpack_msvc_std_copts(),
3351 deps = [
3352 ":tables",
3353 "@FP16",
3354 "@pthreadpool",
3355 ],
3356)
3357
3358xnnpack_cc_library(
3359 name = "neondot_ukernels_test_mode",
3360 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003361 aarch32_copts = [
3362 "-marm",
3363 "-march=armv8.2-a+dotprod",
3364 "-mfpu=neon-fp-armv8",
3365 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003366 aarch32_srcs = NEONDOT_UKERNELS,
3367 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3368 aarch64_srcs = NEONDOT_UKERNELS,
3369 copts = [
3370 "-UNDEBUG",
3371 "-DXNN_TEST_MODE=1",
3372 ],
3373 gcc_copts = xnnpack_gcc_std_copts(),
3374 msvc_copts = xnnpack_msvc_std_copts(),
3375 deps = [
3376 ":tables",
3377 "@FP16",
3378 "@pthreadpool",
3379 ],
3380)
3381
3382xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 name = "sse2_ukernels",
3384 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003385 gcc_copts = xnnpack_gcc_std_copts(),
3386 gcc_x86_copts = ["-msse2"],
3387 msvc_copts = xnnpack_msvc_std_copts(),
3388 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003389 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003390 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003391 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003392 "@FP16",
3393 "@pthreadpool",
3394 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003395)
3396
3397xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003398 name = "sse2_ukernels_test_mode",
3399 hdrs = INTERNAL_HDRS,
3400 copts = [
3401 "-UNDEBUG",
3402 "-DXNN_TEST_MODE=1",
3403 ],
3404 gcc_copts = xnnpack_gcc_std_copts(),
3405 gcc_x86_copts = ["-msse2"],
3406 msvc_copts = xnnpack_msvc_std_copts(),
3407 msvc_x86_32_copts = ["/arch:SSE2"],
3408 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3409 deps = [
3410 ":tables",
3411 "@FP16",
3412 "@pthreadpool",
3413 ],
3414)
3415
3416xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003417 name = "ssse3_ukernels",
3418 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003419 gcc_copts = xnnpack_gcc_std_copts(),
3420 gcc_x86_copts = ["-mssse3"],
3421 msvc_copts = xnnpack_msvc_std_copts(),
3422 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003423 x86_srcs = SSSE3_UKERNELS,
3424 deps = [
3425 ":tables",
3426 "@FP16",
3427 "@pthreadpool",
3428 ],
3429)
3430
3431xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003432 name = "ssse3_ukernels_test_mode",
3433 hdrs = INTERNAL_HDRS,
3434 copts = [
3435 "-UNDEBUG",
3436 "-DXNN_TEST_MODE=1",
3437 ],
3438 gcc_copts = xnnpack_gcc_std_copts(),
3439 gcc_x86_copts = ["-mssse3"],
3440 msvc_copts = xnnpack_msvc_std_copts(),
3441 msvc_x86_32_copts = ["/arch:SSE2"],
3442 x86_srcs = SSSE3_UKERNELS,
3443 deps = [
3444 ":tables",
3445 "@FP16",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003451 name = "sse41_ukernels",
3452 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003453 gcc_copts = xnnpack_gcc_std_copts(),
3454 gcc_x86_copts = ["-msse4.1"],
3455 msvc_copts = xnnpack_msvc_std_copts(),
3456 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003457 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003458 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003459 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003460 "@FP16",
3461 "@pthreadpool",
3462 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003463)
3464
3465xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003466 name = "sse41_ukernels_test_mode",
3467 hdrs = INTERNAL_HDRS,
3468 copts = [
3469 "-UNDEBUG",
3470 "-DXNN_TEST_MODE=1",
3471 ],
3472 gcc_copts = xnnpack_gcc_std_copts(),
3473 gcc_x86_copts = ["-msse4.1"],
3474 msvc_copts = xnnpack_msvc_std_copts(),
3475 msvc_x86_32_copts = ["/arch:SSE2"],
3476 x86_srcs = SSE41_UKERNELS,
3477 deps = [
3478 ":tables",
3479 "@FP16",
3480 "@pthreadpool",
3481 ],
3482)
3483
3484xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485 name = "avx_ukernels",
3486 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003487 gcc_copts = xnnpack_gcc_std_copts(),
3488 gcc_x86_copts = ["-mavx"],
3489 msvc_copts = xnnpack_msvc_std_copts(),
3490 msvc_x86_32_copts = ["/arch:AVX"],
3491 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003492 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003493 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003494 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003495 "@FP16",
3496 "@pthreadpool",
3497 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003498)
3499
3500xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003501 name = "avx_ukernels_test_mode",
3502 hdrs = INTERNAL_HDRS,
3503 copts = [
3504 "-UNDEBUG",
3505 "-DXNN_TEST_MODE=1",
3506 ],
3507 gcc_copts = xnnpack_gcc_std_copts(),
3508 gcc_x86_copts = ["-mavx"],
3509 msvc_copts = xnnpack_msvc_std_copts(),
3510 msvc_x86_32_copts = ["/arch:AVX"],
3511 msvc_x86_64_copts = ["/arch:AVX"],
3512 x86_srcs = AVX_UKERNELS,
3513 deps = [
3514 ":tables",
3515 "@FP16",
3516 "@pthreadpool",
3517 ],
3518)
3519
3520xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003521 name = "xop_ukernels",
3522 hdrs = INTERNAL_HDRS,
3523 gcc_copts = xnnpack_gcc_std_copts(),
3524 gcc_x86_copts = ["-mxop"],
3525 msvc_copts = xnnpack_msvc_std_copts(),
3526 msvc_x86_32_copts = ["/arch:AVX"],
3527 msvc_x86_64_copts = ["/arch:AVX"],
3528 x86_srcs = XOP_UKERNELS,
3529 deps = [
3530 ":tables",
3531 "@FP16",
3532 "@pthreadpool",
3533 ],
3534)
3535
3536xnnpack_cc_library(
3537 name = "xop_ukernels_test_mode",
3538 hdrs = INTERNAL_HDRS,
3539 copts = [
3540 "-UNDEBUG",
3541 "-DXNN_TEST_MODE=1",
3542 ],
3543 gcc_copts = xnnpack_gcc_std_copts(),
3544 gcc_x86_copts = ["-mxop"],
3545 msvc_copts = xnnpack_msvc_std_copts(),
3546 msvc_x86_32_copts = ["/arch:AVX"],
3547 msvc_x86_64_copts = ["/arch:AVX"],
3548 x86_srcs = XOP_UKERNELS,
3549 deps = [
3550 ":tables",
3551 "@FP16",
3552 "@pthreadpool",
3553 ],
3554)
3555
3556xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003557 name = "fma3_ukernels",
3558 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003559 gcc_copts = xnnpack_gcc_std_copts(),
3560 gcc_x86_copts = ["-mfma"],
3561 msvc_copts = xnnpack_msvc_std_copts(),
3562 msvc_x86_32_copts = ["/arch:AVX"],
3563 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003564 x86_srcs = FMA3_UKERNELS,
3565 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003566 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003567 "@FP16",
3568 "@pthreadpool",
3569 ],
3570)
3571
3572xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003573 name = "fma3_ukernels_test_mode",
3574 hdrs = INTERNAL_HDRS,
3575 copts = [
3576 "-UNDEBUG",
3577 "-DXNN_TEST_MODE=1",
3578 ],
3579 gcc_copts = xnnpack_gcc_std_copts(),
3580 gcc_x86_copts = ["-mfma"],
3581 msvc_copts = xnnpack_msvc_std_copts(),
3582 msvc_x86_32_copts = ["/arch:AVX"],
3583 msvc_x86_64_copts = ["/arch:AVX"],
3584 x86_srcs = FMA3_UKERNELS,
3585 deps = [
3586 ":tables",
3587 "@FP16",
3588 "@pthreadpool",
3589 ],
3590)
3591
3592xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003593 name = "avx2_ukernels",
3594 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003595 gcc_copts = xnnpack_gcc_std_copts(),
3596 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003597 "-mfma",
3598 "-mavx2",
3599 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003600 msvc_copts = xnnpack_msvc_std_copts(),
3601 msvc_x86_32_copts = ["/arch:AVX2"],
3602 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003603 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003604 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003605 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003606 "@FP16",
3607 "@pthreadpool",
3608 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003609)
3610
3611xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003612 name = "avx2_ukernels_test_mode",
3613 hdrs = INTERNAL_HDRS,
3614 copts = [
3615 "-UNDEBUG",
3616 "-DXNN_TEST_MODE=1",
3617 ],
3618 gcc_copts = xnnpack_gcc_std_copts(),
3619 gcc_x86_copts = [
3620 "-mfma",
3621 "-mavx2",
3622 ],
3623 msvc_copts = xnnpack_msvc_std_copts(),
3624 msvc_x86_32_copts = ["/arch:AVX2"],
3625 msvc_x86_64_copts = ["/arch:AVX2"],
3626 x86_srcs = AVX2_UKERNELS,
3627 deps = [
3628 ":tables",
3629 "@FP16",
3630 "@pthreadpool",
3631 ],
3632)
3633
3634xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003635 name = "avx512f_ukernels",
3636 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003637 gcc_copts = xnnpack_gcc_std_copts(),
3638 gcc_x86_copts = ["-mavx512f"],
3639 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3640 msvc_copts = xnnpack_msvc_std_copts(),
3641 msvc_x86_32_copts = ["/arch:AVX512"],
3642 msvc_x86_64_copts = ["/arch:AVX512"],
3643 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003645 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003646 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003647 "@FP16",
3648 "@pthreadpool",
3649 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650)
3651
3652xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003653 name = "avx512f_ukernels_test_mode",
3654 hdrs = INTERNAL_HDRS,
3655 copts = [
3656 "-UNDEBUG",
3657 "-DXNN_TEST_MODE=1",
3658 ],
3659 gcc_copts = xnnpack_gcc_std_copts(),
3660 gcc_x86_copts = ["-mavx512f"],
3661 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3662 msvc_copts = xnnpack_msvc_std_copts(),
3663 msvc_x86_32_copts = ["/arch:AVX512"],
3664 msvc_x86_64_copts = ["/arch:AVX512"],
3665 msys_copts = ["-fno-asynchronous-unwind-tables"],
3666 x86_srcs = AVX512F_UKERNELS,
3667 deps = [
3668 ":tables",
3669 "@FP16",
3670 "@pthreadpool",
3671 ],
3672)
3673
3674xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003675 name = "avx512skx_ukernels",
3676 hdrs = INTERNAL_HDRS,
3677 gcc_copts = xnnpack_gcc_std_copts(),
3678 gcc_x86_copts = [
3679 "-mavx512f",
3680 "-mavx512cd",
3681 "-mavx512bw",
3682 "-mavx512dq",
3683 "-mavx512vl",
3684 ],
3685 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3686 msvc_copts = xnnpack_msvc_std_copts(),
3687 msvc_x86_32_copts = ["/arch:AVX512"],
3688 msvc_x86_64_copts = ["/arch:AVX512"],
3689 msys_copts = ["-fno-asynchronous-unwind-tables"],
3690 x86_srcs = AVX512SKX_UKERNELS,
3691 deps = [
3692 ":tables",
3693 "@FP16",
3694 "@pthreadpool",
3695 ],
3696)
3697
3698xnnpack_cc_library(
3699 name = "avx512skx_ukernels_test_mode",
3700 hdrs = INTERNAL_HDRS,
3701 copts = [
3702 "-UNDEBUG",
3703 "-DXNN_TEST_MODE=1",
3704 ],
3705 gcc_copts = xnnpack_gcc_std_copts(),
3706 gcc_x86_copts = [
3707 "-mavx512f",
3708 "-mavx512cd",
3709 "-mavx512bw",
3710 "-mavx512dq",
3711 "-mavx512vl",
3712 ],
3713 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3714 msvc_copts = xnnpack_msvc_std_copts(),
3715 msvc_x86_32_copts = ["/arch:AVX512"],
3716 msvc_x86_64_copts = ["/arch:AVX512"],
3717 msys_copts = ["-fno-asynchronous-unwind-tables"],
3718 x86_srcs = AVX512SKX_UKERNELS,
3719 deps = [
3720 ":tables",
3721 "@FP16",
3722 "@pthreadpool",
3723 ],
3724)
3725
3726xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 name = "asm_ukernels",
3728 hdrs = ["src/xnnpack/assembly.h"],
3729 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003730 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003731 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003732 wasm_srcs = WASM32_ASM_UKERNELS,
3733 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003734)
3735
Marat Dukhan3b59de22020-06-03 20:15:19 -07003736xnnpack_cc_library(
3737 name = "logging_utils",
3738 srcs = LOGGING_SRCS,
3739 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3740 copts = LOGGING_COPTS + [
3741 "-Isrc",
3742 "-Iinclude",
3743 ] + select({
3744 ":debug_build": [],
3745 "//conditions:default": xnnpack_min_size_copts(),
3746 }),
3747 gcc_copts = xnnpack_gcc_std_copts(),
3748 msvc_copts = xnnpack_msvc_std_copts(),
3749 visibility = xnnpack_visibility(),
3750 deps = [
3751 "@FP16",
3752 "@clog",
3753 "@pthreadpool",
3754 ],
3755)
3756
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757xnnpack_aggregate_library(
3758 name = "ukernels",
3759 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 ":neon_ukernels",
3761 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003762 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003763 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764 ":asm_ukernels",
3765 ],
3766 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 ":neon_ukernels",
3768 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003769 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003771 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 ":asm_ukernels",
3773 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003774 generic_deps = [
3775 ":scalar_ukernels",
3776 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003777 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003778 ":psimd_fastmath_ukernels",
3779 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003780 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003781 wasm_deps = [
3782 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003783 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003784 ],
3785 wasmsimd_deps = [
3786 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003787 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003791 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003792 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003794 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003795 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003796 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003797 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003798 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 ],
3800)
3801
Marat Dukhan33fcf782020-05-24 14:27:15 -07003802xnnpack_aggregate_library(
3803 name = "ukernels_test_mode",
3804 aarch32_deps = [
3805 ":neon_ukernels_test_mode",
3806 ":neonfma_ukernels_test_mode",
3807 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003808 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003809 ":asm_ukernels",
3810 ],
3811 aarch64_deps = [
3812 ":neon_ukernels_test_mode",
3813 ":neonfma_ukernels_test_mode",
3814 ":neonv8_ukernels_test_mode",
3815 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003816 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003817 ":asm_ukernels",
3818 ],
3819 generic_deps = [
3820 ":scalar_ukernels_test_mode",
3821 ],
3822 psimd_deps = [
3823 ":psimd_fastmath_ukernels_test_mode",
3824 ":psimd_accmath_ukernels_test_mode",
3825 ],
3826 wasm_deps = [
3827 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003828 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003829 ],
3830 wasmsimd_deps = [
3831 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003832 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003833 ],
3834 x86_deps = [
3835 ":sse2_ukernels_test_mode",
3836 ":ssse3_ukernels_test_mode",
3837 ":sse41_ukernels_test_mode",
3838 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003839 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003840 ":fma3_ukernels_test_mode",
3841 ":avx2_ukernels_test_mode",
3842 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003843 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003844 ],
3845)
3846
Marat Dukhan86df3842020-06-17 14:03:31 -07003847# This target must be compatible with Android NDK r17 and compile with Clang 6.
3848# It is similar to :ukernels target, but excludes microkernels which require
3849# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003850xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003851 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003852 aarch32_deps = [
3853 ":neon_ukernels",
3854 ":neonfma_ukernels",
3855 ":neonv8_ukernels",
3856 ":asm_ukernels",
3857 ],
3858 aarch64_deps = [
3859 ":neon_ukernels",
3860 ":neonfma_ukernels",
3861 ":neonv8_ukernels",
3862 ":asm_ukernels",
3863 ],
3864 generic_deps = [
3865 ":scalar_ukernels",
3866 ],
3867 psimd_deps = [
3868 ":psimd_fastmath_ukernels",
3869 ":psimd_accmath_ukernels",
3870 ],
3871 wasm_deps = [
3872 ":wasm_ukernels",
3873 ],
3874 wasmsimd_deps = [
3875 ":wasm_ukernels",
3876 ],
3877 x86_deps = [
3878 ":sse2_ukernels",
3879 ":ssse3_ukernels",
3880 ":sse41_ukernels",
3881 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003882 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003883 ":fma3_ukernels",
3884 ":avx2_ukernels",
3885 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003886 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003887 ],
3888)
3889
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890xnnpack_cc_library(
3891 name = "im2col",
3892 srcs = ["src/im2col.c"],
3893 hdrs = [
3894 "src/xnnpack/common.h",
3895 "src/xnnpack/im2col.h",
3896 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003897 gcc_copts = xnnpack_gcc_std_copts(),
3898 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899)
3900
3901xnnpack_cc_library(
3902 name = "indirection",
3903 srcs = ["src/indirection.c"],
3904 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003905 gcc_copts = xnnpack_gcc_std_copts(),
3906 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003907 deps = [
3908 "@FP16",
3909 "@FXdiv",
3910 "@pthreadpool",
3911 ],
3912)
3913
3914xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003915 name = "indirection_test_mode",
3916 srcs = ["src/indirection.c"],
3917 hdrs = INTERNAL_HDRS,
3918 copts = [
3919 "-UNDEBUG",
3920 "-DXNN_TEST_MODE=1",
3921 ],
3922 gcc_copts = xnnpack_gcc_std_copts(),
3923 msvc_copts = xnnpack_msvc_std_copts(),
3924 deps = [
3925 "@FP16",
3926 "@FXdiv",
3927 "@pthreadpool",
3928 ],
3929)
3930
3931xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003932 name = "packing",
3933 srcs = ["src/packing.c"],
3934 hdrs = INTERNAL_HDRS,
3935 gcc_copts = xnnpack_gcc_std_copts(),
3936 msvc_copts = xnnpack_msvc_std_copts(),
3937 deps = [
3938 "@FP16",
3939 "@FXdiv",
3940 "@pthreadpool",
3941 ],
3942)
3943
3944xnnpack_cc_library(
3945 name = "packing_test_mode",
3946 srcs = ["src/packing.c"],
3947 hdrs = INTERNAL_HDRS,
3948 copts = [
3949 "-UNDEBUG",
3950 "-DXNN_TEST_MODE=1",
3951 ],
3952 gcc_copts = xnnpack_gcc_std_copts(),
3953 msvc_copts = xnnpack_msvc_std_copts(),
3954 deps = [
3955 "@FP16",
3956 "@FXdiv",
3957 "@pthreadpool",
3958 ],
3959)
3960
3961xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003962 name = "operator_run",
3963 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003964 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003965 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003966 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3967 "//conditions:default": [],
3968 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003969 gcc_copts = xnnpack_gcc_std_copts(),
3970 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003972 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003973 "@FP16",
3974 "@FXdiv",
3975 "@clog",
3976 "@pthreadpool",
3977 ],
3978)
3979
Chao Mei6ddfc602020-05-13 22:29:36 -07003980xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003981 name = "operator_run_test_mode",
3982 srcs = ["src/operator-run.c"],
3983 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3984 copts = LOGGING_COPTS + [
3985 "-UNDEBUG",
3986 "-DXNN_TEST_MODE=1",
3987 ] + select({
3988 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3989 "//conditions:default": [],
3990 }),
3991 gcc_copts = xnnpack_gcc_std_copts(),
3992 msvc_copts = xnnpack_msvc_std_copts(),
3993 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003994 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003995 "@FP16",
3996 "@FXdiv",
3997 "@clog",
3998 "@pthreadpool",
3999 ],
4000)
4001
4002xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004003 name = "memory_planner",
4004 srcs = ["src/memory-planner.c"],
4005 hdrs = INTERNAL_HDRS,
4006 defines = select({
4007 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4008 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4009 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4010 }),
4011 gcc_copts = xnnpack_gcc_std_copts(),
4012 msvc_copts = xnnpack_msvc_std_copts(),
4013 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004014 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004015 "@pthreadpool",
4016 ],
4017)
4018
Marat Dukhan33fcf782020-05-24 14:27:15 -07004019xnnpack_cc_library(
4020 name = "memory_planner_test_mode",
4021 srcs = ["src/memory-planner.c"],
4022 hdrs = INTERNAL_HDRS,
4023 copts = [
4024 "-UNDEBUG",
4025 "-DXNN_TEST_MODE=1",
4026 ],
4027 defines = select({
4028 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4029 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4030 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4031 }),
4032 gcc_copts = xnnpack_gcc_std_copts(),
4033 msvc_copts = xnnpack_msvc_std_copts(),
4034 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004035 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004036 "@pthreadpool",
4037 ],
4038)
4039
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040cc_library(
4041 name = "enable_assembly",
4042 defines = select({
4043 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4044 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004045 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004046 }),
4047)
4048
Marat Dukhan9de90e02020-06-18 16:04:12 -07004049cc_library(
4050 name = "enable_sparse",
4051 defines = select({
4052 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4053 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004054 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004055 }),
4056)
4057
Marat Dukhancf056b22019-10-07 10:26:29 -07004058xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004059 name = "operators",
4060 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004061 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004062 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004063 ],
4064 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004065 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004066 "-Isrc",
4067 "-Iinclude",
4068 ] + select({
4069 ":debug_build": [],
4070 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004071 }) + select({
4072 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4073 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004074 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004075 gcc_copts = xnnpack_gcc_std_copts(),
4076 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004077 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004079 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004080 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081 "@FP16",
4082 "@FXdiv",
4083 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004085 ],
4086)
4087
Marat Dukhan10a38082020-04-17 03:58:35 -07004088xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004089 name = "operators_test_mode",
4090 srcs = OPERATOR_SRCS + [
4091 "src/memory.c",
4092 "src/operator-delete.c",
4093 ],
4094 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4095 copts = LOGGING_COPTS + [
4096 "-Isrc",
4097 "-Iinclude",
4098 "-UNDEBUG",
4099 "-DXNN_TEST_MODE=1",
4100 ] + select({
4101 ":debug_build": [],
4102 "//conditions:default": xnnpack_min_size_copts(),
4103 }) + select({
4104 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4105 "//conditions:default": [],
4106 }),
4107 gcc_copts = xnnpack_gcc_std_copts(),
4108 msvc_copts = xnnpack_msvc_std_copts(),
4109 deps = [
4110 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004111 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004112 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004113 "@FP16",
4114 "@FXdiv",
4115 "@clog",
4116 "@pthreadpool",
4117 ],
4118)
4119
4120xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004121 name = "XNNPACK",
4122 srcs = [
4123 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004124 "src/runtime.c",
4125 "src/subgraph.c",
4126 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004127 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004128 hdrs = ["include/xnnpack.h"],
4129 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004130 "-Isrc",
4131 "-Iinclude",
4132 ] + select({
4133 ":debug_build": [],
4134 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004135 }) + select({
4136 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4137 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004138 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004139 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004140 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004141 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004142 visibility = xnnpack_visibility(),
4143 deps = [
4144 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004145 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004146 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004147 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004148 ":operator_run",
4149 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004150 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004151 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004152 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004153 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004154 ] + select({
4155 ":emscripten": [],
4156 "//conditions:default": ["@cpuinfo"],
4157 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004158)
4159
Marat Dukhan10a38082020-04-17 03:58:35 -07004160xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004161 name = "XNNPACK_test_mode",
4162 srcs = [
4163 "src/init.c",
4164 "src/runtime.c",
4165 "src/subgraph.c",
4166 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004167 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004168 hdrs = ["include/xnnpack.h"],
4169 copts = LOGGING_COPTS + [
4170 "-Isrc",
4171 "-Iinclude",
4172 "-UNDEBUG",
4173 "-DXNN_TEST_MODE=1",
4174 ] + select({
4175 ":debug_build": [],
4176 "//conditions:default": xnnpack_min_size_copts(),
4177 }) + select({
4178 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4179 "//conditions:default": [],
4180 }),
4181 gcc_copts = xnnpack_gcc_std_copts(),
4182 includes = ["include"],
4183 msvc_copts = xnnpack_msvc_std_copts(),
4184 visibility = xnnpack_visibility(),
4185 deps = [
4186 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004187 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004188 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004189 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004190 ":operator_run_test_mode",
4191 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004192 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004193 "@clog",
4194 "@FP16",
4195 "@pthreadpool",
4196 ] + select({
4197 ":emscripten": [],
4198 "//conditions:default": ["@cpuinfo"],
4199 }),
4200)
4201
4202xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004203 name = "xnnpack_f32",
4204 srcs = [
4205 "src/init.c",
4206 "src/runtime.c",
4207 "src/subgraph.c",
4208 "src/tensor.c",
4209 ] + SUBGRAPH_SRCS,
4210 hdrs = ["include/xnnpack.h"],
4211 copts = LOGGING_COPTS + [
4212 "-Isrc",
4213 "-Iinclude",
4214 ] + select({
4215 ":debug_build": [],
4216 "//conditions:default": xnnpack_min_size_copts(),
4217 }) + select({
4218 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4219 "//conditions:default": [],
4220 }),
4221 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004222 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004223 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004224 "XNN_NO_U8_OPERATORS",
4225 "XNN_NO_X8_OPERATORS",
4226 "XNN_NO_F16_OPERATORS",
4227 "XNN_NO_X16_OPERATORS",
4228 ],
4229 gcc_copts = xnnpack_gcc_std_copts(),
4230 includes = ["include"],
4231 msvc_copts = xnnpack_msvc_std_copts(),
4232 visibility = xnnpack_visibility(),
4233 deps = [
4234 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004235 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004236 ":logging_utils",
4237 ":memory_planner",
4238 ":operator_run",
4239 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004240 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004241 "@clog",
4242 "@FP16",
4243 "@pthreadpool",
4244 ] + select({
4245 ":emscripten": [],
4246 "//conditions:default": ["@cpuinfo"],
4247 }),
4248)
4249
4250xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004251 name = "xnnpack_operators_nhwc_f32",
4252 srcs = [
4253 "src/init.c",
4254 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004255 hdrs = ["include/xnnpack.h"],
4256 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004257 "-Isrc",
4258 "-Iinclude",
4259 ] + select({
4260 ":debug_build": [],
4261 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004262 }) + select({
4263 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4264 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004265 }),
4266 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004267 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004268 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004269 "XNN_NO_U8_OPERATORS",
4270 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004271 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004272 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004273 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004274 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004275 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 visibility = xnnpack_visibility(),
4277 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004278 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004279 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280 ":operator_run",
4281 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004282 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004283 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004284 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004285 ] + select({
4286 ":emscripten": [],
4287 "//conditions:default": ["@cpuinfo"],
4288 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
Marat Dukhancf056b22019-10-07 10:26:29 -07004291xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292 name = "bench_utils",
4293 srcs = ["bench/utils.cc"],
4294 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004295 deps = [
4296 "@com_google_benchmark//:benchmark",
4297 "@cpuinfo",
4298 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299)
4300
Frank Barchard7e955972019-10-11 10:34:25 -07004301######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302
4303xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004304 name = "qs8_gemm_bench",
4305 srcs = [
4306 "bench/gemm.h",
4307 "bench/qs8-gemm.cc",
4308 "src/xnnpack/AlignedAllocator.h",
4309 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004310 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4311 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004312)
4313
4314xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004315 name = "qs8_requantization_bench",
4316 srcs = [
4317 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004318 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004319 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004320 ] + MICROKERNEL_BENCHMARK_HDRS,
4321 deps = MICROKERNEL_BENCHMARK_DEPS,
4322)
4323
4324xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004325 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326 srcs = [
4327 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004328 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 "src/xnnpack/AlignedAllocator.h",
4330 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004331 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004332 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333)
4334
4335xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004336 name = "qu8_requantization_bench",
4337 srcs = [
4338 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004339 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004340 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004341 ] + MICROKERNEL_BENCHMARK_HDRS,
4342 deps = MICROKERNEL_BENCHMARK_DEPS,
4343)
4344
4345xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004346 name = "f16_igemm_bench",
4347 srcs = [
4348 "bench/f16-igemm.cc",
4349 "bench/conv.h",
4350 "bench/google/conv.h",
4351 "src/xnnpack/AlignedAllocator.h",
4352 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004353 deps = MICROKERNEL_BENCHMARK_DEPS + [
4354 ":indirection",
4355 ":packing",
4356 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004357)
4358
4359xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360 name = "f16_gemm_bench",
4361 srcs = [
4362 "bench/f16-gemm.cc",
4363 "bench/gemm.h",
4364 "src/xnnpack/AlignedAllocator.h",
4365 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004366 deps = MICROKERNEL_BENCHMARK_DEPS + [
4367 ":packing",
4368 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369)
4370
4371xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004372 name = "f16_spmm_bench",
4373 srcs = [
4374 "bench/f16-spmm.cc",
4375 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004376 "src/xnnpack/AlignedAllocator.h",
4377 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004378 deps = MICROKERNEL_BENCHMARK_DEPS,
4379)
4380
4381xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004382 name = "f32_igemm_bench",
4383 srcs = [
4384 "bench/f32-igemm.cc",
4385 "bench/conv.h",
4386 "src/xnnpack/AlignedAllocator.h",
4387 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004388 deps = MICROKERNEL_BENCHMARK_DEPS + [
4389 ":indirection",
4390 ":packing",
4391 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004392)
4393
4394xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004395 name = "f16_relu_bench",
4396 srcs = [
4397 "bench/f16-relu.cc",
4398 "src/xnnpack/AlignedAllocator.h",
4399 ] + MICROKERNEL_BENCHMARK_HDRS,
4400 deps = MICROKERNEL_BENCHMARK_DEPS,
4401)
4402
4403xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004404 name = "f32_conv_hwc_bench",
4405 srcs = [
4406 "bench/f32-conv-hwc.cc",
4407 "bench/dconv.h",
4408 "src/xnnpack/AlignedAllocator.h",
4409 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004410 deps = MICROKERNEL_BENCHMARK_DEPS + [
4411 ":packing",
4412 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413)
4414
4415xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004416 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004417 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004418 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004419 "bench/dconv.h",
4420 "src/xnnpack/AlignedAllocator.h",
4421 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004422 deps = MICROKERNEL_BENCHMARK_DEPS + [
4423 ":packing",
4424 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004425)
4426
4427xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004428 name = "f16_dwconv_bench",
4429 srcs = [
4430 "bench/f16-dwconv.cc",
4431 "bench/dwconv.h",
4432 "bench/google/dwconv.h",
4433 "src/xnnpack/AlignedAllocator.h",
4434 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004435 deps = MICROKERNEL_BENCHMARK_DEPS + [
4436 ":indirection",
4437 ":packing",
4438 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004439)
4440
4441xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 name = "f32_dwconv_bench",
4443 srcs = [
4444 "bench/f32-dwconv.cc",
4445 "bench/dwconv.h",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004448 deps = MICROKERNEL_BENCHMARK_DEPS + [
4449 ":indirection",
4450 ":packing",
4451 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452)
4453
4454xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004455 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004457 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 "bench/dwconv.h",
4459 "src/xnnpack/AlignedAllocator.h",
4460 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004461 deps = MICROKERNEL_BENCHMARK_DEPS + [
4462 ":indirection",
4463 ":packing",
4464 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465)
4466
4467xnnpack_benchmark(
4468 name = "f32_gemm_bench",
4469 srcs = [
4470 "bench/f32-gemm.cc",
4471 "bench/gemm.h",
4472 "src/xnnpack/AlignedAllocator.h",
4473 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004474 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004475 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476)
4477
4478xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004479 name = "f32_hswish_bench",
4480 srcs = [
4481 "bench/f32-hswish.cc",
4482 "src/xnnpack/AlignedAllocator.h",
4483 ] + MICROKERNEL_BENCHMARK_HDRS,
4484 deps = MICROKERNEL_BENCHMARK_DEPS,
4485)
4486
4487xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004488 name = "f32_raddexpminusmax_bench",
4489 srcs = [
4490 "bench/f32-raddexpminusmax.cc",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + MICROKERNEL_BENCHMARK_HDRS,
4493 deps = MICROKERNEL_BENCHMARK_DEPS,
4494)
4495
4496xnnpack_benchmark(
4497 name = "f32_raddextexp_bench",
4498 srcs = [
4499 "bench/f32-raddextexp.cc",
4500 "src/xnnpack/AlignedAllocator.h",
4501 ] + MICROKERNEL_BENCHMARK_HDRS,
4502 deps = MICROKERNEL_BENCHMARK_DEPS,
4503)
4504
4505xnnpack_benchmark(
4506 name = "f32_raddstoreexpminusmax_bench",
4507 srcs = [
4508 "bench/f32-raddstoreexpminusmax.cc",
4509 "src/xnnpack/AlignedAllocator.h",
4510 ] + MICROKERNEL_BENCHMARK_HDRS,
4511 deps = MICROKERNEL_BENCHMARK_DEPS,
4512)
4513
4514xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004515 name = "f32_relu_bench",
4516 srcs = [
4517 "bench/f32-relu.cc",
4518 "src/xnnpack/AlignedAllocator.h",
4519 ] + MICROKERNEL_BENCHMARK_HDRS,
4520 deps = MICROKERNEL_BENCHMARK_DEPS,
4521)
4522
4523xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524 name = "f32_rmax_bench",
4525 srcs = [
4526 "bench/f32-rmax.cc",
4527 "src/xnnpack/AlignedAllocator.h",
4528 ] + MICROKERNEL_BENCHMARK_HDRS,
4529 deps = MICROKERNEL_BENCHMARK_DEPS,
4530)
4531
4532xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004533 name = "f32_sigmoid_bench",
4534 srcs = [
4535 "bench/f32-sigmoid.cc",
4536 "src/xnnpack/AlignedAllocator.h",
4537 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004538 deps = MICROKERNEL_BENCHMARK_DEPS,
4539)
4540
4541xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004542 name = "f32_spmm_bench",
4543 srcs = [
4544 "bench/f32-spmm.cc",
4545 "bench/gemm.h",
4546 "src/xnnpack/AlignedAllocator.h",
4547 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004548 deps = MICROKERNEL_BENCHMARK_DEPS,
4549)
4550
4551xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004552 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004553 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004554 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004555 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004556 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004557 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004558)
4559
4560xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004561 name = "f32_vscaleexpminusmax_bench",
4562 srcs = [
4563 "bench/f32-vscaleexpminusmax.cc",
4564 "src/xnnpack/AlignedAllocator.h",
4565 ] + MICROKERNEL_BENCHMARK_HDRS,
4566 deps = MICROKERNEL_BENCHMARK_DEPS,
4567)
4568
4569xnnpack_benchmark(
4570 name = "f32_vscaleextexp_bench",
4571 srcs = [
4572 "bench/f32-vscaleextexp.cc",
4573 "src/xnnpack/AlignedAllocator.h",
4574 ] + MICROKERNEL_BENCHMARK_HDRS,
4575 deps = MICROKERNEL_BENCHMARK_DEPS,
4576)
4577
4578xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004579 name = "f32_vsqrt_bench",
4580 srcs = [
4581 "bench/f32-vsqrt.cc",
4582 "src/xnnpack/AlignedAllocator.h",
4583 ] + MICROKERNEL_BENCHMARK_HDRS,
4584 deps = MICROKERNEL_BENCHMARK_DEPS,
4585)
4586
4587xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004588 name = "f32_im2col_gemm_bench",
4589 srcs = [
4590 "bench/f32-im2col-gemm.cc",
4591 "bench/conv.h",
4592 "src/xnnpack/AlignedAllocator.h",
4593 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004594 deps = MICROKERNEL_BENCHMARK_DEPS + [
4595 ":im2col",
4596 ":packing",
4597 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598)
4599
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004600xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004601 name = "rounding_bench",
4602 srcs = [
4603 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004604 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004605 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004606 ] + MICROKERNEL_BENCHMARK_HDRS,
4607 deps = MICROKERNEL_BENCHMARK_DEPS,
4608)
4609
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610########################### Benchmarks for operators ###########################
4611
4612xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004613 name = "average_pooling_bench",
4614 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004615 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004616 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004617 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004618)
4619
4620xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004621 name = "bankers_rounding_bench",
4622 srcs = ["bench/bankers-rounding.cc"],
4623 copts = xnnpack_optional_tflite_copts(),
4624 tags = ["nowin32"],
4625 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4626)
4627
4628xnnpack_benchmark(
4629 name = "ceiling_bench",
4630 srcs = ["bench/ceiling.cc"],
4631 copts = xnnpack_optional_tflite_copts(),
4632 tags = ["nowin32"],
4633 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4634)
4635
4636xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004637 name = "channel_shuffle_bench",
4638 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004639 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004640)
4641
4642xnnpack_benchmark(
4643 name = "convolution_bench",
4644 srcs = ["bench/convolution.cc"],
4645 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004646 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004647 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648)
4649
4650xnnpack_benchmark(
4651 name = "deconvolution_bench",
4652 srcs = ["bench/deconvolution.cc"],
4653 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004654 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004655 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656)
4657
4658xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004659 name = "floor_bench",
4660 srcs = ["bench/floor.cc"],
4661 copts = xnnpack_optional_tflite_copts(),
4662 tags = ["nowin32"],
4663 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4664)
4665
4666xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 name = "global_average_pooling_bench",
4668 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004669 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670)
4671
4672xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004673 name = "hardswish_bench",
4674 srcs = ["bench/hardswish.cc"],
4675 copts = xnnpack_optional_tflite_copts(),
4676 tags = ["nowin32"],
4677 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4678)
4679
4680xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004681 name = "max_pooling_bench",
4682 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004683 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004684)
4685
4686xnnpack_benchmark(
4687 name = "sigmoid_bench",
4688 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004689 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004690 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004691 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692)
4693
4694xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004695 name = "prelu_bench",
4696 srcs = ["bench/prelu.cc"],
4697 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004698 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004699 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004700)
4701
4702xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004703 name = "softmax_bench",
4704 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004705 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004706 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004707 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004708)
4709
Marat Dukhan87727142020-06-24 15:24:10 -07004710xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004711 name = "square_root_bench",
4712 srcs = ["bench/square-root.cc"],
4713 copts = xnnpack_optional_tflite_copts(),
4714 tags = ["nowin32"],
4715 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4716)
4717
4718xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004719 name = "truncation_bench",
4720 srcs = ["bench/truncation.cc"],
4721 deps = OPERATOR_BENCHMARK_DEPS,
4722)
4723
Marat Dukhanc068bb62019-10-04 13:24:39 -07004724############################# End-to-end benchmarks ############################
4725
4726cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004727 name = "fp32_mobilenet_v1",
4728 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004729 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004730 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004731 linkstatic = True,
4732 deps = [
4733 ":XNNPACK",
4734 "@pthreadpool",
4735 ],
4736)
4737
4738cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004739 name = "fp16_mobilenet_v1",
4740 srcs = ["models/fp16-mobilenet-v1.cc"],
4741 hdrs = ["models/models.h"],
4742 copts = xnnpack_std_cxxopts(),
4743 linkstatic = True,
4744 deps = [
4745 ":XNNPACK",
4746 "@FP16",
4747 "@pthreadpool",
4748 ],
4749)
4750
4751cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004752 name = "qs8_mobilenet_v1",
4753 srcs = ["models/qs8-mobilenet-v1.cc"],
4754 hdrs = ["models/models.h"],
4755 copts = xnnpack_std_cxxopts(),
4756 linkstatic = True,
4757 deps = [
4758 ":XNNPACK",
4759 "@pthreadpool",
4760 ],
4761)
4762
4763cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004764 name = "qs8_mobilenet_v2",
4765 srcs = ["models/qs8-mobilenet-v2.cc"],
4766 hdrs = ["models/models.h"],
4767 copts = xnnpack_std_cxxopts(),
4768 linkstatic = True,
4769 deps = [
4770 ":XNNPACK",
4771 "@pthreadpool",
4772 ],
4773)
4774
4775cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004776 name = "fp32_mobilenet_v2",
4777 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004778 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004779 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004780 linkstatic = True,
4781 deps = [
4782 ":XNNPACK",
4783 "@pthreadpool",
4784 ],
4785)
4786
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004787cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004788 name = "fp16_mobilenet_v2",
4789 srcs = ["models/fp16-mobilenet-v2.cc"],
4790 hdrs = ["models/models.h"],
4791 copts = xnnpack_std_cxxopts(),
4792 linkstatic = True,
4793 deps = [
4794 ":XNNPACK",
4795 "@FP16",
4796 "@pthreadpool",
4797 ],
4798)
4799
4800cc_library(
4801 name = "fp32_mobilenet_v3_large",
4802 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004803 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004804 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004805 linkstatic = True,
4806 deps = [
4807 ":XNNPACK",
4808 "@pthreadpool",
4809 ],
4810)
4811
4812cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004813 name = "fp16_mobilenet_v3_large",
4814 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4815 hdrs = ["models/models.h"],
4816 copts = xnnpack_std_cxxopts(),
4817 linkstatic = True,
4818 deps = [
4819 ":XNNPACK",
4820 "@FP16",
4821 "@pthreadpool",
4822 ],
4823)
4824
4825cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004826 name = "fp32_mobilenet_v3_small",
4827 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004828 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004829 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004830 linkstatic = True,
4831 deps = [
4832 ":XNNPACK",
4833 "@pthreadpool",
4834 ],
4835)
4836
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004837cc_library(
4838 name = "fp16_mobilenet_v3_small",
4839 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4840 hdrs = ["models/models.h"],
4841 copts = xnnpack_std_cxxopts(),
4842 linkstatic = True,
4843 deps = [
4844 ":XNNPACK",
4845 "@FP16",
4846 "@pthreadpool",
4847 ],
4848)
4849
Marat Dukhanc068bb62019-10-04 13:24:39 -07004850xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004851 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004852 srcs = [
4853 "bench/f32-dwconv-e2e.cc",
4854 "bench/end2end.h",
4855 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004856 deps = MICROKERNEL_BENCHMARK_DEPS + [
4857 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004858 ":fp32_mobilenet_v1",
4859 ":fp32_mobilenet_v2",
4860 ":fp32_mobilenet_v3_large",
4861 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004862 ],
4863)
4864
4865xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004866 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004867 srcs = [
4868 "bench/f32-gemm-e2e.cc",
4869 "bench/end2end.h",
4870 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004871 deps = MICROKERNEL_BENCHMARK_DEPS + [
4872 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004873 ":fp32_mobilenet_v1",
4874 ":fp32_mobilenet_v2",
4875 ":fp32_mobilenet_v3_large",
4876 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004877 ],
4878)
4879
4880xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004881 name = "end2end_bench",
4882 srcs = ["bench/end2end.cc"],
4883 deps = [
4884 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004885 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004886 ":fp16_mobilenet_v1",
4887 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004888 ":fp16_mobilenet_v3_large",
4889 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004890 ":fp32_mobilenet_v1",
4891 ":fp32_mobilenet_v2",
4892 ":fp32_mobilenet_v3_large",
4893 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004894 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004895 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004896 "@pthreadpool",
4897 ],
4898)
4899
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004900#################### Accuracy evaluation for math functions ####################
4901
4902xnnpack_benchmark(
4903 name = "f32_exp_eval",
4904 srcs = [
4905 "eval/f32-exp.cc",
4906 "src/xnnpack/AlignedAllocator.h",
4907 ] + ACCURACY_EVAL_HDRS,
4908 deps = ACCURACY_EVAL_DEPS,
4909)
4910
Marat Dukhan515c9772019-10-17 18:07:57 -07004911xnnpack_benchmark(
4912 name = "f32_expminus_eval",
4913 srcs = [
4914 "eval/f32-expminus.cc",
4915 "src/xnnpack/AlignedAllocator.h",
4916 ] + ACCURACY_EVAL_HDRS,
4917 deps = ACCURACY_EVAL_DEPS,
4918)
4919
Marat Dukhan98ba4412019-10-23 02:14:28 -07004920xnnpack_benchmark(
4921 name = "f32_extexp_eval",
4922 srcs = [
4923 "eval/f32-extexp.cc",
4924 "src/xnnpack/AlignedAllocator.h",
4925 ] + ACCURACY_EVAL_HDRS,
4926 deps = ACCURACY_EVAL_DEPS,
4927)
4928
Marat Dukhan8853b822020-05-07 12:19:01 -07004929xnnpack_unit_test(
4930 name = "f32_roundne_eval",
4931 srcs = [
4932 "eval/f32-roundne.cc",
4933 "src/xnnpack/AlignedAllocator.h",
4934 "src/xnnpack/math-stubs.h",
4935 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004936 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004937 deps = MICROKERNEL_TEST_DEPS,
4938)
4939
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004940xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004941 name = "f32_roundd_eval",
4942 srcs = [
4943 "eval/f32-roundd.cc",
4944 "src/xnnpack/AlignedAllocator.h",
4945 "src/xnnpack/math-stubs.h",
4946 ] + MICROKERNEL_TEST_HDRS,
4947 automatic = False,
4948 deps = MICROKERNEL_TEST_DEPS,
4949)
4950
4951xnnpack_unit_test(
4952 name = "f32_roundu_eval",
4953 srcs = [
4954 "eval/f32-roundu.cc",
4955 "src/xnnpack/AlignedAllocator.h",
4956 "src/xnnpack/math-stubs.h",
4957 ] + MICROKERNEL_TEST_HDRS,
4958 automatic = False,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004963 name = "f32_roundz_eval",
4964 srcs = [
4965 "eval/f32-roundz.cc",
4966 "src/xnnpack/AlignedAllocator.h",
4967 "src/xnnpack/math-stubs.h",
4968 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004969 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004970 deps = MICROKERNEL_TEST_DEPS,
4971)
4972
Marat Dukhan346a9e52019-11-15 09:06:30 -08004973xnnpack_benchmark(
4974 name = "f32_sigmoid_eval",
4975 srcs = [
4976 "eval/f32-sigmoid.cc",
4977 "src/xnnpack/AlignedAllocator.h",
4978 ] + ACCURACY_EVAL_HDRS,
4979 deps = ACCURACY_EVAL_DEPS,
4980)
4981
Marat Dukhan84000762020-06-29 18:38:43 -07004982xnnpack_benchmark(
4983 name = "f32_sqrt_eval",
4984 srcs = [
4985 "eval/f32-sqrt.cc",
4986 "src/xnnpack/AlignedAllocator.h",
4987 ] + ACCURACY_EVAL_HDRS,
4988 deps = ACCURACY_EVAL_DEPS + [
4989 ":bench_utils",
4990 ],
4991)
4992
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993######################### Unit tests for micro-kernels #########################
4994
4995xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004996 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004998 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 "test/gemm-microkernel-tester.h",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005002 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003)
5004
5005xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005006 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005007 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005008 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005009 "test/spmm-microkernel-tester.h",
5010 "src/xnnpack/AlignedAllocator.h",
5011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016 name = "f32_argmaxpool_test",
5017 srcs = [
5018 "test/f32-argmaxpool.cc",
5019 "test/argmaxpool-microkernel-tester.h",
5020 "src/xnnpack/AlignedAllocator.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005026 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005028 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029 "test/avgpool-microkernel-tester.h",
5030 "src/xnnpack/AlignedAllocator.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005036 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005037 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005038 "test/f32-ibilinear.cc",
5039 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005040 "src/xnnpack/AlignedAllocator.h",
5041 ] + MICROKERNEL_TEST_HDRS,
5042 deps = MICROKERNEL_TEST_DEPS,
5043)
5044
5045xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005046 name = "f32_ibilinear_chw_test",
5047 srcs = [
5048 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005049 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005050 "src/xnnpack/AlignedAllocator.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005056 name = "f16_clamp_test",
5057 srcs = [
5058 "test/f16-clamp.cc",
5059 "test/clamp-microkernel-tester.h",
5060 ] + MICROKERNEL_TEST_HDRS,
5061 deps = MICROKERNEL_TEST_DEPS,
5062)
5063
5064xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 name = "f32_clamp_test",
5066 srcs = [
5067 "test/f32-clamp.cc",
5068 "test/clamp-microkernel-tester.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005074 name = "f32_igemm_test",
5075 srcs = [
5076 "test/f32-igemm.cc",
5077 "test/gemm-microkernel-tester.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005080 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005081)
5082
5083xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005084 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005086 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087 "test/gemm-microkernel-tester.h",
5088 "src/xnnpack/AlignedAllocator.h",
5089 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005090 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091)
5092
5093xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005094 name = "f32_igemm_minmax_test",
5095 srcs = [
5096 "test/f32-igemm-minmax.cc",
5097 "test/gemm-microkernel-tester.h",
5098 "src/xnnpack/AlignedAllocator.h",
5099 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005100 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005101)
5102
5103xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005104 name = "f16_igemm_minmax_test",
5105 srcs = [
5106 "test/f16-igemm-minmax.cc",
5107 "test/gemm-microkernel-tester.h",
5108 "src/xnnpack/AlignedAllocator.h",
5109 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005110 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005111)
5112
5113xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 name = "f32_conv_hwc_test",
5115 srcs = [
5116 "test/f32-conv-hwc.cc",
5117 "test/conv-hwc-microkernel-tester.h",
5118 "src/xnnpack/AlignedAllocator.h",
5119 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005120 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005121)
5122
5123xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005124 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005126 "test/f32-conv-hwc2chw.cc",
5127 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128 "src/xnnpack/AlignedAllocator.h",
5129 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005130 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005131)
5132
5133xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005134 name = "f32_dwconv_test",
5135 srcs = [
5136 "test/f32-dwconv.cc",
5137 "test/dwconv-microkernel-tester.h",
5138 "src/xnnpack/AlignedAllocator.h",
5139 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005140 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005141)
5142
5143xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005144 name = "f16_dwconv_minmax_test",
5145 srcs = [
5146 "test/f16-dwconv-minmax.cc",
5147 "test/dwconv-microkernel-tester.h",
5148 "src/xnnpack/AlignedAllocator.h",
5149 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005150 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005151)
5152
5153xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005154 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005156 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157 "test/dwconv-microkernel-tester.h",
5158 "src/xnnpack/AlignedAllocator.h",
5159 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005160 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161)
5162
5163xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005164 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005166 "test/f32-dwconv2d-chw.cc",
5167 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168 "src/xnnpack/AlignedAllocator.h",
5169 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005170 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171)
5172
5173xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005174 name = "f16_gavgpool_minmax_test",
5175 srcs = [
5176 "test/f16-gavgpool-minmax.cc",
5177 "test/gavgpool-microkernel-tester.h",
5178 "src/xnnpack/AlignedAllocator.h",
5179 ] + MICROKERNEL_TEST_HDRS,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005184 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005186 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 "test/gavgpool-microkernel-tester.h",
5188 "src/xnnpack/AlignedAllocator.h",
5189 ] + MICROKERNEL_TEST_HDRS,
5190 deps = MICROKERNEL_TEST_DEPS,
5191)
5192
5193xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005194 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005196 "test/f32-gavgpool-cw.cc",
5197 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005198 "src/xnnpack/AlignedAllocator.h",
5199 ] + MICROKERNEL_TEST_HDRS,
5200 deps = MICROKERNEL_TEST_DEPS,
5201)
5202
5203xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005204 name = "f32_gemm_test",
5205 srcs = [
5206 "test/f32-gemm.cc",
5207 "test/gemm-microkernel-tester.h",
5208 "src/xnnpack/AlignedAllocator.h",
5209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005210 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005211)
5212
5213xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005214 name = "f32_gemm_relu_test",
5215 srcs = [
5216 "test/f32-gemm-relu.cc",
5217 "test/gemm-microkernel-tester.h",
5218 "src/xnnpack/AlignedAllocator.h",
5219 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005220 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005221)
5222
5223xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005224 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005226 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227 "test/gemm-microkernel-tester.h",
5228 "src/xnnpack/AlignedAllocator.h",
5229 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005230 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231)
5232
5233xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005234 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005236 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 "test/gemm-microkernel-tester.h",
5238 "src/xnnpack/AlignedAllocator.h",
5239 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005240 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005241)
5242
5243xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005244 name = "f16_hswish_test",
5245 srcs = [
5246 "test/f16-hswish.cc",
5247 "test/hswish-microkernel-tester.h",
5248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 name = "f32_hswish_test",
5254 srcs = [
5255 "test/f32-hswish.cc",
5256 "test/hswish-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005262 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005264 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 "test/maxpool-microkernel-tester.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005271 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005273 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274 "test/avgpool-microkernel-tester.h",
5275 "src/xnnpack/AlignedAllocator.h",
5276 ] + MICROKERNEL_TEST_HDRS,
5277 deps = MICROKERNEL_TEST_DEPS,
5278)
5279
5280xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005281 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005283 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 "test/gemm-microkernel-tester.h",
5285 "src/xnnpack/AlignedAllocator.h",
5286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005287 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288)
5289
5290xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005291 name = "f16_prelu_test",
5292 srcs = [
5293 "test/f16-prelu.cc",
5294 "test/prelu-microkernel-tester.h",
5295 "src/xnnpack/AlignedAllocator.h",
5296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005301 name = "f32_prelu_test",
5302 srcs = [
5303 "test/f32-prelu.cc",
5304 "test/prelu-microkernel-tester.h",
5305 "src/xnnpack/AlignedAllocator.h",
5306 ] + MICROKERNEL_TEST_HDRS,
5307 deps = MICROKERNEL_TEST_DEPS,
5308)
5309
5310xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005311 name = "f32_raddexpminusmax_test",
5312 srcs = [
5313 "test/f32-raddexpminusmax.cc",
5314 "test/raddexpminusmax-microkernel-tester.h",
5315 ] + MICROKERNEL_TEST_HDRS,
5316 deps = MICROKERNEL_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005320 name = "f32_raddextexp_test",
5321 srcs = [
5322 "test/f32-raddextexp.cc",
5323 "test/raddextexp-microkernel-tester.h",
5324 ] + MICROKERNEL_TEST_HDRS,
5325 deps = MICROKERNEL_TEST_DEPS,
5326)
5327
5328xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005329 name = "f32_raddstoreexpminusmax_test",
5330 srcs = [
5331 "test/f32-raddstoreexpminusmax.cc",
5332 "test/raddstoreexpminusmax-microkernel-tester.h",
5333 ] + MICROKERNEL_TEST_HDRS,
5334 deps = MICROKERNEL_TEST_DEPS,
5335)
5336
5337xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005338 name = "f32_relu_test",
5339 srcs = [
5340 "test/f32-relu.cc",
5341 "test/vunary-microkernel-tester.h",
5342 ] + MICROKERNEL_TEST_HDRS,
5343 deps = MICROKERNEL_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005347 name = "f32_rmax_test",
5348 srcs = [
5349 "test/f32-rmax.cc",
5350 "test/rmax-microkernel-tester.h",
5351 ] + MICROKERNEL_TEST_HDRS,
5352 deps = MICROKERNEL_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005356 name = "f32_sigmoid_test",
5357 srcs = [
5358 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005359 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005360 ] + MICROKERNEL_TEST_HDRS,
5361 deps = MICROKERNEL_TEST_DEPS,
5362)
5363
5364xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005365 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005367 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005368 "test/spmm-microkernel-tester.h",
5369 "src/xnnpack/AlignedAllocator.h",
5370 ] + MICROKERNEL_TEST_HDRS,
5371 deps = MICROKERNEL_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005375 name = "f16_vadd_minmax_test",
5376 srcs = [
5377 "test/f16-vadd-minmax.cc",
5378 "test/vbinary-microkernel-tester.h",
5379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
5384 name = "f16_vaddc_minmax_test",
5385 srcs = [
5386 "test/f16-vaddc-minmax.cc",
5387 "test/vbinaryc-microkernel-tester.h",
5388 ] + MICROKERNEL_TEST_HDRS,
5389 deps = MICROKERNEL_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
5393 name = "f16_vdiv_minmax_test",
5394 srcs = [
5395 "test/f16-vdiv-minmax.cc",
5396 "test/vbinary-microkernel-tester.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
5402 name = "f16_vdivc_minmax_test",
5403 srcs = [
5404 "test/f16-vdivc-minmax.cc",
5405 "test/vbinaryc-microkernel-tester.h",
5406 ] + MICROKERNEL_TEST_HDRS,
5407 deps = MICROKERNEL_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
5411 name = "f16_vrdivc_minmax_test",
5412 srcs = [
5413 "test/f16-vrdivc-minmax.cc",
5414 "test/vbinaryc-microkernel-tester.h",
5415 ] + MICROKERNEL_TEST_HDRS,
5416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
5419xnnpack_unit_test(
5420 name = "f16_vmax_test",
5421 srcs = [
5422 "test/f16-vmax.cc",
5423 "test/vbinary-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
5429 name = "f16_vmaxc_test",
5430 srcs = [
5431 "test/f16-vmaxc.cc",
5432 "test/vbinaryc-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
5438 name = "f16_vmin_test",
5439 srcs = [
5440 "test/f16-vmin.cc",
5441 "test/vbinary-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
5447 name = "f16_vminc_test",
5448 srcs = [
5449 "test/f16-vminc.cc",
5450 "test/vbinaryc-microkernel-tester.h",
5451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
5456 name = "f16_vmul_minmax_test",
5457 srcs = [
5458 "test/f16-vmul-minmax.cc",
5459 "test/vbinary-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
5465 name = "f16_vmulc_minmax_test",
5466 srcs = [
5467 "test/f16-vmulc-minmax.cc",
5468 "test/vbinaryc-microkernel-tester.h",
5469 ] + MICROKERNEL_TEST_HDRS,
5470 deps = MICROKERNEL_TEST_DEPS,
5471)
5472
5473xnnpack_unit_test(
5474 name = "f16_vsub_minmax_test",
5475 srcs = [
5476 "test/f16-vsub-minmax.cc",
5477 "test/vbinary-microkernel-tester.h",
5478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
5483 name = "f16_vsubc_minmax_test",
5484 srcs = [
5485 "test/f16-vsubc-minmax.cc",
5486 "test/vbinaryc-microkernel-tester.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
5492 name = "f16_vrsubc_minmax_test",
5493 srcs = [
5494 "test/f16-vrsubc-minmax.cc",
5495 "test/vbinaryc-microkernel-tester.h",
5496 ] + MICROKERNEL_TEST_HDRS,
5497 deps = MICROKERNEL_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005501 name = "f32_vabs_test",
5502 srcs = [
5503 "test/f32-vabs.cc",
5504 "test/vunary-microkernel-tester.h",
5505 ] + MICROKERNEL_TEST_HDRS,
5506 deps = MICROKERNEL_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005510 name = "f32_vadd_test",
5511 srcs = [
5512 "test/f32-vadd.cc",
5513 "test/vbinary-microkernel-tester.h",
5514 ] + MICROKERNEL_TEST_HDRS,
5515 deps = MICROKERNEL_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005519 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005520 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005521 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005522 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005523 ] + MICROKERNEL_TEST_HDRS,
5524 deps = MICROKERNEL_TEST_DEPS,
5525)
5526
5527xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005528 name = "f32_vadd_relu_test",
5529 srcs = [
5530 "test/f32-vadd-relu.cc",
5531 "test/vbinary-microkernel-tester.h",
5532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005537 name = "f32_vaddc_test",
5538 srcs = [
5539 "test/f32-vaddc.cc",
5540 "test/vbinaryc-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005546 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005547 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005548 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005549 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005555 name = "f32_vaddc_relu_test",
5556 srcs = [
5557 "test/f32-vaddc-relu.cc",
5558 "test/vbinaryc-microkernel-tester.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005564 name = "f32_vdiv_test",
5565 srcs = [
5566 "test/f32-vdiv.cc",
5567 "test/vbinary-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005573 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005574 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005575 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005576 "test/vbinary-microkernel-tester.h",
5577 ] + MICROKERNEL_TEST_HDRS,
5578 deps = MICROKERNEL_TEST_DEPS,
5579)
5580
5581xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005582 name = "f32_vdiv_relu_test",
5583 srcs = [
5584 "test/f32-vdiv-relu.cc",
5585 "test/vbinary-microkernel-tester.h",
5586 ] + MICROKERNEL_TEST_HDRS,
5587 deps = MICROKERNEL_TEST_DEPS,
5588)
5589
5590xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005591 name = "f32_vdivc_test",
5592 srcs = [
5593 "test/f32-vdivc.cc",
5594 "test/vbinaryc-microkernel-tester.h",
5595 ] + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS,
5597)
5598
5599xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005600 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005601 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005602 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005603 "test/vbinaryc-microkernel-tester.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005609 name = "f32_vdivc_relu_test",
5610 srcs = [
5611 "test/f32-vdivc-relu.cc",
5612 "test/vbinaryc-microkernel-tester.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005618 name = "f32_vrdivc_test",
5619 srcs = [
5620 "test/f32-vrdivc.cc",
5621 "test/vbinaryc-microkernel-tester.h",
5622 ] + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005627 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005628 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005629 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005630 "test/vbinaryc-microkernel-tester.h",
5631 ] + MICROKERNEL_TEST_HDRS,
5632 deps = MICROKERNEL_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005636 name = "f32_vrdivc_relu_test",
5637 srcs = [
5638 "test/f32-vrdivc-relu.cc",
5639 "test/vbinaryc-microkernel-tester.h",
5640 ] + MICROKERNEL_TEST_HDRS,
5641 deps = MICROKERNEL_TEST_DEPS,
5642)
5643
5644xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005645 name = "f32_vmax_test",
5646 srcs = [
5647 "test/f32-vmax.cc",
5648 "test/vbinary-microkernel-tester.h",
5649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
5654 name = "f32_vmaxc_test",
5655 srcs = [
5656 "test/f32-vmaxc.cc",
5657 "test/vbinaryc-microkernel-tester.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
5663 name = "f32_vmin_test",
5664 srcs = [
5665 "test/f32-vmin.cc",
5666 "test/vbinary-microkernel-tester.h",
5667 ] + MICROKERNEL_TEST_HDRS,
5668 deps = MICROKERNEL_TEST_DEPS,
5669)
5670
5671xnnpack_unit_test(
5672 name = "f32_vminc_test",
5673 srcs = [
5674 "test/f32-vminc.cc",
5675 "test/vbinaryc-microkernel-tester.h",
5676 ] + MICROKERNEL_TEST_HDRS,
5677 deps = MICROKERNEL_TEST_DEPS,
5678)
5679
5680xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005681 name = "f32_vmul_test",
5682 srcs = [
5683 "test/f32-vmul.cc",
5684 "test/vbinary-microkernel-tester.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 deps = MICROKERNEL_TEST_DEPS,
5687)
5688
5689xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005690 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005691 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005692 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005693 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005699 name = "f32_vmul_relu_test",
5700 srcs = [
5701 "test/f32-vmul-relu.cc",
5702 "test/vbinary-microkernel-tester.h",
5703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005708 name = "f32_vmulc_test",
5709 srcs = [
5710 "test/f32-vmulc.cc",
5711 "test/vbinaryc-microkernel-tester.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005717 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005718 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005719 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005720 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005721 ] + MICROKERNEL_TEST_HDRS,
5722 deps = MICROKERNEL_TEST_DEPS,
5723)
5724
5725xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005726 name = "f32_vmulc_relu_test",
5727 srcs = [
5728 "test/f32-vmulc-relu.cc",
5729 "test/vbinaryc-microkernel-tester.h",
5730 ] + MICROKERNEL_TEST_HDRS,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
5734xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005735 name = "f16_vmulcaddc_minmax_test",
5736 srcs = [
5737 "test/f16-vmulcaddc-minmax.cc",
5738 "test/vmulcaddc-microkernel-tester.h",
5739 "src/xnnpack/AlignedAllocator.h",
5740 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005741 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005742)
5743
5744xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005745 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005746 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005747 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005748 "test/vmulcaddc-microkernel-tester.h",
5749 "src/xnnpack/AlignedAllocator.h",
5750 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005751 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752)
5753
5754xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005755 name = "f32_vlrelu_test",
5756 srcs = [
5757 "test/f32-vlrelu.cc",
5758 "test/vunary-microkernel-tester.h",
5759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005764 name = "f32_vneg_test",
5765 srcs = [
5766 "test/f32-vneg.cc",
5767 "test/vunary-microkernel-tester.h",
5768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005773 name = "f32_vrndne_test",
5774 srcs = [
5775 "test/f32-vrndne.cc",
5776 "test/vunary-microkernel-tester.h",
5777 ] + MICROKERNEL_TEST_HDRS,
5778 deps = MICROKERNEL_TEST_DEPS,
5779)
5780
5781xnnpack_unit_test(
5782 name = "f32_vrndz_test",
5783 srcs = [
5784 "test/f32-vrndz.cc",
5785 "test/vunary-microkernel-tester.h",
5786 ] + MICROKERNEL_TEST_HDRS,
5787 deps = MICROKERNEL_TEST_DEPS,
5788)
5789
5790xnnpack_unit_test(
5791 name = "f32_vrndu_test",
5792 srcs = [
5793 "test/f32-vrndu.cc",
5794 "test/vunary-microkernel-tester.h",
5795 ] + MICROKERNEL_TEST_HDRS,
5796 deps = MICROKERNEL_TEST_DEPS,
5797)
5798
5799xnnpack_unit_test(
5800 name = "f32_vrndd_test",
5801 srcs = [
5802 "test/f32-vrndd.cc",
5803 "test/vunary-microkernel-tester.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 deps = MICROKERNEL_TEST_DEPS,
5806)
5807
5808xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005809 name = "f32_vscale_test",
5810 srcs = [
5811 "test/f32-vscale.cc",
5812 "test/vscale-microkernel-tester.h",
5813 ] + MICROKERNEL_TEST_HDRS,
5814 deps = MICROKERNEL_TEST_DEPS,
5815)
5816
5817xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005818 name = "f32_vscaleexpminusmax_test",
5819 srcs = [
5820 "test/f32-vscaleexpminusmax.cc",
5821 "test/vscaleexpminusmax-microkernel-tester.h",
5822 ] + MICROKERNEL_TEST_HDRS,
5823 deps = MICROKERNEL_TEST_DEPS,
5824)
5825
5826xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005827 name = "f32_vscaleextexp_test",
5828 srcs = [
5829 "test/f32-vscaleextexp.cc",
5830 "test/vscaleextexp-microkernel-tester.h",
5831 ] + MICROKERNEL_TEST_HDRS,
5832 deps = MICROKERNEL_TEST_DEPS,
5833)
5834
5835xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005836 name = "f32_vsqr_test",
5837 srcs = [
5838 "test/f32-vsqr.cc",
5839 "test/vunary-microkernel-tester.h",
5840 ] + MICROKERNEL_TEST_HDRS,
5841 deps = MICROKERNEL_TEST_DEPS,
5842)
5843
5844xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005845 name = "f32_vsqrdiff_test",
5846 srcs = [
5847 "test/f32-vsqrdiff.cc",
5848 "test/vbinary-microkernel-tester.h",
5849 ] + MICROKERNEL_TEST_HDRS,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
5854 name = "f32_vsqrdiffc_test",
5855 srcs = [
5856 "test/f32-vsqrdiffc.cc",
5857 "test/vbinaryc-microkernel-tester.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005863 name = "f32_vsqrt_test",
5864 srcs = [
5865 "test/f32-vsqrt.cc",
5866 "test/vunary-microkernel-tester.h",
5867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005872 name = "f32_vsub_test",
5873 srcs = [
5874 "test/f32-vsub.cc",
5875 "test/vbinary-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005881 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005882 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005883 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005884 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005890 name = "f32_vsub_relu_test",
5891 srcs = [
5892 "test/f32-vsub-relu.cc",
5893 "test/vbinary-microkernel-tester.h",
5894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005899 name = "f32_vsubc_test",
5900 srcs = [
5901 "test/f32-vsubc.cc",
5902 "test/vbinaryc-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005908 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005909 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005910 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005911 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005912 ] + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS,
5914)
5915
5916xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005917 name = "f32_vsubc_relu_test",
5918 srcs = [
5919 "test/f32-vsubc-relu.cc",
5920 "test/vbinaryc-microkernel-tester.h",
5921 ] + MICROKERNEL_TEST_HDRS,
5922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
5925xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005926 name = "f32_vrsubc_test",
5927 srcs = [
5928 "test/f32-vrsubc.cc",
5929 "test/vbinaryc-microkernel-tester.h",
5930 ] + MICROKERNEL_TEST_HDRS,
5931 deps = MICROKERNEL_TEST_DEPS,
5932)
5933
5934xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005935 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005936 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005937 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005938 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005939 ] + MICROKERNEL_TEST_HDRS,
5940 deps = MICROKERNEL_TEST_DEPS,
5941)
5942
5943xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005944 name = "f32_vrsubc_relu_test",
5945 srcs = [
5946 "test/f32-vrsubc-relu.cc",
5947 "test/vbinaryc-microkernel-tester.h",
5948 ] + MICROKERNEL_TEST_HDRS,
5949 deps = MICROKERNEL_TEST_DEPS,
5950)
5951
5952xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005953 name = "qs8_dwconv_minmax_test",
5954 srcs = [
5955 "test/qs8-dwconv-minmax.cc",
5956 "test/dwconv-microkernel-tester.h",
5957 "src/xnnpack/AlignedAllocator.h",
5958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5959 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5960)
5961
5962xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005963 name = "qs8_gavgpool_minmax_test",
5964 srcs = [
5965 "test/qs8-gavgpool-minmax.cc",
5966 "test/gavgpool-microkernel-tester.h",
5967 "src/xnnpack/AlignedAllocator.h",
5968 ] + MICROKERNEL_TEST_HDRS,
5969 deps = MICROKERNEL_TEST_DEPS,
5970)
5971
5972xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005973 name = "qs8_gemm_minmax_test",
5974 srcs = [
5975 "test/qs8-gemm-minmax.cc",
5976 "test/gemm-microkernel-tester.h",
5977 "src/xnnpack/AlignedAllocator.h",
5978 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5979 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5980)
5981
5982xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005983 name = "qs8_igemm_minmax_test",
5984 srcs = [
5985 "test/qs8-igemm-minmax.cc",
5986 "test/gemm-microkernel-tester.h",
5987 "src/xnnpack/AlignedAllocator.h",
5988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5989 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5990)
5991
5992xnnpack_unit_test(
5993 name = "qs8_requantization_test",
5994 srcs = [
5995 "src/xnnpack/requantization-stubs.h",
5996 "test/qs8-requantization.cc",
5997 "test/requantization-tester.h",
5998 ] + MICROKERNEL_TEST_HDRS,
5999 deps = MICROKERNEL_TEST_DEPS,
6000)
6001
6002xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006003 name = "qs8_vadd_minmax_test",
6004 srcs = [
6005 "test/qs8-vadd-minmax.cc",
6006 "test/vadd-microkernel-tester.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006012 name = "qs8_vaddc_minmax_test",
6013 srcs = [
6014 "test/qs8-vaddc-minmax.cc",
6015 "test/vaddc-microkernel-tester.h",
6016 ] + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006021 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006022 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006023 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006024 "test/avgpool-microkernel-tester.h",
6025 "src/xnnpack/AlignedAllocator.h",
6026 ] + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006031 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006032 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006033 "test/qu8-dwconv-minmax.cc",
6034 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006035 "src/xnnpack/AlignedAllocator.h",
6036 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006037 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038)
6039
6040xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006041 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006042 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006043 "test/qu8-igemm-minmax.cc",
6044 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006045 "src/xnnpack/AlignedAllocator.h",
6046 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006047 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006048)
6049
6050xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006051 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006053 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054 "test/gavgpool-microkernel-tester.h",
6055 "src/xnnpack/AlignedAllocator.h",
6056 ] + MICROKERNEL_TEST_HDRS,
6057 deps = MICROKERNEL_TEST_DEPS,
6058)
6059
6060xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006061 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006062 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006063 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006064 "test/gemm-microkernel-tester.h",
6065 "src/xnnpack/AlignedAllocator.h",
6066 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006067 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068)
6069
6070xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006071 name = "qu8_requantization_test",
6072 srcs = [
6073 "src/xnnpack/requantization-stubs.h",
6074 "test/qu8-requantization.cc",
6075 "test/requantization-tester.h",
6076 ] + MICROKERNEL_TEST_HDRS,
6077 deps = MICROKERNEL_TEST_DEPS,
6078)
6079
6080xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006081 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006083 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006084 "test/vadd-microkernel-tester.h",
6085 ] + MICROKERNEL_TEST_HDRS,
6086 deps = MICROKERNEL_TEST_DEPS,
6087)
6088
6089xnnpack_unit_test(
6090 name = "u8_clamp_test",
6091 srcs = [
6092 "test/u8-clamp.cc",
6093 "test/clamp-microkernel-tester.h",
6094 ] + MICROKERNEL_TEST_HDRS,
6095 deps = MICROKERNEL_TEST_DEPS,
6096)
6097
6098xnnpack_unit_test(
6099 name = "u8_lut32norm_test",
6100 srcs = [
6101 "test/u8-lut32norm.cc",
6102 "test/lut-norm-microkernel-tester.h",
6103 ] + MICROKERNEL_TEST_HDRS,
6104 deps = MICROKERNEL_TEST_DEPS,
6105)
6106
6107xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006108 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006109 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006110 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006111 "test/maxpool-microkernel-tester.h",
6112 ] + MICROKERNEL_TEST_HDRS,
6113 deps = MICROKERNEL_TEST_DEPS,
6114)
6115
6116xnnpack_unit_test(
6117 name = "u8_rmax_test",
6118 srcs = [
6119 "test/u8-rmax.cc",
6120 "test/rmax-microkernel-tester.h",
6121 ] + MICROKERNEL_TEST_HDRS,
6122 deps = MICROKERNEL_TEST_DEPS,
6123)
6124
6125xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006126 name = "x32_fill_test",
6127 srcs = [
6128 "test/x32-fill.cc",
6129 "test/fill-microkernel-tester.h",
6130 ] + MICROKERNEL_TEST_HDRS,
6131 deps = MICROKERNEL_TEST_DEPS,
6132)
6133
6134xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135 name = "x32_packx_test",
6136 srcs = [
6137 "test/x32-packx.cc",
6138 "test/pack-microkernel-tester.h",
6139 "src/xnnpack/AlignedAllocator.h",
6140 ] + MICROKERNEL_TEST_HDRS,
6141 deps = MICROKERNEL_TEST_DEPS,
6142)
6143
6144xnnpack_unit_test(
6145 name = "x32_pad_test",
6146 srcs = [
6147 "test/x32-pad.cc",
6148 "test/pad-microkernel-tester.h",
6149 ] + MICROKERNEL_TEST_HDRS,
6150 deps = MICROKERNEL_TEST_DEPS,
6151)
6152
6153xnnpack_unit_test(
6154 name = "x32_unpool_test",
6155 srcs = [
6156 "test/x32-unpool.cc",
6157 "test/unpool-microkernel-tester.h",
6158 ] + MICROKERNEL_TEST_HDRS,
6159 deps = MICROKERNEL_TEST_DEPS,
6160)
6161
6162xnnpack_unit_test(
6163 name = "x32_zip_test",
6164 srcs = [
6165 "test/x32-zip.cc",
6166 "test/zip-microkernel-tester.h",
6167 ] + MICROKERNEL_TEST_HDRS,
6168 deps = MICROKERNEL_TEST_DEPS,
6169)
6170
6171xnnpack_unit_test(
6172 name = "x8_lut_test",
6173 srcs = [
6174 "test/x8-lut.cc",
6175 "test/lut-microkernel-tester.h",
6176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
6181 name = "x8_zip_test",
6182 srcs = [
6183 "test/x8-zip.cc",
6184 "test/zip-microkernel-tester.h",
6185 ] + MICROKERNEL_TEST_HDRS,
6186 deps = MICROKERNEL_TEST_DEPS,
6187)
6188
Marat Dukhan20c3b922020-03-10 03:45:06 -07006189########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006190
6191xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006192 name = "operator_size_test",
6193 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006194 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006195)
6196
Marat Dukhan20c3b922020-03-10 03:45:06 -07006197xnnpack_binary(
6198 name = "subgraph_size_test",
6199 srcs = ["test/subgraph-size.c"],
6200 deps = [":XNNPACK"],
6201)
6202
6203########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204
6205xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006206 name = "abs_nc_test",
6207 srcs = [
6208 "test/abs-nc.cc",
6209 "test/abs-operator-tester.h",
6210 ],
6211 deps = OPERATOR_TEST_DEPS,
6212)
6213
6214xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006215 name = "add_nd_test",
6216 srcs = [
6217 "test/add-nd.cc",
6218 "test/binary-elementwise-operator-tester.h",
6219 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006220 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006221)
6222
6223xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006224 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006226 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227 "test/argmax-pooling-operator-tester.h",
6228 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006229 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230)
6231
6232xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006233 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006234 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006235 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236 "test/average-pooling-operator-tester.h",
6237 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006238 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239)
6240
6241xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006242 name = "bankers_rounding_nc_test",
6243 srcs = [
6244 "test/bankers-rounding-nc.cc",
6245 "test/bankers-rounding-operator-tester.h",
6246 ],
6247 deps = OPERATOR_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
6251 name = "ceiling_nc_test",
6252 srcs = [
6253 "test/ceiling-nc.cc",
6254 "test/ceiling-operator-tester.h",
6255 ],
6256 deps = OPERATOR_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006260 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006262 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263 "test/channel-shuffle-operator-tester.h",
6264 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006265 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266)
6267
6268xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006269 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006271 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 "test/clamp-operator-tester.h",
6273 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006274 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006275)
6276
6277xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006278 name = "constant_pad_nd_test",
6279 srcs = [
6280 "test/constant-pad-nd.cc",
6281 "test/constant-pad-operator-tester.h",
6282 ],
6283 deps = OPERATOR_TEST_DEPS,
6284)
6285
6286xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006287 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006289 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290 "test/convolution-operator-tester.h",
6291 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006292 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293)
6294
6295xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006296 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006297 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006298 "test/convolution-nchw.cc",
6299 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006300 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006301 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006302)
6303
6304xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006305 name = "copy_nc_test",
6306 srcs = [
6307 "test/copy-nc.cc",
6308 "test/copy-operator-tester.h",
6309 ],
6310 deps = OPERATOR_TEST_DEPS,
6311)
6312
6313xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006314 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006315 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006316 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317 "test/deconvolution-operator-tester.h",
6318 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006319 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006320)
6321
6322xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006323 name = "divide_nd_test",
6324 srcs = [
6325 "test/binary-elementwise-operator-tester.h",
6326 "test/divide-nd.cc",
6327 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006328 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006329)
6330
6331xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006332 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006333 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006334 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006335 "test/fully-connected-operator-tester.h",
6336 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006337 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006338)
6339
6340xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006341 name = "floor_nc_test",
6342 srcs = [
6343 "test/floor-nc.cc",
6344 "test/floor-operator-tester.h",
6345 ],
6346 deps = OPERATOR_TEST_DEPS,
6347)
6348
6349xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006350 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006351 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006352 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006353 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006354 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006355 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006356)
6357
6358xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006359 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006360 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006361 "test/global-average-pooling-ncw.cc",
6362 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006363 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006364 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006365)
6366
6367xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006368 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006369 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006370 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371 "test/hardswish-operator-tester.h",
6372 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006373 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006374)
6375
6376xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006377 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006378 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006379 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380 "test/leaky-relu-operator-tester.h",
6381 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006382 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006383)
6384
6385xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006386 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006387 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006388 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389 "test/max-pooling-operator-tester.h",
6390 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006391 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006392)
6393
6394xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006395 name = "maximum_nd_test",
6396 srcs = [
6397 "test/binary-elementwise-operator-tester.h",
6398 "test/maximum-nd.cc",
6399 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006400 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006401)
6402
6403xnnpack_unit_test(
6404 name = "minimum_nd_test",
6405 srcs = [
6406 "test/binary-elementwise-operator-tester.h",
6407 "test/minimum-nd.cc",
6408 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006409 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006410)
6411
6412xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006413 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006414 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006415 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006416 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006417 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006418 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006419)
6420
6421xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006422 name = "negate_nc_test",
6423 srcs = [
6424 "test/negate-nc.cc",
6425 "test/negate-operator-tester.h",
6426 ],
6427 deps = OPERATOR_TEST_DEPS,
6428)
6429
6430xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006431 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006432 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006433 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006434 "test/prelu-operator-tester.h",
6435 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006436 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006437)
6438
6439xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006440 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006441 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006442 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006443 "test/resize-bilinear-operator-tester.h",
6444 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006445 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006446)
6447
6448xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006449 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006451 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452 "test/sigmoid-operator-tester.h",
6453 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006454 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006455)
6456
6457xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006458 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006459 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006460 "test/softmax-nc.cc",
6461 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006462 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006463 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006464)
6465
6466xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006467 name = "square_nc_test",
6468 srcs = [
6469 "test/square-nc.cc",
6470 "test/square-operator-tester.h",
6471 ],
6472 deps = OPERATOR_TEST_DEPS,
6473)
6474
6475xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006476 name = "square_root_nc_test",
6477 srcs = [
6478 "test/square-root-nc.cc",
6479 "test/square-root-operator-tester.h",
6480 ],
6481 deps = OPERATOR_TEST_DEPS,
6482)
6483
6484xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006485 name = "squared_difference_nd_test",
6486 srcs = [
6487 "test/binary-elementwise-operator-tester.h",
6488 "test/squared-difference-nd.cc",
6489 ],
6490 deps = OPERATOR_TEST_DEPS,
6491)
6492
6493xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006494 name = "subtract_nd_test",
6495 srcs = [
6496 "test/binary-elementwise-operator-tester.h",
6497 "test/subtract-nd.cc",
6498 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006499 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006500)
6501
6502xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006503 name = "truncation_nc_test",
6504 srcs = [
6505 "test/truncation-nc.cc",
6506 "test/truncation-operator-tester.h",
6507 ],
6508 deps = OPERATOR_TEST_DEPS,
6509)
6510
6511xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006512 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006513 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006514 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006515 "test/unpooling-operator-tester.h",
6516 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006517 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006518)
6519
Chao Mei6ddfc602020-05-13 22:29:36 -07006520############################### Misc unit tests ###############################
6521
6522xnnpack_unit_test(
6523 name = "memory_planner_test",
6524 srcs = [
6525 "test/memory-planner-test.cc",
6526 ],
6527 deps = [
6528 ":XNNPACK",
6529 ":memory_planner",
6530 ],
6531)
6532
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006533xnnpack_unit_test(
6534 name = "subgraph_nchw_test",
6535 srcs = [
6536 "src/xnnpack/subgraph.h",
6537 "test/subgraph-nchw.cc",
6538 "test/subgraph-tester.h",
6539 ],
6540 deps = [
6541 ":XNNPACK",
6542 ],
6543)
6544
Marat Dukhan08c4a432019-10-03 09:29:21 -07006545############################# Build configurations #############################
6546
Marat Dukhanb8642352019-10-30 15:43:02 -07006547# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006548config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006549 name = "xnn_enable_assembly_explicit_true",
6550 define_values = {"xnn_enable_assembly": "true"},
6551)
6552
6553# Disables usage of assembly kernels.
6554config_setting(
6555 name = "xnn_enable_assembly_explicit_false",
6556 define_values = {"xnn_enable_assembly": "false"},
6557)
6558
Marat Dukhan9de90e02020-06-18 16:04:12 -07006559# Enables usage of sparse inference.
6560config_setting(
6561 name = "xnn_enable_sparse_explicit_true",
6562 define_values = {"xnn_enable_sparse": "true"},
6563)
6564
6565# Disables usage of sparse inference.
6566config_setting(
6567 name = "xnn_enable_sparse_explicit_false",
6568 define_values = {"xnn_enable_sparse": "false"},
6569)
6570
Marat Dukhan05702cf2020-03-26 15:41:33 -07006571# Disables usage of HMP-aware optimizations.
6572config_setting(
6573 name = "xnn_enable_hmp_explicit_false",
6574 define_values = {"xnn_enable_hmp": "false"},
6575)
6576
Chao Mei6ddfc602020-05-13 22:29:36 -07006577# Enable usage of optimized memory allocation
6578config_setting(
6579 name = "xnn_enable_memopt_explicit_true",
6580 define_values = {"xnnpack_enable_memopt": "true"},
6581)
6582
6583# Disable usage of optimized memory allocation
6584config_setting(
6585 name = "xnn_enable_memopt_explicit_false",
6586 define_values = {"xnnpack_enable_memopt": "false"},
6587)
6588
Marat Dukhanb8642352019-10-30 15:43:02 -07006589# Builds with -c dbg
6590config_setting(
6591 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006592 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006593 "compilation_mode": "dbg",
6594 },
6595)
6596
6597# Builds with -c opt
6598config_setting(
6599 name = "optimized_build",
6600 values = {
6601 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006602 },
6603)
6604
6605config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006606 name = "linux_k8",
6607 values = {"cpu": "k8"},
6608)
6609
6610config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006611 name = "linux_arm",
6612 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006613)
6614
6615config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006616 name = "linux_armeabi",
6617 values = {"cpu": "armeabi"},
6618)
6619
6620config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006621 name = "linux_armhf",
6622 values = {"cpu": "armhf"},
6623)
6624
6625config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006626 name = "linux_armv7a",
6627 values = {"cpu": "armv7a"},
6628)
6629
6630config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006631 name = "linux_aarch64",
6632 values = {"cpu": "aarch64"},
6633)
6634
6635config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006636 name = "android",
6637 values = {"crosstool_top": "//external:android/crosstool"},
6638)
6639
6640config_setting(
6641 name = "android_armv7",
6642 values = {
6643 "crosstool_top": "//external:android/crosstool",
6644 "cpu": "armeabi-v7a",
6645 },
6646)
6647
6648config_setting(
6649 name = "android_arm64",
6650 values = {
6651 "crosstool_top": "//external:android/crosstool",
6652 "cpu": "arm64-v8a",
6653 },
6654)
6655
6656config_setting(
6657 name = "android_x86",
6658 values = {
6659 "crosstool_top": "//external:android/crosstool",
6660 "cpu": "x86",
6661 },
6662)
6663
6664config_setting(
6665 name = "android_x86_64",
6666 values = {
6667 "crosstool_top": "//external:android/crosstool",
6668 "cpu": "x86_64",
6669 },
6670)
6671
6672config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006673 name = "windows_x86_64",
6674 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006675)
6676
6677config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006678 name = "windows_x86_64_clang",
6679 values = {
6680 "compiler": "clang-cl",
6681 "cpu": "x64_windows",
6682 },
6683)
6684
6685config_setting(
6686 name = "windows_x86_64_mingw",
6687 values = {
6688 "compiler": "mingw-gcc",
6689 "cpu": "x64_windows",
6690 },
6691)
6692
6693config_setting(
6694 name = "windows_x86_64_msys",
6695 values = {
6696 "compiler": "msys-gcc",
6697 "cpu": "x64_windows",
6698 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006699)
6700
6701config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006702 name = "macos_x86_64",
6703 values = {
6704 "apple_platform_type": "macos",
6705 "cpu": "darwin",
6706 },
6707)
6708
6709config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006710 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006711 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006712)
6713
6714config_setting(
6715 name = "emscripten_wasm",
6716 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006717 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006718 "cpu": "wasm",
6719 },
6720)
6721
6722config_setting(
6723 name = "emscripten_wasmsimd",
6724 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006725 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006726 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006727 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006728 },
6729)
6730
6731config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006732 name = "ios_armv7",
6733 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006734 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006735 "cpu": "ios_armv7",
6736 },
6737)
6738
6739config_setting(
6740 name = "ios_arm64",
6741 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006742 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006743 "cpu": "ios_arm64",
6744 },
6745)
6746
6747config_setting(
6748 name = "ios_arm64e",
6749 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006750 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006751 "cpu": "ios_arm64e",
6752 },
6753)
6754
6755config_setting(
6756 name = "ios_x86",
6757 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006758 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006759 "cpu": "ios_i386",
6760 },
6761)
6762
6763config_setting(
6764 name = "ios_x86_64",
6765 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006766 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006767 "cpu": "ios_x86_64",
6768 },
6769)
6770
6771config_setting(
6772 name = "watchos_armv7k",
6773 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006774 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006775 "cpu": "watchos_armv7k",
6776 },
6777)
6778
6779config_setting(
6780 name = "watchos_arm64_32",
6781 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006782 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006783 "cpu": "watchos_arm64_32",
6784 },
6785)
6786
6787config_setting(
6788 name = "watchos_x86",
6789 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006790 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006791 "cpu": "watchos_i386",
6792 },
6793)
6794
6795config_setting(
6796 name = "watchos_x86_64",
6797 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006798 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006799 "cpu": "watchos_x86_64",
6800 },
6801)
6802
6803config_setting(
6804 name = "tvos_arm64",
6805 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006806 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006807 "cpu": "tvos_arm64",
6808 },
6809)
6810
6811config_setting(
6812 name = "tvos_x86_64",
6813 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006814 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006815 "cpu": "tvos_x86_64",
6816 },
6817)