blob: 113f32e19cb9d4d41eccc50fc305ce068d98ab86 [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 Dukhan470078a2020-10-23 22:36:52 -07001811 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1812 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1813 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1814 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1815 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1816 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1817 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1818 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1819 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1820 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001821 "src/f32-dwconv2d-chw/3x3s2p1-sse-1x4-acc3.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001822 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001823 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1824 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001825 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1826 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1827 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1828 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1829 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1830 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001831 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1832 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1833 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1834 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1835 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1836 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1837 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001838 "src/f32-hswish/gen/hswish-sse-x4.c",
1839 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001840 "src/f32-ibilinear/gen/sse-c4.c",
1841 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001842 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1843 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1844 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1845 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1846 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1847 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1848 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001849 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1850 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1851 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001852 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001853 "src/f32-prelu/gen/sse-2x4.c",
1854 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001855 "src/f32-relu/gen/sse-x4.c",
1856 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001857 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001858 "src/f32-spmm/gen/4x1-minmax-sse.c",
1859 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001860 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001861 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1862 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1863 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1864 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1865 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1866 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1867 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1868 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001869 "src/f32-vbinary/gen/vmax-sse-x4.c",
1870 "src/f32-vbinary/gen/vmax-sse-x8.c",
1871 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1872 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1873 "src/f32-vbinary/gen/vmin-sse-x4.c",
1874 "src/f32-vbinary/gen/vmin-sse-x8.c",
1875 "src/f32-vbinary/gen/vminc-sse-x4.c",
1876 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001877 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1878 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1879 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1880 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1881 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1882 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1883 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1884 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001885 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1886 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1887 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1888 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001889 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1890 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1891 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1892 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001893 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1894 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001895 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1896 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001897 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1898 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001899 "src/f32-vunary/gen/vabs-sse-x4.c",
1900 "src/f32-vunary/gen/vabs-sse-x8.c",
1901 "src/f32-vunary/gen/vneg-sse-x4.c",
1902 "src/f32-vunary/gen/vneg-sse-x8.c",
1903 "src/f32-vunary/gen/vsqr-sse-x4.c",
1904 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001905 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001906 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001907 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001908 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001909 "src/math/sqrt-sse-hh1mac.c",
1910 "src/math/sqrt-sse-nr1mac.c",
1911 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001912 "src/x32-fill/sse.c",
1913 "src/x32-packx/x4-sse.c",
1914 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915]
1916
1917SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001918 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001919 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08001920 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001921 "src/f32-prelu/gen/sse2-2x4.c",
1922 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001923 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001924 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001926 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1927 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001928 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001929 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1930 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001931 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001932 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1933 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001934 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001935 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1936 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1937 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1938 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1939 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1940 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001941 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001942 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001943 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001944 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001945 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1946 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001947 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1948 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001949 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1950 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001951 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1952 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1953 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1954 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1955 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1956 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1957 "src/math/exp-sse2-lut64-p2.c",
1958 "src/math/exp-sse2-p5.c",
1959 "src/math/expminus-sse2-p5.c",
1960 "src/math/roundd-sse2-cvt.c",
1961 "src/math/roundne-sse2-cvt.c",
1962 "src/math/roundu-sse2-cvt.c",
1963 "src/math/roundz-sse2-cvt.c",
1964 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1965 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1966 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1967 "src/math/sigmoid-sse2-rr2-p5-div.c",
1968 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1969 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001970 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1971 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1972 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001973 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1974 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1975 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001976 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1977 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1978 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001979 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001980 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001981 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001982 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001983 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001984 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001985 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
1986 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001987 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001988 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
1989 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001990 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001991 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
1992 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
1993 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001994 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001995 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001996 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001997 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001998 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001999 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002000 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002001 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002002 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2003 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002004 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002005 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002006 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002007 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2008 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2009 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2010 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002011 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2012 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2013 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2014 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002015 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2016 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002017 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2018 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2019 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2020 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2021 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002022 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002023 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002024 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002025 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002026 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002027 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002028 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002029 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002030 "src/x8-zip/x2-sse2.c",
2031 "src/x8-zip/x3-sse2.c",
2032 "src/x8-zip/x4-sse2.c",
2033 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002034 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002035 "src/x32-zip/x2-sse2.c",
2036 "src/x32-zip/x3-sse2.c",
2037 "src/x32-zip/x4-sse2.c",
2038 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002039]
2040
2041SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002042 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2043 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2044 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002045 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2046 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2047 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2049 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2050 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002051 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002052 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002053 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002054 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002055 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002056 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002057 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2058 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002059 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2061 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002062 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002063 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2064 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2065 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002066 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002067 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002068 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002069 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002070 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002071 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002072 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002073 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002074 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2075 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002076 "src/qs8-requantization/precise-ssse3.c",
2077 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002078 "src/qu8-requantization/precise-ssse3.c",
2079 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080]
2081
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002082SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002083 "src/f32-prelu/gen/sse41-2x4.c",
2084 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002085 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2086 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2087 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2088 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2089 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2090 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002091 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002092 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002093 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002094 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002095 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2096 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002097 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2098 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002099 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2100 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002101 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2102 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2103 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2104 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2105 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2106 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2107 "src/math/roundd-sse41.c",
2108 "src/math/roundne-sse41.c",
2109 "src/math/roundu-sse41.c",
2110 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002111 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2112 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2113 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002114 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2115 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2116 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002117 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2118 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2119 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002120 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002121 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002122 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002123 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002124 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002125 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002126 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2127 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002128 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2130 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002131 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2133 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2134 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002135 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002136 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002137 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002138 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002139 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002140 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002141 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002142 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002143 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2144 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002145 "src/qs8-requantization/fp32-sse4.c",
2146 "src/qs8-requantization/precise-sse4.c",
2147 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002148 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2149 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2150 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2151 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002152 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2153 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2154 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2155 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002156 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2157 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2158 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2159 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002160 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2161 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2162 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2163 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002164 "src/qu8-requantization/precise-sse4.c",
2165 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002166]
2167
Marat Dukhan08c4a432019-10-03 09:29:21 -07002168AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002169 "src/f32-clamp/gen/avx-x8.c",
2170 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002171 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2172 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002173 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2174 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002175 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2176 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2178 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2179 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2180 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2181 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2182 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002183 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002184 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2185 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002186 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002187 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002188 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002189 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2191 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2192 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2193 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2194 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2195 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2196 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2197 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2198 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2199 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2200 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002201 "src/f32-hswish/gen/hswish-avx-x8.c",
2202 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002203 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002204 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2205 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002206 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002207 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002208 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002209 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002210 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2211 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002212 "src/f32-prelu/gen/avx-2x8.c",
2213 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002214 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002215 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002217 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2218 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2219 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2220 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2221 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2222 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2223 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2224 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2225 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2226 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2227 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2228 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2229 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2230 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2231 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2232 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2233 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2234 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2235 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2236 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002237 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2238 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2239 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2240 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2241 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2242 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2243 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2244 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002245 "src/f32-vbinary/gen/vmax-avx-x8.c",
2246 "src/f32-vbinary/gen/vmax-avx-x16.c",
2247 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2248 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2249 "src/f32-vbinary/gen/vmin-avx-x8.c",
2250 "src/f32-vbinary/gen/vmin-avx-x16.c",
2251 "src/f32-vbinary/gen/vminc-avx-x8.c",
2252 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002253 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2254 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2255 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2256 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2257 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2258 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2259 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2260 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002261 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2262 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2263 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2264 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002265 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2266 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2267 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2268 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002269 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2270 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002271 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2272 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2274 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2275 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2276 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2277 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2278 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2279 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002280 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2281 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002282 "src/f32-vunary/gen/vabs-avx-x8.c",
2283 "src/f32-vunary/gen/vabs-avx-x16.c",
2284 "src/f32-vunary/gen/vneg-avx-x8.c",
2285 "src/f32-vunary/gen/vneg-avx-x16.c",
2286 "src/f32-vunary/gen/vsqr-avx-x8.c",
2287 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002288 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002289 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002290 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002291 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2292 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002293]
2294
Marat Dukhan1566fee2020-08-02 21:55:41 -07002295XOP_UKERNELS = [
2296 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002297 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002298 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002299 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002300 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002301 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002302 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2303 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002304 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002305 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2306 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002307 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002308 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2309 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2310 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002311 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002312 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002313 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002314 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002315 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002316 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002317 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002318 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002319 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2320 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002321 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2322 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2323 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2324 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2325 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2326 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2327 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2328 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002329]
2330
Marat Dukhanfda12b82019-11-21 12:27:59 -08002331FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002332 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2333 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002334 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2335 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002336 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2337 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002338 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2339 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2340 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2341 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2342 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2343 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002344 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002345 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2346 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2347 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2348 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002349 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002350 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2351 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002352 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2354 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002355 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2356 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2357 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2359 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2360 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2361 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2362 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2363 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2364 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2365 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2366 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2367 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2368 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2369 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2370 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2371 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002372 "src/f32-hswish/gen/hswish-fma3-x8.c",
2373 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002374 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002375 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2376 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2377 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2378 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002379 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002380 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2381 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002382 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002383 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2384 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002385 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2386 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2387 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002388 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2389 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2390 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2391 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2392 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2393 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2394 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2395 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002396 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002397 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002398 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002399]
2400
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002401AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002402 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2403 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002404 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002405 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002406 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002407 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2408 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002409 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002410 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2411 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2412 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002413 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002414 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2415 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002416 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002417 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002419 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2420 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002422 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2423 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2424 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002425 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002426 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2427 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002429 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002430 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002431 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2432 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002433 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002434 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2435 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2436 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002437 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002438 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2439 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2440 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2441 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2442 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2443 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2444 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2445 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2446 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2447 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2448 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2449 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2450 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2451 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2452 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2453 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2454 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2455 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2456 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2457 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2458 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2459 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2460 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2461 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2462 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2463 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2464 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2465 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2466 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2467 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002468 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2469 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2470 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2471 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2472 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2473 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2474 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2475 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2476 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2477 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2478 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2479 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2480 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2481 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2482 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2483 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2484 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2485 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2486 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2487 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2488 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2489 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2490 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2491 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002492 "src/math/exp-avx2-lut8-p3-perm.c",
2493 "src/math/exp-avx2-lut8-p4-perm.c",
2494 "src/math/exp-avx2-p5.c",
2495 "src/math/expminus-avx2-p5.c",
2496 "src/math/extexp-avx2-p5.c",
2497 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2498 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2499 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2500 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2501 "src/math/sigmoid-avx2-rr1-p5-div.c",
2502 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2503 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2504 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2505 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2506 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2507 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2508 "src/math/sigmoid-avx2-rr2-p5-div.c",
2509 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2510 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002511 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002512 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002513 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2514 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002516 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002517 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002518 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002519 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002520 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002522 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002523 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2524 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2525 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002526 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2527 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2528 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2529 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2530 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2531 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2532 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2533 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002534]
2535
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002537 "src/f32-clamp/gen/avx512f-x16.c",
2538 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002539 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2540 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002541 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2542 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002543 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2544 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002545 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2546 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2547 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2548 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2549 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2550 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002551 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2552 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2553 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2554 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2555 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2556 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002557 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2558 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2559 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2560 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2561 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2562 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002563 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2564 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002565 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2566 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2567 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2568 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2569 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2570 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002571 "src/f32-prelu/gen/avx512f-2x16.c",
2572 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002573 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2574 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002575 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002576 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002577 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002578 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2579 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002580 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002581 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2582 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2583 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002584 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002585 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2586 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002588 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002589 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002590 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2591 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002592 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002593 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2594 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2595 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002597 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2598 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002599 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002600 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002601 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002602 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2603 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002604 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002605 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2606 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2607 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002608 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002609 "src/f32-relu/gen/avx512f-x16.c",
2610 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002612 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2613 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2614 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2615 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2616 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2617 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2618 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2619 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2620 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2621 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2622 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2623 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2624 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2625 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2626 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2627 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002628 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2629 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2630 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2631 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2632 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2633 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2634 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2635 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2636 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2637 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2638 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2639 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2640 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2641 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2642 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2643 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002644 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2645 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2646 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2647 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2648 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2649 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2650 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2651 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2652 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2653 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2654 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2655 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2656 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2657 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2658 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2659 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002660 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2661 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2662 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2663 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2664 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2665 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2666 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2667 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002668 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2669 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2670 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2671 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2672 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2673 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2674 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2675 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002676 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2677 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2678 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2679 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2680 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2681 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2682 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2683 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002684 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2685 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2686 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2687 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002688 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2689 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2690 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2691 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002692 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2693 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002694 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2695 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2696 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2697 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2698 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2699 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2700 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2701 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002702 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002703 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2704 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2705 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2706 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2707 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2708 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2709 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2710 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2711 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2712 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2713 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2714 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2715 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2716 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2717 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2718 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2719 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2720 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2721 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2722 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2723 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2724 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2725 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2726 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002727 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2728 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2729 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2730 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2731 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2732 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2733 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2734 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002735 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2736 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2737 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2738 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2739 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2740 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002741 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2742 "src/math/exp-avx512f-lut16-p3-perm.c",
2743 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2744 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002745 "src/math/exp-avx512f-p5-scalef.c",
2746 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002747 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002748 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002749 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002750 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002751 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002752 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002753 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002754 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002755 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002756 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2757 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2758 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2759 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2760 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2761 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2762 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2763 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2764 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2765 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002766 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002767 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002768 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2769 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2770 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2771 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002772 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002773 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002774 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002775]
2776
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002777AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002778 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2779 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002780 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2781 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2782 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2783 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2784 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2785 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2786 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2787 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2788]
2789
Frank Barchardbcedc082020-08-17 18:00:51 -07002790WASM32_ASM_UKERNELS = [
2791 "src/f32-relu/wasm_shr_x1.S",
2792 "src/f32-relu/wasm_shr_x2.S",
2793 "src/f32-relu/wasm_shr_x4.S",
2794]
2795
Marat Dukhan08c4a432019-10-03 09:29:21 -07002796AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002797 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002798 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2800 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002801 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002802 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002803 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002804 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002805 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2806 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002807 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2808 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2809 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2810 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002811]
2812
2813AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002814 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002815 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002816 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002818 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002819 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002820 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002821 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2822 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2823 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2824 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2825 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2826 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2827 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002828 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2829 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002830 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2831 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2832 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002833 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2834 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002835 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2836 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2837 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2838 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002839 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002840 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2841 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002842 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2843 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2844 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2845 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002846 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002847 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002848 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002849 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002850 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2851 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2852 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2853 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2854 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2855 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2856 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2857 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2858 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2859 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2860 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2861 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2862 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2863 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2864 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2865 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2866 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2867 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2868 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2869 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2870 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2871 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002872 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002873 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002874 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2875 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002876 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2877 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2878 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2879 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2880 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2881 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002882 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2883 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2884 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2885 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002886 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2887 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002888 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002889 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002890 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891]
2892
Marat Dukhan1b354632020-03-23 12:50:22 -07002893INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002894 "src/xnnpack/argmaxpool.h",
2895 "src/xnnpack/avgpool.h",
2896 "src/xnnpack/clamp.h",
2897 "src/xnnpack/common.h",
2898 "src/xnnpack/conv.h",
2899 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002900 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 "src/xnnpack/gavgpool.h",
2902 "src/xnnpack/gemm.h",
2903 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002904 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002906 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002907 "src/xnnpack/lut.h",
2908 "src/xnnpack/math.h",
2909 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002910 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 "src/xnnpack/packx.h",
2912 "src/xnnpack/pad.h",
2913 "src/xnnpack/params.h",
2914 "src/xnnpack/pavgpool.h",
2915 "src/xnnpack/ppmm.h",
2916 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002917 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002918 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002919 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920 "src/xnnpack/rmax.h",
2921 "src/xnnpack/scalar-utils.h",
2922 "src/xnnpack/spmm.h",
2923 "src/xnnpack/unpool.h",
2924 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002925 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002927 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002928 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002929 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002930 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002932]
2933
2934INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002935 "include/xnnpack.h",
2936 "src/xnnpack/allocator.h",
2937 "src/xnnpack/compute.h",
2938 "src/xnnpack/im2col.h",
2939 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002940 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002941 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 "src/xnnpack/operator.h",
2943 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002944 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002946 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002947 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002948]
2949
Marat Dukhan1b354632020-03-23 12:50:22 -07002950ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002951 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952]
2953
Marat Dukhan1b354632020-03-23 12:50:22 -07002954MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002955 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 "include/xnnpack.h",
2957]
2958
Marat Dukhan1b354632020-03-23 12:50:22 -07002959MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002961 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 "src/xnnpack/requantization.h",
2963 "include/xnnpack.h",
2964]
2965
2966OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002968 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969]
2970
2971WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002972 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002973 "src/xnnpack/operator.h",
2974 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002975]
2976
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002977LOGGING_COPTS = select({
2978 # No logging in optimized mode
2979 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2980 # Full logging in debug mode
2981 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2982 # Error-only logging in default (fastbuild) mode
2983 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2984})
2985
Marat Dukhan3b59de22020-06-03 20:15:19 -07002986LOGGING_SRCS = select({
2987 # No logging in optimized mode
2988 ":optimized_build": [],
2989 "//conditions:default": [
2990 "src/operator-strings.c",
2991 "src/subgraph-strings.c",
2992 ],
2993})
2994
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002995LOGGING_HDRS = [
2996 "src/xnnpack/log.h",
2997]
2998
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003000 name = "tables",
3001 srcs = TABLE_SRCS,
3002 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003003 gcc_copts = xnnpack_gcc_std_copts(),
3004 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003005)
3006
3007xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 name = "scalar_ukernels",
3009 srcs = SCALAR_UKERNELS,
3010 hdrs = INTERNAL_HDRS,
3011 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003012 gcc_copts = xnnpack_gcc_std_copts(),
3013 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003015 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 "@FP16",
3017 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003018 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 ],
3020)
3021
3022xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003023 name = "scalar_ukernels_test_mode",
3024 srcs = SCALAR_UKERNELS,
3025 hdrs = INTERNAL_HDRS,
3026 aarch32_copts = ["-marm"],
3027 copts = [
3028 "-UNDEBUG",
3029 "-DXNN_TEST_MODE=1",
3030 ],
3031 gcc_copts = xnnpack_gcc_std_copts(),
3032 msvc_copts = xnnpack_msvc_std_copts(),
3033 deps = [
3034 ":tables",
3035 "@FP16",
3036 "@FXdiv",
3037 "@pthreadpool",
3038 ],
3039)
3040
3041xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003042 name = "wasm_ukernels",
3043 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003044 gcc_copts = xnnpack_gcc_std_copts(),
3045 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003046 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003047 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003048 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003049 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003050 "@FP16",
3051 "@FXdiv",
3052 "@pthreadpool",
3053 ],
3054)
3055
3056xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003057 name = "wasm_ukernels_test_mode",
3058 hdrs = INTERNAL_HDRS,
3059 copts = [
3060 "-UNDEBUG",
3061 "-DXNN_TEST_MODE=1",
3062 ],
3063 gcc_copts = xnnpack_gcc_std_copts(),
3064 msvc_copts = xnnpack_msvc_std_copts(),
3065 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003066 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003067 deps = [
3068 ":tables",
3069 "@FP16",
3070 "@FXdiv",
3071 "@pthreadpool",
3072 ],
3073)
3074
3075xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003076 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077 hdrs = INTERNAL_HDRS,
3078 aarch32_copts = [
3079 "-marm",
3080 "-mfpu=neon",
3081 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003082 gcc_copts = xnnpack_gcc_std_copts(),
3083 gcc_x86_copts = ["-msse2"],
3084 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085 optimized_copts = [
3086 "-O3",
3087 "-ffast-math",
3088 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003089 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003091 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092 "@FP16",
3093 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003094 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 ],
3096)
3097
3098xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003099 name = "psimd_fastmath_ukernels_test_mode",
3100 hdrs = INTERNAL_HDRS,
3101 aarch32_copts = [
3102 "-marm",
3103 "-mfpu=neon",
3104 ],
3105 copts = [
3106 "-UNDEBUG",
3107 "-DXNN_TEST_MODE=1",
3108 ],
3109 gcc_copts = xnnpack_gcc_std_copts(),
3110 gcc_x86_copts = ["-msse2"],
3111 msvc_copts = xnnpack_msvc_std_copts(),
3112 optimized_copts = [
3113 "-O3",
3114 "-ffast-math",
3115 ],
3116 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3117 deps = [
3118 ":tables",
3119 "@FP16",
3120 "@psimd",
3121 "@pthreadpool",
3122 ],
3123)
3124
3125xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003126 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003127 hdrs = INTERNAL_HDRS,
3128 aarch32_copts = [
3129 "-marm",
3130 "-mfpu=neon",
3131 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003132 gcc_copts = xnnpack_gcc_std_copts(),
3133 gcc_x86_copts = ["-msse2"],
3134 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003135 optimized_copts = [
3136 "-O3",
3137 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003138 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003139 deps = [
3140 ":tables",
3141 "@FP16",
3142 "@psimd",
3143 "@pthreadpool",
3144 ],
3145)
3146
3147xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003148 name = "psimd_accmath_ukernels_test_mode",
3149 hdrs = INTERNAL_HDRS,
3150 aarch32_copts = [
3151 "-marm",
3152 "-mfpu=neon",
3153 ],
3154 copts = [
3155 "-UNDEBUG",
3156 "-DXNN_TEST_MODE=1",
3157 ],
3158 gcc_copts = xnnpack_gcc_std_copts(),
3159 gcc_x86_copts = ["-msse2"],
3160 msvc_copts = xnnpack_msvc_std_copts(),
3161 optimized_copts = [
3162 "-O3",
3163 ],
3164 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3165 deps = [
3166 ":tables",
3167 "@FP16",
3168 "@psimd",
3169 "@pthreadpool",
3170 ],
3171)
3172
3173xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 name = "neon_ukernels",
3175 hdrs = INTERNAL_HDRS,
3176 aarch32_copts = [
3177 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003178 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003179 "-mfpu=neon",
3180 ],
3181 aarch32_srcs = NEON_UKERNELS,
3182 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003183 gcc_copts = xnnpack_gcc_std_copts(),
3184 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003185 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003186 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003187 "@FP16",
3188 "@pthreadpool",
3189 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190)
3191
3192xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003193 name = "neon_ukernels_test_mode",
3194 hdrs = INTERNAL_HDRS,
3195 aarch32_copts = [
3196 "-marm",
3197 "-march=armv7-a",
3198 "-mfpu=neon",
3199 ],
3200 aarch32_srcs = NEON_UKERNELS,
3201 aarch64_srcs = NEON_UKERNELS,
3202 copts = [
3203 "-UNDEBUG",
3204 "-DXNN_TEST_MODE=1",
3205 ],
3206 gcc_copts = xnnpack_gcc_std_copts(),
3207 msvc_copts = xnnpack_msvc_std_copts(),
3208 deps = [
3209 ":tables",
3210 "@FP16",
3211 "@pthreadpool",
3212 ],
3213)
3214
3215xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003216 name = "neonfma_ukernels",
3217 hdrs = INTERNAL_HDRS,
3218 aarch32_copts = [
3219 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003220 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221 "-mfpu=neon-vfpv4",
3222 ],
3223 aarch32_srcs = NEONFMA_UKERNELS,
3224 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003225 apple_aarch32_copts = [
3226 "-mcpu=swift",
3227 "-mtune=generic",
3228 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003229 gcc_copts = xnnpack_gcc_std_copts(),
3230 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003231 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003232 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003233 "@FP16",
3234 "@pthreadpool",
3235 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236)
3237
3238xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003239 name = "neonfma_ukernels_test_mode",
3240 hdrs = INTERNAL_HDRS,
3241 aarch32_copts = [
3242 "-marm",
3243 "-march=armv7-a",
3244 "-mfpu=neon-vfpv4",
3245 ],
3246 aarch32_srcs = NEONFMA_UKERNELS,
3247 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003248 apple_aarch32_copts = [
3249 "-mcpu=swift",
3250 "-mtune=generic",
3251 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003252 copts = [
3253 "-UNDEBUG",
3254 "-DXNN_TEST_MODE=1",
3255 ],
3256 gcc_copts = xnnpack_gcc_std_copts(),
3257 msvc_copts = xnnpack_msvc_std_copts(),
3258 deps = [
3259 ":tables",
3260 "@FP16",
3261 "@pthreadpool",
3262 ],
3263)
3264
3265xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003266 name = "neonv8_ukernels",
3267 hdrs = INTERNAL_HDRS,
3268 aarch32_copts = [
3269 "-marm",
3270 "-march=armv8-a",
3271 "-mfpu=neon-fp-armv8",
3272 ],
3273 aarch32_srcs = NEONV8_UKERNELS,
3274 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003275 apple_aarch32_copts = [
3276 "-mcpu=cyclone",
3277 "-mtune=generic",
3278 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003279 gcc_copts = xnnpack_gcc_std_copts(),
3280 msvc_copts = xnnpack_msvc_std_copts(),
3281 deps = [
3282 ":tables",
3283 "@FP16",
3284 "@pthreadpool",
3285 ],
3286)
3287
3288xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003289 name = "neonv8_ukernels_test_mode",
3290 hdrs = INTERNAL_HDRS,
3291 aarch32_copts = [
3292 "-marm",
3293 "-march=armv8-a",
3294 "-mfpu=neon-fp-armv8",
3295 ],
3296 aarch32_srcs = NEONV8_UKERNELS,
3297 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003298 apple_aarch32_copts = [
3299 "-mcpu=cyclone",
3300 "-mtune=generic",
3301 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003302 copts = [
3303 "-UNDEBUG",
3304 "-DXNN_TEST_MODE=1",
3305 ],
3306 gcc_copts = xnnpack_gcc_std_copts(),
3307 msvc_copts = xnnpack_msvc_std_copts(),
3308 deps = [
3309 ":tables",
3310 "@FP16",
3311 "@pthreadpool",
3312 ],
3313)
3314
3315xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003316 name = "neonfp16arith_ukernels",
3317 hdrs = INTERNAL_HDRS,
3318 aarch64_copts = ["-march=armv8.2-a+fp16"],
3319 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003320 gcc_copts = xnnpack_gcc_std_copts(),
3321 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003322 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003323 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003324 "@FP16",
3325 "@pthreadpool",
3326 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003327)
3328
3329xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003330 name = "neonfp16arith_ukernels_test_mode",
3331 hdrs = INTERNAL_HDRS,
3332 aarch64_copts = ["-march=armv8.2-a+fp16"],
3333 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3334 copts = [
3335 "-UNDEBUG",
3336 "-DXNN_TEST_MODE=1",
3337 ],
3338 gcc_copts = xnnpack_gcc_std_copts(),
3339 msvc_copts = xnnpack_msvc_std_copts(),
3340 deps = [
3341 ":tables",
3342 "@FP16",
3343 "@pthreadpool",
3344 ],
3345)
3346
3347xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003348 name = "neondot_ukernels",
3349 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003350 aarch32_copts = [
3351 "-marm",
3352 "-march=armv8.2-a+dotprod",
3353 "-mfpu=neon-fp-armv8",
3354 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003355 aarch32_srcs = NEONDOT_UKERNELS,
3356 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3357 aarch64_srcs = NEONDOT_UKERNELS,
3358 gcc_copts = xnnpack_gcc_std_copts(),
3359 msvc_copts = xnnpack_msvc_std_copts(),
3360 deps = [
3361 ":tables",
3362 "@FP16",
3363 "@pthreadpool",
3364 ],
3365)
3366
3367xnnpack_cc_library(
3368 name = "neondot_ukernels_test_mode",
3369 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003370 aarch32_copts = [
3371 "-marm",
3372 "-march=armv8.2-a+dotprod",
3373 "-mfpu=neon-fp-armv8",
3374 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003375 aarch32_srcs = NEONDOT_UKERNELS,
3376 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3377 aarch64_srcs = NEONDOT_UKERNELS,
3378 copts = [
3379 "-UNDEBUG",
3380 "-DXNN_TEST_MODE=1",
3381 ],
3382 gcc_copts = xnnpack_gcc_std_copts(),
3383 msvc_copts = xnnpack_msvc_std_copts(),
3384 deps = [
3385 ":tables",
3386 "@FP16",
3387 "@pthreadpool",
3388 ],
3389)
3390
3391xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 name = "sse2_ukernels",
3393 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003394 gcc_copts = xnnpack_gcc_std_copts(),
3395 gcc_x86_copts = ["-msse2"],
3396 msvc_copts = xnnpack_msvc_std_copts(),
3397 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003398 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003399 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003400 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003401 "@FP16",
3402 "@pthreadpool",
3403 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404)
3405
3406xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003407 name = "sse2_ukernels_test_mode",
3408 hdrs = INTERNAL_HDRS,
3409 copts = [
3410 "-UNDEBUG",
3411 "-DXNN_TEST_MODE=1",
3412 ],
3413 gcc_copts = xnnpack_gcc_std_copts(),
3414 gcc_x86_copts = ["-msse2"],
3415 msvc_copts = xnnpack_msvc_std_copts(),
3416 msvc_x86_32_copts = ["/arch:SSE2"],
3417 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3418 deps = [
3419 ":tables",
3420 "@FP16",
3421 "@pthreadpool",
3422 ],
3423)
3424
3425xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003426 name = "ssse3_ukernels",
3427 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003428 gcc_copts = xnnpack_gcc_std_copts(),
3429 gcc_x86_copts = ["-mssse3"],
3430 msvc_copts = xnnpack_msvc_std_copts(),
3431 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003432 x86_srcs = SSSE3_UKERNELS,
3433 deps = [
3434 ":tables",
3435 "@FP16",
3436 "@pthreadpool",
3437 ],
3438)
3439
3440xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003441 name = "ssse3_ukernels_test_mode",
3442 hdrs = INTERNAL_HDRS,
3443 copts = [
3444 "-UNDEBUG",
3445 "-DXNN_TEST_MODE=1",
3446 ],
3447 gcc_copts = xnnpack_gcc_std_copts(),
3448 gcc_x86_copts = ["-mssse3"],
3449 msvc_copts = xnnpack_msvc_std_copts(),
3450 msvc_x86_32_copts = ["/arch:SSE2"],
3451 x86_srcs = SSSE3_UKERNELS,
3452 deps = [
3453 ":tables",
3454 "@FP16",
3455 "@pthreadpool",
3456 ],
3457)
3458
3459xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003460 name = "sse41_ukernels",
3461 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003462 gcc_copts = xnnpack_gcc_std_copts(),
3463 gcc_x86_copts = ["-msse4.1"],
3464 msvc_copts = xnnpack_msvc_std_copts(),
3465 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003466 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003467 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003468 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003469 "@FP16",
3470 "@pthreadpool",
3471 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003472)
3473
3474xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003475 name = "sse41_ukernels_test_mode",
3476 hdrs = INTERNAL_HDRS,
3477 copts = [
3478 "-UNDEBUG",
3479 "-DXNN_TEST_MODE=1",
3480 ],
3481 gcc_copts = xnnpack_gcc_std_copts(),
3482 gcc_x86_copts = ["-msse4.1"],
3483 msvc_copts = xnnpack_msvc_std_copts(),
3484 msvc_x86_32_copts = ["/arch:SSE2"],
3485 x86_srcs = SSE41_UKERNELS,
3486 deps = [
3487 ":tables",
3488 "@FP16",
3489 "@pthreadpool",
3490 ],
3491)
3492
3493xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003494 name = "avx_ukernels",
3495 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003496 gcc_copts = xnnpack_gcc_std_copts(),
3497 gcc_x86_copts = ["-mavx"],
3498 msvc_copts = xnnpack_msvc_std_copts(),
3499 msvc_x86_32_copts = ["/arch:AVX"],
3500 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003502 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003503 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003504 "@FP16",
3505 "@pthreadpool",
3506 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003507)
3508
3509xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003510 name = "avx_ukernels_test_mode",
3511 hdrs = INTERNAL_HDRS,
3512 copts = [
3513 "-UNDEBUG",
3514 "-DXNN_TEST_MODE=1",
3515 ],
3516 gcc_copts = xnnpack_gcc_std_copts(),
3517 gcc_x86_copts = ["-mavx"],
3518 msvc_copts = xnnpack_msvc_std_copts(),
3519 msvc_x86_32_copts = ["/arch:AVX"],
3520 msvc_x86_64_copts = ["/arch:AVX"],
3521 x86_srcs = AVX_UKERNELS,
3522 deps = [
3523 ":tables",
3524 "@FP16",
3525 "@pthreadpool",
3526 ],
3527)
3528
3529xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003530 name = "xop_ukernels",
3531 hdrs = INTERNAL_HDRS,
3532 gcc_copts = xnnpack_gcc_std_copts(),
3533 gcc_x86_copts = ["-mxop"],
3534 msvc_copts = xnnpack_msvc_std_copts(),
3535 msvc_x86_32_copts = ["/arch:AVX"],
3536 msvc_x86_64_copts = ["/arch:AVX"],
3537 x86_srcs = XOP_UKERNELS,
3538 deps = [
3539 ":tables",
3540 "@FP16",
3541 "@pthreadpool",
3542 ],
3543)
3544
3545xnnpack_cc_library(
3546 name = "xop_ukernels_test_mode",
3547 hdrs = INTERNAL_HDRS,
3548 copts = [
3549 "-UNDEBUG",
3550 "-DXNN_TEST_MODE=1",
3551 ],
3552 gcc_copts = xnnpack_gcc_std_copts(),
3553 gcc_x86_copts = ["-mxop"],
3554 msvc_copts = xnnpack_msvc_std_copts(),
3555 msvc_x86_32_copts = ["/arch:AVX"],
3556 msvc_x86_64_copts = ["/arch:AVX"],
3557 x86_srcs = XOP_UKERNELS,
3558 deps = [
3559 ":tables",
3560 "@FP16",
3561 "@pthreadpool",
3562 ],
3563)
3564
3565xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003566 name = "fma3_ukernels",
3567 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003568 gcc_copts = xnnpack_gcc_std_copts(),
3569 gcc_x86_copts = ["-mfma"],
3570 msvc_copts = xnnpack_msvc_std_copts(),
3571 msvc_x86_32_copts = ["/arch:AVX"],
3572 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003573 x86_srcs = FMA3_UKERNELS,
3574 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003575 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003576 "@FP16",
3577 "@pthreadpool",
3578 ],
3579)
3580
3581xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003582 name = "fma3_ukernels_test_mode",
3583 hdrs = INTERNAL_HDRS,
3584 copts = [
3585 "-UNDEBUG",
3586 "-DXNN_TEST_MODE=1",
3587 ],
3588 gcc_copts = xnnpack_gcc_std_copts(),
3589 gcc_x86_copts = ["-mfma"],
3590 msvc_copts = xnnpack_msvc_std_copts(),
3591 msvc_x86_32_copts = ["/arch:AVX"],
3592 msvc_x86_64_copts = ["/arch:AVX"],
3593 x86_srcs = FMA3_UKERNELS,
3594 deps = [
3595 ":tables",
3596 "@FP16",
3597 "@pthreadpool",
3598 ],
3599)
3600
3601xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003602 name = "avx2_ukernels",
3603 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003604 gcc_copts = xnnpack_gcc_std_copts(),
3605 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003606 "-mfma",
3607 "-mavx2",
3608 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003609 msvc_copts = xnnpack_msvc_std_copts(),
3610 msvc_x86_32_copts = ["/arch:AVX2"],
3611 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003612 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003613 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003614 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003615 "@FP16",
3616 "@pthreadpool",
3617 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003618)
3619
3620xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003621 name = "avx2_ukernels_test_mode",
3622 hdrs = INTERNAL_HDRS,
3623 copts = [
3624 "-UNDEBUG",
3625 "-DXNN_TEST_MODE=1",
3626 ],
3627 gcc_copts = xnnpack_gcc_std_copts(),
3628 gcc_x86_copts = [
3629 "-mfma",
3630 "-mavx2",
3631 ],
3632 msvc_copts = xnnpack_msvc_std_copts(),
3633 msvc_x86_32_copts = ["/arch:AVX2"],
3634 msvc_x86_64_copts = ["/arch:AVX2"],
3635 x86_srcs = AVX2_UKERNELS,
3636 deps = [
3637 ":tables",
3638 "@FP16",
3639 "@pthreadpool",
3640 ],
3641)
3642
3643xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 name = "avx512f_ukernels",
3645 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003646 gcc_copts = xnnpack_gcc_std_copts(),
3647 gcc_x86_copts = ["-mavx512f"],
3648 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3649 msvc_copts = xnnpack_msvc_std_copts(),
3650 msvc_x86_32_copts = ["/arch:AVX512"],
3651 msvc_x86_64_copts = ["/arch:AVX512"],
3652 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003654 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003655 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003656 "@FP16",
3657 "@pthreadpool",
3658 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659)
3660
3661xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003662 name = "avx512f_ukernels_test_mode",
3663 hdrs = INTERNAL_HDRS,
3664 copts = [
3665 "-UNDEBUG",
3666 "-DXNN_TEST_MODE=1",
3667 ],
3668 gcc_copts = xnnpack_gcc_std_copts(),
3669 gcc_x86_copts = ["-mavx512f"],
3670 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3671 msvc_copts = xnnpack_msvc_std_copts(),
3672 msvc_x86_32_copts = ["/arch:AVX512"],
3673 msvc_x86_64_copts = ["/arch:AVX512"],
3674 msys_copts = ["-fno-asynchronous-unwind-tables"],
3675 x86_srcs = AVX512F_UKERNELS,
3676 deps = [
3677 ":tables",
3678 "@FP16",
3679 "@pthreadpool",
3680 ],
3681)
3682
3683xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003684 name = "avx512skx_ukernels",
3685 hdrs = INTERNAL_HDRS,
3686 gcc_copts = xnnpack_gcc_std_copts(),
3687 gcc_x86_copts = [
3688 "-mavx512f",
3689 "-mavx512cd",
3690 "-mavx512bw",
3691 "-mavx512dq",
3692 "-mavx512vl",
3693 ],
3694 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3695 msvc_copts = xnnpack_msvc_std_copts(),
3696 msvc_x86_32_copts = ["/arch:AVX512"],
3697 msvc_x86_64_copts = ["/arch:AVX512"],
3698 msys_copts = ["-fno-asynchronous-unwind-tables"],
3699 x86_srcs = AVX512SKX_UKERNELS,
3700 deps = [
3701 ":tables",
3702 "@FP16",
3703 "@pthreadpool",
3704 ],
3705)
3706
3707xnnpack_cc_library(
3708 name = "avx512skx_ukernels_test_mode",
3709 hdrs = INTERNAL_HDRS,
3710 copts = [
3711 "-UNDEBUG",
3712 "-DXNN_TEST_MODE=1",
3713 ],
3714 gcc_copts = xnnpack_gcc_std_copts(),
3715 gcc_x86_copts = [
3716 "-mavx512f",
3717 "-mavx512cd",
3718 "-mavx512bw",
3719 "-mavx512dq",
3720 "-mavx512vl",
3721 ],
3722 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3723 msvc_copts = xnnpack_msvc_std_copts(),
3724 msvc_x86_32_copts = ["/arch:AVX512"],
3725 msvc_x86_64_copts = ["/arch:AVX512"],
3726 msys_copts = ["-fno-asynchronous-unwind-tables"],
3727 x86_srcs = AVX512SKX_UKERNELS,
3728 deps = [
3729 ":tables",
3730 "@FP16",
3731 "@pthreadpool",
3732 ],
3733)
3734
3735xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736 name = "asm_ukernels",
3737 hdrs = ["src/xnnpack/assembly.h"],
3738 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003739 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003741 wasm_srcs = WASM32_ASM_UKERNELS,
3742 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743)
3744
Marat Dukhan3b59de22020-06-03 20:15:19 -07003745xnnpack_cc_library(
3746 name = "logging_utils",
3747 srcs = LOGGING_SRCS,
3748 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3749 copts = LOGGING_COPTS + [
3750 "-Isrc",
3751 "-Iinclude",
3752 ] + select({
3753 ":debug_build": [],
3754 "//conditions:default": xnnpack_min_size_copts(),
3755 }),
3756 gcc_copts = xnnpack_gcc_std_copts(),
3757 msvc_copts = xnnpack_msvc_std_copts(),
3758 visibility = xnnpack_visibility(),
3759 deps = [
3760 "@FP16",
3761 "@clog",
3762 "@pthreadpool",
3763 ],
3764)
3765
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766xnnpack_aggregate_library(
3767 name = "ukernels",
3768 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 ":neon_ukernels",
3770 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003771 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003772 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 ":asm_ukernels",
3774 ],
3775 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 ":neon_ukernels",
3777 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003778 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003780 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781 ":asm_ukernels",
3782 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003783 generic_deps = [
3784 ":scalar_ukernels",
3785 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003786 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003787 ":psimd_fastmath_ukernels",
3788 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003790 wasm_deps = [
3791 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003792 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003793 ],
3794 wasmsimd_deps = [
3795 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003796 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003797 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003800 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003801 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003803 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003804 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003805 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003806 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003807 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 ],
3809)
3810
Marat Dukhan33fcf782020-05-24 14:27:15 -07003811xnnpack_aggregate_library(
3812 name = "ukernels_test_mode",
3813 aarch32_deps = [
3814 ":neon_ukernels_test_mode",
3815 ":neonfma_ukernels_test_mode",
3816 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003817 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003818 ":asm_ukernels",
3819 ],
3820 aarch64_deps = [
3821 ":neon_ukernels_test_mode",
3822 ":neonfma_ukernels_test_mode",
3823 ":neonv8_ukernels_test_mode",
3824 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003825 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003826 ":asm_ukernels",
3827 ],
3828 generic_deps = [
3829 ":scalar_ukernels_test_mode",
3830 ],
3831 psimd_deps = [
3832 ":psimd_fastmath_ukernels_test_mode",
3833 ":psimd_accmath_ukernels_test_mode",
3834 ],
3835 wasm_deps = [
3836 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003837 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003838 ],
3839 wasmsimd_deps = [
3840 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003841 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003842 ],
3843 x86_deps = [
3844 ":sse2_ukernels_test_mode",
3845 ":ssse3_ukernels_test_mode",
3846 ":sse41_ukernels_test_mode",
3847 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003848 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003849 ":fma3_ukernels_test_mode",
3850 ":avx2_ukernels_test_mode",
3851 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003852 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003853 ],
3854)
3855
Marat Dukhan86df3842020-06-17 14:03:31 -07003856# This target must be compatible with Android NDK r17 and compile with Clang 6.
3857# It is similar to :ukernels target, but excludes microkernels which require
3858# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003859xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003860 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003861 aarch32_deps = [
3862 ":neon_ukernels",
3863 ":neonfma_ukernels",
3864 ":neonv8_ukernels",
3865 ":asm_ukernels",
3866 ],
3867 aarch64_deps = [
3868 ":neon_ukernels",
3869 ":neonfma_ukernels",
3870 ":neonv8_ukernels",
3871 ":asm_ukernels",
3872 ],
3873 generic_deps = [
3874 ":scalar_ukernels",
3875 ],
3876 psimd_deps = [
3877 ":psimd_fastmath_ukernels",
3878 ":psimd_accmath_ukernels",
3879 ],
3880 wasm_deps = [
3881 ":wasm_ukernels",
3882 ],
3883 wasmsimd_deps = [
3884 ":wasm_ukernels",
3885 ],
3886 x86_deps = [
3887 ":sse2_ukernels",
3888 ":ssse3_ukernels",
3889 ":sse41_ukernels",
3890 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003891 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003892 ":fma3_ukernels",
3893 ":avx2_ukernels",
3894 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003895 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003896 ],
3897)
3898
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899xnnpack_cc_library(
3900 name = "im2col",
3901 srcs = ["src/im2col.c"],
3902 hdrs = [
3903 "src/xnnpack/common.h",
3904 "src/xnnpack/im2col.h",
3905 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003906 gcc_copts = xnnpack_gcc_std_copts(),
3907 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908)
3909
3910xnnpack_cc_library(
3911 name = "indirection",
3912 srcs = ["src/indirection.c"],
3913 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003914 gcc_copts = xnnpack_gcc_std_copts(),
3915 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003916 deps = [
3917 "@FP16",
3918 "@FXdiv",
3919 "@pthreadpool",
3920 ],
3921)
3922
3923xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003924 name = "indirection_test_mode",
3925 srcs = ["src/indirection.c"],
3926 hdrs = INTERNAL_HDRS,
3927 copts = [
3928 "-UNDEBUG",
3929 "-DXNN_TEST_MODE=1",
3930 ],
3931 gcc_copts = xnnpack_gcc_std_copts(),
3932 msvc_copts = xnnpack_msvc_std_copts(),
3933 deps = [
3934 "@FP16",
3935 "@FXdiv",
3936 "@pthreadpool",
3937 ],
3938)
3939
3940xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003941 name = "packing",
3942 srcs = ["src/packing.c"],
3943 hdrs = INTERNAL_HDRS,
3944 gcc_copts = xnnpack_gcc_std_copts(),
3945 msvc_copts = xnnpack_msvc_std_copts(),
3946 deps = [
3947 "@FP16",
3948 "@FXdiv",
3949 "@pthreadpool",
3950 ],
3951)
3952
3953xnnpack_cc_library(
3954 name = "packing_test_mode",
3955 srcs = ["src/packing.c"],
3956 hdrs = INTERNAL_HDRS,
3957 copts = [
3958 "-UNDEBUG",
3959 "-DXNN_TEST_MODE=1",
3960 ],
3961 gcc_copts = xnnpack_gcc_std_copts(),
3962 msvc_copts = xnnpack_msvc_std_copts(),
3963 deps = [
3964 "@FP16",
3965 "@FXdiv",
3966 "@pthreadpool",
3967 ],
3968)
3969
3970xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 name = "operator_run",
3972 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003973 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003974 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003975 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3976 "//conditions:default": [],
3977 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003978 gcc_copts = xnnpack_gcc_std_copts(),
3979 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003980 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003981 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003982 "@FP16",
3983 "@FXdiv",
3984 "@clog",
3985 "@pthreadpool",
3986 ],
3987)
3988
Chao Mei6ddfc602020-05-13 22:29:36 -07003989xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003990 name = "operator_run_test_mode",
3991 srcs = ["src/operator-run.c"],
3992 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3993 copts = LOGGING_COPTS + [
3994 "-UNDEBUG",
3995 "-DXNN_TEST_MODE=1",
3996 ] + select({
3997 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3998 "//conditions:default": [],
3999 }),
4000 gcc_copts = xnnpack_gcc_std_copts(),
4001 msvc_copts = xnnpack_msvc_std_copts(),
4002 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004003 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004004 "@FP16",
4005 "@FXdiv",
4006 "@clog",
4007 "@pthreadpool",
4008 ],
4009)
4010
4011xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004012 name = "memory_planner",
4013 srcs = ["src/memory-planner.c"],
4014 hdrs = INTERNAL_HDRS,
4015 defines = select({
4016 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4017 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4018 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4019 }),
4020 gcc_copts = xnnpack_gcc_std_copts(),
4021 msvc_copts = xnnpack_msvc_std_copts(),
4022 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004023 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004024 "@pthreadpool",
4025 ],
4026)
4027
Marat Dukhan33fcf782020-05-24 14:27:15 -07004028xnnpack_cc_library(
4029 name = "memory_planner_test_mode",
4030 srcs = ["src/memory-planner.c"],
4031 hdrs = INTERNAL_HDRS,
4032 copts = [
4033 "-UNDEBUG",
4034 "-DXNN_TEST_MODE=1",
4035 ],
4036 defines = select({
4037 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4038 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4039 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4040 }),
4041 gcc_copts = xnnpack_gcc_std_copts(),
4042 msvc_copts = xnnpack_msvc_std_copts(),
4043 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004044 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004045 "@pthreadpool",
4046 ],
4047)
4048
Marat Dukhan08c4a432019-10-03 09:29:21 -07004049cc_library(
4050 name = "enable_assembly",
4051 defines = select({
4052 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4053 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004054 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004055 }),
4056)
4057
Marat Dukhan9de90e02020-06-18 16:04:12 -07004058cc_library(
4059 name = "enable_sparse",
4060 defines = select({
4061 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4062 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004063 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004064 }),
4065)
4066
Marat Dukhancf056b22019-10-07 10:26:29 -07004067xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004068 name = "operators",
4069 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004070 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004071 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004072 ],
4073 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004074 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004075 "-Isrc",
4076 "-Iinclude",
4077 ] + select({
4078 ":debug_build": [],
4079 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004080 }) + select({
4081 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4082 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004083 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004084 gcc_copts = xnnpack_gcc_std_copts(),
4085 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004086 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004087 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004088 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004089 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 "@FP16",
4091 "@FXdiv",
4092 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004094 ],
4095)
4096
Marat Dukhan10a38082020-04-17 03:58:35 -07004097xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004098 name = "operators_test_mode",
4099 srcs = OPERATOR_SRCS + [
4100 "src/memory.c",
4101 "src/operator-delete.c",
4102 ],
4103 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4104 copts = LOGGING_COPTS + [
4105 "-Isrc",
4106 "-Iinclude",
4107 "-UNDEBUG",
4108 "-DXNN_TEST_MODE=1",
4109 ] + select({
4110 ":debug_build": [],
4111 "//conditions:default": xnnpack_min_size_copts(),
4112 }) + select({
4113 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4114 "//conditions:default": [],
4115 }),
4116 gcc_copts = xnnpack_gcc_std_copts(),
4117 msvc_copts = xnnpack_msvc_std_copts(),
4118 deps = [
4119 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004120 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004121 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004122 "@FP16",
4123 "@FXdiv",
4124 "@clog",
4125 "@pthreadpool",
4126 ],
4127)
4128
4129xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004130 name = "XNNPACK",
4131 srcs = [
4132 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004133 "src/runtime.c",
4134 "src/subgraph.c",
4135 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004136 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004137 hdrs = ["include/xnnpack.h"],
4138 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004139 "-Isrc",
4140 "-Iinclude",
4141 ] + select({
4142 ":debug_build": [],
4143 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004144 }) + select({
4145 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4146 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004147 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004148 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004149 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004150 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004151 visibility = xnnpack_visibility(),
4152 deps = [
4153 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004154 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004155 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004156 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004157 ":operator_run",
4158 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004159 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004160 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004161 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004162 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004163 ] + select({
4164 ":emscripten": [],
4165 "//conditions:default": ["@cpuinfo"],
4166 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004167)
4168
Marat Dukhan10a38082020-04-17 03:58:35 -07004169xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004170 name = "XNNPACK_test_mode",
4171 srcs = [
4172 "src/init.c",
4173 "src/runtime.c",
4174 "src/subgraph.c",
4175 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004176 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004177 hdrs = ["include/xnnpack.h"],
4178 copts = LOGGING_COPTS + [
4179 "-Isrc",
4180 "-Iinclude",
4181 "-UNDEBUG",
4182 "-DXNN_TEST_MODE=1",
4183 ] + select({
4184 ":debug_build": [],
4185 "//conditions:default": xnnpack_min_size_copts(),
4186 }) + select({
4187 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4188 "//conditions:default": [],
4189 }),
4190 gcc_copts = xnnpack_gcc_std_copts(),
4191 includes = ["include"],
4192 msvc_copts = xnnpack_msvc_std_copts(),
4193 visibility = xnnpack_visibility(),
4194 deps = [
4195 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004196 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004197 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004198 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004199 ":operator_run_test_mode",
4200 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004201 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004202 "@clog",
4203 "@FP16",
4204 "@pthreadpool",
4205 ] + select({
4206 ":emscripten": [],
4207 "//conditions:default": ["@cpuinfo"],
4208 }),
4209)
4210
4211xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004212 name = "xnnpack_f32",
4213 srcs = [
4214 "src/init.c",
4215 "src/runtime.c",
4216 "src/subgraph.c",
4217 "src/tensor.c",
4218 ] + SUBGRAPH_SRCS,
4219 hdrs = ["include/xnnpack.h"],
4220 copts = LOGGING_COPTS + [
4221 "-Isrc",
4222 "-Iinclude",
4223 ] + select({
4224 ":debug_build": [],
4225 "//conditions:default": xnnpack_min_size_copts(),
4226 }) + select({
4227 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4228 "//conditions:default": [],
4229 }),
4230 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004231 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004232 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004233 "XNN_NO_U8_OPERATORS",
4234 "XNN_NO_X8_OPERATORS",
4235 "XNN_NO_F16_OPERATORS",
4236 "XNN_NO_X16_OPERATORS",
4237 ],
4238 gcc_copts = xnnpack_gcc_std_copts(),
4239 includes = ["include"],
4240 msvc_copts = xnnpack_msvc_std_copts(),
4241 visibility = xnnpack_visibility(),
4242 deps = [
4243 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004244 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004245 ":logging_utils",
4246 ":memory_planner",
4247 ":operator_run",
4248 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004249 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004250 "@clog",
4251 "@FP16",
4252 "@pthreadpool",
4253 ] + select({
4254 ":emscripten": [],
4255 "//conditions:default": ["@cpuinfo"],
4256 }),
4257)
4258
4259xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004260 name = "xnnpack_operators_nhwc_f32",
4261 srcs = [
4262 "src/init.c",
4263 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004264 hdrs = ["include/xnnpack.h"],
4265 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004266 "-Isrc",
4267 "-Iinclude",
4268 ] + select({
4269 ":debug_build": [],
4270 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004271 }) + select({
4272 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4273 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004274 }),
4275 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004276 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004277 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004278 "XNN_NO_U8_OPERATORS",
4279 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004280 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004281 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004282 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004284 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 visibility = xnnpack_visibility(),
4286 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004287 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004288 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289 ":operator_run",
4290 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004291 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004292 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004294 ] + select({
4295 ":emscripten": [],
4296 "//conditions:default": ["@cpuinfo"],
4297 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298)
4299
Marat Dukhancf056b22019-10-07 10:26:29 -07004300xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301 name = "bench_utils",
4302 srcs = ["bench/utils.cc"],
4303 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004304 deps = [
4305 "@com_google_benchmark//:benchmark",
4306 "@cpuinfo",
4307 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308)
4309
Frank Barchard7e955972019-10-11 10:34:25 -07004310######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004311
4312xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004313 name = "qs8_gemm_bench",
4314 srcs = [
4315 "bench/gemm.h",
4316 "bench/qs8-gemm.cc",
4317 "src/xnnpack/AlignedAllocator.h",
4318 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004319 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4320 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004321)
4322
4323xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004324 name = "qs8_requantization_bench",
4325 srcs = [
4326 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004327 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004328 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004329 ] + MICROKERNEL_BENCHMARK_HDRS,
4330 deps = MICROKERNEL_BENCHMARK_DEPS,
4331)
4332
4333xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004334 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 srcs = [
4336 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004337 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338 "src/xnnpack/AlignedAllocator.h",
4339 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004340 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004341 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342)
4343
4344xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004345 name = "qu8_requantization_bench",
4346 srcs = [
4347 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004348 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004349 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004350 ] + MICROKERNEL_BENCHMARK_HDRS,
4351 deps = MICROKERNEL_BENCHMARK_DEPS,
4352)
4353
4354xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004355 name = "f16_igemm_bench",
4356 srcs = [
4357 "bench/f16-igemm.cc",
4358 "bench/conv.h",
4359 "bench/google/conv.h",
4360 "src/xnnpack/AlignedAllocator.h",
4361 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004362 deps = MICROKERNEL_BENCHMARK_DEPS + [
4363 ":indirection",
4364 ":packing",
4365 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004366)
4367
4368xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369 name = "f16_gemm_bench",
4370 srcs = [
4371 "bench/f16-gemm.cc",
4372 "bench/gemm.h",
4373 "src/xnnpack/AlignedAllocator.h",
4374 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004375 deps = MICROKERNEL_BENCHMARK_DEPS + [
4376 ":packing",
4377 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378)
4379
4380xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004381 name = "f16_spmm_bench",
4382 srcs = [
4383 "bench/f16-spmm.cc",
4384 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004385 "src/xnnpack/AlignedAllocator.h",
4386 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004387 deps = MICROKERNEL_BENCHMARK_DEPS,
4388)
4389
4390xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 name = "f32_igemm_bench",
4392 srcs = [
4393 "bench/f32-igemm.cc",
4394 "bench/conv.h",
4395 "src/xnnpack/AlignedAllocator.h",
4396 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004397 deps = MICROKERNEL_BENCHMARK_DEPS + [
4398 ":indirection",
4399 ":packing",
4400 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401)
4402
4403xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004404 name = "f16_relu_bench",
4405 srcs = [
4406 "bench/f16-relu.cc",
4407 "src/xnnpack/AlignedAllocator.h",
4408 ] + MICROKERNEL_BENCHMARK_HDRS,
4409 deps = MICROKERNEL_BENCHMARK_DEPS,
4410)
4411
4412xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 name = "f32_conv_hwc_bench",
4414 srcs = [
4415 "bench/f32-conv-hwc.cc",
4416 "bench/dconv.h",
4417 "src/xnnpack/AlignedAllocator.h",
4418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004419 deps = MICROKERNEL_BENCHMARK_DEPS + [
4420 ":packing",
4421 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004422)
4423
4424xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004425 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004426 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004427 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004428 "bench/dconv.h",
4429 "src/xnnpack/AlignedAllocator.h",
4430 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004431 deps = MICROKERNEL_BENCHMARK_DEPS + [
4432 ":packing",
4433 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004434)
4435
4436xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004437 name = "f16_dwconv_bench",
4438 srcs = [
4439 "bench/f16-dwconv.cc",
4440 "bench/dwconv.h",
4441 "bench/google/dwconv.h",
4442 "src/xnnpack/AlignedAllocator.h",
4443 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004444 deps = MICROKERNEL_BENCHMARK_DEPS + [
4445 ":indirection",
4446 ":packing",
4447 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004448)
4449
4450xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 name = "f32_dwconv_bench",
4452 srcs = [
4453 "bench/f32-dwconv.cc",
4454 "bench/dwconv.h",
4455 "src/xnnpack/AlignedAllocator.h",
4456 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004457 deps = MICROKERNEL_BENCHMARK_DEPS + [
4458 ":indirection",
4459 ":packing",
4460 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461)
4462
4463xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004464 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004466 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 "bench/dwconv.h",
4468 "src/xnnpack/AlignedAllocator.h",
4469 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004470 deps = MICROKERNEL_BENCHMARK_DEPS + [
4471 ":indirection",
4472 ":packing",
4473 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474)
4475
4476xnnpack_benchmark(
4477 name = "f32_gemm_bench",
4478 srcs = [
4479 "bench/f32-gemm.cc",
4480 "bench/gemm.h",
4481 "src/xnnpack/AlignedAllocator.h",
4482 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004483 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004484 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485)
4486
4487xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004488 name = "f32_hswish_bench",
4489 srcs = [
4490 "bench/f32-hswish.cc",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + MICROKERNEL_BENCHMARK_HDRS,
4493 deps = MICROKERNEL_BENCHMARK_DEPS,
4494)
4495
4496xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004497 name = "f32_raddexpminusmax_bench",
4498 srcs = [
4499 "bench/f32-raddexpminusmax.cc",
4500 "src/xnnpack/AlignedAllocator.h",
4501 ] + MICROKERNEL_BENCHMARK_HDRS,
4502 deps = MICROKERNEL_BENCHMARK_DEPS,
4503)
4504
4505xnnpack_benchmark(
4506 name = "f32_raddextexp_bench",
4507 srcs = [
4508 "bench/f32-raddextexp.cc",
4509 "src/xnnpack/AlignedAllocator.h",
4510 ] + MICROKERNEL_BENCHMARK_HDRS,
4511 deps = MICROKERNEL_BENCHMARK_DEPS,
4512)
4513
4514xnnpack_benchmark(
4515 name = "f32_raddstoreexpminusmax_bench",
4516 srcs = [
4517 "bench/f32-raddstoreexpminusmax.cc",
4518 "src/xnnpack/AlignedAllocator.h",
4519 ] + MICROKERNEL_BENCHMARK_HDRS,
4520 deps = MICROKERNEL_BENCHMARK_DEPS,
4521)
4522
4523xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004524 name = "f32_relu_bench",
4525 srcs = [
4526 "bench/f32-relu.cc",
4527 "src/xnnpack/AlignedAllocator.h",
4528 ] + MICROKERNEL_BENCHMARK_HDRS,
4529 deps = MICROKERNEL_BENCHMARK_DEPS,
4530)
4531
4532xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004533 name = "f32_rmax_bench",
4534 srcs = [
4535 "bench/f32-rmax.cc",
4536 "src/xnnpack/AlignedAllocator.h",
4537 ] + MICROKERNEL_BENCHMARK_HDRS,
4538 deps = MICROKERNEL_BENCHMARK_DEPS,
4539)
4540
4541xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004542 name = "f32_sigmoid_bench",
4543 srcs = [
4544 "bench/f32-sigmoid.cc",
4545 "src/xnnpack/AlignedAllocator.h",
4546 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004547 deps = MICROKERNEL_BENCHMARK_DEPS,
4548)
4549
4550xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 name = "f32_spmm_bench",
4552 srcs = [
4553 "bench/f32-spmm.cc",
4554 "bench/gemm.h",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004557 deps = MICROKERNEL_BENCHMARK_DEPS,
4558)
4559
4560xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004561 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004562 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004563 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004564 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004565 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004566 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004567)
4568
4569xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004570 name = "f32_vscaleexpminusmax_bench",
4571 srcs = [
4572 "bench/f32-vscaleexpminusmax.cc",
4573 "src/xnnpack/AlignedAllocator.h",
4574 ] + MICROKERNEL_BENCHMARK_HDRS,
4575 deps = MICROKERNEL_BENCHMARK_DEPS,
4576)
4577
4578xnnpack_benchmark(
4579 name = "f32_vscaleextexp_bench",
4580 srcs = [
4581 "bench/f32-vscaleextexp.cc",
4582 "src/xnnpack/AlignedAllocator.h",
4583 ] + MICROKERNEL_BENCHMARK_HDRS,
4584 deps = MICROKERNEL_BENCHMARK_DEPS,
4585)
4586
4587xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004588 name = "f32_vsqrt_bench",
4589 srcs = [
4590 "bench/f32-vsqrt.cc",
4591 "src/xnnpack/AlignedAllocator.h",
4592 ] + MICROKERNEL_BENCHMARK_HDRS,
4593 deps = MICROKERNEL_BENCHMARK_DEPS,
4594)
4595
4596xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004597 name = "f32_im2col_gemm_bench",
4598 srcs = [
4599 "bench/f32-im2col-gemm.cc",
4600 "bench/conv.h",
4601 "src/xnnpack/AlignedAllocator.h",
4602 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004603 deps = MICROKERNEL_BENCHMARK_DEPS + [
4604 ":im2col",
4605 ":packing",
4606 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004607)
4608
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004609xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004610 name = "rounding_bench",
4611 srcs = [
4612 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004613 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004614 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004615 ] + MICROKERNEL_BENCHMARK_HDRS,
4616 deps = MICROKERNEL_BENCHMARK_DEPS,
4617)
4618
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619########################### Benchmarks for operators ###########################
4620
4621xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004622 name = "average_pooling_bench",
4623 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004624 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004625 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004626 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627)
4628
4629xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004630 name = "bankers_rounding_bench",
4631 srcs = ["bench/bankers-rounding.cc"],
4632 copts = xnnpack_optional_tflite_copts(),
4633 tags = ["nowin32"],
4634 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4635)
4636
4637xnnpack_benchmark(
4638 name = "ceiling_bench",
4639 srcs = ["bench/ceiling.cc"],
4640 copts = xnnpack_optional_tflite_copts(),
4641 tags = ["nowin32"],
4642 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4643)
4644
4645xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646 name = "channel_shuffle_bench",
4647 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004648 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004649)
4650
4651xnnpack_benchmark(
4652 name = "convolution_bench",
4653 srcs = ["bench/convolution.cc"],
4654 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004655 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004656 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657)
4658
4659xnnpack_benchmark(
4660 name = "deconvolution_bench",
4661 srcs = ["bench/deconvolution.cc"],
4662 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004663 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004664 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665)
4666
4667xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004668 name = "floor_bench",
4669 srcs = ["bench/floor.cc"],
4670 copts = xnnpack_optional_tflite_copts(),
4671 tags = ["nowin32"],
4672 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4673)
4674
4675xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004676 name = "global_average_pooling_bench",
4677 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004678 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679)
4680
4681xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004682 name = "hardswish_bench",
4683 srcs = ["bench/hardswish.cc"],
4684 copts = xnnpack_optional_tflite_copts(),
4685 tags = ["nowin32"],
4686 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4687)
4688
4689xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004690 name = "max_pooling_bench",
4691 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004692 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004693)
4694
4695xnnpack_benchmark(
4696 name = "sigmoid_bench",
4697 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004698 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004699 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004700 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004701)
4702
4703xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004704 name = "prelu_bench",
4705 srcs = ["bench/prelu.cc"],
4706 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004707 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004708 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004709)
4710
4711xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004712 name = "softmax_bench",
4713 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004714 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004715 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004716 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717)
4718
Marat Dukhan87727142020-06-24 15:24:10 -07004719xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004720 name = "square_root_bench",
4721 srcs = ["bench/square-root.cc"],
4722 copts = xnnpack_optional_tflite_copts(),
4723 tags = ["nowin32"],
4724 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4725)
4726
4727xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004728 name = "truncation_bench",
4729 srcs = ["bench/truncation.cc"],
4730 deps = OPERATOR_BENCHMARK_DEPS,
4731)
4732
Marat Dukhanc068bb62019-10-04 13:24:39 -07004733############################# End-to-end benchmarks ############################
4734
4735cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004736 name = "fp32_mobilenet_v1",
4737 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004738 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004739 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004740 linkstatic = True,
4741 deps = [
4742 ":XNNPACK",
4743 "@pthreadpool",
4744 ],
4745)
4746
4747cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004748 name = "fp16_mobilenet_v1",
4749 srcs = ["models/fp16-mobilenet-v1.cc"],
4750 hdrs = ["models/models.h"],
4751 copts = xnnpack_std_cxxopts(),
4752 linkstatic = True,
4753 deps = [
4754 ":XNNPACK",
4755 "@FP16",
4756 "@pthreadpool",
4757 ],
4758)
4759
4760cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004761 name = "qs8_mobilenet_v1",
4762 srcs = ["models/qs8-mobilenet-v1.cc"],
4763 hdrs = ["models/models.h"],
4764 copts = xnnpack_std_cxxopts(),
4765 linkstatic = True,
4766 deps = [
4767 ":XNNPACK",
4768 "@pthreadpool",
4769 ],
4770)
4771
4772cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004773 name = "qs8_mobilenet_v2",
4774 srcs = ["models/qs8-mobilenet-v2.cc"],
4775 hdrs = ["models/models.h"],
4776 copts = xnnpack_std_cxxopts(),
4777 linkstatic = True,
4778 deps = [
4779 ":XNNPACK",
4780 "@pthreadpool",
4781 ],
4782)
4783
4784cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004785 name = "fp32_mobilenet_v2",
4786 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004787 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004788 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004789 linkstatic = True,
4790 deps = [
4791 ":XNNPACK",
4792 "@pthreadpool",
4793 ],
4794)
4795
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004796cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004797 name = "fp16_mobilenet_v2",
4798 srcs = ["models/fp16-mobilenet-v2.cc"],
4799 hdrs = ["models/models.h"],
4800 copts = xnnpack_std_cxxopts(),
4801 linkstatic = True,
4802 deps = [
4803 ":XNNPACK",
4804 "@FP16",
4805 "@pthreadpool",
4806 ],
4807)
4808
4809cc_library(
4810 name = "fp32_mobilenet_v3_large",
4811 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004812 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004813 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004814 linkstatic = True,
4815 deps = [
4816 ":XNNPACK",
4817 "@pthreadpool",
4818 ],
4819)
4820
4821cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004822 name = "fp16_mobilenet_v3_large",
4823 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4824 hdrs = ["models/models.h"],
4825 copts = xnnpack_std_cxxopts(),
4826 linkstatic = True,
4827 deps = [
4828 ":XNNPACK",
4829 "@FP16",
4830 "@pthreadpool",
4831 ],
4832)
4833
4834cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004835 name = "fp32_mobilenet_v3_small",
4836 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004837 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004838 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004839 linkstatic = True,
4840 deps = [
4841 ":XNNPACK",
4842 "@pthreadpool",
4843 ],
4844)
4845
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004846cc_library(
4847 name = "fp16_mobilenet_v3_small",
4848 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4849 hdrs = ["models/models.h"],
4850 copts = xnnpack_std_cxxopts(),
4851 linkstatic = True,
4852 deps = [
4853 ":XNNPACK",
4854 "@FP16",
4855 "@pthreadpool",
4856 ],
4857)
4858
Marat Dukhanc068bb62019-10-04 13:24:39 -07004859xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004860 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004861 srcs = [
4862 "bench/f32-dwconv-e2e.cc",
4863 "bench/end2end.h",
4864 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004865 deps = MICROKERNEL_BENCHMARK_DEPS + [
4866 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004867 ":fp32_mobilenet_v1",
4868 ":fp32_mobilenet_v2",
4869 ":fp32_mobilenet_v3_large",
4870 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004871 ],
4872)
4873
4874xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004875 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004876 srcs = [
4877 "bench/f32-gemm-e2e.cc",
4878 "bench/end2end.h",
4879 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004880 deps = MICROKERNEL_BENCHMARK_DEPS + [
4881 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004882 ":fp32_mobilenet_v1",
4883 ":fp32_mobilenet_v2",
4884 ":fp32_mobilenet_v3_large",
4885 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004886 ],
4887)
4888
4889xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004890 name = "end2end_bench",
4891 srcs = ["bench/end2end.cc"],
4892 deps = [
4893 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004894 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004895 ":fp16_mobilenet_v1",
4896 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004897 ":fp16_mobilenet_v3_large",
4898 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004899 ":fp32_mobilenet_v1",
4900 ":fp32_mobilenet_v2",
4901 ":fp32_mobilenet_v3_large",
4902 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004903 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004904 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004905 "@pthreadpool",
4906 ],
4907)
4908
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004909#################### Accuracy evaluation for math functions ####################
4910
4911xnnpack_benchmark(
4912 name = "f32_exp_eval",
4913 srcs = [
4914 "eval/f32-exp.cc",
4915 "src/xnnpack/AlignedAllocator.h",
4916 ] + ACCURACY_EVAL_HDRS,
4917 deps = ACCURACY_EVAL_DEPS,
4918)
4919
Marat Dukhan515c9772019-10-17 18:07:57 -07004920xnnpack_benchmark(
4921 name = "f32_expminus_eval",
4922 srcs = [
4923 "eval/f32-expminus.cc",
4924 "src/xnnpack/AlignedAllocator.h",
4925 ] + ACCURACY_EVAL_HDRS,
4926 deps = ACCURACY_EVAL_DEPS,
4927)
4928
Marat Dukhan98ba4412019-10-23 02:14:28 -07004929xnnpack_benchmark(
4930 name = "f32_extexp_eval",
4931 srcs = [
4932 "eval/f32-extexp.cc",
4933 "src/xnnpack/AlignedAllocator.h",
4934 ] + ACCURACY_EVAL_HDRS,
4935 deps = ACCURACY_EVAL_DEPS,
4936)
4937
Marat Dukhan8853b822020-05-07 12:19:01 -07004938xnnpack_unit_test(
4939 name = "f32_roundne_eval",
4940 srcs = [
4941 "eval/f32-roundne.cc",
4942 "src/xnnpack/AlignedAllocator.h",
4943 "src/xnnpack/math-stubs.h",
4944 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004945 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004946 deps = MICROKERNEL_TEST_DEPS,
4947)
4948
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004949xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004950 name = "f32_roundd_eval",
4951 srcs = [
4952 "eval/f32-roundd.cc",
4953 "src/xnnpack/AlignedAllocator.h",
4954 "src/xnnpack/math-stubs.h",
4955 ] + MICROKERNEL_TEST_HDRS,
4956 automatic = False,
4957 deps = MICROKERNEL_TEST_DEPS,
4958)
4959
4960xnnpack_unit_test(
4961 name = "f32_roundu_eval",
4962 srcs = [
4963 "eval/f32-roundu.cc",
4964 "src/xnnpack/AlignedAllocator.h",
4965 "src/xnnpack/math-stubs.h",
4966 ] + MICROKERNEL_TEST_HDRS,
4967 automatic = False,
4968 deps = MICROKERNEL_TEST_DEPS,
4969)
4970
4971xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004972 name = "f32_roundz_eval",
4973 srcs = [
4974 "eval/f32-roundz.cc",
4975 "src/xnnpack/AlignedAllocator.h",
4976 "src/xnnpack/math-stubs.h",
4977 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004978 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004979 deps = MICROKERNEL_TEST_DEPS,
4980)
4981
Marat Dukhan346a9e52019-11-15 09:06:30 -08004982xnnpack_benchmark(
4983 name = "f32_sigmoid_eval",
4984 srcs = [
4985 "eval/f32-sigmoid.cc",
4986 "src/xnnpack/AlignedAllocator.h",
4987 ] + ACCURACY_EVAL_HDRS,
4988 deps = ACCURACY_EVAL_DEPS,
4989)
4990
Marat Dukhan84000762020-06-29 18:38:43 -07004991xnnpack_benchmark(
4992 name = "f32_sqrt_eval",
4993 srcs = [
4994 "eval/f32-sqrt.cc",
4995 "src/xnnpack/AlignedAllocator.h",
4996 ] + ACCURACY_EVAL_HDRS,
4997 deps = ACCURACY_EVAL_DEPS + [
4998 ":bench_utils",
4999 ],
5000)
5001
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002######################### Unit tests for micro-kernels #########################
5003
5004xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005005 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005006 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005007 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008 "test/gemm-microkernel-tester.h",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005011 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012)
5013
5014xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005015 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005016 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005017 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005018 "test/spmm-microkernel-tester.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + MICROKERNEL_TEST_HDRS,
5021 deps = MICROKERNEL_TEST_DEPS,
5022)
5023
5024xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 name = "f32_argmaxpool_test",
5026 srcs = [
5027 "test/f32-argmaxpool.cc",
5028 "test/argmaxpool-microkernel-tester.h",
5029 "src/xnnpack/AlignedAllocator.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005035 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005037 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 "test/avgpool-microkernel-tester.h",
5039 "src/xnnpack/AlignedAllocator.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005045 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005046 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005047 "test/f32-ibilinear.cc",
5048 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005049 "src/xnnpack/AlignedAllocator.h",
5050 ] + MICROKERNEL_TEST_HDRS,
5051 deps = MICROKERNEL_TEST_DEPS,
5052)
5053
5054xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005055 name = "f32_ibilinear_chw_test",
5056 srcs = [
5057 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005058 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005059 "src/xnnpack/AlignedAllocator.h",
5060 ] + MICROKERNEL_TEST_HDRS,
5061 deps = MICROKERNEL_TEST_DEPS,
5062)
5063
5064xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005065 name = "f16_clamp_test",
5066 srcs = [
5067 "test/f16-clamp.cc",
5068 "test/clamp-microkernel-tester.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 name = "f32_clamp_test",
5075 srcs = [
5076 "test/f32-clamp.cc",
5077 "test/clamp-microkernel-tester.h",
5078 ] + MICROKERNEL_TEST_HDRS,
5079 deps = MICROKERNEL_TEST_DEPS,
5080)
5081
5082xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005083 name = "f32_igemm_test",
5084 srcs = [
5085 "test/f32-igemm.cc",
5086 "test/gemm-microkernel-tester.h",
5087 "src/xnnpack/AlignedAllocator.h",
5088 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005089 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005090)
5091
5092xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005093 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005094 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005095 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005096 "test/gemm-microkernel-tester.h",
5097 "src/xnnpack/AlignedAllocator.h",
5098 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005099 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005100)
5101
5102xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005103 name = "f32_igemm_minmax_test",
5104 srcs = [
5105 "test/f32-igemm-minmax.cc",
5106 "test/gemm-microkernel-tester.h",
5107 "src/xnnpack/AlignedAllocator.h",
5108 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005109 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005110)
5111
5112xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005113 name = "f16_igemm_minmax_test",
5114 srcs = [
5115 "test/f16-igemm-minmax.cc",
5116 "test/gemm-microkernel-tester.h",
5117 "src/xnnpack/AlignedAllocator.h",
5118 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005119 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005120)
5121
5122xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123 name = "f32_conv_hwc_test",
5124 srcs = [
5125 "test/f32-conv-hwc.cc",
5126 "test/conv-hwc-microkernel-tester.h",
5127 "src/xnnpack/AlignedAllocator.h",
5128 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005129 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005130)
5131
5132xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005133 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005135 "test/f32-conv-hwc2chw.cc",
5136 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005137 "src/xnnpack/AlignedAllocator.h",
5138 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005139 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005140)
5141
5142xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005143 name = "f32_dwconv_test",
5144 srcs = [
5145 "test/f32-dwconv.cc",
5146 "test/dwconv-microkernel-tester.h",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005149 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005150)
5151
5152xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005153 name = "f16_dwconv_minmax_test",
5154 srcs = [
5155 "test/f16-dwconv-minmax.cc",
5156 "test/dwconv-microkernel-tester.h",
5157 "src/xnnpack/AlignedAllocator.h",
5158 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005159 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005160)
5161
5162xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005163 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005165 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005166 "test/dwconv-microkernel-tester.h",
5167 "src/xnnpack/AlignedAllocator.h",
5168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005169 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005170)
5171
5172xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005173 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005175 "test/f32-dwconv2d-chw.cc",
5176 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 "src/xnnpack/AlignedAllocator.h",
5178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005179 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180)
5181
5182xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005183 name = "f16_gavgpool_minmax_test",
5184 srcs = [
5185 "test/f16-gavgpool-minmax.cc",
5186 "test/gavgpool-microkernel-tester.h",
5187 "src/xnnpack/AlignedAllocator.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005193 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005195 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 "test/gavgpool-microkernel-tester.h",
5197 "src/xnnpack/AlignedAllocator.h",
5198 ] + MICROKERNEL_TEST_HDRS,
5199 deps = MICROKERNEL_TEST_DEPS,
5200)
5201
5202xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005203 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005205 "test/f32-gavgpool-cw.cc",
5206 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 "src/xnnpack/AlignedAllocator.h",
5208 ] + MICROKERNEL_TEST_HDRS,
5209 deps = MICROKERNEL_TEST_DEPS,
5210)
5211
5212xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005213 name = "f32_gemm_test",
5214 srcs = [
5215 "test/f32-gemm.cc",
5216 "test/gemm-microkernel-tester.h",
5217 "src/xnnpack/AlignedAllocator.h",
5218 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005219 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005220)
5221
5222xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005223 name = "f32_gemm_relu_test",
5224 srcs = [
5225 "test/f32-gemm-relu.cc",
5226 "test/gemm-microkernel-tester.h",
5227 "src/xnnpack/AlignedAllocator.h",
5228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005229 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005230)
5231
5232xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005233 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005235 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005236 "test/gemm-microkernel-tester.h",
5237 "src/xnnpack/AlignedAllocator.h",
5238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005239 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240)
5241
5242xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005243 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005245 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 "test/gemm-microkernel-tester.h",
5247 "src/xnnpack/AlignedAllocator.h",
5248 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005249 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005250)
5251
5252xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005253 name = "f16_hswish_test",
5254 srcs = [
5255 "test/f16-hswish.cc",
5256 "test/hswish-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 name = "f32_hswish_test",
5263 srcs = [
5264 "test/f32-hswish.cc",
5265 "test/hswish-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_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005273 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274 "test/maxpool-microkernel-tester.h",
5275 ] + MICROKERNEL_TEST_HDRS,
5276 deps = MICROKERNEL_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005280 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005282 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005283 "test/avgpool-microkernel-tester.h",
5284 "src/xnnpack/AlignedAllocator.h",
5285 ] + MICROKERNEL_TEST_HDRS,
5286 deps = MICROKERNEL_TEST_DEPS,
5287)
5288
5289xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005290 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005292 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 "test/gemm-microkernel-tester.h",
5294 "src/xnnpack/AlignedAllocator.h",
5295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005296 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297)
5298
5299xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005300 name = "f16_prelu_test",
5301 srcs = [
5302 "test/f16-prelu.cc",
5303 "test/prelu-microkernel-tester.h",
5304 "src/xnnpack/AlignedAllocator.h",
5305 ] + MICROKERNEL_TEST_HDRS,
5306 deps = MICROKERNEL_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 name = "f32_prelu_test",
5311 srcs = [
5312 "test/f32-prelu.cc",
5313 "test/prelu-microkernel-tester.h",
5314 "src/xnnpack/AlignedAllocator.h",
5315 ] + MICROKERNEL_TEST_HDRS,
5316 deps = MICROKERNEL_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005320 name = "f32_raddexpminusmax_test",
5321 srcs = [
5322 "test/f32-raddexpminusmax.cc",
5323 "test/raddexpminusmax-microkernel-tester.h",
5324 ] + MICROKERNEL_TEST_HDRS,
5325 deps = MICROKERNEL_TEST_DEPS,
5326)
5327
5328xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005329 name = "f32_raddextexp_test",
5330 srcs = [
5331 "test/f32-raddextexp.cc",
5332 "test/raddextexp-microkernel-tester.h",
5333 ] + MICROKERNEL_TEST_HDRS,
5334 deps = MICROKERNEL_TEST_DEPS,
5335)
5336
5337xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005338 name = "f32_raddstoreexpminusmax_test",
5339 srcs = [
5340 "test/f32-raddstoreexpminusmax.cc",
5341 "test/raddstoreexpminusmax-microkernel-tester.h",
5342 ] + MICROKERNEL_TEST_HDRS,
5343 deps = MICROKERNEL_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005347 name = "f32_relu_test",
5348 srcs = [
5349 "test/f32-relu.cc",
5350 "test/vunary-microkernel-tester.h",
5351 ] + MICROKERNEL_TEST_HDRS,
5352 deps = MICROKERNEL_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005356 name = "f32_rmax_test",
5357 srcs = [
5358 "test/f32-rmax.cc",
5359 "test/rmax-microkernel-tester.h",
5360 ] + MICROKERNEL_TEST_HDRS,
5361 deps = MICROKERNEL_TEST_DEPS,
5362)
5363
5364xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005365 name = "f32_sigmoid_test",
5366 srcs = [
5367 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005368 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005369 ] + MICROKERNEL_TEST_HDRS,
5370 deps = MICROKERNEL_TEST_DEPS,
5371)
5372
5373xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005374 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005376 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005377 "test/spmm-microkernel-tester.h",
5378 "src/xnnpack/AlignedAllocator.h",
5379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005384 name = "f16_vadd_minmax_test",
5385 srcs = [
5386 "test/f16-vadd-minmax.cc",
5387 "test/vbinary-microkernel-tester.h",
5388 ] + MICROKERNEL_TEST_HDRS,
5389 deps = MICROKERNEL_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
5393 name = "f16_vaddc_minmax_test",
5394 srcs = [
5395 "test/f16-vaddc-minmax.cc",
5396 "test/vbinaryc-microkernel-tester.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
5402 name = "f16_vdiv_minmax_test",
5403 srcs = [
5404 "test/f16-vdiv-minmax.cc",
5405 "test/vbinary-microkernel-tester.h",
5406 ] + MICROKERNEL_TEST_HDRS,
5407 deps = MICROKERNEL_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
5411 name = "f16_vdivc_minmax_test",
5412 srcs = [
5413 "test/f16-vdivc-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_vrdivc_minmax_test",
5421 srcs = [
5422 "test/f16-vrdivc-minmax.cc",
5423 "test/vbinaryc-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
5429 name = "f16_vmax_test",
5430 srcs = [
5431 "test/f16-vmax.cc",
5432 "test/vbinary-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
5438 name = "f16_vmaxc_test",
5439 srcs = [
5440 "test/f16-vmaxc.cc",
5441 "test/vbinaryc-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
5447 name = "f16_vmin_test",
5448 srcs = [
5449 "test/f16-vmin.cc",
5450 "test/vbinary-microkernel-tester.h",
5451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
5456 name = "f16_vminc_test",
5457 srcs = [
5458 "test/f16-vminc.cc",
5459 "test/vbinaryc-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
5465 name = "f16_vmul_minmax_test",
5466 srcs = [
5467 "test/f16-vmul-minmax.cc",
5468 "test/vbinary-microkernel-tester.h",
5469 ] + MICROKERNEL_TEST_HDRS,
5470 deps = MICROKERNEL_TEST_DEPS,
5471)
5472
5473xnnpack_unit_test(
5474 name = "f16_vmulc_minmax_test",
5475 srcs = [
5476 "test/f16-vmulc-minmax.cc",
5477 "test/vbinaryc-microkernel-tester.h",
5478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
5483 name = "f16_vsub_minmax_test",
5484 srcs = [
5485 "test/f16-vsub-minmax.cc",
5486 "test/vbinary-microkernel-tester.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
5492 name = "f16_vsubc_minmax_test",
5493 srcs = [
5494 "test/f16-vsubc-minmax.cc",
5495 "test/vbinaryc-microkernel-tester.h",
5496 ] + MICROKERNEL_TEST_HDRS,
5497 deps = MICROKERNEL_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
5501 name = "f16_vrsubc_minmax_test",
5502 srcs = [
5503 "test/f16-vrsubc-minmax.cc",
5504 "test/vbinaryc-microkernel-tester.h",
5505 ] + MICROKERNEL_TEST_HDRS,
5506 deps = MICROKERNEL_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005510 name = "f32_vabs_test",
5511 srcs = [
5512 "test/f32-vabs.cc",
5513 "test/vunary-microkernel-tester.h",
5514 ] + MICROKERNEL_TEST_HDRS,
5515 deps = MICROKERNEL_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005519 name = "f32_vadd_test",
5520 srcs = [
5521 "test/f32-vadd.cc",
5522 "test/vbinary-microkernel-tester.h",
5523 ] + MICROKERNEL_TEST_HDRS,
5524 deps = MICROKERNEL_TEST_DEPS,
5525)
5526
5527xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005528 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005529 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005530 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005531 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005537 name = "f32_vadd_relu_test",
5538 srcs = [
5539 "test/f32-vadd-relu.cc",
5540 "test/vbinary-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005546 name = "f32_vaddc_test",
5547 srcs = [
5548 "test/f32-vaddc.cc",
5549 "test/vbinaryc-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005555 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005556 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005557 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005558 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005564 name = "f32_vaddc_relu_test",
5565 srcs = [
5566 "test/f32-vaddc-relu.cc",
5567 "test/vbinaryc-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005573 name = "f32_vdiv_test",
5574 srcs = [
5575 "test/f32-vdiv.cc",
5576 "test/vbinary-microkernel-tester.h",
5577 ] + MICROKERNEL_TEST_HDRS,
5578 deps = MICROKERNEL_TEST_DEPS,
5579)
5580
5581xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005582 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005583 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005584 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005585 "test/vbinary-microkernel-tester.h",
5586 ] + MICROKERNEL_TEST_HDRS,
5587 deps = MICROKERNEL_TEST_DEPS,
5588)
5589
5590xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005591 name = "f32_vdiv_relu_test",
5592 srcs = [
5593 "test/f32-vdiv-relu.cc",
5594 "test/vbinary-microkernel-tester.h",
5595 ] + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS,
5597)
5598
5599xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005600 name = "f32_vdivc_test",
5601 srcs = [
5602 "test/f32-vdivc.cc",
5603 "test/vbinaryc-microkernel-tester.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005609 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005610 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005611 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005612 "test/vbinaryc-microkernel-tester.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005618 name = "f32_vdivc_relu_test",
5619 srcs = [
5620 "test/f32-vdivc-relu.cc",
5621 "test/vbinaryc-microkernel-tester.h",
5622 ] + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005627 name = "f32_vrdivc_test",
5628 srcs = [
5629 "test/f32-vrdivc.cc",
5630 "test/vbinaryc-microkernel-tester.h",
5631 ] + MICROKERNEL_TEST_HDRS,
5632 deps = MICROKERNEL_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005636 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005637 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005638 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005639 "test/vbinaryc-microkernel-tester.h",
5640 ] + MICROKERNEL_TEST_HDRS,
5641 deps = MICROKERNEL_TEST_DEPS,
5642)
5643
5644xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005645 name = "f32_vrdivc_relu_test",
5646 srcs = [
5647 "test/f32-vrdivc-relu.cc",
5648 "test/vbinaryc-microkernel-tester.h",
5649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005654 name = "f32_vmax_test",
5655 srcs = [
5656 "test/f32-vmax.cc",
5657 "test/vbinary-microkernel-tester.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
5663 name = "f32_vmaxc_test",
5664 srcs = [
5665 "test/f32-vmaxc.cc",
5666 "test/vbinaryc-microkernel-tester.h",
5667 ] + MICROKERNEL_TEST_HDRS,
5668 deps = MICROKERNEL_TEST_DEPS,
5669)
5670
5671xnnpack_unit_test(
5672 name = "f32_vmin_test",
5673 srcs = [
5674 "test/f32-vmin.cc",
5675 "test/vbinary-microkernel-tester.h",
5676 ] + MICROKERNEL_TEST_HDRS,
5677 deps = MICROKERNEL_TEST_DEPS,
5678)
5679
5680xnnpack_unit_test(
5681 name = "f32_vminc_test",
5682 srcs = [
5683 "test/f32-vminc.cc",
5684 "test/vbinaryc-microkernel-tester.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 deps = MICROKERNEL_TEST_DEPS,
5687)
5688
5689xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005690 name = "f32_vmul_test",
5691 srcs = [
5692 "test/f32-vmul.cc",
5693 "test/vbinary-microkernel-tester.h",
5694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005699 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005700 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005701 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005702 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005708 name = "f32_vmul_relu_test",
5709 srcs = [
5710 "test/f32-vmul-relu.cc",
5711 "test/vbinary-microkernel-tester.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005717 name = "f32_vmulc_test",
5718 srcs = [
5719 "test/f32-vmulc.cc",
5720 "test/vbinaryc-microkernel-tester.h",
5721 ] + MICROKERNEL_TEST_HDRS,
5722 deps = MICROKERNEL_TEST_DEPS,
5723)
5724
5725xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005726 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005727 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005728 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005729 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005730 ] + MICROKERNEL_TEST_HDRS,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
5734xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005735 name = "f32_vmulc_relu_test",
5736 srcs = [
5737 "test/f32-vmulc-relu.cc",
5738 "test/vbinaryc-microkernel-tester.h",
5739 ] + MICROKERNEL_TEST_HDRS,
5740 deps = MICROKERNEL_TEST_DEPS,
5741)
5742
5743xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005744 name = "f16_vmulcaddc_minmax_test",
5745 srcs = [
5746 "test/f16-vmulcaddc-minmax.cc",
5747 "test/vmulcaddc-microkernel-tester.h",
5748 "src/xnnpack/AlignedAllocator.h",
5749 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005750 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005751)
5752
5753xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005754 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005755 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005756 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005757 "test/vmulcaddc-microkernel-tester.h",
5758 "src/xnnpack/AlignedAllocator.h",
5759 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005760 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005761)
5762
5763xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005764 name = "f32_vlrelu_test",
5765 srcs = [
5766 "test/f32-vlrelu.cc",
5767 "test/vunary-microkernel-tester.h",
5768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005773 name = "f32_vneg_test",
5774 srcs = [
5775 "test/f32-vneg.cc",
5776 "test/vunary-microkernel-tester.h",
5777 ] + MICROKERNEL_TEST_HDRS,
5778 deps = MICROKERNEL_TEST_DEPS,
5779)
5780
5781xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005782 name = "f32_vrndne_test",
5783 srcs = [
5784 "test/f32-vrndne.cc",
5785 "test/vunary-microkernel-tester.h",
5786 ] + MICROKERNEL_TEST_HDRS,
5787 deps = MICROKERNEL_TEST_DEPS,
5788)
5789
5790xnnpack_unit_test(
5791 name = "f32_vrndz_test",
5792 srcs = [
5793 "test/f32-vrndz.cc",
5794 "test/vunary-microkernel-tester.h",
5795 ] + MICROKERNEL_TEST_HDRS,
5796 deps = MICROKERNEL_TEST_DEPS,
5797)
5798
5799xnnpack_unit_test(
5800 name = "f32_vrndu_test",
5801 srcs = [
5802 "test/f32-vrndu.cc",
5803 "test/vunary-microkernel-tester.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 deps = MICROKERNEL_TEST_DEPS,
5806)
5807
5808xnnpack_unit_test(
5809 name = "f32_vrndd_test",
5810 srcs = [
5811 "test/f32-vrndd.cc",
5812 "test/vunary-microkernel-tester.h",
5813 ] + MICROKERNEL_TEST_HDRS,
5814 deps = MICROKERNEL_TEST_DEPS,
5815)
5816
5817xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005818 name = "f32_vscale_test",
5819 srcs = [
5820 "test/f32-vscale.cc",
5821 "test/vscale-microkernel-tester.h",
5822 ] + MICROKERNEL_TEST_HDRS,
5823 deps = MICROKERNEL_TEST_DEPS,
5824)
5825
5826xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005827 name = "f32_vscaleexpminusmax_test",
5828 srcs = [
5829 "test/f32-vscaleexpminusmax.cc",
5830 "test/vscaleexpminusmax-microkernel-tester.h",
5831 ] + MICROKERNEL_TEST_HDRS,
5832 deps = MICROKERNEL_TEST_DEPS,
5833)
5834
5835xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005836 name = "f32_vscaleextexp_test",
5837 srcs = [
5838 "test/f32-vscaleextexp.cc",
5839 "test/vscaleextexp-microkernel-tester.h",
5840 ] + MICROKERNEL_TEST_HDRS,
5841 deps = MICROKERNEL_TEST_DEPS,
5842)
5843
5844xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005845 name = "f32_vsqr_test",
5846 srcs = [
5847 "test/f32-vsqr.cc",
5848 "test/vunary-microkernel-tester.h",
5849 ] + MICROKERNEL_TEST_HDRS,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005854 name = "f32_vsqrdiff_test",
5855 srcs = [
5856 "test/f32-vsqrdiff.cc",
5857 "test/vbinary-microkernel-tester.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
5863 name = "f32_vsqrdiffc_test",
5864 srcs = [
5865 "test/f32-vsqrdiffc.cc",
5866 "test/vbinaryc-microkernel-tester.h",
5867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005872 name = "f32_vsqrt_test",
5873 srcs = [
5874 "test/f32-vsqrt.cc",
5875 "test/vunary-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005881 name = "f32_vsub_test",
5882 srcs = [
5883 "test/f32-vsub.cc",
5884 "test/vbinary-microkernel-tester.h",
5885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005890 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005891 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005892 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005893 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005899 name = "f32_vsub_relu_test",
5900 srcs = [
5901 "test/f32-vsub-relu.cc",
5902 "test/vbinary-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005908 name = "f32_vsubc_test",
5909 srcs = [
5910 "test/f32-vsubc.cc",
5911 "test/vbinaryc-microkernel-tester.h",
5912 ] + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS,
5914)
5915
5916xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005917 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005918 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005919 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005920 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005921 ] + MICROKERNEL_TEST_HDRS,
5922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
5925xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005926 name = "f32_vsubc_relu_test",
5927 srcs = [
5928 "test/f32-vsubc-relu.cc",
5929 "test/vbinaryc-microkernel-tester.h",
5930 ] + MICROKERNEL_TEST_HDRS,
5931 deps = MICROKERNEL_TEST_DEPS,
5932)
5933
5934xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005935 name = "f32_vrsubc_test",
5936 srcs = [
5937 "test/f32-vrsubc.cc",
5938 "test/vbinaryc-microkernel-tester.h",
5939 ] + MICROKERNEL_TEST_HDRS,
5940 deps = MICROKERNEL_TEST_DEPS,
5941)
5942
5943xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005944 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005945 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005946 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005947 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005948 ] + MICROKERNEL_TEST_HDRS,
5949 deps = MICROKERNEL_TEST_DEPS,
5950)
5951
5952xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005953 name = "f32_vrsubc_relu_test",
5954 srcs = [
5955 "test/f32-vrsubc-relu.cc",
5956 "test/vbinaryc-microkernel-tester.h",
5957 ] + MICROKERNEL_TEST_HDRS,
5958 deps = MICROKERNEL_TEST_DEPS,
5959)
5960
5961xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005962 name = "qs8_dwconv_minmax_test",
5963 srcs = [
5964 "test/qs8-dwconv-minmax.cc",
5965 "test/dwconv-microkernel-tester.h",
5966 "src/xnnpack/AlignedAllocator.h",
5967 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5968 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5969)
5970
5971xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005972 name = "qs8_gavgpool_minmax_test",
5973 srcs = [
5974 "test/qs8-gavgpool-minmax.cc",
5975 "test/gavgpool-microkernel-tester.h",
5976 "src/xnnpack/AlignedAllocator.h",
5977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005982 name = "qs8_gemm_minmax_test",
5983 srcs = [
5984 "test/qs8-gemm-minmax.cc",
5985 "test/gemm-microkernel-tester.h",
5986 "src/xnnpack/AlignedAllocator.h",
5987 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5988 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5989)
5990
5991xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005992 name = "qs8_igemm_minmax_test",
5993 srcs = [
5994 "test/qs8-igemm-minmax.cc",
5995 "test/gemm-microkernel-tester.h",
5996 "src/xnnpack/AlignedAllocator.h",
5997 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5998 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5999)
6000
6001xnnpack_unit_test(
6002 name = "qs8_requantization_test",
6003 srcs = [
6004 "src/xnnpack/requantization-stubs.h",
6005 "test/qs8-requantization.cc",
6006 "test/requantization-tester.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006012 name = "qs8_vadd_minmax_test",
6013 srcs = [
6014 "test/qs8-vadd-minmax.cc",
6015 "test/vadd-microkernel-tester.h",
6016 ] + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006021 name = "qs8_vaddc_minmax_test",
6022 srcs = [
6023 "test/qs8-vaddc-minmax.cc",
6024 "test/vaddc-microkernel-tester.h",
6025 ] + MICROKERNEL_TEST_HDRS,
6026 deps = MICROKERNEL_TEST_DEPS,
6027)
6028
6029xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006030 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006031 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006032 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006033 "test/avgpool-microkernel-tester.h",
6034 "src/xnnpack/AlignedAllocator.h",
6035 ] + MICROKERNEL_TEST_HDRS,
6036 deps = MICROKERNEL_TEST_DEPS,
6037)
6038
6039xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006040 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006041 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006042 "test/qu8-dwconv-minmax.cc",
6043 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006044 "src/xnnpack/AlignedAllocator.h",
6045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006046 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006047)
6048
6049xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006050 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006051 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006052 "test/qu8-igemm-minmax.cc",
6053 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054 "src/xnnpack/AlignedAllocator.h",
6055 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006056 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006057)
6058
6059xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006060 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006062 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063 "test/gavgpool-microkernel-tester.h",
6064 "src/xnnpack/AlignedAllocator.h",
6065 ] + MICROKERNEL_TEST_HDRS,
6066 deps = MICROKERNEL_TEST_DEPS,
6067)
6068
6069xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006070 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006072 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006073 "test/gemm-microkernel-tester.h",
6074 "src/xnnpack/AlignedAllocator.h",
6075 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006076 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006077)
6078
6079xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006080 name = "qu8_requantization_test",
6081 srcs = [
6082 "src/xnnpack/requantization-stubs.h",
6083 "test/qu8-requantization.cc",
6084 "test/requantization-tester.h",
6085 ] + MICROKERNEL_TEST_HDRS,
6086 deps = MICROKERNEL_TEST_DEPS,
6087)
6088
6089xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006090 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006091 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006092 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093 "test/vadd-microkernel-tester.h",
6094 ] + MICROKERNEL_TEST_HDRS,
6095 deps = MICROKERNEL_TEST_DEPS,
6096)
6097
6098xnnpack_unit_test(
6099 name = "u8_clamp_test",
6100 srcs = [
6101 "test/u8-clamp.cc",
6102 "test/clamp-microkernel-tester.h",
6103 ] + MICROKERNEL_TEST_HDRS,
6104 deps = MICROKERNEL_TEST_DEPS,
6105)
6106
6107xnnpack_unit_test(
6108 name = "u8_lut32norm_test",
6109 srcs = [
6110 "test/u8-lut32norm.cc",
6111 "test/lut-norm-microkernel-tester.h",
6112 ] + MICROKERNEL_TEST_HDRS,
6113 deps = MICROKERNEL_TEST_DEPS,
6114)
6115
6116xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006117 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006118 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006119 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006120 "test/maxpool-microkernel-tester.h",
6121 ] + MICROKERNEL_TEST_HDRS,
6122 deps = MICROKERNEL_TEST_DEPS,
6123)
6124
6125xnnpack_unit_test(
6126 name = "u8_rmax_test",
6127 srcs = [
6128 "test/u8-rmax.cc",
6129 "test/rmax-microkernel-tester.h",
6130 ] + MICROKERNEL_TEST_HDRS,
6131 deps = MICROKERNEL_TEST_DEPS,
6132)
6133
6134xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006135 name = "x32_fill_test",
6136 srcs = [
6137 "test/x32-fill.cc",
6138 "test/fill-microkernel-tester.h",
6139 ] + MICROKERNEL_TEST_HDRS,
6140 deps = MICROKERNEL_TEST_DEPS,
6141)
6142
6143xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 name = "x32_packx_test",
6145 srcs = [
6146 "test/x32-packx.cc",
6147 "test/pack-microkernel-tester.h",
6148 "src/xnnpack/AlignedAllocator.h",
6149 ] + MICROKERNEL_TEST_HDRS,
6150 deps = MICROKERNEL_TEST_DEPS,
6151)
6152
6153xnnpack_unit_test(
6154 name = "x32_pad_test",
6155 srcs = [
6156 "test/x32-pad.cc",
6157 "test/pad-microkernel-tester.h",
6158 ] + MICROKERNEL_TEST_HDRS,
6159 deps = MICROKERNEL_TEST_DEPS,
6160)
6161
6162xnnpack_unit_test(
6163 name = "x32_unpool_test",
6164 srcs = [
6165 "test/x32-unpool.cc",
6166 "test/unpool-microkernel-tester.h",
6167 ] + MICROKERNEL_TEST_HDRS,
6168 deps = MICROKERNEL_TEST_DEPS,
6169)
6170
6171xnnpack_unit_test(
6172 name = "x32_zip_test",
6173 srcs = [
6174 "test/x32-zip.cc",
6175 "test/zip-microkernel-tester.h",
6176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
6181 name = "x8_lut_test",
6182 srcs = [
6183 "test/x8-lut.cc",
6184 "test/lut-microkernel-tester.h",
6185 ] + MICROKERNEL_TEST_HDRS,
6186 deps = MICROKERNEL_TEST_DEPS,
6187)
6188
6189xnnpack_unit_test(
6190 name = "x8_zip_test",
6191 srcs = [
6192 "test/x8-zip.cc",
6193 "test/zip-microkernel-tester.h",
6194 ] + MICROKERNEL_TEST_HDRS,
6195 deps = MICROKERNEL_TEST_DEPS,
6196)
6197
Marat Dukhan20c3b922020-03-10 03:45:06 -07006198########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199
6200xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006201 name = "operator_size_test",
6202 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006203 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204)
6205
Marat Dukhan20c3b922020-03-10 03:45:06 -07006206xnnpack_binary(
6207 name = "subgraph_size_test",
6208 srcs = ["test/subgraph-size.c"],
6209 deps = [":XNNPACK"],
6210)
6211
6212########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213
6214xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006215 name = "abs_nc_test",
6216 srcs = [
6217 "test/abs-nc.cc",
6218 "test/abs-operator-tester.h",
6219 ],
6220 deps = OPERATOR_TEST_DEPS,
6221)
6222
6223xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006224 name = "add_nd_test",
6225 srcs = [
6226 "test/add-nd.cc",
6227 "test/binary-elementwise-operator-tester.h",
6228 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006229 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006230)
6231
6232xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006233 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006234 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006235 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236 "test/argmax-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 Dukhanefc47b82019-11-18 09:25:38 -08006242 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006244 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245 "test/average-pooling-operator-tester.h",
6246 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006247 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006248)
6249
6250xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006251 name = "bankers_rounding_nc_test",
6252 srcs = [
6253 "test/bankers-rounding-nc.cc",
6254 "test/bankers-rounding-operator-tester.h",
6255 ],
6256 deps = OPERATOR_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
6260 name = "ceiling_nc_test",
6261 srcs = [
6262 "test/ceiling-nc.cc",
6263 "test/ceiling-operator-tester.h",
6264 ],
6265 deps = OPERATOR_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006269 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006271 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 "test/channel-shuffle-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 Dukhanefc47b82019-11-18 09:25:38 -08006278 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006279 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006280 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281 "test/clamp-operator-tester.h",
6282 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006283 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284)
6285
6286xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006287 name = "constant_pad_nd_test",
6288 srcs = [
6289 "test/constant-pad-nd.cc",
6290 "test/constant-pad-operator-tester.h",
6291 ],
6292 deps = OPERATOR_TEST_DEPS,
6293)
6294
6295xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006296 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006297 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006298 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299 "test/convolution-operator-tester.h",
6300 ],
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 Dukhanefc47b82019-11-18 09:25:38 -08006305 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006307 "test/convolution-nchw.cc",
6308 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006310 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006311)
6312
6313xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006314 name = "copy_nc_test",
6315 srcs = [
6316 "test/copy-nc.cc",
6317 "test/copy-operator-tester.h",
6318 ],
6319 deps = OPERATOR_TEST_DEPS,
6320)
6321
6322xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006323 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006325 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006326 "test/deconvolution-operator-tester.h",
6327 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006328 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006329)
6330
6331xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006332 name = "divide_nd_test",
6333 srcs = [
6334 "test/binary-elementwise-operator-tester.h",
6335 "test/divide-nd.cc",
6336 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006337 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006338)
6339
6340xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006341 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006342 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006343 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006344 "test/fully-connected-operator-tester.h",
6345 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006346 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006347)
6348
6349xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006350 name = "floor_nc_test",
6351 srcs = [
6352 "test/floor-nc.cc",
6353 "test/floor-operator-tester.h",
6354 ],
6355 deps = OPERATOR_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006359 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006360 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006361 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006363 ] + OPERATOR_TEST_PARAMS_HDRS,
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 = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006369 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006370 "test/global-average-pooling-ncw.cc",
6371 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006372 ],
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 = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006378 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006379 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380 "test/hardswish-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 = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006387 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006388 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389 "test/leaky-relu-operator-tester.h",
6390 ],
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 Dukhanefc47b82019-11-18 09:25:38 -08006395 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006396 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006397 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398 "test/max-pooling-operator-tester.h",
6399 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006400 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006401)
6402
6403xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006404 name = "maximum_nd_test",
6405 srcs = [
6406 "test/binary-elementwise-operator-tester.h",
6407 "test/maximum-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(
6413 name = "minimum_nd_test",
6414 srcs = [
6415 "test/binary-elementwise-operator-tester.h",
6416 "test/minimum-nd.cc",
6417 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006418 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006419)
6420
6421xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006422 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006423 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006424 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006425 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006426 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006427 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006428)
6429
6430xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006431 name = "negate_nc_test",
6432 srcs = [
6433 "test/negate-nc.cc",
6434 "test/negate-operator-tester.h",
6435 ],
6436 deps = OPERATOR_TEST_DEPS,
6437)
6438
6439xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006440 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006441 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006442 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006443 "test/prelu-operator-tester.h",
6444 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006445 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006446)
6447
6448xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006449 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006450 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006451 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006452 "test/resize-bilinear-operator-tester.h",
6453 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006454 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006455)
6456
6457xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006458 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006459 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006460 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006461 "test/sigmoid-operator-tester.h",
6462 ],
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 Dukhanfd8e6892020-01-27 15:25:25 -08006467 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006468 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006469 "test/softmax-nc.cc",
6470 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006472 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006473)
6474
6475xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006476 name = "square_nc_test",
6477 srcs = [
6478 "test/square-nc.cc",
6479 "test/square-operator-tester.h",
6480 ],
6481 deps = OPERATOR_TEST_DEPS,
6482)
6483
6484xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006485 name = "square_root_nc_test",
6486 srcs = [
6487 "test/square-root-nc.cc",
6488 "test/square-root-operator-tester.h",
6489 ],
6490 deps = OPERATOR_TEST_DEPS,
6491)
6492
6493xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006494 name = "squared_difference_nd_test",
6495 srcs = [
6496 "test/binary-elementwise-operator-tester.h",
6497 "test/squared-difference-nd.cc",
6498 ],
6499 deps = OPERATOR_TEST_DEPS,
6500)
6501
6502xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006503 name = "subtract_nd_test",
6504 srcs = [
6505 "test/binary-elementwise-operator-tester.h",
6506 "test/subtract-nd.cc",
6507 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006508 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006509)
6510
6511xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006512 name = "truncation_nc_test",
6513 srcs = [
6514 "test/truncation-nc.cc",
6515 "test/truncation-operator-tester.h",
6516 ],
6517 deps = OPERATOR_TEST_DEPS,
6518)
6519
6520xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006521 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006522 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006523 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006524 "test/unpooling-operator-tester.h",
6525 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006526 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006527)
6528
Chao Mei6ddfc602020-05-13 22:29:36 -07006529############################### Misc unit tests ###############################
6530
6531xnnpack_unit_test(
6532 name = "memory_planner_test",
6533 srcs = [
6534 "test/memory-planner-test.cc",
6535 ],
6536 deps = [
6537 ":XNNPACK",
6538 ":memory_planner",
6539 ],
6540)
6541
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006542xnnpack_unit_test(
6543 name = "subgraph_nchw_test",
6544 srcs = [
6545 "src/xnnpack/subgraph.h",
6546 "test/subgraph-nchw.cc",
6547 "test/subgraph-tester.h",
6548 ],
6549 deps = [
6550 ":XNNPACK",
6551 ],
6552)
6553
Marat Dukhan08c4a432019-10-03 09:29:21 -07006554############################# Build configurations #############################
6555
Marat Dukhanb8642352019-10-30 15:43:02 -07006556# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006557config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006558 name = "xnn_enable_assembly_explicit_true",
6559 define_values = {"xnn_enable_assembly": "true"},
6560)
6561
6562# Disables usage of assembly kernels.
6563config_setting(
6564 name = "xnn_enable_assembly_explicit_false",
6565 define_values = {"xnn_enable_assembly": "false"},
6566)
6567
Marat Dukhan9de90e02020-06-18 16:04:12 -07006568# Enables usage of sparse inference.
6569config_setting(
6570 name = "xnn_enable_sparse_explicit_true",
6571 define_values = {"xnn_enable_sparse": "true"},
6572)
6573
6574# Disables usage of sparse inference.
6575config_setting(
6576 name = "xnn_enable_sparse_explicit_false",
6577 define_values = {"xnn_enable_sparse": "false"},
6578)
6579
Marat Dukhan05702cf2020-03-26 15:41:33 -07006580# Disables usage of HMP-aware optimizations.
6581config_setting(
6582 name = "xnn_enable_hmp_explicit_false",
6583 define_values = {"xnn_enable_hmp": "false"},
6584)
6585
Chao Mei6ddfc602020-05-13 22:29:36 -07006586# Enable usage of optimized memory allocation
6587config_setting(
6588 name = "xnn_enable_memopt_explicit_true",
6589 define_values = {"xnnpack_enable_memopt": "true"},
6590)
6591
6592# Disable usage of optimized memory allocation
6593config_setting(
6594 name = "xnn_enable_memopt_explicit_false",
6595 define_values = {"xnnpack_enable_memopt": "false"},
6596)
6597
Marat Dukhanb8642352019-10-30 15:43:02 -07006598# Builds with -c dbg
6599config_setting(
6600 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006601 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006602 "compilation_mode": "dbg",
6603 },
6604)
6605
6606# Builds with -c opt
6607config_setting(
6608 name = "optimized_build",
6609 values = {
6610 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006611 },
6612)
6613
6614config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006615 name = "linux_k8",
6616 values = {"cpu": "k8"},
6617)
6618
6619config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006620 name = "linux_arm",
6621 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006622)
6623
6624config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006625 name = "linux_armeabi",
6626 values = {"cpu": "armeabi"},
6627)
6628
6629config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006630 name = "linux_armhf",
6631 values = {"cpu": "armhf"},
6632)
6633
6634config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006635 name = "linux_armv7a",
6636 values = {"cpu": "armv7a"},
6637)
6638
6639config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006640 name = "linux_aarch64",
6641 values = {"cpu": "aarch64"},
6642)
6643
6644config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006645 name = "android",
6646 values = {"crosstool_top": "//external:android/crosstool"},
6647)
6648
6649config_setting(
6650 name = "android_armv7",
6651 values = {
6652 "crosstool_top": "//external:android/crosstool",
6653 "cpu": "armeabi-v7a",
6654 },
6655)
6656
6657config_setting(
6658 name = "android_arm64",
6659 values = {
6660 "crosstool_top": "//external:android/crosstool",
6661 "cpu": "arm64-v8a",
6662 },
6663)
6664
6665config_setting(
6666 name = "android_x86",
6667 values = {
6668 "crosstool_top": "//external:android/crosstool",
6669 "cpu": "x86",
6670 },
6671)
6672
6673config_setting(
6674 name = "android_x86_64",
6675 values = {
6676 "crosstool_top": "//external:android/crosstool",
6677 "cpu": "x86_64",
6678 },
6679)
6680
6681config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006682 name = "windows_x86_64",
6683 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006684)
6685
6686config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006687 name = "windows_x86_64_clang",
6688 values = {
6689 "compiler": "clang-cl",
6690 "cpu": "x64_windows",
6691 },
6692)
6693
6694config_setting(
6695 name = "windows_x86_64_mingw",
6696 values = {
6697 "compiler": "mingw-gcc",
6698 "cpu": "x64_windows",
6699 },
6700)
6701
6702config_setting(
6703 name = "windows_x86_64_msys",
6704 values = {
6705 "compiler": "msys-gcc",
6706 "cpu": "x64_windows",
6707 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006708)
6709
6710config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006711 name = "macos_x86_64",
6712 values = {
6713 "apple_platform_type": "macos",
6714 "cpu": "darwin",
6715 },
6716)
6717
6718config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006719 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006720 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006721)
6722
6723config_setting(
6724 name = "emscripten_wasm",
6725 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006726 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006727 "cpu": "wasm",
6728 },
6729)
6730
6731config_setting(
6732 name = "emscripten_wasmsimd",
6733 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006734 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006735 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006736 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006737 },
6738)
6739
6740config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006741 name = "ios_armv7",
6742 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006743 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006744 "cpu": "ios_armv7",
6745 },
6746)
6747
6748config_setting(
6749 name = "ios_arm64",
6750 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006751 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006752 "cpu": "ios_arm64",
6753 },
6754)
6755
6756config_setting(
6757 name = "ios_arm64e",
6758 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006759 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006760 "cpu": "ios_arm64e",
6761 },
6762)
6763
6764config_setting(
6765 name = "ios_x86",
6766 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006767 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006768 "cpu": "ios_i386",
6769 },
6770)
6771
6772config_setting(
6773 name = "ios_x86_64",
6774 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006775 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006776 "cpu": "ios_x86_64",
6777 },
6778)
6779
6780config_setting(
6781 name = "watchos_armv7k",
6782 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006783 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006784 "cpu": "watchos_armv7k",
6785 },
6786)
6787
6788config_setting(
6789 name = "watchos_arm64_32",
6790 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006791 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006792 "cpu": "watchos_arm64_32",
6793 },
6794)
6795
6796config_setting(
6797 name = "watchos_x86",
6798 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006799 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006800 "cpu": "watchos_i386",
6801 },
6802)
6803
6804config_setting(
6805 name = "watchos_x86_64",
6806 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006807 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006808 "cpu": "watchos_x86_64",
6809 },
6810)
6811
6812config_setting(
6813 name = "tvos_arm64",
6814 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006815 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006816 "cpu": "tvos_arm64",
6817 },
6818)
6819
6820config_setting(
6821 name = "tvos_x86_64",
6822 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006823 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006824 "cpu": "tvos_x86_64",
6825 },
6826)