blob: 8e6348bd56ee40c0c252b6755d5aa72193d298cd [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 Dukhan98f2eeb2020-10-23 23:13:41 -07002042 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
2043 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2044 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2045 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2046 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2047 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
2048 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2049 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2050 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
2051 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002052 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2053 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2054 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002055 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2056 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2057 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002058 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2059 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2060 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002061 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002062 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002063 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002064 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002065 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002066 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002067 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2068 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002069 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002070 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2071 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002072 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002073 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2074 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2075 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002076 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002077 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002078 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002079 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002080 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002081 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002082 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002083 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002084 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2085 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002086 "src/qs8-requantization/precise-ssse3.c",
2087 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002088 "src/qu8-requantization/precise-ssse3.c",
2089 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090]
2091
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002092SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002093 "src/f32-prelu/gen/sse41-2x4.c",
2094 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002095 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2096 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2097 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2098 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2099 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2100 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002101 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002102 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002103 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002104 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002105 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2106 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002107 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2108 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002109 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2110 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002111 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2112 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2113 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2114 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2115 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2116 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2117 "src/math/roundd-sse41.c",
2118 "src/math/roundne-sse41.c",
2119 "src/math/roundu-sse41.c",
2120 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002121 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2122 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2123 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002124 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2125 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2126 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002127 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2128 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2129 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002130 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002131 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002132 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002133 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002134 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002135 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2137 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002138 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002139 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2140 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002141 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002142 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2143 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2144 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002145 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002146 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002147 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002148 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002149 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002150 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002151 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002152 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002153 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2154 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002155 "src/qs8-requantization/fp32-sse4.c",
2156 "src/qs8-requantization/precise-sse4.c",
2157 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002158 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2159 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2160 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2161 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002162 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2163 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2164 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2165 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002166 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2167 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2168 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2169 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002170 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2171 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2172 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2173 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002174 "src/qu8-requantization/precise-sse4.c",
2175 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002176]
2177
Marat Dukhan08c4a432019-10-03 09:29:21 -07002178AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002179 "src/f32-clamp/gen/avx-x8.c",
2180 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002181 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2182 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002183 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2184 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002185 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2186 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002187 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2188 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2189 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2190 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2191 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2192 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002193 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002194 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2195 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002196 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002197 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002198 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002199 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2201 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2202 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2203 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2204 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2205 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2206 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2207 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2208 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2209 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2210 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002211 "src/f32-hswish/gen/hswish-avx-x8.c",
2212 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002213 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002214 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2215 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002216 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002217 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002218 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002219 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002220 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2221 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002222 "src/f32-prelu/gen/avx-2x8.c",
2223 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002224 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002225 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002226 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002227 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2228 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2229 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2230 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2231 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2232 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2233 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2234 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2235 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2236 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2237 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2238 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2239 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2240 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2241 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2242 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2243 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2244 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2245 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2246 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002247 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2248 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2249 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2250 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2251 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2252 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2253 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2254 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002255 "src/f32-vbinary/gen/vmax-avx-x8.c",
2256 "src/f32-vbinary/gen/vmax-avx-x16.c",
2257 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2258 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2259 "src/f32-vbinary/gen/vmin-avx-x8.c",
2260 "src/f32-vbinary/gen/vmin-avx-x16.c",
2261 "src/f32-vbinary/gen/vminc-avx-x8.c",
2262 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002263 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2264 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2265 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2266 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2267 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2268 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2269 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2270 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002271 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2272 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2273 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2274 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002275 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2276 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2277 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2278 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002279 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2280 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002281 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2282 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2284 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2285 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2286 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2287 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2288 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2289 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002290 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2291 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002292 "src/f32-vunary/gen/vabs-avx-x8.c",
2293 "src/f32-vunary/gen/vabs-avx-x16.c",
2294 "src/f32-vunary/gen/vneg-avx-x8.c",
2295 "src/f32-vunary/gen/vneg-avx-x16.c",
2296 "src/f32-vunary/gen/vsqr-avx-x8.c",
2297 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002298 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002299 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002300 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002301 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2302 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002303]
2304
Marat Dukhan1566fee2020-08-02 21:55:41 -07002305XOP_UKERNELS = [
2306 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002307 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002308 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002309 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002310 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002311 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002312 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2313 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002314 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002315 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2316 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002317 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002318 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2319 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2320 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002321 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002322 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002323 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002324 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002325 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002326 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002327 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002328 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002329 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2330 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002331 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2332 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2333 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2334 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2335 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2336 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2337 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2338 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002339]
2340
Marat Dukhanfda12b82019-11-21 12:27:59 -08002341FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002342 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2343 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002344 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2345 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002346 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2347 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002348 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2349 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2350 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2351 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2352 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2353 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002354 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002355 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2356 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2357 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2358 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002359 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002360 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2361 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002362 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002363 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2364 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2366 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2367 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002368 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2369 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2370 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2371 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2372 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2373 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2374 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2375 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2376 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2377 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2378 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2379 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2380 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2381 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002382 "src/f32-hswish/gen/hswish-fma3-x8.c",
2383 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002384 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002385 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2386 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2387 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2388 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002389 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002390 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2391 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002392 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002393 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2394 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002395 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2396 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2397 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002398 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2399 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2400 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2401 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2402 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2403 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2404 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2405 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002406 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002407 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002408 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002409]
2410
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002411AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002412 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2413 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002414 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002415 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002416 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002417 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2418 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002419 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002420 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2421 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2422 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002423 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002424 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2425 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002426 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002427 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002429 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2430 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002431 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002432 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2433 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2434 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002436 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2437 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002439 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002440 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002441 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2442 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002443 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002444 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2445 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2446 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002447 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002448 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2449 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2450 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2451 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2452 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2453 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2454 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2455 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2456 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2457 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2458 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2459 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2460 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2461 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2462 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2463 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2464 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2465 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2466 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2467 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2468 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2469 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2470 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2471 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2472 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2473 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2474 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2475 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2476 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2477 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002478 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2479 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2480 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2481 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2482 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2483 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2484 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2485 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2486 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2487 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2488 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2489 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2490 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2491 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2492 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2493 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2494 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2495 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2496 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2497 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2498 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2499 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2500 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2501 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002502 "src/math/exp-avx2-lut8-p3-perm.c",
2503 "src/math/exp-avx2-lut8-p4-perm.c",
2504 "src/math/exp-avx2-p5.c",
2505 "src/math/expminus-avx2-p5.c",
2506 "src/math/extexp-avx2-p5.c",
2507 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2508 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2509 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2510 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2511 "src/math/sigmoid-avx2-rr1-p5-div.c",
2512 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2513 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2514 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2515 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2516 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2517 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2518 "src/math/sigmoid-avx2-rr2-p5-div.c",
2519 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2520 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002521 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002522 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002523 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2524 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002525 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002526 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002527 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002528 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002530 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002531 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002532 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002533 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2534 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2535 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002536 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2537 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2538 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2539 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2540 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2541 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2542 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2543 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002544]
2545
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002547 "src/f32-clamp/gen/avx512f-x16.c",
2548 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002549 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2550 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002551 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2552 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002553 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2554 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002555 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2556 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2557 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2558 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2559 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2560 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002561 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2562 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2563 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2564 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2565 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2566 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002567 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2568 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2569 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2570 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2571 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2572 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002573 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2574 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002575 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2576 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2577 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2578 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2579 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2580 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002581 "src/f32-prelu/gen/avx512f-2x16.c",
2582 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002583 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2584 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002585 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002586 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002588 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2589 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002590 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2592 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2593 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002594 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002595 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2596 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002597 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002598 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002599 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002600 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2601 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002603 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2604 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2605 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002606 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002607 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2608 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002609 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002610 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002611 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002612 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2613 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002615 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2616 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2617 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002619 "src/f32-relu/gen/avx512f-x16.c",
2620 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002622 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2623 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2624 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2625 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2626 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2627 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2628 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2629 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2630 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2631 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2632 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2633 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2634 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2635 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2636 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2637 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002638 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2639 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2640 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2641 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2642 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2643 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2644 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2645 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2646 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2647 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2648 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2649 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2650 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2651 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2652 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2653 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002654 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2655 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2656 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2657 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2658 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2659 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2660 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2661 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2662 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2663 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2664 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2665 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2666 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2667 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2668 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2669 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002670 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2671 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2672 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2673 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2674 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2675 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2676 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2677 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002678 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2679 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2680 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2681 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2682 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2683 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2684 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2685 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002686 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2687 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2688 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2689 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2690 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2691 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2692 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2693 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002694 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2695 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2696 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2697 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002698 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2699 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2700 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2701 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002702 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2703 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002704 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2705 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2706 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2707 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2708 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2709 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2710 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2711 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002712 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002713 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2714 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2715 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2716 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2717 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2718 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2719 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2720 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2721 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2722 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2723 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2724 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2725 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2726 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2727 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2728 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2729 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2730 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2731 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2732 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2733 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2734 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2735 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2736 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002737 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2738 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2739 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2740 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2741 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2742 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2743 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2744 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002745 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2746 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2747 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2748 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2749 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2750 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002751 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2752 "src/math/exp-avx512f-lut16-p3-perm.c",
2753 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2754 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002755 "src/math/exp-avx512f-p5-scalef.c",
2756 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002757 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002758 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002759 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002760 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002761 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002762 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002763 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002764 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002765 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2767 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2768 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2769 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2770 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2771 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2772 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2773 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2774 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2775 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002776 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002777 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002778 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2779 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2780 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2781 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002782 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002783 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002784 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002785]
2786
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002787AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002788 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2789 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002790 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2791 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2792 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2793 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2794 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2795 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2796 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2797 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2798]
2799
Frank Barchardbcedc082020-08-17 18:00:51 -07002800WASM32_ASM_UKERNELS = [
2801 "src/f32-relu/wasm_shr_x1.S",
2802 "src/f32-relu/wasm_shr_x2.S",
2803 "src/f32-relu/wasm_shr_x4.S",
2804]
2805
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002807 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002808 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002809 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2810 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002811 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002812 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002813 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002814 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002815 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2816 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002817 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2818 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2819 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2820 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821]
2822
2823AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002824 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002825 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002826 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002827 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002828 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002829 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002830 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002831 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2832 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2833 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2834 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2835 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2836 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2837 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002838 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2839 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002840 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2841 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2842 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002843 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2844 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002845 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2846 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2847 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2848 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002849 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002850 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2851 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002852 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2853 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2854 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2855 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002856 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002857 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002858 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002859 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002860 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2861 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2862 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2863 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2864 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2865 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2866 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2867 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2868 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2869 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2870 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2871 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2872 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2873 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2874 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2875 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2876 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2877 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2878 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2879 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2880 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2881 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002882 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002883 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002884 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2885 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002886 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2887 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2888 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2889 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2890 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2891 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002892 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2893 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2894 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2895 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002896 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2897 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002898 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002899 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002900 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901]
2902
Marat Dukhan1b354632020-03-23 12:50:22 -07002903INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002904 "src/xnnpack/argmaxpool.h",
2905 "src/xnnpack/avgpool.h",
2906 "src/xnnpack/clamp.h",
2907 "src/xnnpack/common.h",
2908 "src/xnnpack/conv.h",
2909 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002910 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 "src/xnnpack/gavgpool.h",
2912 "src/xnnpack/gemm.h",
2913 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002914 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002915 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002916 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917 "src/xnnpack/lut.h",
2918 "src/xnnpack/math.h",
2919 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002920 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 "src/xnnpack/packx.h",
2922 "src/xnnpack/pad.h",
2923 "src/xnnpack/params.h",
2924 "src/xnnpack/pavgpool.h",
2925 "src/xnnpack/ppmm.h",
2926 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002927 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002928 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002929 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 "src/xnnpack/rmax.h",
2931 "src/xnnpack/scalar-utils.h",
2932 "src/xnnpack/spmm.h",
2933 "src/xnnpack/unpool.h",
2934 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002935 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002937 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002938 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002939 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002940 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002942]
2943
2944INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 "include/xnnpack.h",
2946 "src/xnnpack/allocator.h",
2947 "src/xnnpack/compute.h",
2948 "src/xnnpack/im2col.h",
2949 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002950 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002951 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952 "src/xnnpack/operator.h",
2953 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002954 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002956 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002957 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002958]
2959
Marat Dukhan1b354632020-03-23 12:50:22 -07002960ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002961 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962]
2963
Marat Dukhan1b354632020-03-23 12:50:22 -07002964MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002965 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 "include/xnnpack.h",
2967]
2968
Marat Dukhan1b354632020-03-23 12:50:22 -07002969MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002971 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002972 "src/xnnpack/requantization.h",
2973 "include/xnnpack.h",
2974]
2975
2976OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002978 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979]
2980
2981WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002983 "src/xnnpack/operator.h",
2984 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985]
2986
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002987LOGGING_COPTS = select({
2988 # No logging in optimized mode
2989 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2990 # Full logging in debug mode
2991 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2992 # Error-only logging in default (fastbuild) mode
2993 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2994})
2995
Marat Dukhan3b59de22020-06-03 20:15:19 -07002996LOGGING_SRCS = select({
2997 # No logging in optimized mode
2998 ":optimized_build": [],
2999 "//conditions:default": [
3000 "src/operator-strings.c",
3001 "src/subgraph-strings.c",
3002 ],
3003})
3004
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003005LOGGING_HDRS = [
3006 "src/xnnpack/log.h",
3007]
3008
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003010 name = "tables",
3011 srcs = TABLE_SRCS,
3012 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003013 gcc_copts = xnnpack_gcc_std_copts(),
3014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003015)
3016
3017xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 name = "scalar_ukernels",
3019 srcs = SCALAR_UKERNELS,
3020 hdrs = INTERNAL_HDRS,
3021 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003022 gcc_copts = xnnpack_gcc_std_copts(),
3023 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003025 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 "@FP16",
3027 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003028 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 ],
3030)
3031
3032xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003033 name = "scalar_ukernels_test_mode",
3034 srcs = SCALAR_UKERNELS,
3035 hdrs = INTERNAL_HDRS,
3036 aarch32_copts = ["-marm"],
3037 copts = [
3038 "-UNDEBUG",
3039 "-DXNN_TEST_MODE=1",
3040 ],
3041 gcc_copts = xnnpack_gcc_std_copts(),
3042 msvc_copts = xnnpack_msvc_std_copts(),
3043 deps = [
3044 ":tables",
3045 "@FP16",
3046 "@FXdiv",
3047 "@pthreadpool",
3048 ],
3049)
3050
3051xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003052 name = "wasm_ukernels",
3053 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003054 gcc_copts = xnnpack_gcc_std_copts(),
3055 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003056 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003057 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003058 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003059 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003060 "@FP16",
3061 "@FXdiv",
3062 "@pthreadpool",
3063 ],
3064)
3065
3066xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003067 name = "wasm_ukernels_test_mode",
3068 hdrs = INTERNAL_HDRS,
3069 copts = [
3070 "-UNDEBUG",
3071 "-DXNN_TEST_MODE=1",
3072 ],
3073 gcc_copts = xnnpack_gcc_std_copts(),
3074 msvc_copts = xnnpack_msvc_std_copts(),
3075 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003076 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003077 deps = [
3078 ":tables",
3079 "@FP16",
3080 "@FXdiv",
3081 "@pthreadpool",
3082 ],
3083)
3084
3085xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003086 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087 hdrs = INTERNAL_HDRS,
3088 aarch32_copts = [
3089 "-marm",
3090 "-mfpu=neon",
3091 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003092 gcc_copts = xnnpack_gcc_std_copts(),
3093 gcc_x86_copts = ["-msse2"],
3094 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 optimized_copts = [
3096 "-O3",
3097 "-ffast-math",
3098 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003099 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003100 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003101 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102 "@FP16",
3103 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003104 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 ],
3106)
3107
3108xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003109 name = "psimd_fastmath_ukernels_test_mode",
3110 hdrs = INTERNAL_HDRS,
3111 aarch32_copts = [
3112 "-marm",
3113 "-mfpu=neon",
3114 ],
3115 copts = [
3116 "-UNDEBUG",
3117 "-DXNN_TEST_MODE=1",
3118 ],
3119 gcc_copts = xnnpack_gcc_std_copts(),
3120 gcc_x86_copts = ["-msse2"],
3121 msvc_copts = xnnpack_msvc_std_copts(),
3122 optimized_copts = [
3123 "-O3",
3124 "-ffast-math",
3125 ],
3126 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3127 deps = [
3128 ":tables",
3129 "@FP16",
3130 "@psimd",
3131 "@pthreadpool",
3132 ],
3133)
3134
3135xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003136 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003137 hdrs = INTERNAL_HDRS,
3138 aarch32_copts = [
3139 "-marm",
3140 "-mfpu=neon",
3141 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003142 gcc_copts = xnnpack_gcc_std_copts(),
3143 gcc_x86_copts = ["-msse2"],
3144 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003145 optimized_copts = [
3146 "-O3",
3147 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003148 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003149 deps = [
3150 ":tables",
3151 "@FP16",
3152 "@psimd",
3153 "@pthreadpool",
3154 ],
3155)
3156
3157xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003158 name = "psimd_accmath_ukernels_test_mode",
3159 hdrs = INTERNAL_HDRS,
3160 aarch32_copts = [
3161 "-marm",
3162 "-mfpu=neon",
3163 ],
3164 copts = [
3165 "-UNDEBUG",
3166 "-DXNN_TEST_MODE=1",
3167 ],
3168 gcc_copts = xnnpack_gcc_std_copts(),
3169 gcc_x86_copts = ["-msse2"],
3170 msvc_copts = xnnpack_msvc_std_copts(),
3171 optimized_copts = [
3172 "-O3",
3173 ],
3174 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3175 deps = [
3176 ":tables",
3177 "@FP16",
3178 "@psimd",
3179 "@pthreadpool",
3180 ],
3181)
3182
3183xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 name = "neon_ukernels",
3185 hdrs = INTERNAL_HDRS,
3186 aarch32_copts = [
3187 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003188 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 "-mfpu=neon",
3190 ],
3191 aarch32_srcs = NEON_UKERNELS,
3192 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003193 gcc_copts = xnnpack_gcc_std_copts(),
3194 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003195 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003196 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003197 "@FP16",
3198 "@pthreadpool",
3199 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200)
3201
3202xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003203 name = "neon_ukernels_test_mode",
3204 hdrs = INTERNAL_HDRS,
3205 aarch32_copts = [
3206 "-marm",
3207 "-march=armv7-a",
3208 "-mfpu=neon",
3209 ],
3210 aarch32_srcs = NEON_UKERNELS,
3211 aarch64_srcs = NEON_UKERNELS,
3212 copts = [
3213 "-UNDEBUG",
3214 "-DXNN_TEST_MODE=1",
3215 ],
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 deps = [
3219 ":tables",
3220 "@FP16",
3221 "@pthreadpool",
3222 ],
3223)
3224
3225xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226 name = "neonfma_ukernels",
3227 hdrs = INTERNAL_HDRS,
3228 aarch32_copts = [
3229 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003230 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 "-mfpu=neon-vfpv4",
3232 ],
3233 aarch32_srcs = NEONFMA_UKERNELS,
3234 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003235 apple_aarch32_copts = [
3236 "-mcpu=swift",
3237 "-mtune=generic",
3238 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003239 gcc_copts = xnnpack_gcc_std_copts(),
3240 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003241 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003242 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003243 "@FP16",
3244 "@pthreadpool",
3245 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246)
3247
3248xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003249 name = "neonfma_ukernels_test_mode",
3250 hdrs = INTERNAL_HDRS,
3251 aarch32_copts = [
3252 "-marm",
3253 "-march=armv7-a",
3254 "-mfpu=neon-vfpv4",
3255 ],
3256 aarch32_srcs = NEONFMA_UKERNELS,
3257 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003258 apple_aarch32_copts = [
3259 "-mcpu=swift",
3260 "-mtune=generic",
3261 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003262 copts = [
3263 "-UNDEBUG",
3264 "-DXNN_TEST_MODE=1",
3265 ],
3266 gcc_copts = xnnpack_gcc_std_copts(),
3267 msvc_copts = xnnpack_msvc_std_copts(),
3268 deps = [
3269 ":tables",
3270 "@FP16",
3271 "@pthreadpool",
3272 ],
3273)
3274
3275xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003276 name = "neonv8_ukernels",
3277 hdrs = INTERNAL_HDRS,
3278 aarch32_copts = [
3279 "-marm",
3280 "-march=armv8-a",
3281 "-mfpu=neon-fp-armv8",
3282 ],
3283 aarch32_srcs = NEONV8_UKERNELS,
3284 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003285 apple_aarch32_copts = [
3286 "-mcpu=cyclone",
3287 "-mtune=generic",
3288 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003289 gcc_copts = xnnpack_gcc_std_copts(),
3290 msvc_copts = xnnpack_msvc_std_copts(),
3291 deps = [
3292 ":tables",
3293 "@FP16",
3294 "@pthreadpool",
3295 ],
3296)
3297
3298xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003299 name = "neonv8_ukernels_test_mode",
3300 hdrs = INTERNAL_HDRS,
3301 aarch32_copts = [
3302 "-marm",
3303 "-march=armv8-a",
3304 "-mfpu=neon-fp-armv8",
3305 ],
3306 aarch32_srcs = NEONV8_UKERNELS,
3307 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003308 apple_aarch32_copts = [
3309 "-mcpu=cyclone",
3310 "-mtune=generic",
3311 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003312 copts = [
3313 "-UNDEBUG",
3314 "-DXNN_TEST_MODE=1",
3315 ],
3316 gcc_copts = xnnpack_gcc_std_copts(),
3317 msvc_copts = xnnpack_msvc_std_copts(),
3318 deps = [
3319 ":tables",
3320 "@FP16",
3321 "@pthreadpool",
3322 ],
3323)
3324
3325xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003326 name = "neonfp16arith_ukernels",
3327 hdrs = INTERNAL_HDRS,
3328 aarch64_copts = ["-march=armv8.2-a+fp16"],
3329 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003330 gcc_copts = xnnpack_gcc_std_copts(),
3331 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003332 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003333 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003334 "@FP16",
3335 "@pthreadpool",
3336 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337)
3338
3339xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003340 name = "neonfp16arith_ukernels_test_mode",
3341 hdrs = INTERNAL_HDRS,
3342 aarch64_copts = ["-march=armv8.2-a+fp16"],
3343 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3344 copts = [
3345 "-UNDEBUG",
3346 "-DXNN_TEST_MODE=1",
3347 ],
3348 gcc_copts = xnnpack_gcc_std_copts(),
3349 msvc_copts = xnnpack_msvc_std_copts(),
3350 deps = [
3351 ":tables",
3352 "@FP16",
3353 "@pthreadpool",
3354 ],
3355)
3356
3357xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003358 name = "neondot_ukernels",
3359 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003360 aarch32_copts = [
3361 "-marm",
3362 "-march=armv8.2-a+dotprod",
3363 "-mfpu=neon-fp-armv8",
3364 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003365 aarch32_srcs = NEONDOT_UKERNELS,
3366 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3367 aarch64_srcs = NEONDOT_UKERNELS,
3368 gcc_copts = xnnpack_gcc_std_copts(),
3369 msvc_copts = xnnpack_msvc_std_copts(),
3370 deps = [
3371 ":tables",
3372 "@FP16",
3373 "@pthreadpool",
3374 ],
3375)
3376
3377xnnpack_cc_library(
3378 name = "neondot_ukernels_test_mode",
3379 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003380 aarch32_copts = [
3381 "-marm",
3382 "-march=armv8.2-a+dotprod",
3383 "-mfpu=neon-fp-armv8",
3384 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003385 aarch32_srcs = NEONDOT_UKERNELS,
3386 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3387 aarch64_srcs = NEONDOT_UKERNELS,
3388 copts = [
3389 "-UNDEBUG",
3390 "-DXNN_TEST_MODE=1",
3391 ],
3392 gcc_copts = xnnpack_gcc_std_copts(),
3393 msvc_copts = xnnpack_msvc_std_copts(),
3394 deps = [
3395 ":tables",
3396 "@FP16",
3397 "@pthreadpool",
3398 ],
3399)
3400
3401xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 name = "sse2_ukernels",
3403 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003404 gcc_copts = xnnpack_gcc_std_copts(),
3405 gcc_x86_copts = ["-msse2"],
3406 msvc_copts = xnnpack_msvc_std_copts(),
3407 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003408 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003409 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003410 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003411 "@FP16",
3412 "@pthreadpool",
3413 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414)
3415
3416xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003417 name = "sse2_ukernels_test_mode",
3418 hdrs = INTERNAL_HDRS,
3419 copts = [
3420 "-UNDEBUG",
3421 "-DXNN_TEST_MODE=1",
3422 ],
3423 gcc_copts = xnnpack_gcc_std_copts(),
3424 gcc_x86_copts = ["-msse2"],
3425 msvc_copts = xnnpack_msvc_std_copts(),
3426 msvc_x86_32_copts = ["/arch:SSE2"],
3427 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3428 deps = [
3429 ":tables",
3430 "@FP16",
3431 "@pthreadpool",
3432 ],
3433)
3434
3435xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003436 name = "ssse3_ukernels",
3437 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003438 gcc_copts = xnnpack_gcc_std_copts(),
3439 gcc_x86_copts = ["-mssse3"],
3440 msvc_copts = xnnpack_msvc_std_copts(),
3441 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003442 x86_srcs = SSSE3_UKERNELS,
3443 deps = [
3444 ":tables",
3445 "@FP16",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003451 name = "ssse3_ukernels_test_mode",
3452 hdrs = INTERNAL_HDRS,
3453 copts = [
3454 "-UNDEBUG",
3455 "-DXNN_TEST_MODE=1",
3456 ],
3457 gcc_copts = xnnpack_gcc_std_copts(),
3458 gcc_x86_copts = ["-mssse3"],
3459 msvc_copts = xnnpack_msvc_std_copts(),
3460 msvc_x86_32_copts = ["/arch:SSE2"],
3461 x86_srcs = SSSE3_UKERNELS,
3462 deps = [
3463 ":tables",
3464 "@FP16",
3465 "@pthreadpool",
3466 ],
3467)
3468
3469xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003470 name = "sse41_ukernels",
3471 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003472 gcc_copts = xnnpack_gcc_std_copts(),
3473 gcc_x86_copts = ["-msse4.1"],
3474 msvc_copts = xnnpack_msvc_std_copts(),
3475 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003476 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003477 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003478 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003479 "@FP16",
3480 "@pthreadpool",
3481 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003482)
3483
3484xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003485 name = "sse41_ukernels_test_mode",
3486 hdrs = INTERNAL_HDRS,
3487 copts = [
3488 "-UNDEBUG",
3489 "-DXNN_TEST_MODE=1",
3490 ],
3491 gcc_copts = xnnpack_gcc_std_copts(),
3492 gcc_x86_copts = ["-msse4.1"],
3493 msvc_copts = xnnpack_msvc_std_copts(),
3494 msvc_x86_32_copts = ["/arch:SSE2"],
3495 x86_srcs = SSE41_UKERNELS,
3496 deps = [
3497 ":tables",
3498 "@FP16",
3499 "@pthreadpool",
3500 ],
3501)
3502
3503xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504 name = "avx_ukernels",
3505 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003506 gcc_copts = xnnpack_gcc_std_copts(),
3507 gcc_x86_copts = ["-mavx"],
3508 msvc_copts = xnnpack_msvc_std_copts(),
3509 msvc_x86_32_copts = ["/arch:AVX"],
3510 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003512 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003513 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003514 "@FP16",
3515 "@pthreadpool",
3516 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517)
3518
3519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003520 name = "avx_ukernels_test_mode",
3521 hdrs = INTERNAL_HDRS,
3522 copts = [
3523 "-UNDEBUG",
3524 "-DXNN_TEST_MODE=1",
3525 ],
3526 gcc_copts = xnnpack_gcc_std_copts(),
3527 gcc_x86_copts = ["-mavx"],
3528 msvc_copts = xnnpack_msvc_std_copts(),
3529 msvc_x86_32_copts = ["/arch:AVX"],
3530 msvc_x86_64_copts = ["/arch:AVX"],
3531 x86_srcs = AVX_UKERNELS,
3532 deps = [
3533 ":tables",
3534 "@FP16",
3535 "@pthreadpool",
3536 ],
3537)
3538
3539xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003540 name = "xop_ukernels",
3541 hdrs = INTERNAL_HDRS,
3542 gcc_copts = xnnpack_gcc_std_copts(),
3543 gcc_x86_copts = ["-mxop"],
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 msvc_x86_32_copts = ["/arch:AVX"],
3546 msvc_x86_64_copts = ["/arch:AVX"],
3547 x86_srcs = XOP_UKERNELS,
3548 deps = [
3549 ":tables",
3550 "@FP16",
3551 "@pthreadpool",
3552 ],
3553)
3554
3555xnnpack_cc_library(
3556 name = "xop_ukernels_test_mode",
3557 hdrs = INTERNAL_HDRS,
3558 copts = [
3559 "-UNDEBUG",
3560 "-DXNN_TEST_MODE=1",
3561 ],
3562 gcc_copts = xnnpack_gcc_std_copts(),
3563 gcc_x86_copts = ["-mxop"],
3564 msvc_copts = xnnpack_msvc_std_copts(),
3565 msvc_x86_32_copts = ["/arch:AVX"],
3566 msvc_x86_64_copts = ["/arch:AVX"],
3567 x86_srcs = XOP_UKERNELS,
3568 deps = [
3569 ":tables",
3570 "@FP16",
3571 "@pthreadpool",
3572 ],
3573)
3574
3575xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003576 name = "fma3_ukernels",
3577 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003578 gcc_copts = xnnpack_gcc_std_copts(),
3579 gcc_x86_copts = ["-mfma"],
3580 msvc_copts = xnnpack_msvc_std_copts(),
3581 msvc_x86_32_copts = ["/arch:AVX"],
3582 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003583 x86_srcs = FMA3_UKERNELS,
3584 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003585 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003586 "@FP16",
3587 "@pthreadpool",
3588 ],
3589)
3590
3591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "fma3_ukernels_test_mode",
3593 hdrs = INTERNAL_HDRS,
3594 copts = [
3595 "-UNDEBUG",
3596 "-DXNN_TEST_MODE=1",
3597 ],
3598 gcc_copts = xnnpack_gcc_std_copts(),
3599 gcc_x86_copts = ["-mfma"],
3600 msvc_copts = xnnpack_msvc_std_copts(),
3601 msvc_x86_32_copts = ["/arch:AVX"],
3602 msvc_x86_64_copts = ["/arch:AVX"],
3603 x86_srcs = FMA3_UKERNELS,
3604 deps = [
3605 ":tables",
3606 "@FP16",
3607 "@pthreadpool",
3608 ],
3609)
3610
3611xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003612 name = "avx2_ukernels",
3613 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003614 gcc_copts = xnnpack_gcc_std_copts(),
3615 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003616 "-mfma",
3617 "-mavx2",
3618 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003619 msvc_copts = xnnpack_msvc_std_copts(),
3620 msvc_x86_32_copts = ["/arch:AVX2"],
3621 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003622 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003623 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003624 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003625 "@FP16",
3626 "@pthreadpool",
3627 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003628)
3629
3630xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003631 name = "avx2_ukernels_test_mode",
3632 hdrs = INTERNAL_HDRS,
3633 copts = [
3634 "-UNDEBUG",
3635 "-DXNN_TEST_MODE=1",
3636 ],
3637 gcc_copts = xnnpack_gcc_std_copts(),
3638 gcc_x86_copts = [
3639 "-mfma",
3640 "-mavx2",
3641 ],
3642 msvc_copts = xnnpack_msvc_std_copts(),
3643 msvc_x86_32_copts = ["/arch:AVX2"],
3644 msvc_x86_64_copts = ["/arch:AVX2"],
3645 x86_srcs = AVX2_UKERNELS,
3646 deps = [
3647 ":tables",
3648 "@FP16",
3649 "@pthreadpool",
3650 ],
3651)
3652
3653xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654 name = "avx512f_ukernels",
3655 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003656 gcc_copts = xnnpack_gcc_std_copts(),
3657 gcc_x86_copts = ["-mavx512f"],
3658 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3659 msvc_copts = xnnpack_msvc_std_copts(),
3660 msvc_x86_32_copts = ["/arch:AVX512"],
3661 msvc_x86_64_copts = ["/arch:AVX512"],
3662 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003664 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003665 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003666 "@FP16",
3667 "@pthreadpool",
3668 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669)
3670
3671xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003672 name = "avx512f_ukernels_test_mode",
3673 hdrs = INTERNAL_HDRS,
3674 copts = [
3675 "-UNDEBUG",
3676 "-DXNN_TEST_MODE=1",
3677 ],
3678 gcc_copts = xnnpack_gcc_std_copts(),
3679 gcc_x86_copts = ["-mavx512f"],
3680 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3681 msvc_copts = xnnpack_msvc_std_copts(),
3682 msvc_x86_32_copts = ["/arch:AVX512"],
3683 msvc_x86_64_copts = ["/arch:AVX512"],
3684 msys_copts = ["-fno-asynchronous-unwind-tables"],
3685 x86_srcs = AVX512F_UKERNELS,
3686 deps = [
3687 ":tables",
3688 "@FP16",
3689 "@pthreadpool",
3690 ],
3691)
3692
3693xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003694 name = "avx512skx_ukernels",
3695 hdrs = INTERNAL_HDRS,
3696 gcc_copts = xnnpack_gcc_std_copts(),
3697 gcc_x86_copts = [
3698 "-mavx512f",
3699 "-mavx512cd",
3700 "-mavx512bw",
3701 "-mavx512dq",
3702 "-mavx512vl",
3703 ],
3704 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3705 msvc_copts = xnnpack_msvc_std_copts(),
3706 msvc_x86_32_copts = ["/arch:AVX512"],
3707 msvc_x86_64_copts = ["/arch:AVX512"],
3708 msys_copts = ["-fno-asynchronous-unwind-tables"],
3709 x86_srcs = AVX512SKX_UKERNELS,
3710 deps = [
3711 ":tables",
3712 "@FP16",
3713 "@pthreadpool",
3714 ],
3715)
3716
3717xnnpack_cc_library(
3718 name = "avx512skx_ukernels_test_mode",
3719 hdrs = INTERNAL_HDRS,
3720 copts = [
3721 "-UNDEBUG",
3722 "-DXNN_TEST_MODE=1",
3723 ],
3724 gcc_copts = xnnpack_gcc_std_copts(),
3725 gcc_x86_copts = [
3726 "-mavx512f",
3727 "-mavx512cd",
3728 "-mavx512bw",
3729 "-mavx512dq",
3730 "-mavx512vl",
3731 ],
3732 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3733 msvc_copts = xnnpack_msvc_std_copts(),
3734 msvc_x86_32_copts = ["/arch:AVX512"],
3735 msvc_x86_64_copts = ["/arch:AVX512"],
3736 msys_copts = ["-fno-asynchronous-unwind-tables"],
3737 x86_srcs = AVX512SKX_UKERNELS,
3738 deps = [
3739 ":tables",
3740 "@FP16",
3741 "@pthreadpool",
3742 ],
3743)
3744
3745xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746 name = "asm_ukernels",
3747 hdrs = ["src/xnnpack/assembly.h"],
3748 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003749 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003751 wasm_srcs = WASM32_ASM_UKERNELS,
3752 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753)
3754
Marat Dukhan3b59de22020-06-03 20:15:19 -07003755xnnpack_cc_library(
3756 name = "logging_utils",
3757 srcs = LOGGING_SRCS,
3758 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3759 copts = LOGGING_COPTS + [
3760 "-Isrc",
3761 "-Iinclude",
3762 ] + select({
3763 ":debug_build": [],
3764 "//conditions:default": xnnpack_min_size_copts(),
3765 }),
3766 gcc_copts = xnnpack_gcc_std_copts(),
3767 msvc_copts = xnnpack_msvc_std_copts(),
3768 visibility = xnnpack_visibility(),
3769 deps = [
3770 "@FP16",
3771 "@clog",
3772 "@pthreadpool",
3773 ],
3774)
3775
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776xnnpack_aggregate_library(
3777 name = "ukernels",
3778 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779 ":neon_ukernels",
3780 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003781 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003782 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 ":asm_ukernels",
3784 ],
3785 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786 ":neon_ukernels",
3787 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003788 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003790 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791 ":asm_ukernels",
3792 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003793 generic_deps = [
3794 ":scalar_ukernels",
3795 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003796 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003797 ":psimd_fastmath_ukernels",
3798 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003799 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003800 wasm_deps = [
3801 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003802 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003803 ],
3804 wasmsimd_deps = [
3805 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003806 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003807 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003809 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003810 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003811 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003812 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003813 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003814 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003815 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003817 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003818 ],
3819)
3820
Marat Dukhan33fcf782020-05-24 14:27:15 -07003821xnnpack_aggregate_library(
3822 name = "ukernels_test_mode",
3823 aarch32_deps = [
3824 ":neon_ukernels_test_mode",
3825 ":neonfma_ukernels_test_mode",
3826 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003827 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003828 ":asm_ukernels",
3829 ],
3830 aarch64_deps = [
3831 ":neon_ukernels_test_mode",
3832 ":neonfma_ukernels_test_mode",
3833 ":neonv8_ukernels_test_mode",
3834 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003835 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003836 ":asm_ukernels",
3837 ],
3838 generic_deps = [
3839 ":scalar_ukernels_test_mode",
3840 ],
3841 psimd_deps = [
3842 ":psimd_fastmath_ukernels_test_mode",
3843 ":psimd_accmath_ukernels_test_mode",
3844 ],
3845 wasm_deps = [
3846 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003847 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003848 ],
3849 wasmsimd_deps = [
3850 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003851 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003852 ],
3853 x86_deps = [
3854 ":sse2_ukernels_test_mode",
3855 ":ssse3_ukernels_test_mode",
3856 ":sse41_ukernels_test_mode",
3857 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003858 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003859 ":fma3_ukernels_test_mode",
3860 ":avx2_ukernels_test_mode",
3861 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003862 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003863 ],
3864)
3865
Marat Dukhan86df3842020-06-17 14:03:31 -07003866# This target must be compatible with Android NDK r17 and compile with Clang 6.
3867# It is similar to :ukernels target, but excludes microkernels which require
3868# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003869xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003870 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003871 aarch32_deps = [
3872 ":neon_ukernels",
3873 ":neonfma_ukernels",
3874 ":neonv8_ukernels",
3875 ":asm_ukernels",
3876 ],
3877 aarch64_deps = [
3878 ":neon_ukernels",
3879 ":neonfma_ukernels",
3880 ":neonv8_ukernels",
3881 ":asm_ukernels",
3882 ],
3883 generic_deps = [
3884 ":scalar_ukernels",
3885 ],
3886 psimd_deps = [
3887 ":psimd_fastmath_ukernels",
3888 ":psimd_accmath_ukernels",
3889 ],
3890 wasm_deps = [
3891 ":wasm_ukernels",
3892 ],
3893 wasmsimd_deps = [
3894 ":wasm_ukernels",
3895 ],
3896 x86_deps = [
3897 ":sse2_ukernels",
3898 ":ssse3_ukernels",
3899 ":sse41_ukernels",
3900 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003901 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003902 ":fma3_ukernels",
3903 ":avx2_ukernels",
3904 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003905 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003906 ],
3907)
3908
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909xnnpack_cc_library(
3910 name = "im2col",
3911 srcs = ["src/im2col.c"],
3912 hdrs = [
3913 "src/xnnpack/common.h",
3914 "src/xnnpack/im2col.h",
3915 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003916 gcc_copts = xnnpack_gcc_std_copts(),
3917 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003918)
3919
3920xnnpack_cc_library(
3921 name = "indirection",
3922 srcs = ["src/indirection.c"],
3923 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003924 gcc_copts = xnnpack_gcc_std_copts(),
3925 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003926 deps = [
3927 "@FP16",
3928 "@FXdiv",
3929 "@pthreadpool",
3930 ],
3931)
3932
3933xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003934 name = "indirection_test_mode",
3935 srcs = ["src/indirection.c"],
3936 hdrs = INTERNAL_HDRS,
3937 copts = [
3938 "-UNDEBUG",
3939 "-DXNN_TEST_MODE=1",
3940 ],
3941 gcc_copts = xnnpack_gcc_std_copts(),
3942 msvc_copts = xnnpack_msvc_std_copts(),
3943 deps = [
3944 "@FP16",
3945 "@FXdiv",
3946 "@pthreadpool",
3947 ],
3948)
3949
3950xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003951 name = "packing",
3952 srcs = ["src/packing.c"],
3953 hdrs = INTERNAL_HDRS,
3954 gcc_copts = xnnpack_gcc_std_copts(),
3955 msvc_copts = xnnpack_msvc_std_copts(),
3956 deps = [
3957 "@FP16",
3958 "@FXdiv",
3959 "@pthreadpool",
3960 ],
3961)
3962
3963xnnpack_cc_library(
3964 name = "packing_test_mode",
3965 srcs = ["src/packing.c"],
3966 hdrs = INTERNAL_HDRS,
3967 copts = [
3968 "-UNDEBUG",
3969 "-DXNN_TEST_MODE=1",
3970 ],
3971 gcc_copts = xnnpack_gcc_std_copts(),
3972 msvc_copts = xnnpack_msvc_std_copts(),
3973 deps = [
3974 "@FP16",
3975 "@FXdiv",
3976 "@pthreadpool",
3977 ],
3978)
3979
3980xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 name = "operator_run",
3982 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003983 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003984 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003985 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3986 "//conditions:default": [],
3987 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003988 gcc_copts = xnnpack_gcc_std_copts(),
3989 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003991 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003992 "@FP16",
3993 "@FXdiv",
3994 "@clog",
3995 "@pthreadpool",
3996 ],
3997)
3998
Chao Mei6ddfc602020-05-13 22:29:36 -07003999xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004000 name = "operator_run_test_mode",
4001 srcs = ["src/operator-run.c"],
4002 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4003 copts = LOGGING_COPTS + [
4004 "-UNDEBUG",
4005 "-DXNN_TEST_MODE=1",
4006 ] + select({
4007 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4008 "//conditions:default": [],
4009 }),
4010 gcc_copts = xnnpack_gcc_std_copts(),
4011 msvc_copts = xnnpack_msvc_std_copts(),
4012 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004013 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004014 "@FP16",
4015 "@FXdiv",
4016 "@clog",
4017 "@pthreadpool",
4018 ],
4019)
4020
4021xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004022 name = "memory_planner",
4023 srcs = ["src/memory-planner.c"],
4024 hdrs = INTERNAL_HDRS,
4025 defines = select({
4026 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4027 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4028 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4029 }),
4030 gcc_copts = xnnpack_gcc_std_copts(),
4031 msvc_copts = xnnpack_msvc_std_copts(),
4032 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004033 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004034 "@pthreadpool",
4035 ],
4036)
4037
Marat Dukhan33fcf782020-05-24 14:27:15 -07004038xnnpack_cc_library(
4039 name = "memory_planner_test_mode",
4040 srcs = ["src/memory-planner.c"],
4041 hdrs = INTERNAL_HDRS,
4042 copts = [
4043 "-UNDEBUG",
4044 "-DXNN_TEST_MODE=1",
4045 ],
4046 defines = select({
4047 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4048 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4049 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4050 }),
4051 gcc_copts = xnnpack_gcc_std_copts(),
4052 msvc_copts = xnnpack_msvc_std_copts(),
4053 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004054 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004055 "@pthreadpool",
4056 ],
4057)
4058
Marat Dukhan08c4a432019-10-03 09:29:21 -07004059cc_library(
4060 name = "enable_assembly",
4061 defines = select({
4062 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4063 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004064 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004065 }),
4066)
4067
Marat Dukhan9de90e02020-06-18 16:04:12 -07004068cc_library(
4069 name = "enable_sparse",
4070 defines = select({
4071 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4072 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004073 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004074 }),
4075)
4076
Marat Dukhancf056b22019-10-07 10:26:29 -07004077xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078 name = "operators",
4079 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004080 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004082 ],
4083 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004084 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004085 "-Isrc",
4086 "-Iinclude",
4087 ] + select({
4088 ":debug_build": [],
4089 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004090 }) + select({
4091 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4092 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004094 gcc_copts = xnnpack_gcc_std_copts(),
4095 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004098 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004099 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004100 "@FP16",
4101 "@FXdiv",
4102 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004103 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004104 ],
4105)
4106
Marat Dukhan10a38082020-04-17 03:58:35 -07004107xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004108 name = "operators_test_mode",
4109 srcs = OPERATOR_SRCS + [
4110 "src/memory.c",
4111 "src/operator-delete.c",
4112 ],
4113 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4114 copts = LOGGING_COPTS + [
4115 "-Isrc",
4116 "-Iinclude",
4117 "-UNDEBUG",
4118 "-DXNN_TEST_MODE=1",
4119 ] + select({
4120 ":debug_build": [],
4121 "//conditions:default": xnnpack_min_size_copts(),
4122 }) + select({
4123 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4124 "//conditions:default": [],
4125 }),
4126 gcc_copts = xnnpack_gcc_std_copts(),
4127 msvc_copts = xnnpack_msvc_std_copts(),
4128 deps = [
4129 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004130 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004131 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004132 "@FP16",
4133 "@FXdiv",
4134 "@clog",
4135 "@pthreadpool",
4136 ],
4137)
4138
4139xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004140 name = "XNNPACK",
4141 srcs = [
4142 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004143 "src/runtime.c",
4144 "src/subgraph.c",
4145 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004146 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004147 hdrs = ["include/xnnpack.h"],
4148 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004149 "-Isrc",
4150 "-Iinclude",
4151 ] + select({
4152 ":debug_build": [],
4153 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004154 }) + select({
4155 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4156 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004157 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004158 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004159 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004160 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004161 visibility = xnnpack_visibility(),
4162 deps = [
4163 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004164 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004165 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004166 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004167 ":operator_run",
4168 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004169 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004170 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004171 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004172 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004173 ] + select({
4174 ":emscripten": [],
4175 "//conditions:default": ["@cpuinfo"],
4176 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177)
4178
Marat Dukhan10a38082020-04-17 03:58:35 -07004179xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004180 name = "XNNPACK_test_mode",
4181 srcs = [
4182 "src/init.c",
4183 "src/runtime.c",
4184 "src/subgraph.c",
4185 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004186 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004187 hdrs = ["include/xnnpack.h"],
4188 copts = LOGGING_COPTS + [
4189 "-Isrc",
4190 "-Iinclude",
4191 "-UNDEBUG",
4192 "-DXNN_TEST_MODE=1",
4193 ] + select({
4194 ":debug_build": [],
4195 "//conditions:default": xnnpack_min_size_copts(),
4196 }) + select({
4197 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4198 "//conditions:default": [],
4199 }),
4200 gcc_copts = xnnpack_gcc_std_copts(),
4201 includes = ["include"],
4202 msvc_copts = xnnpack_msvc_std_copts(),
4203 visibility = xnnpack_visibility(),
4204 deps = [
4205 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004206 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004207 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004208 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004209 ":operator_run_test_mode",
4210 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004211 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004212 "@clog",
4213 "@FP16",
4214 "@pthreadpool",
4215 ] + select({
4216 ":emscripten": [],
4217 "//conditions:default": ["@cpuinfo"],
4218 }),
4219)
4220
4221xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004222 name = "xnnpack_f32",
4223 srcs = [
4224 "src/init.c",
4225 "src/runtime.c",
4226 "src/subgraph.c",
4227 "src/tensor.c",
4228 ] + SUBGRAPH_SRCS,
4229 hdrs = ["include/xnnpack.h"],
4230 copts = LOGGING_COPTS + [
4231 "-Isrc",
4232 "-Iinclude",
4233 ] + select({
4234 ":debug_build": [],
4235 "//conditions:default": xnnpack_min_size_copts(),
4236 }) + select({
4237 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4238 "//conditions:default": [],
4239 }),
4240 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004241 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004242 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004243 "XNN_NO_U8_OPERATORS",
4244 "XNN_NO_X8_OPERATORS",
4245 "XNN_NO_F16_OPERATORS",
4246 "XNN_NO_X16_OPERATORS",
4247 ],
4248 gcc_copts = xnnpack_gcc_std_copts(),
4249 includes = ["include"],
4250 msvc_copts = xnnpack_msvc_std_copts(),
4251 visibility = xnnpack_visibility(),
4252 deps = [
4253 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004254 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004255 ":logging_utils",
4256 ":memory_planner",
4257 ":operator_run",
4258 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004259 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004260 "@clog",
4261 "@FP16",
4262 "@pthreadpool",
4263 ] + select({
4264 ":emscripten": [],
4265 "//conditions:default": ["@cpuinfo"],
4266 }),
4267)
4268
4269xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004270 name = "xnnpack_operators_nhwc_f32",
4271 srcs = [
4272 "src/init.c",
4273 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004274 hdrs = ["include/xnnpack.h"],
4275 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004276 "-Isrc",
4277 "-Iinclude",
4278 ] + select({
4279 ":debug_build": [],
4280 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004281 }) + select({
4282 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4283 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004284 }),
4285 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004286 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004287 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004288 "XNN_NO_U8_OPERATORS",
4289 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004290 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004291 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004292 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004294 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004295 visibility = xnnpack_visibility(),
4296 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004297 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004298 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 ":operator_run",
4300 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004301 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004302 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004304 ] + select({
4305 ":emscripten": [],
4306 "//conditions:default": ["@cpuinfo"],
4307 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308)
4309
Marat Dukhancf056b22019-10-07 10:26:29 -07004310xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004311 name = "bench_utils",
4312 srcs = ["bench/utils.cc"],
4313 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004314 deps = [
4315 "@com_google_benchmark//:benchmark",
4316 "@cpuinfo",
4317 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318)
4319
Frank Barchard7e955972019-10-11 10:34:25 -07004320######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004321
4322xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004323 name = "qs8_gemm_bench",
4324 srcs = [
4325 "bench/gemm.h",
4326 "bench/qs8-gemm.cc",
4327 "src/xnnpack/AlignedAllocator.h",
4328 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004329 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4330 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004331)
4332
4333xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004334 name = "qs8_requantization_bench",
4335 srcs = [
4336 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004337 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004338 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004339 ] + MICROKERNEL_BENCHMARK_HDRS,
4340 deps = MICROKERNEL_BENCHMARK_DEPS,
4341)
4342
4343xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004344 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 srcs = [
4346 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004347 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348 "src/xnnpack/AlignedAllocator.h",
4349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004350 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004351 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352)
4353
4354xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004355 name = "qu8_requantization_bench",
4356 srcs = [
4357 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004358 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004359 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004360 ] + MICROKERNEL_BENCHMARK_HDRS,
4361 deps = MICROKERNEL_BENCHMARK_DEPS,
4362)
4363
4364xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004365 name = "f16_igemm_bench",
4366 srcs = [
4367 "bench/f16-igemm.cc",
4368 "bench/conv.h",
4369 "bench/google/conv.h",
4370 "src/xnnpack/AlignedAllocator.h",
4371 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004372 deps = MICROKERNEL_BENCHMARK_DEPS + [
4373 ":indirection",
4374 ":packing",
4375 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004376)
4377
4378xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 name = "f16_gemm_bench",
4380 srcs = [
4381 "bench/f16-gemm.cc",
4382 "bench/gemm.h",
4383 "src/xnnpack/AlignedAllocator.h",
4384 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004385 deps = MICROKERNEL_BENCHMARK_DEPS + [
4386 ":packing",
4387 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388)
4389
4390xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004391 name = "f16_spmm_bench",
4392 srcs = [
4393 "bench/f16-spmm.cc",
4394 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004395 "src/xnnpack/AlignedAllocator.h",
4396 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004397 deps = MICROKERNEL_BENCHMARK_DEPS,
4398)
4399
4400xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401 name = "f32_igemm_bench",
4402 srcs = [
4403 "bench/f32-igemm.cc",
4404 "bench/conv.h",
4405 "src/xnnpack/AlignedAllocator.h",
4406 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004407 deps = MICROKERNEL_BENCHMARK_DEPS + [
4408 ":indirection",
4409 ":packing",
4410 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004411)
4412
4413xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004414 name = "f16_relu_bench",
4415 srcs = [
4416 "bench/f16-relu.cc",
4417 "src/xnnpack/AlignedAllocator.h",
4418 ] + MICROKERNEL_BENCHMARK_HDRS,
4419 deps = MICROKERNEL_BENCHMARK_DEPS,
4420)
4421
4422xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423 name = "f32_conv_hwc_bench",
4424 srcs = [
4425 "bench/f32-conv-hwc.cc",
4426 "bench/dconv.h",
4427 "src/xnnpack/AlignedAllocator.h",
4428 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004429 deps = MICROKERNEL_BENCHMARK_DEPS + [
4430 ":packing",
4431 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432)
4433
4434xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004435 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004436 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004437 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004438 "bench/dconv.h",
4439 "src/xnnpack/AlignedAllocator.h",
4440 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004441 deps = MICROKERNEL_BENCHMARK_DEPS + [
4442 ":packing",
4443 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004444)
4445
4446xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004447 name = "f16_dwconv_bench",
4448 srcs = [
4449 "bench/f16-dwconv.cc",
4450 "bench/dwconv.h",
4451 "bench/google/dwconv.h",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004454 deps = MICROKERNEL_BENCHMARK_DEPS + [
4455 ":indirection",
4456 ":packing",
4457 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004458)
4459
4460xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 name = "f32_dwconv_bench",
4462 srcs = [
4463 "bench/f32-dwconv.cc",
4464 "bench/dwconv.h",
4465 "src/xnnpack/AlignedAllocator.h",
4466 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004467 deps = MICROKERNEL_BENCHMARK_DEPS + [
4468 ":indirection",
4469 ":packing",
4470 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471)
4472
4473xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004474 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004476 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477 "bench/dwconv.h",
4478 "src/xnnpack/AlignedAllocator.h",
4479 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004480 deps = MICROKERNEL_BENCHMARK_DEPS + [
4481 ":indirection",
4482 ":packing",
4483 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004484)
4485
4486xnnpack_benchmark(
4487 name = "f32_gemm_bench",
4488 srcs = [
4489 "bench/f32-gemm.cc",
4490 "bench/gemm.h",
4491 "src/xnnpack/AlignedAllocator.h",
4492 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004493 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004494 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495)
4496
4497xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004498 name = "f32_hswish_bench",
4499 srcs = [
4500 "bench/f32-hswish.cc",
4501 "src/xnnpack/AlignedAllocator.h",
4502 ] + MICROKERNEL_BENCHMARK_HDRS,
4503 deps = MICROKERNEL_BENCHMARK_DEPS,
4504)
4505
4506xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004507 name = "f32_raddexpminusmax_bench",
4508 srcs = [
4509 "bench/f32-raddexpminusmax.cc",
4510 "src/xnnpack/AlignedAllocator.h",
4511 ] + MICROKERNEL_BENCHMARK_HDRS,
4512 deps = MICROKERNEL_BENCHMARK_DEPS,
4513)
4514
4515xnnpack_benchmark(
4516 name = "f32_raddextexp_bench",
4517 srcs = [
4518 "bench/f32-raddextexp.cc",
4519 "src/xnnpack/AlignedAllocator.h",
4520 ] + MICROKERNEL_BENCHMARK_HDRS,
4521 deps = MICROKERNEL_BENCHMARK_DEPS,
4522)
4523
4524xnnpack_benchmark(
4525 name = "f32_raddstoreexpminusmax_bench",
4526 srcs = [
4527 "bench/f32-raddstoreexpminusmax.cc",
4528 "src/xnnpack/AlignedAllocator.h",
4529 ] + MICROKERNEL_BENCHMARK_HDRS,
4530 deps = MICROKERNEL_BENCHMARK_DEPS,
4531)
4532
4533xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004534 name = "f32_relu_bench",
4535 srcs = [
4536 "bench/f32-relu.cc",
4537 "src/xnnpack/AlignedAllocator.h",
4538 ] + MICROKERNEL_BENCHMARK_HDRS,
4539 deps = MICROKERNEL_BENCHMARK_DEPS,
4540)
4541
4542xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 name = "f32_rmax_bench",
4544 srcs = [
4545 "bench/f32-rmax.cc",
4546 "src/xnnpack/AlignedAllocator.h",
4547 ] + MICROKERNEL_BENCHMARK_HDRS,
4548 deps = MICROKERNEL_BENCHMARK_DEPS,
4549)
4550
4551xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004552 name = "f32_sigmoid_bench",
4553 srcs = [
4554 "bench/f32-sigmoid.cc",
4555 "src/xnnpack/AlignedAllocator.h",
4556 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004557 deps = MICROKERNEL_BENCHMARK_DEPS,
4558)
4559
4560xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004561 name = "f32_spmm_bench",
4562 srcs = [
4563 "bench/f32-spmm.cc",
4564 "bench/gemm.h",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567 deps = MICROKERNEL_BENCHMARK_DEPS,
4568)
4569
4570xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004571 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004572 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004573 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004574 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004575 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004576 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004577)
4578
4579xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004580 name = "f32_vscaleexpminusmax_bench",
4581 srcs = [
4582 "bench/f32-vscaleexpminusmax.cc",
4583 "src/xnnpack/AlignedAllocator.h",
4584 ] + MICROKERNEL_BENCHMARK_HDRS,
4585 deps = MICROKERNEL_BENCHMARK_DEPS,
4586)
4587
4588xnnpack_benchmark(
4589 name = "f32_vscaleextexp_bench",
4590 srcs = [
4591 "bench/f32-vscaleextexp.cc",
4592 "src/xnnpack/AlignedAllocator.h",
4593 ] + MICROKERNEL_BENCHMARK_HDRS,
4594 deps = MICROKERNEL_BENCHMARK_DEPS,
4595)
4596
4597xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004598 name = "f32_vsqrt_bench",
4599 srcs = [
4600 "bench/f32-vsqrt.cc",
4601 "src/xnnpack/AlignedAllocator.h",
4602 ] + MICROKERNEL_BENCHMARK_HDRS,
4603 deps = MICROKERNEL_BENCHMARK_DEPS,
4604)
4605
4606xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004607 name = "f32_im2col_gemm_bench",
4608 srcs = [
4609 "bench/f32-im2col-gemm.cc",
4610 "bench/conv.h",
4611 "src/xnnpack/AlignedAllocator.h",
4612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004613 deps = MICROKERNEL_BENCHMARK_DEPS + [
4614 ":im2col",
4615 ":packing",
4616 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004617)
4618
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004619xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004620 name = "rounding_bench",
4621 srcs = [
4622 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004623 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004624 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004625 ] + MICROKERNEL_BENCHMARK_HDRS,
4626 deps = MICROKERNEL_BENCHMARK_DEPS,
4627)
4628
Marat Dukhan08c4a432019-10-03 09:29:21 -07004629########################### Benchmarks for operators ###########################
4630
4631xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 name = "average_pooling_bench",
4633 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004634 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004635 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004636 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004637)
4638
4639xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004640 name = "bankers_rounding_bench",
4641 srcs = ["bench/bankers-rounding.cc"],
4642 copts = xnnpack_optional_tflite_copts(),
4643 tags = ["nowin32"],
4644 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4645)
4646
4647xnnpack_benchmark(
4648 name = "ceiling_bench",
4649 srcs = ["bench/ceiling.cc"],
4650 copts = xnnpack_optional_tflite_copts(),
4651 tags = ["nowin32"],
4652 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4653)
4654
4655xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656 name = "channel_shuffle_bench",
4657 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004658 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659)
4660
4661xnnpack_benchmark(
4662 name = "convolution_bench",
4663 srcs = ["bench/convolution.cc"],
4664 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004665 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004666 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667)
4668
4669xnnpack_benchmark(
4670 name = "deconvolution_bench",
4671 srcs = ["bench/deconvolution.cc"],
4672 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004673 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004674 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
4677xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004678 name = "floor_bench",
4679 srcs = ["bench/floor.cc"],
4680 copts = xnnpack_optional_tflite_copts(),
4681 tags = ["nowin32"],
4682 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4683)
4684
4685xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004686 name = "global_average_pooling_bench",
4687 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004688 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004689)
4690
4691xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004692 name = "hardswish_bench",
4693 srcs = ["bench/hardswish.cc"],
4694 copts = xnnpack_optional_tflite_copts(),
4695 tags = ["nowin32"],
4696 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4697)
4698
4699xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004700 name = "max_pooling_bench",
4701 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004702 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004703)
4704
4705xnnpack_benchmark(
4706 name = "sigmoid_bench",
4707 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004708 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004709 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004710 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711)
4712
4713xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004714 name = "prelu_bench",
4715 srcs = ["bench/prelu.cc"],
4716 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004717 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004718 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004719)
4720
4721xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004722 name = "softmax_bench",
4723 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004724 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004725 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004726 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004727)
4728
Marat Dukhan87727142020-06-24 15:24:10 -07004729xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004730 name = "square_root_bench",
4731 srcs = ["bench/square-root.cc"],
4732 copts = xnnpack_optional_tflite_copts(),
4733 tags = ["nowin32"],
4734 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4735)
4736
4737xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004738 name = "truncation_bench",
4739 srcs = ["bench/truncation.cc"],
4740 deps = OPERATOR_BENCHMARK_DEPS,
4741)
4742
Marat Dukhanc068bb62019-10-04 13:24:39 -07004743############################# End-to-end benchmarks ############################
4744
4745cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004746 name = "fp32_mobilenet_v1",
4747 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004748 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004749 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004750 linkstatic = True,
4751 deps = [
4752 ":XNNPACK",
4753 "@pthreadpool",
4754 ],
4755)
4756
4757cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004758 name = "fp16_mobilenet_v1",
4759 srcs = ["models/fp16-mobilenet-v1.cc"],
4760 hdrs = ["models/models.h"],
4761 copts = xnnpack_std_cxxopts(),
4762 linkstatic = True,
4763 deps = [
4764 ":XNNPACK",
4765 "@FP16",
4766 "@pthreadpool",
4767 ],
4768)
4769
4770cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004771 name = "qs8_mobilenet_v1",
4772 srcs = ["models/qs8-mobilenet-v1.cc"],
4773 hdrs = ["models/models.h"],
4774 copts = xnnpack_std_cxxopts(),
4775 linkstatic = True,
4776 deps = [
4777 ":XNNPACK",
4778 "@pthreadpool",
4779 ],
4780)
4781
4782cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004783 name = "qs8_mobilenet_v2",
4784 srcs = ["models/qs8-mobilenet-v2.cc"],
4785 hdrs = ["models/models.h"],
4786 copts = xnnpack_std_cxxopts(),
4787 linkstatic = True,
4788 deps = [
4789 ":XNNPACK",
4790 "@pthreadpool",
4791 ],
4792)
4793
4794cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004795 name = "fp32_mobilenet_v2",
4796 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004797 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004798 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004799 linkstatic = True,
4800 deps = [
4801 ":XNNPACK",
4802 "@pthreadpool",
4803 ],
4804)
4805
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004806cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004807 name = "fp16_mobilenet_v2",
4808 srcs = ["models/fp16-mobilenet-v2.cc"],
4809 hdrs = ["models/models.h"],
4810 copts = xnnpack_std_cxxopts(),
4811 linkstatic = True,
4812 deps = [
4813 ":XNNPACK",
4814 "@FP16",
4815 "@pthreadpool",
4816 ],
4817)
4818
4819cc_library(
4820 name = "fp32_mobilenet_v3_large",
4821 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004822 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004823 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004824 linkstatic = True,
4825 deps = [
4826 ":XNNPACK",
4827 "@pthreadpool",
4828 ],
4829)
4830
4831cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004832 name = "fp16_mobilenet_v3_large",
4833 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4834 hdrs = ["models/models.h"],
4835 copts = xnnpack_std_cxxopts(),
4836 linkstatic = True,
4837 deps = [
4838 ":XNNPACK",
4839 "@FP16",
4840 "@pthreadpool",
4841 ],
4842)
4843
4844cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004845 name = "fp32_mobilenet_v3_small",
4846 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004847 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004848 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004849 linkstatic = True,
4850 deps = [
4851 ":XNNPACK",
4852 "@pthreadpool",
4853 ],
4854)
4855
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004856cc_library(
4857 name = "fp16_mobilenet_v3_small",
4858 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4859 hdrs = ["models/models.h"],
4860 copts = xnnpack_std_cxxopts(),
4861 linkstatic = True,
4862 deps = [
4863 ":XNNPACK",
4864 "@FP16",
4865 "@pthreadpool",
4866 ],
4867)
4868
Marat Dukhanc068bb62019-10-04 13:24:39 -07004869xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004870 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004871 srcs = [
4872 "bench/f32-dwconv-e2e.cc",
4873 "bench/end2end.h",
4874 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004875 deps = MICROKERNEL_BENCHMARK_DEPS + [
4876 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004877 ":fp32_mobilenet_v1",
4878 ":fp32_mobilenet_v2",
4879 ":fp32_mobilenet_v3_large",
4880 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004881 ],
4882)
4883
4884xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004885 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004886 srcs = [
4887 "bench/f32-gemm-e2e.cc",
4888 "bench/end2end.h",
4889 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004890 deps = MICROKERNEL_BENCHMARK_DEPS + [
4891 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004892 ":fp32_mobilenet_v1",
4893 ":fp32_mobilenet_v2",
4894 ":fp32_mobilenet_v3_large",
4895 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004896 ],
4897)
4898
4899xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004900 name = "end2end_bench",
4901 srcs = ["bench/end2end.cc"],
4902 deps = [
4903 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004904 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004905 ":fp16_mobilenet_v1",
4906 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004907 ":fp16_mobilenet_v3_large",
4908 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004909 ":fp32_mobilenet_v1",
4910 ":fp32_mobilenet_v2",
4911 ":fp32_mobilenet_v3_large",
4912 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004913 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004914 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004915 "@pthreadpool",
4916 ],
4917)
4918
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004919#################### Accuracy evaluation for math functions ####################
4920
4921xnnpack_benchmark(
4922 name = "f32_exp_eval",
4923 srcs = [
4924 "eval/f32-exp.cc",
4925 "src/xnnpack/AlignedAllocator.h",
4926 ] + ACCURACY_EVAL_HDRS,
4927 deps = ACCURACY_EVAL_DEPS,
4928)
4929
Marat Dukhan515c9772019-10-17 18:07:57 -07004930xnnpack_benchmark(
4931 name = "f32_expminus_eval",
4932 srcs = [
4933 "eval/f32-expminus.cc",
4934 "src/xnnpack/AlignedAllocator.h",
4935 ] + ACCURACY_EVAL_HDRS,
4936 deps = ACCURACY_EVAL_DEPS,
4937)
4938
Marat Dukhan98ba4412019-10-23 02:14:28 -07004939xnnpack_benchmark(
4940 name = "f32_extexp_eval",
4941 srcs = [
4942 "eval/f32-extexp.cc",
4943 "src/xnnpack/AlignedAllocator.h",
4944 ] + ACCURACY_EVAL_HDRS,
4945 deps = ACCURACY_EVAL_DEPS,
4946)
4947
Marat Dukhan8853b822020-05-07 12:19:01 -07004948xnnpack_unit_test(
4949 name = "f32_roundne_eval",
4950 srcs = [
4951 "eval/f32-roundne.cc",
4952 "src/xnnpack/AlignedAllocator.h",
4953 "src/xnnpack/math-stubs.h",
4954 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004955 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004956 deps = MICROKERNEL_TEST_DEPS,
4957)
4958
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004959xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004960 name = "f32_roundd_eval",
4961 srcs = [
4962 "eval/f32-roundd.cc",
4963 "src/xnnpack/AlignedAllocator.h",
4964 "src/xnnpack/math-stubs.h",
4965 ] + MICROKERNEL_TEST_HDRS,
4966 automatic = False,
4967 deps = MICROKERNEL_TEST_DEPS,
4968)
4969
4970xnnpack_unit_test(
4971 name = "f32_roundu_eval",
4972 srcs = [
4973 "eval/f32-roundu.cc",
4974 "src/xnnpack/AlignedAllocator.h",
4975 "src/xnnpack/math-stubs.h",
4976 ] + MICROKERNEL_TEST_HDRS,
4977 automatic = False,
4978 deps = MICROKERNEL_TEST_DEPS,
4979)
4980
4981xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004982 name = "f32_roundz_eval",
4983 srcs = [
4984 "eval/f32-roundz.cc",
4985 "src/xnnpack/AlignedAllocator.h",
4986 "src/xnnpack/math-stubs.h",
4987 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004988 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004989 deps = MICROKERNEL_TEST_DEPS,
4990)
4991
Marat Dukhan346a9e52019-11-15 09:06:30 -08004992xnnpack_benchmark(
4993 name = "f32_sigmoid_eval",
4994 srcs = [
4995 "eval/f32-sigmoid.cc",
4996 "src/xnnpack/AlignedAllocator.h",
4997 ] + ACCURACY_EVAL_HDRS,
4998 deps = ACCURACY_EVAL_DEPS,
4999)
5000
Marat Dukhan84000762020-06-29 18:38:43 -07005001xnnpack_benchmark(
5002 name = "f32_sqrt_eval",
5003 srcs = [
5004 "eval/f32-sqrt.cc",
5005 "src/xnnpack/AlignedAllocator.h",
5006 ] + ACCURACY_EVAL_HDRS,
5007 deps = ACCURACY_EVAL_DEPS + [
5008 ":bench_utils",
5009 ],
5010)
5011
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012######################### Unit tests for micro-kernels #########################
5013
5014xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005015 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005017 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018 "test/gemm-microkernel-tester.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005021 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005022)
5023
5024xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005025 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005026 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005027 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005028 "test/spmm-microkernel-tester.h",
5029 "src/xnnpack/AlignedAllocator.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 name = "f32_argmaxpool_test",
5036 srcs = [
5037 "test/f32-argmaxpool.cc",
5038 "test/argmaxpool-microkernel-tester.h",
5039 "src/xnnpack/AlignedAllocator.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005045 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005047 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005048 "test/avgpool-microkernel-tester.h",
5049 "src/xnnpack/AlignedAllocator.h",
5050 ] + MICROKERNEL_TEST_HDRS,
5051 deps = MICROKERNEL_TEST_DEPS,
5052)
5053
5054xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005055 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005056 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005057 "test/f32-ibilinear.cc",
5058 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005059 "src/xnnpack/AlignedAllocator.h",
5060 ] + MICROKERNEL_TEST_HDRS,
5061 deps = MICROKERNEL_TEST_DEPS,
5062)
5063
5064xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005065 name = "f32_ibilinear_chw_test",
5066 srcs = [
5067 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005068 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005069 "src/xnnpack/AlignedAllocator.h",
5070 ] + MICROKERNEL_TEST_HDRS,
5071 deps = MICROKERNEL_TEST_DEPS,
5072)
5073
5074xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005075 name = "f16_clamp_test",
5076 srcs = [
5077 "test/f16-clamp.cc",
5078 "test/clamp-microkernel-tester.h",
5079 ] + MICROKERNEL_TEST_HDRS,
5080 deps = MICROKERNEL_TEST_DEPS,
5081)
5082
5083xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 name = "f32_clamp_test",
5085 srcs = [
5086 "test/f32-clamp.cc",
5087 "test/clamp-microkernel-tester.h",
5088 ] + MICROKERNEL_TEST_HDRS,
5089 deps = MICROKERNEL_TEST_DEPS,
5090)
5091
5092xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005093 name = "f32_igemm_test",
5094 srcs = [
5095 "test/f32-igemm.cc",
5096 "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 Dukhan163a7e62020-04-09 04:19:26 -07005100)
5101
5102xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005103 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005104 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005105 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005106 "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 Dukhan08c4a432019-10-03 09:29:21 -07005110)
5111
5112xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005113 name = "f32_igemm_minmax_test",
5114 srcs = [
5115 "test/f32-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"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005120)
5121
5122xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005123 name = "f16_igemm_minmax_test",
5124 srcs = [
5125 "test/f16-igemm-minmax.cc",
5126 "test/gemm-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"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005130)
5131
5132xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005133 name = "f32_conv_hwc_test",
5134 srcs = [
5135 "test/f32-conv-hwc.cc",
5136 "test/conv-hwc-microkernel-tester.h",
5137 "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 Dukhan1f29b802020-05-15 23:46:39 -07005143 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005145 "test/f32-conv-hwc2chw.cc",
5146 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005147 "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 Dukhan08c4a432019-10-03 09:29:21 -07005150)
5151
5152xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005153 name = "f32_dwconv_test",
5154 srcs = [
5155 "test/f32-dwconv.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"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005160)
5161
5162xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005163 name = "f16_dwconv_minmax_test",
5164 srcs = [
5165 "test/f16-dwconv-minmax.cc",
5166 "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"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005170)
5171
5172xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005173 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005175 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005176 "test/dwconv-microkernel-tester.h",
5177 "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(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005183 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005184 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005185 "test/f32-dwconv2d-chw.cc",
5186 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 "src/xnnpack/AlignedAllocator.h",
5188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005189 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005193 name = "f16_gavgpool_minmax_test",
5194 srcs = [
5195 "test/f16-gavgpool-minmax.cc",
5196 "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 Dukhan99936602020-04-11 16:47:01 -07005203 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005205 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 "test/gavgpool-microkernel-tester.h",
5207 "src/xnnpack/AlignedAllocator.h",
5208 ] + MICROKERNEL_TEST_HDRS,
5209 deps = MICROKERNEL_TEST_DEPS,
5210)
5211
5212xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005213 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005214 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005215 "test/f32-gavgpool-cw.cc",
5216 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 "src/xnnpack/AlignedAllocator.h",
5218 ] + MICROKERNEL_TEST_HDRS,
5219 deps = MICROKERNEL_TEST_DEPS,
5220)
5221
5222xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005223 name = "f32_gemm_test",
5224 srcs = [
5225 "test/f32-gemm.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 Dukhan163a7e62020-04-09 04:19:26 -07005230)
5231
5232xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005233 name = "f32_gemm_relu_test",
5234 srcs = [
5235 "test/f32-gemm-relu.cc",
5236 "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 Dukhan467f6362020-05-22 23:21:55 -07005240)
5241
5242xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005243 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005245 "test/f32-gemm-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(
Marat Dukhan1c587112020-04-08 20:04:28 -07005253 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005255 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256 "test/gemm-microkernel-tester.h",
5257 "src/xnnpack/AlignedAllocator.h",
5258 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005259 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260)
5261
5262xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005263 name = "f16_hswish_test",
5264 srcs = [
5265 "test/f16-hswish.cc",
5266 "test/hswish-microkernel-tester.h",
5267 ] + MICROKERNEL_TEST_HDRS,
5268 deps = MICROKERNEL_TEST_DEPS,
5269)
5270
5271xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 name = "f32_hswish_test",
5273 srcs = [
5274 "test/f32-hswish.cc",
5275 "test/hswish-microkernel-tester.h",
5276 ] + MICROKERNEL_TEST_HDRS,
5277 deps = MICROKERNEL_TEST_DEPS,
5278)
5279
5280xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005281 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005283 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 "test/maxpool-microkernel-tester.h",
5285 ] + MICROKERNEL_TEST_HDRS,
5286 deps = MICROKERNEL_TEST_DEPS,
5287)
5288
5289xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005290 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005292 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 "test/avgpool-microkernel-tester.h",
5294 "src/xnnpack/AlignedAllocator.h",
5295 ] + MICROKERNEL_TEST_HDRS,
5296 deps = MICROKERNEL_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005300 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005301 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005302 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 "test/gemm-microkernel-tester.h",
5304 "src/xnnpack/AlignedAllocator.h",
5305 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005306 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307)
5308
5309xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005310 name = "f16_prelu_test",
5311 srcs = [
5312 "test/f16-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 Dukhan08c4a432019-10-03 09:29:21 -07005320 name = "f32_prelu_test",
5321 srcs = [
5322 "test/f32-prelu.cc",
5323 "test/prelu-microkernel-tester.h",
5324 "src/xnnpack/AlignedAllocator.h",
5325 ] + MICROKERNEL_TEST_HDRS,
5326 deps = MICROKERNEL_TEST_DEPS,
5327)
5328
5329xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005330 name = "f32_raddexpminusmax_test",
5331 srcs = [
5332 "test/f32-raddexpminusmax.cc",
5333 "test/raddexpminusmax-microkernel-tester.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005339 name = "f32_raddextexp_test",
5340 srcs = [
5341 "test/f32-raddextexp.cc",
5342 "test/raddextexp-microkernel-tester.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005348 name = "f32_raddstoreexpminusmax_test",
5349 srcs = [
5350 "test/f32-raddstoreexpminusmax.cc",
5351 "test/raddstoreexpminusmax-microkernel-tester.h",
5352 ] + MICROKERNEL_TEST_HDRS,
5353 deps = MICROKERNEL_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005357 name = "f32_relu_test",
5358 srcs = [
5359 "test/f32-relu.cc",
5360 "test/vunary-microkernel-tester.h",
5361 ] + MICROKERNEL_TEST_HDRS,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
5365xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 name = "f32_rmax_test",
5367 srcs = [
5368 "test/f32-rmax.cc",
5369 "test/rmax-microkernel-tester.h",
5370 ] + MICROKERNEL_TEST_HDRS,
5371 deps = MICROKERNEL_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005375 name = "f32_sigmoid_test",
5376 srcs = [
5377 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005378 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005384 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005386 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005387 "test/spmm-microkernel-tester.h",
5388 "src/xnnpack/AlignedAllocator.h",
5389 ] + MICROKERNEL_TEST_HDRS,
5390 deps = MICROKERNEL_TEST_DEPS,
5391)
5392
5393xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005394 name = "f16_vadd_minmax_test",
5395 srcs = [
5396 "test/f16-vadd-minmax.cc",
5397 "test/vbinary-microkernel-tester.h",
5398 ] + MICROKERNEL_TEST_HDRS,
5399 deps = MICROKERNEL_TEST_DEPS,
5400)
5401
5402xnnpack_unit_test(
5403 name = "f16_vaddc_minmax_test",
5404 srcs = [
5405 "test/f16-vaddc-minmax.cc",
5406 "test/vbinaryc-microkernel-tester.h",
5407 ] + MICROKERNEL_TEST_HDRS,
5408 deps = MICROKERNEL_TEST_DEPS,
5409)
5410
5411xnnpack_unit_test(
5412 name = "f16_vdiv_minmax_test",
5413 srcs = [
5414 "test/f16-vdiv-minmax.cc",
5415 "test/vbinary-microkernel-tester.h",
5416 ] + MICROKERNEL_TEST_HDRS,
5417 deps = MICROKERNEL_TEST_DEPS,
5418)
5419
5420xnnpack_unit_test(
5421 name = "f16_vdivc_minmax_test",
5422 srcs = [
5423 "test/f16-vdivc-minmax.cc",
5424 "test/vbinaryc-microkernel-tester.h",
5425 ] + MICROKERNEL_TEST_HDRS,
5426 deps = MICROKERNEL_TEST_DEPS,
5427)
5428
5429xnnpack_unit_test(
5430 name = "f16_vrdivc_minmax_test",
5431 srcs = [
5432 "test/f16-vrdivc-minmax.cc",
5433 "test/vbinaryc-microkernel-tester.h",
5434 ] + MICROKERNEL_TEST_HDRS,
5435 deps = MICROKERNEL_TEST_DEPS,
5436)
5437
5438xnnpack_unit_test(
5439 name = "f16_vmax_test",
5440 srcs = [
5441 "test/f16-vmax.cc",
5442 "test/vbinary-microkernel-tester.h",
5443 ] + MICROKERNEL_TEST_HDRS,
5444 deps = MICROKERNEL_TEST_DEPS,
5445)
5446
5447xnnpack_unit_test(
5448 name = "f16_vmaxc_test",
5449 srcs = [
5450 "test/f16-vmaxc.cc",
5451 "test/vbinaryc-microkernel-tester.h",
5452 ] + MICROKERNEL_TEST_HDRS,
5453 deps = MICROKERNEL_TEST_DEPS,
5454)
5455
5456xnnpack_unit_test(
5457 name = "f16_vmin_test",
5458 srcs = [
5459 "test/f16-vmin.cc",
5460 "test/vbinary-microkernel-tester.h",
5461 ] + MICROKERNEL_TEST_HDRS,
5462 deps = MICROKERNEL_TEST_DEPS,
5463)
5464
5465xnnpack_unit_test(
5466 name = "f16_vminc_test",
5467 srcs = [
5468 "test/f16-vminc.cc",
5469 "test/vbinaryc-microkernel-tester.h",
5470 ] + MICROKERNEL_TEST_HDRS,
5471 deps = MICROKERNEL_TEST_DEPS,
5472)
5473
5474xnnpack_unit_test(
5475 name = "f16_vmul_minmax_test",
5476 srcs = [
5477 "test/f16-vmul-minmax.cc",
5478 "test/vbinary-microkernel-tester.h",
5479 ] + MICROKERNEL_TEST_HDRS,
5480 deps = MICROKERNEL_TEST_DEPS,
5481)
5482
5483xnnpack_unit_test(
5484 name = "f16_vmulc_minmax_test",
5485 srcs = [
5486 "test/f16-vmulc-minmax.cc",
5487 "test/vbinaryc-microkernel-tester.h",
5488 ] + MICROKERNEL_TEST_HDRS,
5489 deps = MICROKERNEL_TEST_DEPS,
5490)
5491
5492xnnpack_unit_test(
5493 name = "f16_vsub_minmax_test",
5494 srcs = [
5495 "test/f16-vsub-minmax.cc",
5496 "test/vbinary-microkernel-tester.h",
5497 ] + MICROKERNEL_TEST_HDRS,
5498 deps = MICROKERNEL_TEST_DEPS,
5499)
5500
5501xnnpack_unit_test(
5502 name = "f16_vsubc_minmax_test",
5503 srcs = [
5504 "test/f16-vsubc-minmax.cc",
5505 "test/vbinaryc-microkernel-tester.h",
5506 ] + MICROKERNEL_TEST_HDRS,
5507 deps = MICROKERNEL_TEST_DEPS,
5508)
5509
5510xnnpack_unit_test(
5511 name = "f16_vrsubc_minmax_test",
5512 srcs = [
5513 "test/f16-vrsubc-minmax.cc",
5514 "test/vbinaryc-microkernel-tester.h",
5515 ] + MICROKERNEL_TEST_HDRS,
5516 deps = MICROKERNEL_TEST_DEPS,
5517)
5518
5519xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005520 name = "f32_vabs_test",
5521 srcs = [
5522 "test/f32-vabs.cc",
5523 "test/vunary-microkernel-tester.h",
5524 ] + MICROKERNEL_TEST_HDRS,
5525 deps = MICROKERNEL_TEST_DEPS,
5526)
5527
5528xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005529 name = "f32_vadd_test",
5530 srcs = [
5531 "test/f32-vadd.cc",
5532 "test/vbinary-microkernel-tester.h",
5533 ] + MICROKERNEL_TEST_HDRS,
5534 deps = MICROKERNEL_TEST_DEPS,
5535)
5536
5537xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005538 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005539 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005540 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005541 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005542 ] + MICROKERNEL_TEST_HDRS,
5543 deps = MICROKERNEL_TEST_DEPS,
5544)
5545
5546xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005547 name = "f32_vadd_relu_test",
5548 srcs = [
5549 "test/f32-vadd-relu.cc",
5550 "test/vbinary-microkernel-tester.h",
5551 ] + MICROKERNEL_TEST_HDRS,
5552 deps = MICROKERNEL_TEST_DEPS,
5553)
5554
5555xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005556 name = "f32_vaddc_test",
5557 srcs = [
5558 "test/f32-vaddc.cc",
5559 "test/vbinaryc-microkernel-tester.h",
5560 ] + MICROKERNEL_TEST_HDRS,
5561 deps = MICROKERNEL_TEST_DEPS,
5562)
5563
5564xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005565 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005566 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005567 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005568 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005569 ] + MICROKERNEL_TEST_HDRS,
5570 deps = MICROKERNEL_TEST_DEPS,
5571)
5572
5573xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005574 name = "f32_vaddc_relu_test",
5575 srcs = [
5576 "test/f32-vaddc-relu.cc",
5577 "test/vbinaryc-microkernel-tester.h",
5578 ] + MICROKERNEL_TEST_HDRS,
5579 deps = MICROKERNEL_TEST_DEPS,
5580)
5581
5582xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005583 name = "f32_vdiv_test",
5584 srcs = [
5585 "test/f32-vdiv.cc",
5586 "test/vbinary-microkernel-tester.h",
5587 ] + MICROKERNEL_TEST_HDRS,
5588 deps = MICROKERNEL_TEST_DEPS,
5589)
5590
5591xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005592 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005593 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005594 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005595 "test/vbinary-microkernel-tester.h",
5596 ] + MICROKERNEL_TEST_HDRS,
5597 deps = MICROKERNEL_TEST_DEPS,
5598)
5599
5600xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005601 name = "f32_vdiv_relu_test",
5602 srcs = [
5603 "test/f32-vdiv-relu.cc",
5604 "test/vbinary-microkernel-tester.h",
5605 ] + MICROKERNEL_TEST_HDRS,
5606 deps = MICROKERNEL_TEST_DEPS,
5607)
5608
5609xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005610 name = "f32_vdivc_test",
5611 srcs = [
5612 "test/f32-vdivc.cc",
5613 "test/vbinaryc-microkernel-tester.h",
5614 ] + MICROKERNEL_TEST_HDRS,
5615 deps = MICROKERNEL_TEST_DEPS,
5616)
5617
5618xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005619 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005620 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005621 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005622 "test/vbinaryc-microkernel-tester.h",
5623 ] + MICROKERNEL_TEST_HDRS,
5624 deps = MICROKERNEL_TEST_DEPS,
5625)
5626
5627xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005628 name = "f32_vdivc_relu_test",
5629 srcs = [
5630 "test/f32-vdivc-relu.cc",
5631 "test/vbinaryc-microkernel-tester.h",
5632 ] + MICROKERNEL_TEST_HDRS,
5633 deps = MICROKERNEL_TEST_DEPS,
5634)
5635
5636xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005637 name = "f32_vrdivc_test",
5638 srcs = [
5639 "test/f32-vrdivc.cc",
5640 "test/vbinaryc-microkernel-tester.h",
5641 ] + MICROKERNEL_TEST_HDRS,
5642 deps = MICROKERNEL_TEST_DEPS,
5643)
5644
5645xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005646 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005647 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005648 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005649 "test/vbinaryc-microkernel-tester.h",
5650 ] + MICROKERNEL_TEST_HDRS,
5651 deps = MICROKERNEL_TEST_DEPS,
5652)
5653
5654xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005655 name = "f32_vrdivc_relu_test",
5656 srcs = [
5657 "test/f32-vrdivc-relu.cc",
5658 "test/vbinaryc-microkernel-tester.h",
5659 ] + MICROKERNEL_TEST_HDRS,
5660 deps = MICROKERNEL_TEST_DEPS,
5661)
5662
5663xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005664 name = "f32_vmax_test",
5665 srcs = [
5666 "test/f32-vmax.cc",
5667 "test/vbinary-microkernel-tester.h",
5668 ] + MICROKERNEL_TEST_HDRS,
5669 deps = MICROKERNEL_TEST_DEPS,
5670)
5671
5672xnnpack_unit_test(
5673 name = "f32_vmaxc_test",
5674 srcs = [
5675 "test/f32-vmaxc.cc",
5676 "test/vbinaryc-microkernel-tester.h",
5677 ] + MICROKERNEL_TEST_HDRS,
5678 deps = MICROKERNEL_TEST_DEPS,
5679)
5680
5681xnnpack_unit_test(
5682 name = "f32_vmin_test",
5683 srcs = [
5684 "test/f32-vmin.cc",
5685 "test/vbinary-microkernel-tester.h",
5686 ] + MICROKERNEL_TEST_HDRS,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
5691 name = "f32_vminc_test",
5692 srcs = [
5693 "test/f32-vminc.cc",
5694 "test/vbinaryc-microkernel-tester.h",
5695 ] + MICROKERNEL_TEST_HDRS,
5696 deps = MICROKERNEL_TEST_DEPS,
5697)
5698
5699xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005700 name = "f32_vmul_test",
5701 srcs = [
5702 "test/f32-vmul.cc",
5703 "test/vbinary-microkernel-tester.h",
5704 ] + MICROKERNEL_TEST_HDRS,
5705 deps = MICROKERNEL_TEST_DEPS,
5706)
5707
5708xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005709 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005710 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005711 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005712 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005713 ] + MICROKERNEL_TEST_HDRS,
5714 deps = MICROKERNEL_TEST_DEPS,
5715)
5716
5717xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005718 name = "f32_vmul_relu_test",
5719 srcs = [
5720 "test/f32-vmul-relu.cc",
5721 "test/vbinary-microkernel-tester.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005727 name = "f32_vmulc_test",
5728 srcs = [
5729 "test/f32-vmulc.cc",
5730 "test/vbinaryc-microkernel-tester.h",
5731 ] + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS,
5733)
5734
5735xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005736 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005737 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005738 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005739 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005740 ] + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS,
5742)
5743
5744xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005745 name = "f32_vmulc_relu_test",
5746 srcs = [
5747 "test/f32-vmulc-relu.cc",
5748 "test/vbinaryc-microkernel-tester.h",
5749 ] + MICROKERNEL_TEST_HDRS,
5750 deps = MICROKERNEL_TEST_DEPS,
5751)
5752
5753xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005754 name = "f16_vmulcaddc_minmax_test",
5755 srcs = [
5756 "test/f16-vmulcaddc-minmax.cc",
5757 "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"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005761)
5762
5763xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005764 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005765 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005766 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005767 "test/vmulcaddc-microkernel-tester.h",
5768 "src/xnnpack/AlignedAllocator.h",
5769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005770 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005771)
5772
5773xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005774 name = "f32_vlrelu_test",
5775 srcs = [
5776 "test/f32-vlrelu.cc",
5777 "test/vunary-microkernel-tester.h",
5778 ] + MICROKERNEL_TEST_HDRS,
5779 deps = MICROKERNEL_TEST_DEPS,
5780)
5781
5782xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005783 name = "f32_vneg_test",
5784 srcs = [
5785 "test/f32-vneg.cc",
5786 "test/vunary-microkernel-tester.h",
5787 ] + MICROKERNEL_TEST_HDRS,
5788 deps = MICROKERNEL_TEST_DEPS,
5789)
5790
5791xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005792 name = "f32_vrndne_test",
5793 srcs = [
5794 "test/f32-vrndne.cc",
5795 "test/vunary-microkernel-tester.h",
5796 ] + MICROKERNEL_TEST_HDRS,
5797 deps = MICROKERNEL_TEST_DEPS,
5798)
5799
5800xnnpack_unit_test(
5801 name = "f32_vrndz_test",
5802 srcs = [
5803 "test/f32-vrndz.cc",
5804 "test/vunary-microkernel-tester.h",
5805 ] + MICROKERNEL_TEST_HDRS,
5806 deps = MICROKERNEL_TEST_DEPS,
5807)
5808
5809xnnpack_unit_test(
5810 name = "f32_vrndu_test",
5811 srcs = [
5812 "test/f32-vrndu.cc",
5813 "test/vunary-microkernel-tester.h",
5814 ] + MICROKERNEL_TEST_HDRS,
5815 deps = MICROKERNEL_TEST_DEPS,
5816)
5817
5818xnnpack_unit_test(
5819 name = "f32_vrndd_test",
5820 srcs = [
5821 "test/f32-vrndd.cc",
5822 "test/vunary-microkernel-tester.h",
5823 ] + MICROKERNEL_TEST_HDRS,
5824 deps = MICROKERNEL_TEST_DEPS,
5825)
5826
5827xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005828 name = "f32_vscale_test",
5829 srcs = [
5830 "test/f32-vscale.cc",
5831 "test/vscale-microkernel-tester.h",
5832 ] + MICROKERNEL_TEST_HDRS,
5833 deps = MICROKERNEL_TEST_DEPS,
5834)
5835
5836xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005837 name = "f32_vscaleexpminusmax_test",
5838 srcs = [
5839 "test/f32-vscaleexpminusmax.cc",
5840 "test/vscaleexpminusmax-microkernel-tester.h",
5841 ] + MICROKERNEL_TEST_HDRS,
5842 deps = MICROKERNEL_TEST_DEPS,
5843)
5844
5845xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005846 name = "f32_vscaleextexp_test",
5847 srcs = [
5848 "test/f32-vscaleextexp.cc",
5849 "test/vscaleextexp-microkernel-tester.h",
5850 ] + MICROKERNEL_TEST_HDRS,
5851 deps = MICROKERNEL_TEST_DEPS,
5852)
5853
5854xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005855 name = "f32_vsqr_test",
5856 srcs = [
5857 "test/f32-vsqr.cc",
5858 "test/vunary-microkernel-tester.h",
5859 ] + MICROKERNEL_TEST_HDRS,
5860 deps = MICROKERNEL_TEST_DEPS,
5861)
5862
5863xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005864 name = "f32_vsqrdiff_test",
5865 srcs = [
5866 "test/f32-vsqrdiff.cc",
5867 "test/vbinary-microkernel-tester.h",
5868 ] + MICROKERNEL_TEST_HDRS,
5869 deps = MICROKERNEL_TEST_DEPS,
5870)
5871
5872xnnpack_unit_test(
5873 name = "f32_vsqrdiffc_test",
5874 srcs = [
5875 "test/f32-vsqrdiffc.cc",
5876 "test/vbinaryc-microkernel-tester.h",
5877 ] + MICROKERNEL_TEST_HDRS,
5878 deps = MICROKERNEL_TEST_DEPS,
5879)
5880
5881xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005882 name = "f32_vsqrt_test",
5883 srcs = [
5884 "test/f32-vsqrt.cc",
5885 "test/vunary-microkernel-tester.h",
5886 ] + MICROKERNEL_TEST_HDRS,
5887 deps = MICROKERNEL_TEST_DEPS,
5888)
5889
5890xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005891 name = "f32_vsub_test",
5892 srcs = [
5893 "test/f32-vsub.cc",
5894 "test/vbinary-microkernel-tester.h",
5895 ] + MICROKERNEL_TEST_HDRS,
5896 deps = MICROKERNEL_TEST_DEPS,
5897)
5898
5899xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005900 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005901 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005902 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005903 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005904 ] + MICROKERNEL_TEST_HDRS,
5905 deps = MICROKERNEL_TEST_DEPS,
5906)
5907
5908xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005909 name = "f32_vsub_relu_test",
5910 srcs = [
5911 "test/f32-vsub-relu.cc",
5912 "test/vbinary-microkernel-tester.h",
5913 ] + MICROKERNEL_TEST_HDRS,
5914 deps = MICROKERNEL_TEST_DEPS,
5915)
5916
5917xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005918 name = "f32_vsubc_test",
5919 srcs = [
5920 "test/f32-vsubc.cc",
5921 "test/vbinaryc-microkernel-tester.h",
5922 ] + MICROKERNEL_TEST_HDRS,
5923 deps = MICROKERNEL_TEST_DEPS,
5924)
5925
5926xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005927 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005928 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005929 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005930 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005931 ] + MICROKERNEL_TEST_HDRS,
5932 deps = MICROKERNEL_TEST_DEPS,
5933)
5934
5935xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005936 name = "f32_vsubc_relu_test",
5937 srcs = [
5938 "test/f32-vsubc-relu.cc",
5939 "test/vbinaryc-microkernel-tester.h",
5940 ] + MICROKERNEL_TEST_HDRS,
5941 deps = MICROKERNEL_TEST_DEPS,
5942)
5943
5944xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005945 name = "f32_vrsubc_test",
5946 srcs = [
5947 "test/f32-vrsubc.cc",
5948 "test/vbinaryc-microkernel-tester.h",
5949 ] + MICROKERNEL_TEST_HDRS,
5950 deps = MICROKERNEL_TEST_DEPS,
5951)
5952
5953xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005954 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005955 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005956 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005957 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005958 ] + MICROKERNEL_TEST_HDRS,
5959 deps = MICROKERNEL_TEST_DEPS,
5960)
5961
5962xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005963 name = "f32_vrsubc_relu_test",
5964 srcs = [
5965 "test/f32-vrsubc-relu.cc",
5966 "test/vbinaryc-microkernel-tester.h",
5967 ] + MICROKERNEL_TEST_HDRS,
5968 deps = MICROKERNEL_TEST_DEPS,
5969)
5970
5971xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005972 name = "qs8_dwconv_minmax_test",
5973 srcs = [
5974 "test/qs8-dwconv-minmax.cc",
5975 "test/dwconv-microkernel-tester.h",
5976 "src/xnnpack/AlignedAllocator.h",
5977 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5979)
5980
5981xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005982 name = "qs8_gavgpool_minmax_test",
5983 srcs = [
5984 "test/qs8-gavgpool-minmax.cc",
5985 "test/gavgpool-microkernel-tester.h",
5986 "src/xnnpack/AlignedAllocator.h",
5987 ] + MICROKERNEL_TEST_HDRS,
5988 deps = MICROKERNEL_TEST_DEPS,
5989)
5990
5991xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005992 name = "qs8_gemm_minmax_test",
5993 srcs = [
5994 "test/qs8-gemm-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(
Marat Dukhanf9480682020-07-31 14:50:24 -07006002 name = "qs8_igemm_minmax_test",
6003 srcs = [
6004 "test/qs8-igemm-minmax.cc",
6005 "test/gemm-microkernel-tester.h",
6006 "src/xnnpack/AlignedAllocator.h",
6007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6009)
6010
6011xnnpack_unit_test(
6012 name = "qs8_requantization_test",
6013 srcs = [
6014 "src/xnnpack/requantization-stubs.h",
6015 "test/qs8-requantization.cc",
6016 "test/requantization-tester.h",
6017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
6021xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006022 name = "qs8_vadd_minmax_test",
6023 srcs = [
6024 "test/qs8-vadd-minmax.cc",
6025 "test/vadd-microkernel-tester.h",
6026 ] + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006031 name = "qs8_vaddc_minmax_test",
6032 srcs = [
6033 "test/qs8-vaddc-minmax.cc",
6034 "test/vaddc-microkernel-tester.h",
6035 ] + MICROKERNEL_TEST_HDRS,
6036 deps = MICROKERNEL_TEST_DEPS,
6037)
6038
6039xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006040 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006041 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006042 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006043 "test/avgpool-microkernel-tester.h",
6044 "src/xnnpack/AlignedAllocator.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006050 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006051 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006052 "test/qu8-dwconv-minmax.cc",
6053 "test/dwconv-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 Dukhanf62bbdc2020-08-04 13:59:04 -07006060 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006062 "test/qu8-igemm-minmax.cc",
6063 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006064 "src/xnnpack/AlignedAllocator.h",
6065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006066 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006067)
6068
6069xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006070 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006072 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006073 "test/gavgpool-microkernel-tester.h",
6074 "src/xnnpack/AlignedAllocator.h",
6075 ] + MICROKERNEL_TEST_HDRS,
6076 deps = MICROKERNEL_TEST_DEPS,
6077)
6078
6079xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006080 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006081 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006082 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006083 "test/gemm-microkernel-tester.h",
6084 "src/xnnpack/AlignedAllocator.h",
6085 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006086 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006087)
6088
6089xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006090 name = "qu8_requantization_test",
6091 srcs = [
6092 "src/xnnpack/requantization-stubs.h",
6093 "test/qu8-requantization.cc",
6094 "test/requantization-tester.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006100 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006102 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006103 "test/vadd-microkernel-tester.h",
6104 ] + MICROKERNEL_TEST_HDRS,
6105 deps = MICROKERNEL_TEST_DEPS,
6106)
6107
6108xnnpack_unit_test(
6109 name = "u8_clamp_test",
6110 srcs = [
6111 "test/u8-clamp.cc",
6112 "test/clamp-microkernel-tester.h",
6113 ] + MICROKERNEL_TEST_HDRS,
6114 deps = MICROKERNEL_TEST_DEPS,
6115)
6116
6117xnnpack_unit_test(
6118 name = "u8_lut32norm_test",
6119 srcs = [
6120 "test/u8-lut32norm.cc",
6121 "test/lut-norm-microkernel-tester.h",
6122 ] + MICROKERNEL_TEST_HDRS,
6123 deps = MICROKERNEL_TEST_DEPS,
6124)
6125
6126xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006127 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006128 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006129 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006130 "test/maxpool-microkernel-tester.h",
6131 ] + MICROKERNEL_TEST_HDRS,
6132 deps = MICROKERNEL_TEST_DEPS,
6133)
6134
6135xnnpack_unit_test(
6136 name = "u8_rmax_test",
6137 srcs = [
6138 "test/u8-rmax.cc",
6139 "test/rmax-microkernel-tester.h",
6140 ] + MICROKERNEL_TEST_HDRS,
6141 deps = MICROKERNEL_TEST_DEPS,
6142)
6143
6144xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006145 name = "x32_fill_test",
6146 srcs = [
6147 "test/x32-fill.cc",
6148 "test/fill-microkernel-tester.h",
6149 ] + MICROKERNEL_TEST_HDRS,
6150 deps = MICROKERNEL_TEST_DEPS,
6151)
6152
6153xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154 name = "x32_packx_test",
6155 srcs = [
6156 "test/x32-packx.cc",
6157 "test/pack-microkernel-tester.h",
6158 "src/xnnpack/AlignedAllocator.h",
6159 ] + MICROKERNEL_TEST_HDRS,
6160 deps = MICROKERNEL_TEST_DEPS,
6161)
6162
6163xnnpack_unit_test(
6164 name = "x32_pad_test",
6165 srcs = [
6166 "test/x32-pad.cc",
6167 "test/pad-microkernel-tester.h",
6168 ] + MICROKERNEL_TEST_HDRS,
6169 deps = MICROKERNEL_TEST_DEPS,
6170)
6171
6172xnnpack_unit_test(
6173 name = "x32_unpool_test",
6174 srcs = [
6175 "test/x32-unpool.cc",
6176 "test/unpool-microkernel-tester.h",
6177 ] + MICROKERNEL_TEST_HDRS,
6178 deps = MICROKERNEL_TEST_DEPS,
6179)
6180
6181xnnpack_unit_test(
6182 name = "x32_zip_test",
6183 srcs = [
6184 "test/x32-zip.cc",
6185 "test/zip-microkernel-tester.h",
6186 ] + MICROKERNEL_TEST_HDRS,
6187 deps = MICROKERNEL_TEST_DEPS,
6188)
6189
6190xnnpack_unit_test(
6191 name = "x8_lut_test",
6192 srcs = [
6193 "test/x8-lut.cc",
6194 "test/lut-microkernel-tester.h",
6195 ] + MICROKERNEL_TEST_HDRS,
6196 deps = MICROKERNEL_TEST_DEPS,
6197)
6198
6199xnnpack_unit_test(
6200 name = "x8_zip_test",
6201 srcs = [
6202 "test/x8-zip.cc",
6203 "test/zip-microkernel-tester.h",
6204 ] + MICROKERNEL_TEST_HDRS,
6205 deps = MICROKERNEL_TEST_DEPS,
6206)
6207
Marat Dukhan20c3b922020-03-10 03:45:06 -07006208########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209
6210xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006211 name = "operator_size_test",
6212 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006213 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214)
6215
Marat Dukhan20c3b922020-03-10 03:45:06 -07006216xnnpack_binary(
6217 name = "subgraph_size_test",
6218 srcs = ["test/subgraph-size.c"],
6219 deps = [":XNNPACK"],
6220)
6221
6222########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223
6224xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006225 name = "abs_nc_test",
6226 srcs = [
6227 "test/abs-nc.cc",
6228 "test/abs-operator-tester.h",
6229 ],
6230 deps = OPERATOR_TEST_DEPS,
6231)
6232
6233xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006234 name = "add_nd_test",
6235 srcs = [
6236 "test/add-nd.cc",
6237 "test/binary-elementwise-operator-tester.h",
6238 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006239 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006240)
6241
6242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006243 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006245 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246 "test/argmax-pooling-operator-tester.h",
6247 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006248 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006249)
6250
6251xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006252 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006254 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006255 "test/average-pooling-operator-tester.h",
6256 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006257 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258)
6259
6260xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006261 name = "bankers_rounding_nc_test",
6262 srcs = [
6263 "test/bankers-rounding-nc.cc",
6264 "test/bankers-rounding-operator-tester.h",
6265 ],
6266 deps = OPERATOR_TEST_DEPS,
6267)
6268
6269xnnpack_unit_test(
6270 name = "ceiling_nc_test",
6271 srcs = [
6272 "test/ceiling-nc.cc",
6273 "test/ceiling-operator-tester.h",
6274 ],
6275 deps = OPERATOR_TEST_DEPS,
6276)
6277
6278xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006279 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006280 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006281 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006282 "test/channel-shuffle-operator-tester.h",
6283 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006284 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285)
6286
6287xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006288 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006289 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006290 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291 "test/clamp-operator-tester.h",
6292 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006293 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294)
6295
6296xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006297 name = "constant_pad_nd_test",
6298 srcs = [
6299 "test/constant-pad-nd.cc",
6300 "test/constant-pad-operator-tester.h",
6301 ],
6302 deps = OPERATOR_TEST_DEPS,
6303)
6304
6305xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006306 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006307 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006308 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309 "test/convolution-operator-tester.h",
6310 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006311 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006312)
6313
6314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006315 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006317 "test/convolution-nchw.cc",
6318 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006319 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006320 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006321)
6322
6323xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006324 name = "copy_nc_test",
6325 srcs = [
6326 "test/copy-nc.cc",
6327 "test/copy-operator-tester.h",
6328 ],
6329 deps = OPERATOR_TEST_DEPS,
6330)
6331
6332xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006333 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006334 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006335 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006336 "test/deconvolution-operator-tester.h",
6337 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006338 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006339)
6340
6341xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006342 name = "divide_nd_test",
6343 srcs = [
6344 "test/binary-elementwise-operator-tester.h",
6345 "test/divide-nd.cc",
6346 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006347 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006348)
6349
6350xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006351 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006352 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006353 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006354 "test/fully-connected-operator-tester.h",
6355 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006356 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357)
6358
6359xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006360 name = "floor_nc_test",
6361 srcs = [
6362 "test/floor-nc.cc",
6363 "test/floor-operator-tester.h",
6364 ],
6365 deps = OPERATOR_TEST_DEPS,
6366)
6367
6368xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006369 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006370 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006371 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006372 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006373 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006374 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006375)
6376
6377xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006378 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006379 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006380 "test/global-average-pooling-ncw.cc",
6381 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006382 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006383 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006384)
6385
6386xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006387 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006388 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006389 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006390 "test/hardswish-operator-tester.h",
6391 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006392 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006393)
6394
6395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006396 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006397 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006398 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006399 "test/leaky-relu-operator-tester.h",
6400 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006401 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006402)
6403
6404xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006405 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006406 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006407 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006408 "test/max-pooling-operator-tester.h",
6409 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006410 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006411)
6412
6413xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006414 name = "maximum_nd_test",
6415 srcs = [
6416 "test/binary-elementwise-operator-tester.h",
6417 "test/maximum-nd.cc",
6418 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006419 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006420)
6421
6422xnnpack_unit_test(
6423 name = "minimum_nd_test",
6424 srcs = [
6425 "test/binary-elementwise-operator-tester.h",
6426 "test/minimum-nd.cc",
6427 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006428 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006429)
6430
6431xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006432 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006433 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006434 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006435 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006436 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006437 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006438)
6439
6440xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006441 name = "negate_nc_test",
6442 srcs = [
6443 "test/negate-nc.cc",
6444 "test/negate-operator-tester.h",
6445 ],
6446 deps = OPERATOR_TEST_DEPS,
6447)
6448
6449xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006450 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006451 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006452 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006453 "test/prelu-operator-tester.h",
6454 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006455 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006456)
6457
6458xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006459 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006460 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006461 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006462 "test/resize-bilinear-operator-tester.h",
6463 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006464 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006465)
6466
6467xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006468 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006469 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006470 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471 "test/sigmoid-operator-tester.h",
6472 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006473 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006474)
6475
6476xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006477 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006478 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006479 "test/softmax-nc.cc",
6480 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483)
6484
6485xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006486 name = "square_nc_test",
6487 srcs = [
6488 "test/square-nc.cc",
6489 "test/square-operator-tester.h",
6490 ],
6491 deps = OPERATOR_TEST_DEPS,
6492)
6493
6494xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006495 name = "square_root_nc_test",
6496 srcs = [
6497 "test/square-root-nc.cc",
6498 "test/square-root-operator-tester.h",
6499 ],
6500 deps = OPERATOR_TEST_DEPS,
6501)
6502
6503xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006504 name = "squared_difference_nd_test",
6505 srcs = [
6506 "test/binary-elementwise-operator-tester.h",
6507 "test/squared-difference-nd.cc",
6508 ],
6509 deps = OPERATOR_TEST_DEPS,
6510)
6511
6512xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006513 name = "subtract_nd_test",
6514 srcs = [
6515 "test/binary-elementwise-operator-tester.h",
6516 "test/subtract-nd.cc",
6517 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006518 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006519)
6520
6521xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006522 name = "truncation_nc_test",
6523 srcs = [
6524 "test/truncation-nc.cc",
6525 "test/truncation-operator-tester.h",
6526 ],
6527 deps = OPERATOR_TEST_DEPS,
6528)
6529
6530xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006531 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006533 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534 "test/unpooling-operator-tester.h",
6535 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006536 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006537)
6538
Chao Mei6ddfc602020-05-13 22:29:36 -07006539############################### Misc unit tests ###############################
6540
6541xnnpack_unit_test(
6542 name = "memory_planner_test",
6543 srcs = [
6544 "test/memory-planner-test.cc",
6545 ],
6546 deps = [
6547 ":XNNPACK",
6548 ":memory_planner",
6549 ],
6550)
6551
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006552xnnpack_unit_test(
6553 name = "subgraph_nchw_test",
6554 srcs = [
6555 "src/xnnpack/subgraph.h",
6556 "test/subgraph-nchw.cc",
6557 "test/subgraph-tester.h",
6558 ],
6559 deps = [
6560 ":XNNPACK",
6561 ],
6562)
6563
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564############################# Build configurations #############################
6565
Marat Dukhanb8642352019-10-30 15:43:02 -07006566# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006567config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006568 name = "xnn_enable_assembly_explicit_true",
6569 define_values = {"xnn_enable_assembly": "true"},
6570)
6571
6572# Disables usage of assembly kernels.
6573config_setting(
6574 name = "xnn_enable_assembly_explicit_false",
6575 define_values = {"xnn_enable_assembly": "false"},
6576)
6577
Marat Dukhan9de90e02020-06-18 16:04:12 -07006578# Enables usage of sparse inference.
6579config_setting(
6580 name = "xnn_enable_sparse_explicit_true",
6581 define_values = {"xnn_enable_sparse": "true"},
6582)
6583
6584# Disables usage of sparse inference.
6585config_setting(
6586 name = "xnn_enable_sparse_explicit_false",
6587 define_values = {"xnn_enable_sparse": "false"},
6588)
6589
Marat Dukhan05702cf2020-03-26 15:41:33 -07006590# Disables usage of HMP-aware optimizations.
6591config_setting(
6592 name = "xnn_enable_hmp_explicit_false",
6593 define_values = {"xnn_enable_hmp": "false"},
6594)
6595
Chao Mei6ddfc602020-05-13 22:29:36 -07006596# Enable usage of optimized memory allocation
6597config_setting(
6598 name = "xnn_enable_memopt_explicit_true",
6599 define_values = {"xnnpack_enable_memopt": "true"},
6600)
6601
6602# Disable usage of optimized memory allocation
6603config_setting(
6604 name = "xnn_enable_memopt_explicit_false",
6605 define_values = {"xnnpack_enable_memopt": "false"},
6606)
6607
Marat Dukhanb8642352019-10-30 15:43:02 -07006608# Builds with -c dbg
6609config_setting(
6610 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006611 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006612 "compilation_mode": "dbg",
6613 },
6614)
6615
6616# Builds with -c opt
6617config_setting(
6618 name = "optimized_build",
6619 values = {
6620 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006621 },
6622)
6623
6624config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006625 name = "linux_k8",
6626 values = {"cpu": "k8"},
6627)
6628
6629config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006630 name = "linux_arm",
6631 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006632)
6633
6634config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006635 name = "linux_armeabi",
6636 values = {"cpu": "armeabi"},
6637)
6638
6639config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006640 name = "linux_armhf",
6641 values = {"cpu": "armhf"},
6642)
6643
6644config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006645 name = "linux_armv7a",
6646 values = {"cpu": "armv7a"},
6647)
6648
6649config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006650 name = "linux_aarch64",
6651 values = {"cpu": "aarch64"},
6652)
6653
6654config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006655 name = "android",
6656 values = {"crosstool_top": "//external:android/crosstool"},
6657)
6658
6659config_setting(
6660 name = "android_armv7",
6661 values = {
6662 "crosstool_top": "//external:android/crosstool",
6663 "cpu": "armeabi-v7a",
6664 },
6665)
6666
6667config_setting(
6668 name = "android_arm64",
6669 values = {
6670 "crosstool_top": "//external:android/crosstool",
6671 "cpu": "arm64-v8a",
6672 },
6673)
6674
6675config_setting(
6676 name = "android_x86",
6677 values = {
6678 "crosstool_top": "//external:android/crosstool",
6679 "cpu": "x86",
6680 },
6681)
6682
6683config_setting(
6684 name = "android_x86_64",
6685 values = {
6686 "crosstool_top": "//external:android/crosstool",
6687 "cpu": "x86_64",
6688 },
6689)
6690
6691config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006692 name = "windows_x86_64",
6693 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006694)
6695
6696config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006697 name = "windows_x86_64_clang",
6698 values = {
6699 "compiler": "clang-cl",
6700 "cpu": "x64_windows",
6701 },
6702)
6703
6704config_setting(
6705 name = "windows_x86_64_mingw",
6706 values = {
6707 "compiler": "mingw-gcc",
6708 "cpu": "x64_windows",
6709 },
6710)
6711
6712config_setting(
6713 name = "windows_x86_64_msys",
6714 values = {
6715 "compiler": "msys-gcc",
6716 "cpu": "x64_windows",
6717 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006718)
6719
6720config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006721 name = "macos_x86_64",
6722 values = {
6723 "apple_platform_type": "macos",
6724 "cpu": "darwin",
6725 },
6726)
6727
6728config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006729 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006730 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006731)
6732
6733config_setting(
6734 name = "emscripten_wasm",
6735 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006736 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006737 "cpu": "wasm",
6738 },
6739)
6740
6741config_setting(
6742 name = "emscripten_wasmsimd",
6743 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006744 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006745 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006746 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006747 },
6748)
6749
6750config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006751 name = "ios_armv7",
6752 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006753 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006754 "cpu": "ios_armv7",
6755 },
6756)
6757
6758config_setting(
6759 name = "ios_arm64",
6760 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006761 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006762 "cpu": "ios_arm64",
6763 },
6764)
6765
6766config_setting(
6767 name = "ios_arm64e",
6768 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006769 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006770 "cpu": "ios_arm64e",
6771 },
6772)
6773
6774config_setting(
6775 name = "ios_x86",
6776 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006777 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006778 "cpu": "ios_i386",
6779 },
6780)
6781
6782config_setting(
6783 name = "ios_x86_64",
6784 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006785 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006786 "cpu": "ios_x86_64",
6787 },
6788)
6789
6790config_setting(
6791 name = "watchos_armv7k",
6792 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006793 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006794 "cpu": "watchos_armv7k",
6795 },
6796)
6797
6798config_setting(
6799 name = "watchos_arm64_32",
6800 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006801 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006802 "cpu": "watchos_arm64_32",
6803 },
6804)
6805
6806config_setting(
6807 name = "watchos_x86",
6808 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006809 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006810 "cpu": "watchos_i386",
6811 },
6812)
6813
6814config_setting(
6815 name = "watchos_x86_64",
6816 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006817 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006818 "cpu": "watchos_x86_64",
6819 },
6820)
6821
6822config_setting(
6823 name = "tvos_arm64",
6824 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006825 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006826 "cpu": "tvos_arm64",
6827 },
6828)
6829
6830config_setting(
6831 name = "tvos_x86_64",
6832 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006833 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006834 "cpu": "tvos_x86_64",
6835 },
6836)