blob: 1fe18fceaa2ff7d715c1e42e4ec79eda8e954493 [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 Dukhan91249d22020-10-24 12:02:51 -0700153 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
154 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
155 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
156 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -0700163 "src/f32-dwconv2d-chw/3x3s2p1-scalar-1x1-acc3.c",
164 "src/f32-dwconv2d-chw/5x5p2-scalar-1x1-acc5.c",
165 "src/f32-dwconv2d-chw/5x5s2p2-scalar-1x1-acc5.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700166 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700167 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
168 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700169 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
170 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700172 "src/f32-gemm/gen/1x4-minmax-scalar.c",
173 "src/f32-gemm/gen/1x4-relu-scalar.c",
174 "src/f32-gemm/gen/1x4-scalar.c",
175 "src/f32-gemm/gen/2x4-minmax-scalar.c",
176 "src/f32-gemm/gen/2x4-relu-scalar.c",
177 "src/f32-gemm/gen/2x4-scalar.c",
178 "src/f32-gemm/gen/4x2-minmax-scalar.c",
179 "src/f32-gemm/gen/4x2-relu-scalar.c",
180 "src/f32-gemm/gen/4x2-scalar.c",
181 "src/f32-gemm/gen/4x4-minmax-scalar.c",
182 "src/f32-gemm/gen/4x4-relu-scalar.c",
183 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700184 "src/f32-hswish/gen/hswish-scalar-x1.c",
185 "src/f32-hswish/gen/hswish-scalar-x2.c",
186 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700187 "src/f32-ibilinear-chw/gen/scalar-p1.c",
188 "src/f32-ibilinear-chw/gen/scalar-p2.c",
189 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700190 "src/f32-ibilinear/gen/scalar-c1.c",
191 "src/f32-ibilinear/gen/scalar-c2.c",
192 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700194 "src/f32-igemm/gen/1x4-relu-scalar.c",
195 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700196 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700197 "src/f32-igemm/gen/2x4-relu-scalar.c",
198 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700199 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700200 "src/f32-igemm/gen/4x2-relu-scalar.c",
201 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700202 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700203 "src/f32-igemm/gen/4x4-relu-scalar.c",
204 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700205 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
206 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
207 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700208 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
209 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
210 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
211 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800212 "src/f32-prelu/gen/scalar-2x1.c",
213 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800214 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800215 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700216 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800217 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
218 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700219 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800220 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800221 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700222 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800223 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
224 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700225 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700226 "src/f32-relu/gen/scalar-x1.c",
227 "src/f32-relu/gen/scalar-x2.c",
228 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700229 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800231 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
232 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
233 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700234 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
235 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
236 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800237 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
238 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
239 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700240 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
241 "src/f32-spmm/gen/1x1-minmax-scalar.c",
242 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
243 "src/f32-spmm/gen/2x1-minmax-scalar.c",
244 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
245 "src/f32-spmm/gen/4x1-minmax-scalar.c",
246 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
247 "src/f32-spmm/gen/8x1-minmax-scalar.c",
248 "src/f32-spmm/gen/8x2-minmax-scalar.c",
249 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700250 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
251 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
252 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700253 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700254 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
255 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
256 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700257 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700258 "src/f32-vbinary/gen/vadd-scalar-x1.c",
259 "src/f32-vbinary/gen/vadd-scalar-x2.c",
260 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700261 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700262 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
263 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
264 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700265 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700266 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
267 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
268 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700269 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700270 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
271 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
272 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700273 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700274 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700277 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700278 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
279 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
280 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700281 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700282 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
283 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
284 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700285 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700289 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700290 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
291 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
292 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700293 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700294 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
295 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
296 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700297 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800298 "src/f32-vbinary/gen/vmax-scalar-x1.c",
299 "src/f32-vbinary/gen/vmax-scalar-x2.c",
300 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700301 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800302 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
303 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
304 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700305 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800306 "src/f32-vbinary/gen/vmin-scalar-x1.c",
307 "src/f32-vbinary/gen/vmin-scalar-x2.c",
308 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700309 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800310 "src/f32-vbinary/gen/vminc-scalar-x1.c",
311 "src/f32-vbinary/gen/vminc-scalar-x2.c",
312 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700313 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700314 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
315 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
316 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700317 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700318 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
319 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
320 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700321 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700322 "src/f32-vbinary/gen/vmul-scalar-x1.c",
323 "src/f32-vbinary/gen/vmul-scalar-x2.c",
324 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700325 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700326 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
327 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
328 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700329 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700330 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
331 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
332 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700333 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700334 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
335 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
336 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700337 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700338 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
339 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
340 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700341 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700342 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
343 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
344 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700345 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700346 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
347 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
348 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700349 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700350 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
351 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
352 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700353 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700354 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
355 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
356 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700357 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700358 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
359 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
360 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700361 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700362 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
363 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
364 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700365 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700366 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
367 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
368 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700369 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700370 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
371 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
372 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700373 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700374 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
375 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
376 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700377 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700378 "src/f32-vbinary/gen/vsub-scalar-x1.c",
379 "src/f32-vbinary/gen/vsub-scalar-x2.c",
380 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700381 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700382 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
383 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
384 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700385 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700386 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
387 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
388 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700389 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700390 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
391 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
392 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700393 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700394 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
395 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
396 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
398 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
399 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700400 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
401 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
402 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700403 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
404 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
405 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
406 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
407 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
408 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
409 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
410 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
411 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700412 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
413 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
414 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700415 "src/f32-vunary/gen/vabs-scalar-x1.c",
416 "src/f32-vunary/gen/vabs-scalar-x2.c",
417 "src/f32-vunary/gen/vabs-scalar-x4.c",
418 "src/f32-vunary/gen/vneg-scalar-x1.c",
419 "src/f32-vunary/gen/vneg-scalar-x2.c",
420 "src/f32-vunary/gen/vneg-scalar-x4.c",
421 "src/f32-vunary/gen/vsqr-scalar-x1.c",
422 "src/f32-vunary/gen/vsqr-scalar-x2.c",
423 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800424 "src/math/expminus-scalar-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700425 "src/math/expminus-scalar-lut2048-p1.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800426 "src/math/expminus-scalar-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700427 "src/math/roundd-scalar-addsub.c",
428 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700429 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700430 "src/math/roundne-scalar-addsub.c",
431 "src/math/roundne-scalar-nearbyint.c",
432 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700433 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700434 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700435 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700436 "src/math/roundz-scalar-addsub.c",
437 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700438 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700439 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700440 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700441 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700442 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
443 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700444 "src/qs8-requantization/fp32-scalar-lrintf.c",
445 "src/qs8-requantization/fp32-scalar-magic.c",
446 "src/qs8-requantization/precise-scalar-signed64.c",
447 "src/qs8-requantization/precise-scalar-unsigned32.c",
448 "src/qs8-requantization/precise-scalar-unsigned64.c",
449 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700450 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
451 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
452 "src/qu8-dwconv/up1x9-minmax-scalar.c",
453 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
454 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700455 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700456 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
457 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700458 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700459 "src/qu8-requantization/fp32-scalar-lrintf.c",
460 "src/qu8-requantization/fp32-scalar-magic.c",
461 "src/qu8-requantization/precise-scalar-signed64.c",
462 "src/qu8-requantization/precise-scalar-unsigned32.c",
463 "src/qu8-requantization/precise-scalar-unsigned64.c",
464 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700465 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700466 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700467 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700469 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700470 "src/x8-lut/scalar.c",
471 "src/x8-zip/x2-scalar.c",
472 "src/x8-zip/x3-scalar.c",
473 "src/x8-zip/x4-scalar.c",
474 "src/x8-zip/xm-scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700475 "src/x32-fill/scalar-float.c",
476 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700477 "src/x32-packx/x2-scalar.c",
478 "src/x32-packx/x3-scalar.c",
479 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700480 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700481 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700482 "src/x32-unpool/scalar.c",
483 "src/x32-zip/x2-scalar.c",
484 "src/x32-zip/x3-scalar.c",
485 "src/x32-zip/x4-scalar.c",
486 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700487]
488
Marat Dukhan436ebe62019-12-04 15:10:12 -0800489WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700490 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
491 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700492 "src/f32-clamp/gen/wasm-x1.c",
493 "src/f32-clamp/gen/wasm-x2.c",
494 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700495 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700497 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700501 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
502 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
506 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
508 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700509 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
510 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700511 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
512 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700513 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
514 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700515 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
516 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700517 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
518 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
520 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700521 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
522 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
523 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
524 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700525 "src/f32-gemm/gen/1x4-relu-wasm.c",
526 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700527 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700528 "src/f32-gemm/gen/2x4-relu-wasm.c",
529 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700530 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-gemm/gen/4x2-relu-wasm.c",
532 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700534 "src/f32-gemm/gen/4x4-relu-wasm.c",
535 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700536 "src/f32-hswish/gen/hswish-wasm-x1.c",
537 "src/f32-hswish/gen/hswish-wasm-x2.c",
538 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700539 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700540 "src/f32-igemm/gen/1x4-relu-wasm.c",
541 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700542 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700543 "src/f32-igemm/gen/2x4-relu-wasm.c",
544 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700546 "src/f32-igemm/gen/4x2-relu-wasm.c",
547 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700548 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700549 "src/f32-igemm/gen/4x4-relu-wasm.c",
550 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700551 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
552 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
553 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700554 "src/f32-prelu/gen/wasm-2x1.c",
555 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700556 "src/f32-relu/gen/wasm-x1.c",
557 "src/f32-relu/gen/wasm-x2.c",
558 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700559 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700560 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
561 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
562 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700563 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700564 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
565 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
566 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700567 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700568 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
569 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
570 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
571 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700572 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
573 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
574 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700575 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700576 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
577 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
578 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
579 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700580 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
581 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
582 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700583 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
585 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
586 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
587 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700588 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
589 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
590 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700591 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800592 "src/f32-vbinary/gen/vmax-wasm-x1.c",
593 "src/f32-vbinary/gen/vmax-wasm-x2.c",
594 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700595 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800596 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
597 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
598 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700599 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800600 "src/f32-vbinary/gen/vmin-wasm-x1.c",
601 "src/f32-vbinary/gen/vmin-wasm-x2.c",
602 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700603 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800604 "src/f32-vbinary/gen/vminc-wasm-x1.c",
605 "src/f32-vbinary/gen/vminc-wasm-x2.c",
606 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700607 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700608 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
609 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
610 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700611 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700612 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
613 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
614 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700615 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700616 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
617 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
618 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
619 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700620 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
621 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
622 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700623 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700624 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
625 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
626 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
627 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
630 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700631 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700632 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
633 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
634 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
635 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700636 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
637 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
638 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700639 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700640 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
641 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
642 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
643 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700644 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
645 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
646 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700647 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700648 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
649 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
650 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
651 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700652 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
653 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
654 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700655 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700656 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
657 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
658 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700659 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
660 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
661 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800662]
663
Marat Dukhan290055c2020-06-09 12:24:29 -0700664WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700665 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
666 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
667 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700668 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
669 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
670 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
671 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700672 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
673 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
674 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
675 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700676 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700677 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700678 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700679 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700680 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
681 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700682 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700683 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700685 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
686 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
687 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
688 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
689 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
691 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700692 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700693 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700694 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700695 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
696 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700697 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700699 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700700 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
701 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
702 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
703 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
704 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
705 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
706 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
707 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700708 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
709 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
710 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
711 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700712 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
714 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
715 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
716 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
717 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
719 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
720 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
721 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
722 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
723 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
725 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
726 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
727 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
728 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
729 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
730 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
731 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
732 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
733 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
734 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
736 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
737 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
738 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
739 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
740 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
742 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
743 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
744 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
745 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
746 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
747 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
748 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
749 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
751 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
752 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
753 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
754 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
755 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
757 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
758 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
759 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
760 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
761 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
762 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
763 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
764 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
765 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
767 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
768 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
769 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
770 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
771 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
772 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
773 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
774 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
776 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
777 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
778 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
779 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
780 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700782 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
783 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
784 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700785 "src/f32-ibilinear/gen/wasmsimd-c4.c",
786 "src/f32-ibilinear/gen/wasmsimd-c8.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700787 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
788 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700789 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
790 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
792 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
793 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
794 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
795 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
796 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
797 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
798 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
799 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
800 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
802 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
804 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
805 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
806 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
807 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
808 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
809 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
810 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
811 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
812 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
813 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
814 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
815 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
816 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
817 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
818 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
819 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
820 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
821 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
822 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
823 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
824 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
825 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
826 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
827 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
828 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700829 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
830 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700831 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
832 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
833 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
834 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700835 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
836 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700837 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
838 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
839 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700840 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
841 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700842 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
843 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
844 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
845 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
846 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
847 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
848 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700849 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
850 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700851 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
852 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
853 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
854 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700855 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700856 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700857 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700858 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
859 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700860 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700861 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
862 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700863 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700864 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
865 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700866 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700867 "src/f32-relu/gen/wasmsimd-x4.c",
868 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700869 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700870 "src/f32-rmax/wasmsimd-arm.c",
871 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700872 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
873 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
874 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
875 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
876 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
877 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700878 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
879 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
880 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
881 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
882 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
883 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
884 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll2.c",
885 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll4.c",
886 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
887 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll2.c",
888 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll4.c",
889 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
890 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll2.c",
891 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll4.c",
892 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
893 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll2.c",
894 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll4.c",
895 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
896 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll2.c",
897 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll4.c",
898 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
899 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll2.c",
900 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll4.c",
901 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700902 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
903 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700904 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700905 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
906 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700907 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700908 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
909 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700910 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700911 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
912 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700913 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700914 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
915 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700916 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700917 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
918 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700919 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700920 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
921 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700922 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700923 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
924 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700925 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700926 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
927 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700928 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700929 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
930 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700931 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700932 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
933 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700934 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700935 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
936 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700937 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700938 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
939 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700940 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700941 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
942 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700943 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700944 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
945 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700946 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700947 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
948 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700949 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700950 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
951 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700952 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700953 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
954 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700955 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700956 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
957 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700958 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700959 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
960 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700961 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700962 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
963 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700964 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700965 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
966 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700967 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700968 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
969 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700970 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700971 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
972 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700973 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700974 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
975 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700976 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700977 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
978 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700979 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700980 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
981 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700982 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700983 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
984 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700985 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700986 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
987 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700988 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700989 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
990 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700991 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700992 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
993 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700994 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700995 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
996 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700997 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700998 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
999 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001000 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001001 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1002 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001003 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001004 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1005 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001006 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001007 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1008 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001009 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001010 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1011 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001012 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001013 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1014 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001015 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001016 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1017 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001018 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001019 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1020 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001021 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001022 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1023 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001024 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001025 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1026 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001027 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001028 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1029 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001030 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001031 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1032 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001033 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001034 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1035 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001036 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001037 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1038 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001039 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001040 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1041 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001042 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001043 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1044 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001045 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001046 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1047 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001048 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001049 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1050 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001051 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001052 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1053 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1054 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1055 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001056 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001057 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001058 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001059 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001060 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1061 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1062 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1063 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001064 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1065 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1066 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1067 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1068 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1069 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1070 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1071 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1072 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1073 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001074 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1075 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001076 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1077 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1078 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1079 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1080 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1081 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001082 "src/math/roundd-wasmsimd-addsub.c",
1083 "src/math/roundd-wasmsimd-cvt.c",
1084 "src/math/roundne-wasmsimd-addsub.c",
1085 "src/math/roundu-wasmsimd-addsub.c",
1086 "src/math/roundu-wasmsimd-cvt.c",
1087 "src/math/roundz-wasmsimd-addsub.c",
1088 "src/math/roundz-wasmsimd-cvt.c",
1089 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1090 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001091 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1092 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1093 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001094 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1095 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1096 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001097 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1098 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1099 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001100 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001101 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001102 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001103 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1104 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001105 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001106 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1107 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001108 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1109 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001110 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001111 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001112 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001113 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001114 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001116 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001117 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1118 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1119 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1120 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1121 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1122 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1123 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1124 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001125 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001126 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001127 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001128 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001129 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001130 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001131 "src/x32-zip/x2-wasmsimd.c",
1132 "src/x32-zip/x3-wasmsimd.c",
1133 "src/x32-zip/x4-wasmsimd.c",
1134 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001135]
1136
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001137PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001138 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001139 "src/f32-dwconv2d-chw/3x3p1-psimd-1x4-acc3.c",
1140 "src/f32-dwconv2d-chw/3x3s2p1-psimd-1x4-acc3.c",
1141 "src/f32-dwconv2d-chw/5x5p2-psimd-3x4.c",
1142 "src/f32-dwconv2d-chw/5x5s2p2-psimd-1x4-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143]
1144
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001145PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001146 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001147 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001148 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001149 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001150]
1151
Marat Dukhan08c4a432019-10-03 09:29:21 -07001152# ISA-specific micro-kernels
1153NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001154 "src/f32-argmaxpool/4x-neon-c4.c",
1155 "src/f32-argmaxpool/9p8x-neon-c4.c",
1156 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001157 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1158 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001159 "src/f32-clamp/gen/neon-x4.c",
1160 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001161 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001162 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001163 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001164 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001165 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001166 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001167 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001168 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001169 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001170 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001171 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001172 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001173 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001174 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1175 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1176 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1177 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1178 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001179 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001180 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001181 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
1182 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1183 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1184 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1185 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1186 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
1187 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1188 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1189 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
1190 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001191 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001192 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1193 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001194 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001195 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1196 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001197 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001198 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1199 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1200 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1201 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1202 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001203 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1204 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001205 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1206 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001207 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1208 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001209 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1210 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1211 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1212 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1213 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1214 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1215 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1216 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1217 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1218 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1219 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1220 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1221 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1222 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1223 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1224 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001225 "src/f32-hswish/gen/hswish-neon-x4.c",
1226 "src/f32-hswish/gen/hswish-neon-x8.c",
1227 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001228 "src/f32-ibilinear/gen/neon-c4.c",
1229 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001230 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001231 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001232 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001233 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1234 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001235 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001236 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1237 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1238 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1239 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001240 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1241 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001242 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1243 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001244 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1245 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001246 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1247 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1248 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001249 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1250 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001251 "src/f32-prelu/gen/neon-1x4.c",
1252 "src/f32-prelu/gen/neon-1x8.c",
1253 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001254 "src/f32-prelu/gen/neon-2x4.c",
1255 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001256 "src/f32-prelu/gen/neon-2x16.c",
1257 "src/f32-prelu/gen/neon-4x4.c",
1258 "src/f32-prelu/gen/neon-4x8.c",
1259 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001260 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001261 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001262 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001263 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1264 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001265 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001266 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1267 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001268 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001269 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1270 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001271 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1272 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1273 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1274 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1275 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1276 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1277 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1278 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1279 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1280 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1281 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1282 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1283 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001284 "src/f32-relu/gen/neon-x4.c",
1285 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001286 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001287 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001288 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1289 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1290 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1291 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1292 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1293 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1294 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1295 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1296 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1297 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1298 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1299 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001300 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1301 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1302 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1303 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1304 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1305 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001306 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1307 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1308 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1309 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001310 "src/f32-vbinary/gen/vmax-neon-x4.c",
1311 "src/f32-vbinary/gen/vmax-neon-x8.c",
1312 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1313 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1314 "src/f32-vbinary/gen/vmin-neon-x4.c",
1315 "src/f32-vbinary/gen/vmin-neon-x8.c",
1316 "src/f32-vbinary/gen/vminc-neon-x4.c",
1317 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001318 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1319 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1320 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1321 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1322 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1323 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001324 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1325 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1326 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1327 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001328 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1329 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1330 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1331 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001332 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1333 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001334 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1335 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001336 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1337 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001338 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1339 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1340 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1341 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1342 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1343 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001344 "src/f32-vunary/gen/vabs-neon-x4.c",
1345 "src/f32-vunary/gen/vabs-neon-x8.c",
1346 "src/f32-vunary/gen/vneg-neon-x4.c",
1347 "src/f32-vunary/gen/vneg-neon-x8.c",
1348 "src/f32-vunary/gen/vsqr-neon-x4.c",
1349 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001350 "src/math/roundd-neon-addsub.c",
1351 "src/math/roundd-neon-cvt.c",
1352 "src/math/roundne-neon-addsub.c",
1353 "src/math/roundu-neon-addsub.c",
1354 "src/math/roundu-neon-cvt.c",
1355 "src/math/roundz-neon-addsub.c",
1356 "src/math/roundz-neon-cvt.c",
1357 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1358 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1359 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1360 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1361 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1362 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1363 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1364 "src/math/sqrt-neon-nr1rsqrts.c",
1365 "src/math/sqrt-neon-nr2rsqrts.c",
1366 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001367 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1368 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1369 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1370 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1374 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001375 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1377 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1378 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001379 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001380 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001381 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001382 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1383 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001384 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001385 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001386 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001387 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001389 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001390 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1391 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1392 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1393 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1394 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1395 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1396 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1397 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001398 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1399 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1400 "src/qu8-dwconv/up8x9-minmax-neon.c",
1401 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1402 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1403 "src/qu8-gemm/4x8-minmax-neon.c",
1404 "src/qu8-gemm/8x8-minmax-neon.c",
1405 "src/qu8-igemm/4x8-minmax-neon.c",
1406 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001407 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001408 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001409 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001410 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001411 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001412 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001414 "src/x8-zip/x2-neon.c",
1415 "src/x8-zip/x3-neon.c",
1416 "src/x8-zip/x4-neon.c",
1417 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001418 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001419 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001420 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001421 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001422 "src/x32-zip/x2-neon.c",
1423 "src/x32-zip/x3-neon.c",
1424 "src/x32-zip/x4-neon.c",
1425 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001426]
1427
1428NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1430 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1431 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1432 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1433 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1434 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1435 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1436 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1437 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1438 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1439 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1440 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1441 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1442 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1443 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1444 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1445 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1446 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1447 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1448 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1449 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1450 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1451 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1452 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1453 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1454 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1455 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1456 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1457 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1458 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001459 "src/f32-ibilinear/gen/neonfma-c4.c",
1460 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001461 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001462 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001463 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001464 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1465 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001466 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1467 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001468 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1469 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001470 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1471 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001472 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001473 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001474 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001475 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1476 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001477 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001478 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1479 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001480 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001481 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1482 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001483 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1484 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1485 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1486 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1487 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1488 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1489 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1490 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1491 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1492 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1493 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1494 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1495 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001496 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001502 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001508 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1510 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1511 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1512 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1513 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001514 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1515 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1516 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1517 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1518 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1519 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001520 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1521 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1522 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1523 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1524 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1525 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001526 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1527 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1528 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1529 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1530 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1531 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1533 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1534 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1535 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1536 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1537 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1538 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1539 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1540 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1541 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1542 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1543 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1544 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1545 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1546 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1547 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1548 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1549 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001550 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1551 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001552 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1553 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1554 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1555 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1556 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1557 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1558 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1559 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1560 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1561 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1562 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1563 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1564 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1565 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1566 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1567 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1568 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1569 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1570 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1571 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001572 "src/math/exp-neonfma-lut64-p2.c",
1573 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001574 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001575 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001576 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001577 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1578 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1579 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001580 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1581 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1582 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001583 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1584 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1585 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001586 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1587 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1588 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001589 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1590 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1591 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001592 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1593 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1594 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001595 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001596 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001597 "src/math/sqrt-neonfma-nr2fma.c",
1598 "src/math/sqrt-neonfma-nr2fma1adj.c",
1599 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001600]
1601
1602AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001603 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001604 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001605 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001606 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001607 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001608 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001609 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001610 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001611 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001612 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
1613 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1614 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1615 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1616 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1617 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
1618 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1619 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
1621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001622 "src/f32-dwconv2d-chw/3x3s2p1-neonfma-1x4-acc3.c",
1623 "src/f32-dwconv2d-chw/5x5p2-neonfma-3x4.c",
1624 "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001625 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1626 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1627 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1628 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1629 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1630 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1631 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1632 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1633 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1634 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1635 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1636 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1637 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1638 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1639 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1640 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1641 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1642 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1643 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1644 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001645 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1656 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001657 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1658 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1659 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1660 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1661 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1662 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001663 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1664 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1665 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1666 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1667 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1668 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1669 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1670 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1671 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1672 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001673 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1674 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1675 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1676 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1677 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1678 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1679 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1680 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1681 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1682 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1683 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1684 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1685 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1686 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001687 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1688 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001689 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001690 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001691 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001692 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001694 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001695]
1696
Marat Dukhan8853b822020-05-07 12:19:01 -07001697NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001698 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1699 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001700 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1701 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1702 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1703 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1704 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1705 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001706 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001707 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001708 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001709 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001710]
1711
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001713 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001714 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001715 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1716 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1717 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1718 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001719 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1720 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1721 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1722 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1723 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1724 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1725 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1726 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001727 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1728 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001729 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1736 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1737 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1738 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1739 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1740 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1741 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1742 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1743 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1744 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001745 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001746 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1747 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1748 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1749 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1750 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1751 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1752 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1753 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1754 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001755 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001756 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001757 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001758 "src/f16-relu/gen/neonfp16arith-x16.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001759 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001760 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001761 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001762 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001763 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001764 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001765 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001766 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1767 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1768 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1769 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1770 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1771 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1772 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1773 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1774 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1775 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1776 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1777 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1778 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1779 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1780 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1781 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1782 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1783 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1784 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1785 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1786 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1787 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1788 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1789 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1790 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1791 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1792 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1793 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1794 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1795 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1796 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797]
1798
Benoit Jacoba9644732020-08-13 12:48:55 -07001799NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001800 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001801 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001802 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001803 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001804 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001805 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001806 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001807 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001808 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001809 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001810 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001811 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001812 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001813 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001814 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001815 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001816 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001817 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001818]
1819
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001821 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1822 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001823 "src/f32-clamp/gen/sse-x4.c",
1824 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001825 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1826 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001827 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1828 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1829 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1830 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001831 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1832 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001833 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1834 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1835 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1836 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001837 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1838 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001839 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1840 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1841 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1842 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1843 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1844 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1845 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1846 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1847 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1848 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001849 "src/f32-dwconv2d-chw/3x3s2p1-sse-1x4-acc3.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001850 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001851 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1852 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001853 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1854 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1855 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1856 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1857 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1858 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001859 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1860 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1861 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1862 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1863 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1864 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1865 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001866 "src/f32-hswish/gen/hswish-sse-x4.c",
1867 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001868 "src/f32-ibilinear/gen/sse-c4.c",
1869 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001870 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1871 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1872 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1873 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1874 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1875 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1876 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001877 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1878 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1879 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001880 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001881 "src/f32-prelu/gen/sse-2x4.c",
1882 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001883 "src/f32-relu/gen/sse-x4.c",
1884 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001885 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001886 "src/f32-spmm/gen/4x1-minmax-sse.c",
1887 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001888 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001889 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1890 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1891 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1892 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1893 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1894 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1895 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1896 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001897 "src/f32-vbinary/gen/vmax-sse-x4.c",
1898 "src/f32-vbinary/gen/vmax-sse-x8.c",
1899 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1900 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1901 "src/f32-vbinary/gen/vmin-sse-x4.c",
1902 "src/f32-vbinary/gen/vmin-sse-x8.c",
1903 "src/f32-vbinary/gen/vminc-sse-x4.c",
1904 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001905 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1906 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1907 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1908 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1909 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1910 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1911 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1912 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001913 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1914 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1915 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1916 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001917 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1918 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1919 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1920 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001921 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1922 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001923 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1924 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001925 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1926 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001927 "src/f32-vunary/gen/vabs-sse-x4.c",
1928 "src/f32-vunary/gen/vabs-sse-x8.c",
1929 "src/f32-vunary/gen/vneg-sse-x4.c",
1930 "src/f32-vunary/gen/vneg-sse-x8.c",
1931 "src/f32-vunary/gen/vsqr-sse-x4.c",
1932 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001933 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001934 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001935 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001936 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001937 "src/math/sqrt-sse-hh1mac.c",
1938 "src/math/sqrt-sse-nr1mac.c",
1939 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001940 "src/x32-fill/sse.c",
1941 "src/x32-packx/x4-sse.c",
1942 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001943]
1944
1945SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001946 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001947 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08001948 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001949 "src/f32-prelu/gen/sse2-2x4.c",
1950 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001951 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001952 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001953 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001954 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1955 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001957 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1958 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001959 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001960 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1961 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001962 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001963 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1964 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1965 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1966 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1967 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1968 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001969 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001970 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001971 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001972 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001973 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1974 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001975 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1976 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001977 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1978 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001979 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1980 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1981 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1982 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1983 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1984 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1985 "src/math/exp-sse2-lut64-p2.c",
1986 "src/math/exp-sse2-p5.c",
1987 "src/math/expminus-sse2-p5.c",
1988 "src/math/roundd-sse2-cvt.c",
1989 "src/math/roundne-sse2-cvt.c",
1990 "src/math/roundu-sse2-cvt.c",
1991 "src/math/roundz-sse2-cvt.c",
1992 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1993 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1994 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1995 "src/math/sigmoid-sse2-rr2-p5-div.c",
1996 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1997 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001998 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1999 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2000 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002001 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2002 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2003 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002004 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2005 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2006 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002007 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002008 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002009 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002010 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002011 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002012 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002013 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2014 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002015 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002016 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2017 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002018 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002019 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2020 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2021 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002022 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002023 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002024 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002025 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002026 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002027 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002028 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002029 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002030 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2031 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002032 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002033 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002034 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002035 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2036 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2037 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2038 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002039 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2040 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2041 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2042 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002043 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2044 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002045 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2046 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2047 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2048 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2049 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002050 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002051 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002052 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002053 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002054 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002055 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002056 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002058 "src/x8-zip/x2-sse2.c",
2059 "src/x8-zip/x3-sse2.c",
2060 "src/x8-zip/x4-sse2.c",
2061 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002062 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002063 "src/x32-zip/x2-sse2.c",
2064 "src/x32-zip/x3-sse2.c",
2065 "src/x32-zip/x4-sse2.c",
2066 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002067]
2068
2069SSSE3_UKERNELS = [
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2073 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2074 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2075 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
2076 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2077 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2078 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002080 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2081 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2082 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002083 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2084 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2085 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002086 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2087 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2088 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002089 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002090 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002091 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002092 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002093 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002094 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002095 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2096 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002097 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002098 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2099 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002100 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002101 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2102 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2103 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002104 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002105 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002106 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002107 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002108 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002109 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002110 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002111 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002112 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2113 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002114 "src/qs8-requantization/precise-ssse3.c",
2115 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002116 "src/qu8-requantization/precise-ssse3.c",
2117 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118]
2119
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002120SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002121 "src/f32-prelu/gen/sse41-2x4.c",
2122 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002123 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2124 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2125 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2126 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2127 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2128 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002129 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002130 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002131 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002132 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002133 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2134 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002135 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2136 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002137 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2138 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002139 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2140 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2141 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2142 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2143 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2144 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2145 "src/math/roundd-sse41.c",
2146 "src/math/roundne-sse41.c",
2147 "src/math/roundu-sse41.c",
2148 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002149 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2150 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2151 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002152 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2153 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2154 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002155 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2156 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2157 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002158 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002159 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002160 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002161 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002162 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002163 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2165 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002166 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002167 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2168 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002169 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002170 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2171 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2172 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002173 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002174 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002175 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002176 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002178 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002180 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002181 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2182 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002183 "src/qs8-requantization/fp32-sse4.c",
2184 "src/qs8-requantization/precise-sse4.c",
2185 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002186 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2187 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2188 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2189 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002190 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2191 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2192 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2193 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002194 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2195 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2196 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2197 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002198 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2199 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2200 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2201 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002202 "src/qu8-requantization/precise-sse4.c",
2203 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002204]
2205
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002207 "src/f32-clamp/gen/avx-x8.c",
2208 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002209 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2210 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002211 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2212 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002213 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2214 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002215 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2216 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2217 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2218 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2219 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2220 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002221 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002222 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2223 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002224 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002225 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002226 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002227 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002228 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2229 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2230 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2231 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2232 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2233 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2234 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2235 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2236 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2237 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2238 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002239 "src/f32-hswish/gen/hswish-avx-x8.c",
2240 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002241 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002242 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2243 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002244 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002245 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002246 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002247 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002248 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2249 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002250 "src/f32-prelu/gen/avx-2x8.c",
2251 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002252 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002253 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002254 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002255 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2256 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2257 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2258 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2259 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2260 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2261 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2262 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2263 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2264 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2265 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2266 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2267 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2268 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2269 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2270 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2271 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2272 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2273 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2274 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002275 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2276 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2277 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2278 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2279 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2280 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2281 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2282 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002283 "src/f32-vbinary/gen/vmax-avx-x8.c",
2284 "src/f32-vbinary/gen/vmax-avx-x16.c",
2285 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2286 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2287 "src/f32-vbinary/gen/vmin-avx-x8.c",
2288 "src/f32-vbinary/gen/vmin-avx-x16.c",
2289 "src/f32-vbinary/gen/vminc-avx-x8.c",
2290 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002291 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2292 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2293 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2294 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2295 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2296 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2297 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2298 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002299 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2300 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2301 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2302 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002303 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2304 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2305 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2306 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002307 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2308 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002309 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2310 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002311 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2312 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2313 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2314 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2315 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2316 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2317 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002318 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2319 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002320 "src/f32-vunary/gen/vabs-avx-x8.c",
2321 "src/f32-vunary/gen/vabs-avx-x16.c",
2322 "src/f32-vunary/gen/vneg-avx-x8.c",
2323 "src/f32-vunary/gen/vneg-avx-x16.c",
2324 "src/f32-vunary/gen/vsqr-avx-x8.c",
2325 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002326 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002327 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002328 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002329 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2330 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002331]
2332
Marat Dukhan1566fee2020-08-02 21:55:41 -07002333XOP_UKERNELS = [
2334 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002335 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002336 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002337 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002338 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002339 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002340 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2341 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002342 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002343 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2344 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002345 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002346 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2347 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2348 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002349 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002350 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002351 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002352 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002354 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002355 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002356 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002357 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2358 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002359 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2360 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2361 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2362 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2363 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2364 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2365 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2366 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002367]
2368
Marat Dukhanfda12b82019-11-21 12:27:59 -08002369FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2371 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002372 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2373 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002374 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2375 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002376 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2377 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2378 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2379 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2380 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2381 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002382 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002383 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2384 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2385 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2386 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002387 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002388 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2389 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002390 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002391 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2392 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002393 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2394 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2395 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002396 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2397 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2398 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2399 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2400 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2401 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2402 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2403 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2404 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2405 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2406 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2407 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2408 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2409 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002410 "src/f32-hswish/gen/hswish-fma3-x8.c",
2411 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002412 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002413 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2414 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2415 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2416 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002417 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2419 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002420 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2422 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002423 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2424 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2425 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002426 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2427 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2428 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2429 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2430 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2431 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2432 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2433 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002434 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002435 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002436 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002437]
2438
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002439AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002440 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2441 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002442 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002443 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002444 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002445 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2446 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002447 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002448 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2449 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2450 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002451 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002452 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2453 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002455 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002456 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002457 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2458 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002459 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002460 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2461 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2462 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002463 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002464 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2465 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002466 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002467 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002468 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002469 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2470 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002471 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002472 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2473 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2474 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002475 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002476 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2477 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2478 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2479 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2480 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2481 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2482 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2483 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2484 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2485 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2486 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2487 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2488 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2489 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2490 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2491 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2492 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2493 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2494 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2495 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2496 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2497 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2498 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2499 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2500 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2501 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2502 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2503 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2504 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2505 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002506 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2507 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2508 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2509 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2510 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2511 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2512 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2513 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2514 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2515 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2516 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2517 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2518 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2519 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2520 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2521 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2522 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2523 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2524 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2525 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2526 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2527 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2528 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2529 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/math/exp-avx2-lut8-p3-perm.c",
2531 "src/math/exp-avx2-lut8-p4-perm.c",
2532 "src/math/exp-avx2-p5.c",
2533 "src/math/expminus-avx2-p5.c",
2534 "src/math/extexp-avx2-p5.c",
2535 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2536 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2537 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2538 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2539 "src/math/sigmoid-avx2-rr1-p5-div.c",
2540 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2541 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2542 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2543 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2544 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2545 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2546 "src/math/sigmoid-avx2-rr2-p5-div.c",
2547 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2548 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002549 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002551 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2552 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002553 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002554 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002555 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002556 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002557 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002558 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002559 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002560 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002561 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2562 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2563 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002564 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2565 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2566 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2567 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2568 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2569 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2570 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2571 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002572]
2573
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002575 "src/f32-clamp/gen/avx512f-x16.c",
2576 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002577 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2578 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002579 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2580 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002581 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2582 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002583 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2584 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2585 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2586 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2587 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2588 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002589 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2590 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2591 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2592 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2593 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2594 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2596 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2597 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2598 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2599 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2600 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002601 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2602 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002603 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2604 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2605 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2606 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2607 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2608 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002609 "src/f32-prelu/gen/avx512f-2x16.c",
2610 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002611 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2612 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002613 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002614 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002615 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002616 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2617 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002619 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2620 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2621 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002622 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002623 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2624 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002625 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002626 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002627 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002628 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2629 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002631 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2632 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2633 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002634 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002635 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2636 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002638 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002639 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002640 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2641 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002643 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2644 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2645 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002646 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002647 "src/f32-relu/gen/avx512f-x16.c",
2648 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002649 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002650 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2651 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2652 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2653 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2654 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2655 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2656 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2657 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2658 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2659 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2660 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2661 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2662 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2663 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2664 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2665 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002666 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2667 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2668 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2669 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2670 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2671 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2672 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2673 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2674 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2675 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2676 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2677 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2678 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2679 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2680 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2681 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002682 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2683 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2684 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2685 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2686 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2687 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2688 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2689 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2690 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2691 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2692 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2693 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2694 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2695 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2696 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2697 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002698 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2699 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2700 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2701 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2702 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2703 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2704 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2705 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002706 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2707 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2708 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2709 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2710 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2711 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2712 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2713 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002714 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2715 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2716 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2717 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2718 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2719 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2720 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2721 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002722 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2723 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2724 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2725 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002726 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2727 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2728 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2729 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002730 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2731 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002732 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2733 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2734 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2735 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2736 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2737 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2738 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2739 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002740 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002741 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2742 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2743 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2744 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2745 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2746 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2747 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2748 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2749 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2750 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2751 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2752 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2753 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2754 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2755 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2756 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2757 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2758 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2759 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2760 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2761 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2762 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2763 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2764 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002765 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2766 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2767 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2768 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2769 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2770 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2771 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2772 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002773 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2774 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2775 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2776 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2777 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2778 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002779 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2780 "src/math/exp-avx512f-lut16-p3-perm.c",
2781 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2782 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002783 "src/math/exp-avx512f-p5-scalef.c",
2784 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002785 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002786 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002787 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002789 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002790 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002791 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002792 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002793 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002794 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2795 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2796 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2797 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2798 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2799 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2800 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2801 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2802 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2803 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002804 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002805 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002806 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2807 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2808 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2809 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002810 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002811 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002813]
2814
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002815AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002816 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2817 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002818 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2819 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2820 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2821 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2822 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2823 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2824 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2825 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2826]
2827
Frank Barchardbcedc082020-08-17 18:00:51 -07002828WASM32_ASM_UKERNELS = [
2829 "src/f32-relu/wasm_shr_x1.S",
2830 "src/f32-relu/wasm_shr_x2.S",
2831 "src/f32-relu/wasm_shr_x4.S",
2832]
2833
Marat Dukhan08c4a432019-10-03 09:29:21 -07002834AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002835 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002836 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002837 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2838 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002839 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002840 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002841 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002842 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002843 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2844 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002845 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2846 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2847 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2848 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002849]
2850
2851AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002852 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002853 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002854 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002855 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002856 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002857 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002858 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002859 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2860 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2861 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2862 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2863 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2864 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2865 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002866 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2867 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002868 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2869 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2870 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002871 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2872 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002873 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2874 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2875 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2876 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002877 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002878 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2879 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002880 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2881 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2882 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2883 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002884 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002885 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002886 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002887 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002888 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2889 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2890 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2891 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2892 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2893 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2894 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2895 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2896 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2897 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2898 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2899 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2900 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2901 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2902 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2903 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2904 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2905 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2906 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2907 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2908 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2909 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002910 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002911 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002912 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2913 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002914 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2915 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2916 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2917 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2918 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2919 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002920 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2921 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2922 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2923 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002924 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2925 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002926 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002927 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002928 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929]
2930
Marat Dukhan1b354632020-03-23 12:50:22 -07002931INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 "src/xnnpack/argmaxpool.h",
2933 "src/xnnpack/avgpool.h",
2934 "src/xnnpack/clamp.h",
2935 "src/xnnpack/common.h",
2936 "src/xnnpack/conv.h",
2937 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002938 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 "src/xnnpack/gavgpool.h",
2940 "src/xnnpack/gemm.h",
2941 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002942 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002944 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 "src/xnnpack/lut.h",
2946 "src/xnnpack/math.h",
2947 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002948 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 "src/xnnpack/packx.h",
2950 "src/xnnpack/pad.h",
2951 "src/xnnpack/params.h",
2952 "src/xnnpack/pavgpool.h",
2953 "src/xnnpack/ppmm.h",
2954 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002955 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002956 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002957 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 "src/xnnpack/rmax.h",
2959 "src/xnnpack/scalar-utils.h",
2960 "src/xnnpack/spmm.h",
2961 "src/xnnpack/unpool.h",
2962 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002963 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002965 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002966 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002967 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002968 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002970]
2971
2972INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973 "include/xnnpack.h",
2974 "src/xnnpack/allocator.h",
2975 "src/xnnpack/compute.h",
2976 "src/xnnpack/im2col.h",
2977 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002978 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002979 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002980 "src/xnnpack/operator.h",
2981 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002982 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002983 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002984 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002985 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002986]
2987
Marat Dukhan1b354632020-03-23 12:50:22 -07002988ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002989 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990]
2991
Marat Dukhan1b354632020-03-23 12:50:22 -07002992MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002993 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 "include/xnnpack.h",
2995]
2996
Marat Dukhan1b354632020-03-23 12:50:22 -07002997MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002998 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002999 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 "src/xnnpack/requantization.h",
3001 "include/xnnpack.h",
3002]
3003
3004OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003006 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007]
3008
3009WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003011 "src/xnnpack/operator.h",
3012 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013]
3014
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003015LOGGING_COPTS = select({
3016 # No logging in optimized mode
3017 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3018 # Full logging in debug mode
3019 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3020 # Error-only logging in default (fastbuild) mode
3021 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3022})
3023
Marat Dukhan3b59de22020-06-03 20:15:19 -07003024LOGGING_SRCS = select({
3025 # No logging in optimized mode
3026 ":optimized_build": [],
3027 "//conditions:default": [
3028 "src/operator-strings.c",
3029 "src/subgraph-strings.c",
3030 ],
3031})
3032
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003033LOGGING_HDRS = [
3034 "src/xnnpack/log.h",
3035]
3036
Marat Dukhan08c4a432019-10-03 09:29:21 -07003037xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003038 name = "tables",
3039 srcs = TABLE_SRCS,
3040 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003041 gcc_copts = xnnpack_gcc_std_copts(),
3042 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003043)
3044
3045xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003046 name = "scalar_ukernels",
3047 srcs = SCALAR_UKERNELS,
3048 hdrs = INTERNAL_HDRS,
3049 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003050 gcc_copts = xnnpack_gcc_std_copts(),
3051 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003053 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 "@FP16",
3055 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003056 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003057 ],
3058)
3059
3060xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003061 name = "scalar_ukernels_test_mode",
3062 srcs = SCALAR_UKERNELS,
3063 hdrs = INTERNAL_HDRS,
3064 aarch32_copts = ["-marm"],
3065 copts = [
3066 "-UNDEBUG",
3067 "-DXNN_TEST_MODE=1",
3068 ],
3069 gcc_copts = xnnpack_gcc_std_copts(),
3070 msvc_copts = xnnpack_msvc_std_copts(),
3071 deps = [
3072 ":tables",
3073 "@FP16",
3074 "@FXdiv",
3075 "@pthreadpool",
3076 ],
3077)
3078
3079xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003080 name = "wasm_ukernels",
3081 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003082 gcc_copts = xnnpack_gcc_std_copts(),
3083 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003084 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003085 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003086 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003087 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003088 "@FP16",
3089 "@FXdiv",
3090 "@pthreadpool",
3091 ],
3092)
3093
3094xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003095 name = "wasm_ukernels_test_mode",
3096 hdrs = INTERNAL_HDRS,
3097 copts = [
3098 "-UNDEBUG",
3099 "-DXNN_TEST_MODE=1",
3100 ],
3101 gcc_copts = xnnpack_gcc_std_copts(),
3102 msvc_copts = xnnpack_msvc_std_copts(),
3103 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003104 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003105 deps = [
3106 ":tables",
3107 "@FP16",
3108 "@FXdiv",
3109 "@pthreadpool",
3110 ],
3111)
3112
3113xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003114 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 hdrs = INTERNAL_HDRS,
3116 aarch32_copts = [
3117 "-marm",
3118 "-mfpu=neon",
3119 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003120 gcc_copts = xnnpack_gcc_std_copts(),
3121 gcc_x86_copts = ["-msse2"],
3122 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123 optimized_copts = [
3124 "-O3",
3125 "-ffast-math",
3126 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003127 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003129 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 "@FP16",
3131 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003132 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003133 ],
3134)
3135
3136xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003137 name = "psimd_fastmath_ukernels_test_mode",
3138 hdrs = INTERNAL_HDRS,
3139 aarch32_copts = [
3140 "-marm",
3141 "-mfpu=neon",
3142 ],
3143 copts = [
3144 "-UNDEBUG",
3145 "-DXNN_TEST_MODE=1",
3146 ],
3147 gcc_copts = xnnpack_gcc_std_copts(),
3148 gcc_x86_copts = ["-msse2"],
3149 msvc_copts = xnnpack_msvc_std_copts(),
3150 optimized_copts = [
3151 "-O3",
3152 "-ffast-math",
3153 ],
3154 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3155 deps = [
3156 ":tables",
3157 "@FP16",
3158 "@psimd",
3159 "@pthreadpool",
3160 ],
3161)
3162
3163xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003164 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003165 hdrs = INTERNAL_HDRS,
3166 aarch32_copts = [
3167 "-marm",
3168 "-mfpu=neon",
3169 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003170 gcc_copts = xnnpack_gcc_std_copts(),
3171 gcc_x86_copts = ["-msse2"],
3172 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003173 optimized_copts = [
3174 "-O3",
3175 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003176 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003177 deps = [
3178 ":tables",
3179 "@FP16",
3180 "@psimd",
3181 "@pthreadpool",
3182 ],
3183)
3184
3185xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003186 name = "psimd_accmath_ukernels_test_mode",
3187 hdrs = INTERNAL_HDRS,
3188 aarch32_copts = [
3189 "-marm",
3190 "-mfpu=neon",
3191 ],
3192 copts = [
3193 "-UNDEBUG",
3194 "-DXNN_TEST_MODE=1",
3195 ],
3196 gcc_copts = xnnpack_gcc_std_copts(),
3197 gcc_x86_copts = ["-msse2"],
3198 msvc_copts = xnnpack_msvc_std_copts(),
3199 optimized_copts = [
3200 "-O3",
3201 ],
3202 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3203 deps = [
3204 ":tables",
3205 "@FP16",
3206 "@psimd",
3207 "@pthreadpool",
3208 ],
3209)
3210
3211xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 name = "neon_ukernels",
3213 hdrs = INTERNAL_HDRS,
3214 aarch32_copts = [
3215 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003216 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217 "-mfpu=neon",
3218 ],
3219 aarch32_srcs = NEON_UKERNELS,
3220 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003221 gcc_copts = xnnpack_gcc_std_copts(),
3222 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003223 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003224 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003225 "@FP16",
3226 "@pthreadpool",
3227 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228)
3229
3230xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003231 name = "neon_ukernels_test_mode",
3232 hdrs = INTERNAL_HDRS,
3233 aarch32_copts = [
3234 "-marm",
3235 "-march=armv7-a",
3236 "-mfpu=neon",
3237 ],
3238 aarch32_srcs = NEON_UKERNELS,
3239 aarch64_srcs = NEON_UKERNELS,
3240 copts = [
3241 "-UNDEBUG",
3242 "-DXNN_TEST_MODE=1",
3243 ],
3244 gcc_copts = xnnpack_gcc_std_copts(),
3245 msvc_copts = xnnpack_msvc_std_copts(),
3246 deps = [
3247 ":tables",
3248 "@FP16",
3249 "@pthreadpool",
3250 ],
3251)
3252
3253xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 name = "neonfma_ukernels",
3255 hdrs = INTERNAL_HDRS,
3256 aarch32_copts = [
3257 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003258 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259 "-mfpu=neon-vfpv4",
3260 ],
3261 aarch32_srcs = NEONFMA_UKERNELS,
3262 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003263 apple_aarch32_copts = [
3264 "-mcpu=swift",
3265 "-mtune=generic",
3266 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003267 gcc_copts = xnnpack_gcc_std_copts(),
3268 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003269 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003270 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003271 "@FP16",
3272 "@pthreadpool",
3273 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003274)
3275
3276xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003277 name = "neonfma_ukernels_test_mode",
3278 hdrs = INTERNAL_HDRS,
3279 aarch32_copts = [
3280 "-marm",
3281 "-march=armv7-a",
3282 "-mfpu=neon-vfpv4",
3283 ],
3284 aarch32_srcs = NEONFMA_UKERNELS,
3285 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003286 apple_aarch32_copts = [
3287 "-mcpu=swift",
3288 "-mtune=generic",
3289 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003290 copts = [
3291 "-UNDEBUG",
3292 "-DXNN_TEST_MODE=1",
3293 ],
3294 gcc_copts = xnnpack_gcc_std_copts(),
3295 msvc_copts = xnnpack_msvc_std_copts(),
3296 deps = [
3297 ":tables",
3298 "@FP16",
3299 "@pthreadpool",
3300 ],
3301)
3302
3303xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003304 name = "neonv8_ukernels",
3305 hdrs = INTERNAL_HDRS,
3306 aarch32_copts = [
3307 "-marm",
3308 "-march=armv8-a",
3309 "-mfpu=neon-fp-armv8",
3310 ],
3311 aarch32_srcs = NEONV8_UKERNELS,
3312 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003313 apple_aarch32_copts = [
3314 "-mcpu=cyclone",
3315 "-mtune=generic",
3316 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003317 gcc_copts = xnnpack_gcc_std_copts(),
3318 msvc_copts = xnnpack_msvc_std_copts(),
3319 deps = [
3320 ":tables",
3321 "@FP16",
3322 "@pthreadpool",
3323 ],
3324)
3325
3326xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003327 name = "neonv8_ukernels_test_mode",
3328 hdrs = INTERNAL_HDRS,
3329 aarch32_copts = [
3330 "-marm",
3331 "-march=armv8-a",
3332 "-mfpu=neon-fp-armv8",
3333 ],
3334 aarch32_srcs = NEONV8_UKERNELS,
3335 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003336 apple_aarch32_copts = [
3337 "-mcpu=cyclone",
3338 "-mtune=generic",
3339 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003340 copts = [
3341 "-UNDEBUG",
3342 "-DXNN_TEST_MODE=1",
3343 ],
3344 gcc_copts = xnnpack_gcc_std_copts(),
3345 msvc_copts = xnnpack_msvc_std_copts(),
3346 deps = [
3347 ":tables",
3348 "@FP16",
3349 "@pthreadpool",
3350 ],
3351)
3352
3353xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003354 name = "neonfp16arith_ukernels",
3355 hdrs = INTERNAL_HDRS,
3356 aarch64_copts = ["-march=armv8.2-a+fp16"],
3357 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003358 gcc_copts = xnnpack_gcc_std_copts(),
3359 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003360 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003361 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003362 "@FP16",
3363 "@pthreadpool",
3364 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365)
3366
3367xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003368 name = "neonfp16arith_ukernels_test_mode",
3369 hdrs = INTERNAL_HDRS,
3370 aarch64_copts = ["-march=armv8.2-a+fp16"],
3371 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3372 copts = [
3373 "-UNDEBUG",
3374 "-DXNN_TEST_MODE=1",
3375 ],
3376 gcc_copts = xnnpack_gcc_std_copts(),
3377 msvc_copts = xnnpack_msvc_std_copts(),
3378 deps = [
3379 ":tables",
3380 "@FP16",
3381 "@pthreadpool",
3382 ],
3383)
3384
3385xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003386 name = "neondot_ukernels",
3387 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003388 aarch32_copts = [
3389 "-marm",
3390 "-march=armv8.2-a+dotprod",
3391 "-mfpu=neon-fp-armv8",
3392 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003393 aarch32_srcs = NEONDOT_UKERNELS,
3394 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3395 aarch64_srcs = NEONDOT_UKERNELS,
3396 gcc_copts = xnnpack_gcc_std_copts(),
3397 msvc_copts = xnnpack_msvc_std_copts(),
3398 deps = [
3399 ":tables",
3400 "@FP16",
3401 "@pthreadpool",
3402 ],
3403)
3404
3405xnnpack_cc_library(
3406 name = "neondot_ukernels_test_mode",
3407 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003408 aarch32_copts = [
3409 "-marm",
3410 "-march=armv8.2-a+dotprod",
3411 "-mfpu=neon-fp-armv8",
3412 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003413 aarch32_srcs = NEONDOT_UKERNELS,
3414 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3415 aarch64_srcs = NEONDOT_UKERNELS,
3416 copts = [
3417 "-UNDEBUG",
3418 "-DXNN_TEST_MODE=1",
3419 ],
3420 gcc_copts = xnnpack_gcc_std_copts(),
3421 msvc_copts = xnnpack_msvc_std_copts(),
3422 deps = [
3423 ":tables",
3424 "@FP16",
3425 "@pthreadpool",
3426 ],
3427)
3428
3429xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 name = "sse2_ukernels",
3431 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003432 gcc_copts = xnnpack_gcc_std_copts(),
3433 gcc_x86_copts = ["-msse2"],
3434 msvc_copts = xnnpack_msvc_std_copts(),
3435 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003437 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003438 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003439 "@FP16",
3440 "@pthreadpool",
3441 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003442)
3443
3444xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003445 name = "sse2_ukernels_test_mode",
3446 hdrs = INTERNAL_HDRS,
3447 copts = [
3448 "-UNDEBUG",
3449 "-DXNN_TEST_MODE=1",
3450 ],
3451 gcc_copts = xnnpack_gcc_std_copts(),
3452 gcc_x86_copts = ["-msse2"],
3453 msvc_copts = xnnpack_msvc_std_copts(),
3454 msvc_x86_32_copts = ["/arch:SSE2"],
3455 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3456 deps = [
3457 ":tables",
3458 "@FP16",
3459 "@pthreadpool",
3460 ],
3461)
3462
3463xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003464 name = "ssse3_ukernels",
3465 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003466 gcc_copts = xnnpack_gcc_std_copts(),
3467 gcc_x86_copts = ["-mssse3"],
3468 msvc_copts = xnnpack_msvc_std_copts(),
3469 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003470 x86_srcs = SSSE3_UKERNELS,
3471 deps = [
3472 ":tables",
3473 "@FP16",
3474 "@pthreadpool",
3475 ],
3476)
3477
3478xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003479 name = "ssse3_ukernels_test_mode",
3480 hdrs = INTERNAL_HDRS,
3481 copts = [
3482 "-UNDEBUG",
3483 "-DXNN_TEST_MODE=1",
3484 ],
3485 gcc_copts = xnnpack_gcc_std_copts(),
3486 gcc_x86_copts = ["-mssse3"],
3487 msvc_copts = xnnpack_msvc_std_copts(),
3488 msvc_x86_32_copts = ["/arch:SSE2"],
3489 x86_srcs = SSSE3_UKERNELS,
3490 deps = [
3491 ":tables",
3492 "@FP16",
3493 "@pthreadpool",
3494 ],
3495)
3496
3497xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003498 name = "sse41_ukernels",
3499 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003500 gcc_copts = xnnpack_gcc_std_copts(),
3501 gcc_x86_copts = ["-msse4.1"],
3502 msvc_copts = xnnpack_msvc_std_copts(),
3503 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003504 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003505 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003506 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003507 "@FP16",
3508 "@pthreadpool",
3509 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003510)
3511
3512xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003513 name = "sse41_ukernels_test_mode",
3514 hdrs = INTERNAL_HDRS,
3515 copts = [
3516 "-UNDEBUG",
3517 "-DXNN_TEST_MODE=1",
3518 ],
3519 gcc_copts = xnnpack_gcc_std_copts(),
3520 gcc_x86_copts = ["-msse4.1"],
3521 msvc_copts = xnnpack_msvc_std_copts(),
3522 msvc_x86_32_copts = ["/arch:SSE2"],
3523 x86_srcs = SSE41_UKERNELS,
3524 deps = [
3525 ":tables",
3526 "@FP16",
3527 "@pthreadpool",
3528 ],
3529)
3530
3531xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003532 name = "avx_ukernels",
3533 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003534 gcc_copts = xnnpack_gcc_std_copts(),
3535 gcc_x86_copts = ["-mavx"],
3536 msvc_copts = xnnpack_msvc_std_copts(),
3537 msvc_x86_32_copts = ["/arch:AVX"],
3538 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003540 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003541 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003542 "@FP16",
3543 "@pthreadpool",
3544 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003545)
3546
3547xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003548 name = "avx_ukernels_test_mode",
3549 hdrs = INTERNAL_HDRS,
3550 copts = [
3551 "-UNDEBUG",
3552 "-DXNN_TEST_MODE=1",
3553 ],
3554 gcc_copts = xnnpack_gcc_std_copts(),
3555 gcc_x86_copts = ["-mavx"],
3556 msvc_copts = xnnpack_msvc_std_copts(),
3557 msvc_x86_32_copts = ["/arch:AVX"],
3558 msvc_x86_64_copts = ["/arch:AVX"],
3559 x86_srcs = AVX_UKERNELS,
3560 deps = [
3561 ":tables",
3562 "@FP16",
3563 "@pthreadpool",
3564 ],
3565)
3566
3567xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003568 name = "xop_ukernels",
3569 hdrs = INTERNAL_HDRS,
3570 gcc_copts = xnnpack_gcc_std_copts(),
3571 gcc_x86_copts = ["-mxop"],
3572 msvc_copts = xnnpack_msvc_std_copts(),
3573 msvc_x86_32_copts = ["/arch:AVX"],
3574 msvc_x86_64_copts = ["/arch:AVX"],
3575 x86_srcs = XOP_UKERNELS,
3576 deps = [
3577 ":tables",
3578 "@FP16",
3579 "@pthreadpool",
3580 ],
3581)
3582
3583xnnpack_cc_library(
3584 name = "xop_ukernels_test_mode",
3585 hdrs = INTERNAL_HDRS,
3586 copts = [
3587 "-UNDEBUG",
3588 "-DXNN_TEST_MODE=1",
3589 ],
3590 gcc_copts = xnnpack_gcc_std_copts(),
3591 gcc_x86_copts = ["-mxop"],
3592 msvc_copts = xnnpack_msvc_std_copts(),
3593 msvc_x86_32_copts = ["/arch:AVX"],
3594 msvc_x86_64_copts = ["/arch:AVX"],
3595 x86_srcs = XOP_UKERNELS,
3596 deps = [
3597 ":tables",
3598 "@FP16",
3599 "@pthreadpool",
3600 ],
3601)
3602
3603xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003604 name = "fma3_ukernels",
3605 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003606 gcc_copts = xnnpack_gcc_std_copts(),
3607 gcc_x86_copts = ["-mfma"],
3608 msvc_copts = xnnpack_msvc_std_copts(),
3609 msvc_x86_32_copts = ["/arch:AVX"],
3610 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003611 x86_srcs = FMA3_UKERNELS,
3612 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003613 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003614 "@FP16",
3615 "@pthreadpool",
3616 ],
3617)
3618
3619xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003620 name = "fma3_ukernels_test_mode",
3621 hdrs = INTERNAL_HDRS,
3622 copts = [
3623 "-UNDEBUG",
3624 "-DXNN_TEST_MODE=1",
3625 ],
3626 gcc_copts = xnnpack_gcc_std_copts(),
3627 gcc_x86_copts = ["-mfma"],
3628 msvc_copts = xnnpack_msvc_std_copts(),
3629 msvc_x86_32_copts = ["/arch:AVX"],
3630 msvc_x86_64_copts = ["/arch:AVX"],
3631 x86_srcs = FMA3_UKERNELS,
3632 deps = [
3633 ":tables",
3634 "@FP16",
3635 "@pthreadpool",
3636 ],
3637)
3638
3639xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003640 name = "avx2_ukernels",
3641 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003642 gcc_copts = xnnpack_gcc_std_copts(),
3643 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003644 "-mfma",
3645 "-mavx2",
3646 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003647 msvc_copts = xnnpack_msvc_std_copts(),
3648 msvc_x86_32_copts = ["/arch:AVX2"],
3649 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003650 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003651 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003652 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003653 "@FP16",
3654 "@pthreadpool",
3655 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003656)
3657
3658xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003659 name = "avx2_ukernels_test_mode",
3660 hdrs = INTERNAL_HDRS,
3661 copts = [
3662 "-UNDEBUG",
3663 "-DXNN_TEST_MODE=1",
3664 ],
3665 gcc_copts = xnnpack_gcc_std_copts(),
3666 gcc_x86_copts = [
3667 "-mfma",
3668 "-mavx2",
3669 ],
3670 msvc_copts = xnnpack_msvc_std_copts(),
3671 msvc_x86_32_copts = ["/arch:AVX2"],
3672 msvc_x86_64_copts = ["/arch:AVX2"],
3673 x86_srcs = AVX2_UKERNELS,
3674 deps = [
3675 ":tables",
3676 "@FP16",
3677 "@pthreadpool",
3678 ],
3679)
3680
3681xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682 name = "avx512f_ukernels",
3683 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003684 gcc_copts = xnnpack_gcc_std_copts(),
3685 gcc_x86_copts = ["-mavx512f"],
3686 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3687 msvc_copts = xnnpack_msvc_std_copts(),
3688 msvc_x86_32_copts = ["/arch:AVX512"],
3689 msvc_x86_64_copts = ["/arch:AVX512"],
3690 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003691 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003692 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003693 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003694 "@FP16",
3695 "@pthreadpool",
3696 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697)
3698
3699xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003700 name = "avx512f_ukernels_test_mode",
3701 hdrs = INTERNAL_HDRS,
3702 copts = [
3703 "-UNDEBUG",
3704 "-DXNN_TEST_MODE=1",
3705 ],
3706 gcc_copts = xnnpack_gcc_std_copts(),
3707 gcc_x86_copts = ["-mavx512f"],
3708 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3709 msvc_copts = xnnpack_msvc_std_copts(),
3710 msvc_x86_32_copts = ["/arch:AVX512"],
3711 msvc_x86_64_copts = ["/arch:AVX512"],
3712 msys_copts = ["-fno-asynchronous-unwind-tables"],
3713 x86_srcs = AVX512F_UKERNELS,
3714 deps = [
3715 ":tables",
3716 "@FP16",
3717 "@pthreadpool",
3718 ],
3719)
3720
3721xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003722 name = "avx512skx_ukernels",
3723 hdrs = INTERNAL_HDRS,
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(
3746 name = "avx512skx_ukernels_test_mode",
3747 hdrs = INTERNAL_HDRS,
3748 copts = [
3749 "-UNDEBUG",
3750 "-DXNN_TEST_MODE=1",
3751 ],
3752 gcc_copts = xnnpack_gcc_std_copts(),
3753 gcc_x86_copts = [
3754 "-mavx512f",
3755 "-mavx512cd",
3756 "-mavx512bw",
3757 "-mavx512dq",
3758 "-mavx512vl",
3759 ],
3760 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3761 msvc_copts = xnnpack_msvc_std_copts(),
3762 msvc_x86_32_copts = ["/arch:AVX512"],
3763 msvc_x86_64_copts = ["/arch:AVX512"],
3764 msys_copts = ["-fno-asynchronous-unwind-tables"],
3765 x86_srcs = AVX512SKX_UKERNELS,
3766 deps = [
3767 ":tables",
3768 "@FP16",
3769 "@pthreadpool",
3770 ],
3771)
3772
3773xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774 name = "asm_ukernels",
3775 hdrs = ["src/xnnpack/assembly.h"],
3776 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003777 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003778 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003779 wasm_srcs = WASM32_ASM_UKERNELS,
3780 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781)
3782
Marat Dukhan3b59de22020-06-03 20:15:19 -07003783xnnpack_cc_library(
3784 name = "logging_utils",
3785 srcs = LOGGING_SRCS,
3786 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3787 copts = LOGGING_COPTS + [
3788 "-Isrc",
3789 "-Iinclude",
3790 ] + select({
3791 ":debug_build": [],
3792 "//conditions:default": xnnpack_min_size_copts(),
3793 }),
3794 gcc_copts = xnnpack_gcc_std_copts(),
3795 msvc_copts = xnnpack_msvc_std_copts(),
3796 visibility = xnnpack_visibility(),
3797 deps = [
3798 "@FP16",
3799 "@clog",
3800 "@pthreadpool",
3801 ],
3802)
3803
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804xnnpack_aggregate_library(
3805 name = "ukernels",
3806 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003807 ":neon_ukernels",
3808 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003809 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003810 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003811 ":asm_ukernels",
3812 ],
3813 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 ":neon_ukernels",
3815 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003816 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003817 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003818 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819 ":asm_ukernels",
3820 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003821 generic_deps = [
3822 ":scalar_ukernels",
3823 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003824 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003825 ":psimd_fastmath_ukernels",
3826 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003828 wasm_deps = [
3829 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003830 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003831 ],
3832 wasmsimd_deps = [
3833 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003834 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003835 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003837 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003838 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003839 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003840 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003841 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003842 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003843 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003845 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846 ],
3847)
3848
Marat Dukhan33fcf782020-05-24 14:27:15 -07003849xnnpack_aggregate_library(
3850 name = "ukernels_test_mode",
3851 aarch32_deps = [
3852 ":neon_ukernels_test_mode",
3853 ":neonfma_ukernels_test_mode",
3854 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003855 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003856 ":asm_ukernels",
3857 ],
3858 aarch64_deps = [
3859 ":neon_ukernels_test_mode",
3860 ":neonfma_ukernels_test_mode",
3861 ":neonv8_ukernels_test_mode",
3862 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003863 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003864 ":asm_ukernels",
3865 ],
3866 generic_deps = [
3867 ":scalar_ukernels_test_mode",
3868 ],
3869 psimd_deps = [
3870 ":psimd_fastmath_ukernels_test_mode",
3871 ":psimd_accmath_ukernels_test_mode",
3872 ],
3873 wasm_deps = [
3874 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003875 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003876 ],
3877 wasmsimd_deps = [
3878 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003879 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003880 ],
3881 x86_deps = [
3882 ":sse2_ukernels_test_mode",
3883 ":ssse3_ukernels_test_mode",
3884 ":sse41_ukernels_test_mode",
3885 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003886 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003887 ":fma3_ukernels_test_mode",
3888 ":avx2_ukernels_test_mode",
3889 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003890 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003891 ],
3892)
3893
Marat Dukhan86df3842020-06-17 14:03:31 -07003894# This target must be compatible with Android NDK r17 and compile with Clang 6.
3895# It is similar to :ukernels target, but excludes microkernels which require
3896# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003897xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003898 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003899 aarch32_deps = [
3900 ":neon_ukernels",
3901 ":neonfma_ukernels",
3902 ":neonv8_ukernels",
3903 ":asm_ukernels",
3904 ],
3905 aarch64_deps = [
3906 ":neon_ukernels",
3907 ":neonfma_ukernels",
3908 ":neonv8_ukernels",
3909 ":asm_ukernels",
3910 ],
3911 generic_deps = [
3912 ":scalar_ukernels",
3913 ],
3914 psimd_deps = [
3915 ":psimd_fastmath_ukernels",
3916 ":psimd_accmath_ukernels",
3917 ],
3918 wasm_deps = [
3919 ":wasm_ukernels",
3920 ],
3921 wasmsimd_deps = [
3922 ":wasm_ukernels",
3923 ],
3924 x86_deps = [
3925 ":sse2_ukernels",
3926 ":ssse3_ukernels",
3927 ":sse41_ukernels",
3928 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003929 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003930 ":fma3_ukernels",
3931 ":avx2_ukernels",
3932 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003933 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003934 ],
3935)
3936
Marat Dukhan08c4a432019-10-03 09:29:21 -07003937xnnpack_cc_library(
3938 name = "im2col",
3939 srcs = ["src/im2col.c"],
3940 hdrs = [
3941 "src/xnnpack/common.h",
3942 "src/xnnpack/im2col.h",
3943 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003944 gcc_copts = xnnpack_gcc_std_copts(),
3945 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946)
3947
3948xnnpack_cc_library(
3949 name = "indirection",
3950 srcs = ["src/indirection.c"],
3951 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003952 gcc_copts = xnnpack_gcc_std_copts(),
3953 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 deps = [
3955 "@FP16",
3956 "@FXdiv",
3957 "@pthreadpool",
3958 ],
3959)
3960
3961xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003962 name = "indirection_test_mode",
3963 srcs = ["src/indirection.c"],
3964 hdrs = INTERNAL_HDRS,
3965 copts = [
3966 "-UNDEBUG",
3967 "-DXNN_TEST_MODE=1",
3968 ],
3969 gcc_copts = xnnpack_gcc_std_copts(),
3970 msvc_copts = xnnpack_msvc_std_copts(),
3971 deps = [
3972 "@FP16",
3973 "@FXdiv",
3974 "@pthreadpool",
3975 ],
3976)
3977
3978xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003979 name = "packing",
3980 srcs = ["src/packing.c"],
3981 hdrs = INTERNAL_HDRS,
3982 gcc_copts = xnnpack_gcc_std_copts(),
3983 msvc_copts = xnnpack_msvc_std_copts(),
3984 deps = [
3985 "@FP16",
3986 "@FXdiv",
3987 "@pthreadpool",
3988 ],
3989)
3990
3991xnnpack_cc_library(
3992 name = "packing_test_mode",
3993 srcs = ["src/packing.c"],
3994 hdrs = INTERNAL_HDRS,
3995 copts = [
3996 "-UNDEBUG",
3997 "-DXNN_TEST_MODE=1",
3998 ],
3999 gcc_copts = xnnpack_gcc_std_copts(),
4000 msvc_copts = xnnpack_msvc_std_copts(),
4001 deps = [
4002 "@FP16",
4003 "@FXdiv",
4004 "@pthreadpool",
4005 ],
4006)
4007
4008xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009 name = "operator_run",
4010 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004011 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004012 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004013 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4014 "//conditions:default": [],
4015 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004016 gcc_copts = xnnpack_gcc_std_copts(),
4017 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004018 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004019 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004020 "@FP16",
4021 "@FXdiv",
4022 "@clog",
4023 "@pthreadpool",
4024 ],
4025)
4026
Chao Mei6ddfc602020-05-13 22:29:36 -07004027xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004028 name = "operator_run_test_mode",
4029 srcs = ["src/operator-run.c"],
4030 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4031 copts = LOGGING_COPTS + [
4032 "-UNDEBUG",
4033 "-DXNN_TEST_MODE=1",
4034 ] + select({
4035 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4036 "//conditions:default": [],
4037 }),
4038 gcc_copts = xnnpack_gcc_std_copts(),
4039 msvc_copts = xnnpack_msvc_std_copts(),
4040 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004041 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004042 "@FP16",
4043 "@FXdiv",
4044 "@clog",
4045 "@pthreadpool",
4046 ],
4047)
4048
4049xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004050 name = "memory_planner",
4051 srcs = ["src/memory-planner.c"],
4052 hdrs = INTERNAL_HDRS,
4053 defines = select({
4054 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4055 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4056 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4057 }),
4058 gcc_copts = xnnpack_gcc_std_copts(),
4059 msvc_copts = xnnpack_msvc_std_copts(),
4060 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004061 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004062 "@pthreadpool",
4063 ],
4064)
4065
Marat Dukhan33fcf782020-05-24 14:27:15 -07004066xnnpack_cc_library(
4067 name = "memory_planner_test_mode",
4068 srcs = ["src/memory-planner.c"],
4069 hdrs = INTERNAL_HDRS,
4070 copts = [
4071 "-UNDEBUG",
4072 "-DXNN_TEST_MODE=1",
4073 ],
4074 defines = select({
4075 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4076 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4077 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4078 }),
4079 gcc_copts = xnnpack_gcc_std_copts(),
4080 msvc_copts = xnnpack_msvc_std_copts(),
4081 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004082 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004083 "@pthreadpool",
4084 ],
4085)
4086
Marat Dukhan08c4a432019-10-03 09:29:21 -07004087cc_library(
4088 name = "enable_assembly",
4089 defines = select({
4090 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4091 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004092 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093 }),
4094)
4095
Marat Dukhan9de90e02020-06-18 16:04:12 -07004096cc_library(
4097 name = "enable_sparse",
4098 defines = select({
4099 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4100 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004101 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004102 }),
4103)
4104
Marat Dukhancf056b22019-10-07 10:26:29 -07004105xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106 name = "operators",
4107 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004108 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004109 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004110 ],
4111 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004112 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004113 "-Isrc",
4114 "-Iinclude",
4115 ] + select({
4116 ":debug_build": [],
4117 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004118 }) + select({
4119 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4120 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004122 gcc_copts = xnnpack_gcc_std_copts(),
4123 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004124 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004125 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004126 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004127 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004128 "@FP16",
4129 "@FXdiv",
4130 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004132 ],
4133)
4134
Marat Dukhan10a38082020-04-17 03:58:35 -07004135xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004136 name = "operators_test_mode",
4137 srcs = OPERATOR_SRCS + [
4138 "src/memory.c",
4139 "src/operator-delete.c",
4140 ],
4141 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4142 copts = LOGGING_COPTS + [
4143 "-Isrc",
4144 "-Iinclude",
4145 "-UNDEBUG",
4146 "-DXNN_TEST_MODE=1",
4147 ] + select({
4148 ":debug_build": [],
4149 "//conditions:default": xnnpack_min_size_copts(),
4150 }) + select({
4151 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4152 "//conditions:default": [],
4153 }),
4154 gcc_copts = xnnpack_gcc_std_copts(),
4155 msvc_copts = xnnpack_msvc_std_copts(),
4156 deps = [
4157 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004158 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004159 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004160 "@FP16",
4161 "@FXdiv",
4162 "@clog",
4163 "@pthreadpool",
4164 ],
4165)
4166
4167xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004168 name = "XNNPACK",
4169 srcs = [
4170 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004171 "src/runtime.c",
4172 "src/subgraph.c",
4173 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004174 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004175 hdrs = ["include/xnnpack.h"],
4176 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004177 "-Isrc",
4178 "-Iinclude",
4179 ] + select({
4180 ":debug_build": [],
4181 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004182 }) + select({
4183 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4184 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004185 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004186 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004187 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004188 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004189 visibility = xnnpack_visibility(),
4190 deps = [
4191 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004192 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004193 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004194 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004195 ":operator_run",
4196 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004197 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004198 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004199 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004200 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004201 ] + select({
4202 ":emscripten": [],
4203 "//conditions:default": ["@cpuinfo"],
4204 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205)
4206
Marat Dukhan10a38082020-04-17 03:58:35 -07004207xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004208 name = "XNNPACK_test_mode",
4209 srcs = [
4210 "src/init.c",
4211 "src/runtime.c",
4212 "src/subgraph.c",
4213 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004214 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004215 hdrs = ["include/xnnpack.h"],
4216 copts = LOGGING_COPTS + [
4217 "-Isrc",
4218 "-Iinclude",
4219 "-UNDEBUG",
4220 "-DXNN_TEST_MODE=1",
4221 ] + select({
4222 ":debug_build": [],
4223 "//conditions:default": xnnpack_min_size_copts(),
4224 }) + select({
4225 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4226 "//conditions:default": [],
4227 }),
4228 gcc_copts = xnnpack_gcc_std_copts(),
4229 includes = ["include"],
4230 msvc_copts = xnnpack_msvc_std_copts(),
4231 visibility = xnnpack_visibility(),
4232 deps = [
4233 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004234 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004235 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004236 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004237 ":operator_run_test_mode",
4238 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004239 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004240 "@clog",
4241 "@FP16",
4242 "@pthreadpool",
4243 ] + select({
4244 ":emscripten": [],
4245 "//conditions:default": ["@cpuinfo"],
4246 }),
4247)
4248
4249xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004250 name = "xnnpack_f32",
4251 srcs = [
4252 "src/init.c",
4253 "src/runtime.c",
4254 "src/subgraph.c",
4255 "src/tensor.c",
4256 ] + SUBGRAPH_SRCS,
4257 hdrs = ["include/xnnpack.h"],
4258 copts = LOGGING_COPTS + [
4259 "-Isrc",
4260 "-Iinclude",
4261 ] + select({
4262 ":debug_build": [],
4263 "//conditions:default": xnnpack_min_size_copts(),
4264 }) + select({
4265 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4266 "//conditions:default": [],
4267 }),
4268 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004269 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004270 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004271 "XNN_NO_U8_OPERATORS",
4272 "XNN_NO_X8_OPERATORS",
4273 "XNN_NO_F16_OPERATORS",
4274 "XNN_NO_X16_OPERATORS",
4275 ],
4276 gcc_copts = xnnpack_gcc_std_copts(),
4277 includes = ["include"],
4278 msvc_copts = xnnpack_msvc_std_copts(),
4279 visibility = xnnpack_visibility(),
4280 deps = [
4281 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004282 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004283 ":logging_utils",
4284 ":memory_planner",
4285 ":operator_run",
4286 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004287 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004288 "@clog",
4289 "@FP16",
4290 "@pthreadpool",
4291 ] + select({
4292 ":emscripten": [],
4293 "//conditions:default": ["@cpuinfo"],
4294 }),
4295)
4296
4297xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004298 name = "xnnpack_operators_nhwc_f32",
4299 srcs = [
4300 "src/init.c",
4301 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004302 hdrs = ["include/xnnpack.h"],
4303 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004304 "-Isrc",
4305 "-Iinclude",
4306 ] + select({
4307 ":debug_build": [],
4308 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004309 }) + select({
4310 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4311 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004312 }),
4313 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004314 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004315 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004316 "XNN_NO_U8_OPERATORS",
4317 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004318 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004319 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004320 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004321 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004322 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004323 visibility = xnnpack_visibility(),
4324 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004325 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004326 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327 ":operator_run",
4328 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004329 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004330 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004331 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004332 ] + select({
4333 ":emscripten": [],
4334 "//conditions:default": ["@cpuinfo"],
4335 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004336)
4337
Marat Dukhancf056b22019-10-07 10:26:29 -07004338xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004339 name = "bench_utils",
4340 srcs = ["bench/utils.cc"],
4341 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004342 deps = [
4343 "@com_google_benchmark//:benchmark",
4344 "@cpuinfo",
4345 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346)
4347
Frank Barchard7e955972019-10-11 10:34:25 -07004348######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349
4350xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004351 name = "qs8_gemm_bench",
4352 srcs = [
4353 "bench/gemm.h",
4354 "bench/qs8-gemm.cc",
4355 "src/xnnpack/AlignedAllocator.h",
4356 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004357 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4358 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004359)
4360
4361xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004362 name = "qs8_requantization_bench",
4363 srcs = [
4364 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004365 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004366 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004367 ] + MICROKERNEL_BENCHMARK_HDRS,
4368 deps = MICROKERNEL_BENCHMARK_DEPS,
4369)
4370
4371xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004372 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004373 srcs = [
4374 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004375 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 "src/xnnpack/AlignedAllocator.h",
4377 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004378 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004379 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380)
4381
4382xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004383 name = "qu8_requantization_bench",
4384 srcs = [
4385 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004386 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004387 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004388 ] + MICROKERNEL_BENCHMARK_HDRS,
4389 deps = MICROKERNEL_BENCHMARK_DEPS,
4390)
4391
4392xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004393 name = "f16_igemm_bench",
4394 srcs = [
4395 "bench/f16-igemm.cc",
4396 "bench/conv.h",
4397 "bench/google/conv.h",
4398 "src/xnnpack/AlignedAllocator.h",
4399 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004400 deps = MICROKERNEL_BENCHMARK_DEPS + [
4401 ":indirection",
4402 ":packing",
4403 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004404)
4405
4406xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004407 name = "f16_gemm_bench",
4408 srcs = [
4409 "bench/f16-gemm.cc",
4410 "bench/gemm.h",
4411 "src/xnnpack/AlignedAllocator.h",
4412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004413 deps = MICROKERNEL_BENCHMARK_DEPS + [
4414 ":packing",
4415 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416)
4417
4418xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004419 name = "f16_spmm_bench",
4420 srcs = [
4421 "bench/f16-spmm.cc",
4422 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004423 "src/xnnpack/AlignedAllocator.h",
4424 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004425 deps = MICROKERNEL_BENCHMARK_DEPS,
4426)
4427
4428xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 name = "f32_igemm_bench",
4430 srcs = [
4431 "bench/f32-igemm.cc",
4432 "bench/conv.h",
4433 "src/xnnpack/AlignedAllocator.h",
4434 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004435 deps = MICROKERNEL_BENCHMARK_DEPS + [
4436 ":indirection",
4437 ":packing",
4438 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439)
4440
4441xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004442 name = "f16_relu_bench",
4443 srcs = [
4444 "bench/f16-relu.cc",
4445 "src/xnnpack/AlignedAllocator.h",
4446 ] + MICROKERNEL_BENCHMARK_HDRS,
4447 deps = MICROKERNEL_BENCHMARK_DEPS,
4448)
4449
4450xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 name = "f32_conv_hwc_bench",
4452 srcs = [
4453 "bench/f32-conv-hwc.cc",
4454 "bench/dconv.h",
4455 "src/xnnpack/AlignedAllocator.h",
4456 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004457 deps = MICROKERNEL_BENCHMARK_DEPS + [
4458 ":packing",
4459 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460)
4461
4462xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004463 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004464 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004465 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004466 "bench/dconv.h",
4467 "src/xnnpack/AlignedAllocator.h",
4468 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004469 deps = MICROKERNEL_BENCHMARK_DEPS + [
4470 ":packing",
4471 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004472)
4473
4474xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004475 name = "f16_dwconv_bench",
4476 srcs = [
4477 "bench/f16-dwconv.cc",
4478 "bench/dwconv.h",
4479 "bench/google/dwconv.h",
4480 "src/xnnpack/AlignedAllocator.h",
4481 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004482 deps = MICROKERNEL_BENCHMARK_DEPS + [
4483 ":indirection",
4484 ":packing",
4485 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004486)
4487
4488xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004489 name = "f32_dwconv_bench",
4490 srcs = [
4491 "bench/f32-dwconv.cc",
4492 "bench/dwconv.h",
4493 "src/xnnpack/AlignedAllocator.h",
4494 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004495 deps = MICROKERNEL_BENCHMARK_DEPS + [
4496 ":indirection",
4497 ":packing",
4498 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499)
4500
4501xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004502 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004504 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505 "bench/dwconv.h",
4506 "src/xnnpack/AlignedAllocator.h",
4507 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004508 deps = MICROKERNEL_BENCHMARK_DEPS + [
4509 ":indirection",
4510 ":packing",
4511 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512)
4513
4514xnnpack_benchmark(
4515 name = "f32_gemm_bench",
4516 srcs = [
4517 "bench/f32-gemm.cc",
4518 "bench/gemm.h",
4519 "src/xnnpack/AlignedAllocator.h",
4520 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004521 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004522 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004523)
4524
4525xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004526 name = "f32_hswish_bench",
4527 srcs = [
4528 "bench/f32-hswish.cc",
4529 "src/xnnpack/AlignedAllocator.h",
4530 ] + MICROKERNEL_BENCHMARK_HDRS,
4531 deps = MICROKERNEL_BENCHMARK_DEPS,
4532)
4533
4534xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004535 name = "f32_raddexpminusmax_bench",
4536 srcs = [
4537 "bench/f32-raddexpminusmax.cc",
4538 "src/xnnpack/AlignedAllocator.h",
4539 ] + MICROKERNEL_BENCHMARK_HDRS,
4540 deps = MICROKERNEL_BENCHMARK_DEPS,
4541)
4542
4543xnnpack_benchmark(
4544 name = "f32_raddextexp_bench",
4545 srcs = [
4546 "bench/f32-raddextexp.cc",
4547 "src/xnnpack/AlignedAllocator.h",
4548 ] + MICROKERNEL_BENCHMARK_HDRS,
4549 deps = MICROKERNEL_BENCHMARK_DEPS,
4550)
4551
4552xnnpack_benchmark(
4553 name = "f32_raddstoreexpminusmax_bench",
4554 srcs = [
4555 "bench/f32-raddstoreexpminusmax.cc",
4556 "src/xnnpack/AlignedAllocator.h",
4557 ] + MICROKERNEL_BENCHMARK_HDRS,
4558 deps = MICROKERNEL_BENCHMARK_DEPS,
4559)
4560
4561xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004562 name = "f32_relu_bench",
4563 srcs = [
4564 "bench/f32-relu.cc",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_BENCHMARK_HDRS,
4567 deps = MICROKERNEL_BENCHMARK_DEPS,
4568)
4569
4570xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 name = "f32_rmax_bench",
4572 srcs = [
4573 "bench/f32-rmax.cc",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_BENCHMARK_HDRS,
4576 deps = MICROKERNEL_BENCHMARK_DEPS,
4577)
4578
4579xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004580 name = "f32_sigmoid_bench",
4581 srcs = [
4582 "bench/f32-sigmoid.cc",
4583 "src/xnnpack/AlignedAllocator.h",
4584 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004585 deps = MICROKERNEL_BENCHMARK_DEPS,
4586)
4587
4588xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004589 name = "f32_spmm_bench",
4590 srcs = [
4591 "bench/f32-spmm.cc",
4592 "bench/gemm.h",
4593 "src/xnnpack/AlignedAllocator.h",
4594 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004595 deps = MICROKERNEL_BENCHMARK_DEPS,
4596)
4597
4598xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004599 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004600 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004601 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004602 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004603 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004604 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004605)
4606
4607xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004608 name = "f32_vscaleexpminusmax_bench",
4609 srcs = [
4610 "bench/f32-vscaleexpminusmax.cc",
4611 "src/xnnpack/AlignedAllocator.h",
4612 ] + MICROKERNEL_BENCHMARK_HDRS,
4613 deps = MICROKERNEL_BENCHMARK_DEPS,
4614)
4615
4616xnnpack_benchmark(
4617 name = "f32_vscaleextexp_bench",
4618 srcs = [
4619 "bench/f32-vscaleextexp.cc",
4620 "src/xnnpack/AlignedAllocator.h",
4621 ] + MICROKERNEL_BENCHMARK_HDRS,
4622 deps = MICROKERNEL_BENCHMARK_DEPS,
4623)
4624
4625xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004626 name = "f32_vsqrt_bench",
4627 srcs = [
4628 "bench/f32-vsqrt.cc",
4629 "src/xnnpack/AlignedAllocator.h",
4630 ] + MICROKERNEL_BENCHMARK_HDRS,
4631 deps = MICROKERNEL_BENCHMARK_DEPS,
4632)
4633
4634xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004635 name = "f32_im2col_gemm_bench",
4636 srcs = [
4637 "bench/f32-im2col-gemm.cc",
4638 "bench/conv.h",
4639 "src/xnnpack/AlignedAllocator.h",
4640 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004641 deps = MICROKERNEL_BENCHMARK_DEPS + [
4642 ":im2col",
4643 ":packing",
4644 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645)
4646
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004647xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004648 name = "rounding_bench",
4649 srcs = [
4650 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004651 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004652 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004653 ] + MICROKERNEL_BENCHMARK_HDRS,
4654 deps = MICROKERNEL_BENCHMARK_DEPS,
4655)
4656
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657########################### Benchmarks for operators ###########################
4658
4659xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 name = "average_pooling_bench",
4661 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004662 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004663 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004664 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665)
4666
4667xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004668 name = "bankers_rounding_bench",
4669 srcs = ["bench/bankers-rounding.cc"],
4670 copts = xnnpack_optional_tflite_copts(),
4671 tags = ["nowin32"],
4672 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4673)
4674
4675xnnpack_benchmark(
4676 name = "ceiling_bench",
4677 srcs = ["bench/ceiling.cc"],
4678 copts = xnnpack_optional_tflite_copts(),
4679 tags = ["nowin32"],
4680 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4681)
4682
4683xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004684 name = "channel_shuffle_bench",
4685 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004686 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004687)
4688
4689xnnpack_benchmark(
4690 name = "convolution_bench",
4691 srcs = ["bench/convolution.cc"],
4692 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004693 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004694 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004695)
4696
4697xnnpack_benchmark(
4698 name = "deconvolution_bench",
4699 srcs = ["bench/deconvolution.cc"],
4700 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004701 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004702 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004703)
4704
4705xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004706 name = "floor_bench",
4707 srcs = ["bench/floor.cc"],
4708 copts = xnnpack_optional_tflite_copts(),
4709 tags = ["nowin32"],
4710 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4711)
4712
4713xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004714 name = "global_average_pooling_bench",
4715 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004716 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717)
4718
4719xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004720 name = "hardswish_bench",
4721 srcs = ["bench/hardswish.cc"],
4722 copts = xnnpack_optional_tflite_copts(),
4723 tags = ["nowin32"],
4724 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4725)
4726
4727xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 name = "max_pooling_bench",
4729 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004730 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004731)
4732
4733xnnpack_benchmark(
4734 name = "sigmoid_bench",
4735 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004736 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004737 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004738 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739)
4740
4741xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004742 name = "prelu_bench",
4743 srcs = ["bench/prelu.cc"],
4744 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004745 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004746 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004747)
4748
4749xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004750 name = "softmax_bench",
4751 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004752 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004753 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004754 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755)
4756
Marat Dukhan87727142020-06-24 15:24:10 -07004757xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004758 name = "square_root_bench",
4759 srcs = ["bench/square-root.cc"],
4760 copts = xnnpack_optional_tflite_copts(),
4761 tags = ["nowin32"],
4762 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4763)
4764
4765xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004766 name = "truncation_bench",
4767 srcs = ["bench/truncation.cc"],
4768 deps = OPERATOR_BENCHMARK_DEPS,
4769)
4770
Marat Dukhanc068bb62019-10-04 13:24:39 -07004771############################# End-to-end benchmarks ############################
4772
4773cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004774 name = "fp32_mobilenet_v1",
4775 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004776 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004777 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004778 linkstatic = True,
4779 deps = [
4780 ":XNNPACK",
4781 "@pthreadpool",
4782 ],
4783)
4784
4785cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004786 name = "fp16_mobilenet_v1",
4787 srcs = ["models/fp16-mobilenet-v1.cc"],
4788 hdrs = ["models/models.h"],
4789 copts = xnnpack_std_cxxopts(),
4790 linkstatic = True,
4791 deps = [
4792 ":XNNPACK",
4793 "@FP16",
4794 "@pthreadpool",
4795 ],
4796)
4797
4798cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004799 name = "qs8_mobilenet_v1",
4800 srcs = ["models/qs8-mobilenet-v1.cc"],
4801 hdrs = ["models/models.h"],
4802 copts = xnnpack_std_cxxopts(),
4803 linkstatic = True,
4804 deps = [
4805 ":XNNPACK",
4806 "@pthreadpool",
4807 ],
4808)
4809
4810cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004811 name = "qs8_mobilenet_v2",
4812 srcs = ["models/qs8-mobilenet-v2.cc"],
4813 hdrs = ["models/models.h"],
4814 copts = xnnpack_std_cxxopts(),
4815 linkstatic = True,
4816 deps = [
4817 ":XNNPACK",
4818 "@pthreadpool",
4819 ],
4820)
4821
4822cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004823 name = "fp32_mobilenet_v2",
4824 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004825 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004826 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004827 linkstatic = True,
4828 deps = [
4829 ":XNNPACK",
4830 "@pthreadpool",
4831 ],
4832)
4833
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004834cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004835 name = "fp16_mobilenet_v2",
4836 srcs = ["models/fp16-mobilenet-v2.cc"],
4837 hdrs = ["models/models.h"],
4838 copts = xnnpack_std_cxxopts(),
4839 linkstatic = True,
4840 deps = [
4841 ":XNNPACK",
4842 "@FP16",
4843 "@pthreadpool",
4844 ],
4845)
4846
4847cc_library(
4848 name = "fp32_mobilenet_v3_large",
4849 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004850 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004851 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004852 linkstatic = True,
4853 deps = [
4854 ":XNNPACK",
4855 "@pthreadpool",
4856 ],
4857)
4858
4859cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004860 name = "fp16_mobilenet_v3_large",
4861 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4862 hdrs = ["models/models.h"],
4863 copts = xnnpack_std_cxxopts(),
4864 linkstatic = True,
4865 deps = [
4866 ":XNNPACK",
4867 "@FP16",
4868 "@pthreadpool",
4869 ],
4870)
4871
4872cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004873 name = "fp32_mobilenet_v3_small",
4874 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004875 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004876 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004877 linkstatic = True,
4878 deps = [
4879 ":XNNPACK",
4880 "@pthreadpool",
4881 ],
4882)
4883
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004884cc_library(
4885 name = "fp16_mobilenet_v3_small",
4886 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4887 hdrs = ["models/models.h"],
4888 copts = xnnpack_std_cxxopts(),
4889 linkstatic = True,
4890 deps = [
4891 ":XNNPACK",
4892 "@FP16",
4893 "@pthreadpool",
4894 ],
4895)
4896
Marat Dukhanc068bb62019-10-04 13:24:39 -07004897xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004898 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004899 srcs = [
4900 "bench/f32-dwconv-e2e.cc",
4901 "bench/end2end.h",
4902 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004903 deps = MICROKERNEL_BENCHMARK_DEPS + [
4904 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004905 ":fp32_mobilenet_v1",
4906 ":fp32_mobilenet_v2",
4907 ":fp32_mobilenet_v3_large",
4908 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004909 ],
4910)
4911
4912xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004913 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004914 srcs = [
4915 "bench/f32-gemm-e2e.cc",
4916 "bench/end2end.h",
4917 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004918 deps = MICROKERNEL_BENCHMARK_DEPS + [
4919 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004920 ":fp32_mobilenet_v1",
4921 ":fp32_mobilenet_v2",
4922 ":fp32_mobilenet_v3_large",
4923 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004924 ],
4925)
4926
4927xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004928 name = "end2end_bench",
4929 srcs = ["bench/end2end.cc"],
4930 deps = [
4931 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004932 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004933 ":fp16_mobilenet_v1",
4934 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004935 ":fp16_mobilenet_v3_large",
4936 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004937 ":fp32_mobilenet_v1",
4938 ":fp32_mobilenet_v2",
4939 ":fp32_mobilenet_v3_large",
4940 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004941 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004942 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004943 "@pthreadpool",
4944 ],
4945)
4946
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004947#################### Accuracy evaluation for math functions ####################
4948
4949xnnpack_benchmark(
4950 name = "f32_exp_eval",
4951 srcs = [
4952 "eval/f32-exp.cc",
4953 "src/xnnpack/AlignedAllocator.h",
4954 ] + ACCURACY_EVAL_HDRS,
4955 deps = ACCURACY_EVAL_DEPS,
4956)
4957
Marat Dukhan515c9772019-10-17 18:07:57 -07004958xnnpack_benchmark(
4959 name = "f32_expminus_eval",
4960 srcs = [
4961 "eval/f32-expminus.cc",
4962 "src/xnnpack/AlignedAllocator.h",
4963 ] + ACCURACY_EVAL_HDRS,
4964 deps = ACCURACY_EVAL_DEPS,
4965)
4966
Marat Dukhan98ba4412019-10-23 02:14:28 -07004967xnnpack_benchmark(
4968 name = "f32_extexp_eval",
4969 srcs = [
4970 "eval/f32-extexp.cc",
4971 "src/xnnpack/AlignedAllocator.h",
4972 ] + ACCURACY_EVAL_HDRS,
4973 deps = ACCURACY_EVAL_DEPS,
4974)
4975
Marat Dukhan8853b822020-05-07 12:19:01 -07004976xnnpack_unit_test(
4977 name = "f32_roundne_eval",
4978 srcs = [
4979 "eval/f32-roundne.cc",
4980 "src/xnnpack/AlignedAllocator.h",
4981 "src/xnnpack/math-stubs.h",
4982 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004983 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004984 deps = MICROKERNEL_TEST_DEPS,
4985)
4986
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004987xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004988 name = "f32_roundd_eval",
4989 srcs = [
4990 "eval/f32-roundd.cc",
4991 "src/xnnpack/AlignedAllocator.h",
4992 "src/xnnpack/math-stubs.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 automatic = False,
4995 deps = MICROKERNEL_TEST_DEPS,
4996)
4997
4998xnnpack_unit_test(
4999 name = "f32_roundu_eval",
5000 srcs = [
5001 "eval/f32-roundu.cc",
5002 "src/xnnpack/AlignedAllocator.h",
5003 "src/xnnpack/math-stubs.h",
5004 ] + MICROKERNEL_TEST_HDRS,
5005 automatic = False,
5006 deps = MICROKERNEL_TEST_DEPS,
5007)
5008
5009xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005010 name = "f32_roundz_eval",
5011 srcs = [
5012 "eval/f32-roundz.cc",
5013 "src/xnnpack/AlignedAllocator.h",
5014 "src/xnnpack/math-stubs.h",
5015 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005016 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005017 deps = MICROKERNEL_TEST_DEPS,
5018)
5019
Marat Dukhan346a9e52019-11-15 09:06:30 -08005020xnnpack_benchmark(
5021 name = "f32_sigmoid_eval",
5022 srcs = [
5023 "eval/f32-sigmoid.cc",
5024 "src/xnnpack/AlignedAllocator.h",
5025 ] + ACCURACY_EVAL_HDRS,
5026 deps = ACCURACY_EVAL_DEPS,
5027)
5028
Marat Dukhan84000762020-06-29 18:38:43 -07005029xnnpack_benchmark(
5030 name = "f32_sqrt_eval",
5031 srcs = [
5032 "eval/f32-sqrt.cc",
5033 "src/xnnpack/AlignedAllocator.h",
5034 ] + ACCURACY_EVAL_HDRS,
5035 deps = ACCURACY_EVAL_DEPS + [
5036 ":bench_utils",
5037 ],
5038)
5039
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040######################### Unit tests for micro-kernels #########################
5041
5042xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005043 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005044 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005045 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 "test/gemm-microkernel-tester.h",
5047 "src/xnnpack/AlignedAllocator.h",
5048 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005049 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050)
5051
5052xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005053 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005054 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005055 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005056 "test/spmm-microkernel-tester.h",
5057 "src/xnnpack/AlignedAllocator.h",
5058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063 name = "f32_argmaxpool_test",
5064 srcs = [
5065 "test/f32-argmaxpool.cc",
5066 "test/argmaxpool-microkernel-tester.h",
5067 "src/xnnpack/AlignedAllocator.h",
5068 ] + MICROKERNEL_TEST_HDRS,
5069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
5072xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005073 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005075 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076 "test/avgpool-microkernel-tester.h",
5077 "src/xnnpack/AlignedAllocator.h",
5078 ] + MICROKERNEL_TEST_HDRS,
5079 deps = MICROKERNEL_TEST_DEPS,
5080)
5081
5082xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005083 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005084 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005085 "test/f32-ibilinear.cc",
5086 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005087 "src/xnnpack/AlignedAllocator.h",
5088 ] + MICROKERNEL_TEST_HDRS,
5089 deps = MICROKERNEL_TEST_DEPS,
5090)
5091
5092xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005093 name = "f32_ibilinear_chw_test",
5094 srcs = [
5095 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005096 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005097 "src/xnnpack/AlignedAllocator.h",
5098 ] + MICROKERNEL_TEST_HDRS,
5099 deps = MICROKERNEL_TEST_DEPS,
5100)
5101
5102xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005103 name = "f16_clamp_test",
5104 srcs = [
5105 "test/f16-clamp.cc",
5106 "test/clamp-microkernel-tester.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005112 name = "f32_clamp_test",
5113 srcs = [
5114 "test/f32-clamp.cc",
5115 "test/clamp-microkernel-tester.h",
5116 ] + MICROKERNEL_TEST_HDRS,
5117 deps = MICROKERNEL_TEST_DEPS,
5118)
5119
5120xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005121 name = "f32_igemm_test",
5122 srcs = [
5123 "test/f32-igemm.cc",
5124 "test/gemm-microkernel-tester.h",
5125 "src/xnnpack/AlignedAllocator.h",
5126 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005127 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005128)
5129
5130xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005131 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005132 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005133 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 "test/gemm-microkernel-tester.h",
5135 "src/xnnpack/AlignedAllocator.h",
5136 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005137 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005138)
5139
5140xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005141 name = "f32_igemm_minmax_test",
5142 srcs = [
5143 "test/f32-igemm-minmax.cc",
5144 "test/gemm-microkernel-tester.h",
5145 "src/xnnpack/AlignedAllocator.h",
5146 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005147 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005148)
5149
5150xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005151 name = "f16_igemm_minmax_test",
5152 srcs = [
5153 "test/f16-igemm-minmax.cc",
5154 "test/gemm-microkernel-tester.h",
5155 "src/xnnpack/AlignedAllocator.h",
5156 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005157 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005158)
5159
5160xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161 name = "f32_conv_hwc_test",
5162 srcs = [
5163 "test/f32-conv-hwc.cc",
5164 "test/conv-hwc-microkernel-tester.h",
5165 "src/xnnpack/AlignedAllocator.h",
5166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005167 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168)
5169
5170xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005171 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005172 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005173 "test/f32-conv-hwc2chw.cc",
5174 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005175 "src/xnnpack/AlignedAllocator.h",
5176 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005177 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005178)
5179
5180xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005181 name = "f32_dwconv_test",
5182 srcs = [
5183 "test/f32-dwconv.cc",
5184 "test/dwconv-microkernel-tester.h",
5185 "src/xnnpack/AlignedAllocator.h",
5186 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005187 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005188)
5189
5190xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005191 name = "f16_dwconv_minmax_test",
5192 srcs = [
5193 "test/f16-dwconv-minmax.cc",
5194 "test/dwconv-microkernel-tester.h",
5195 "src/xnnpack/AlignedAllocator.h",
5196 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005197 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005198)
5199
5200xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005201 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005202 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005203 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 "test/dwconv-microkernel-tester.h",
5205 "src/xnnpack/AlignedAllocator.h",
5206 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005207 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208)
5209
5210xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005211 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005213 "test/f32-dwconv2d-chw.cc",
5214 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215 "src/xnnpack/AlignedAllocator.h",
5216 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005217 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005218)
5219
5220xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005221 name = "f16_gavgpool_minmax_test",
5222 srcs = [
5223 "test/f16-gavgpool-minmax.cc",
5224 "test/gavgpool-microkernel-tester.h",
5225 "src/xnnpack/AlignedAllocator.h",
5226 ] + MICROKERNEL_TEST_HDRS,
5227 deps = MICROKERNEL_TEST_DEPS,
5228)
5229
5230xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005231 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005232 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005233 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234 "test/gavgpool-microkernel-tester.h",
5235 "src/xnnpack/AlignedAllocator.h",
5236 ] + MICROKERNEL_TEST_HDRS,
5237 deps = MICROKERNEL_TEST_DEPS,
5238)
5239
5240xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005241 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005243 "test/f32-gavgpool-cw.cc",
5244 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005245 "src/xnnpack/AlignedAllocator.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
5250xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005251 name = "f32_gemm_test",
5252 srcs = [
5253 "test/f32-gemm.cc",
5254 "test/gemm-microkernel-tester.h",
5255 "src/xnnpack/AlignedAllocator.h",
5256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005257 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005258)
5259
5260xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005261 name = "f32_gemm_relu_test",
5262 srcs = [
5263 "test/f32-gemm-relu.cc",
5264 "test/gemm-microkernel-tester.h",
5265 "src/xnnpack/AlignedAllocator.h",
5266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005267 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005268)
5269
5270xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005271 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005273 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274 "test/gemm-microkernel-tester.h",
5275 "src/xnnpack/AlignedAllocator.h",
5276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005277 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278)
5279
5280xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005281 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005283 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 "test/gemm-microkernel-tester.h",
5285 "src/xnnpack/AlignedAllocator.h",
5286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005287 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288)
5289
5290xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005291 name = "f16_hswish_test",
5292 srcs = [
5293 "test/f16-hswish.cc",
5294 "test/hswish-microkernel-tester.h",
5295 ] + MICROKERNEL_TEST_HDRS,
5296 deps = MICROKERNEL_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 name = "f32_hswish_test",
5301 srcs = [
5302 "test/f32-hswish.cc",
5303 "test/hswish-microkernel-tester.h",
5304 ] + MICROKERNEL_TEST_HDRS,
5305 deps = MICROKERNEL_TEST_DEPS,
5306)
5307
5308xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005309 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005311 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 "test/maxpool-microkernel-tester.h",
5313 ] + MICROKERNEL_TEST_HDRS,
5314 deps = MICROKERNEL_TEST_DEPS,
5315)
5316
5317xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005318 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005320 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321 "test/avgpool-microkernel-tester.h",
5322 "src/xnnpack/AlignedAllocator.h",
5323 ] + MICROKERNEL_TEST_HDRS,
5324 deps = MICROKERNEL_TEST_DEPS,
5325)
5326
5327xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005328 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005330 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 "test/gemm-microkernel-tester.h",
5332 "src/xnnpack/AlignedAllocator.h",
5333 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005334 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335)
5336
5337xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005338 name = "f16_prelu_test",
5339 srcs = [
5340 "test/f16-prelu.cc",
5341 "test/prelu-microkernel-tester.h",
5342 "src/xnnpack/AlignedAllocator.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348 name = "f32_prelu_test",
5349 srcs = [
5350 "test/f32-prelu.cc",
5351 "test/prelu-microkernel-tester.h",
5352 "src/xnnpack/AlignedAllocator.h",
5353 ] + MICROKERNEL_TEST_HDRS,
5354 deps = MICROKERNEL_TEST_DEPS,
5355)
5356
5357xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005358 name = "f32_raddexpminusmax_test",
5359 srcs = [
5360 "test/f32-raddexpminusmax.cc",
5361 "test/raddexpminusmax-microkernel-tester.h",
5362 ] + MICROKERNEL_TEST_HDRS,
5363 deps = MICROKERNEL_TEST_DEPS,
5364)
5365
5366xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005367 name = "f32_raddextexp_test",
5368 srcs = [
5369 "test/f32-raddextexp.cc",
5370 "test/raddextexp-microkernel-tester.h",
5371 ] + MICROKERNEL_TEST_HDRS,
5372 deps = MICROKERNEL_TEST_DEPS,
5373)
5374
5375xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005376 name = "f32_raddstoreexpminusmax_test",
5377 srcs = [
5378 "test/f32-raddstoreexpminusmax.cc",
5379 "test/raddstoreexpminusmax-microkernel-tester.h",
5380 ] + MICROKERNEL_TEST_HDRS,
5381 deps = MICROKERNEL_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005385 name = "f32_relu_test",
5386 srcs = [
5387 "test/f32-relu.cc",
5388 "test/vunary-microkernel-tester.h",
5389 ] + MICROKERNEL_TEST_HDRS,
5390 deps = MICROKERNEL_TEST_DEPS,
5391)
5392
5393xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 name = "f32_rmax_test",
5395 srcs = [
5396 "test/f32-rmax.cc",
5397 "test/rmax-microkernel-tester.h",
5398 ] + MICROKERNEL_TEST_HDRS,
5399 deps = MICROKERNEL_TEST_DEPS,
5400)
5401
5402xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005403 name = "f32_sigmoid_test",
5404 srcs = [
5405 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005406 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005407 ] + MICROKERNEL_TEST_HDRS,
5408 deps = MICROKERNEL_TEST_DEPS,
5409)
5410
5411xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005412 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005414 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 "test/spmm-microkernel-tester.h",
5416 "src/xnnpack/AlignedAllocator.h",
5417 ] + MICROKERNEL_TEST_HDRS,
5418 deps = MICROKERNEL_TEST_DEPS,
5419)
5420
5421xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005422 name = "f16_vadd_minmax_test",
5423 srcs = [
5424 "test/f16-vadd-minmax.cc",
5425 "test/vbinary-microkernel-tester.h",
5426 ] + MICROKERNEL_TEST_HDRS,
5427 deps = MICROKERNEL_TEST_DEPS,
5428)
5429
5430xnnpack_unit_test(
5431 name = "f16_vaddc_minmax_test",
5432 srcs = [
5433 "test/f16-vaddc-minmax.cc",
5434 "test/vbinaryc-microkernel-tester.h",
5435 ] + MICROKERNEL_TEST_HDRS,
5436 deps = MICROKERNEL_TEST_DEPS,
5437)
5438
5439xnnpack_unit_test(
5440 name = "f16_vdiv_minmax_test",
5441 srcs = [
5442 "test/f16-vdiv-minmax.cc",
5443 "test/vbinary-microkernel-tester.h",
5444 ] + MICROKERNEL_TEST_HDRS,
5445 deps = MICROKERNEL_TEST_DEPS,
5446)
5447
5448xnnpack_unit_test(
5449 name = "f16_vdivc_minmax_test",
5450 srcs = [
5451 "test/f16-vdivc-minmax.cc",
5452 "test/vbinaryc-microkernel-tester.h",
5453 ] + MICROKERNEL_TEST_HDRS,
5454 deps = MICROKERNEL_TEST_DEPS,
5455)
5456
5457xnnpack_unit_test(
5458 name = "f16_vrdivc_minmax_test",
5459 srcs = [
5460 "test/f16-vrdivc-minmax.cc",
5461 "test/vbinaryc-microkernel-tester.h",
5462 ] + MICROKERNEL_TEST_HDRS,
5463 deps = MICROKERNEL_TEST_DEPS,
5464)
5465
5466xnnpack_unit_test(
5467 name = "f16_vmax_test",
5468 srcs = [
5469 "test/f16-vmax.cc",
5470 "test/vbinary-microkernel-tester.h",
5471 ] + MICROKERNEL_TEST_HDRS,
5472 deps = MICROKERNEL_TEST_DEPS,
5473)
5474
5475xnnpack_unit_test(
5476 name = "f16_vmaxc_test",
5477 srcs = [
5478 "test/f16-vmaxc.cc",
5479 "test/vbinaryc-microkernel-tester.h",
5480 ] + MICROKERNEL_TEST_HDRS,
5481 deps = MICROKERNEL_TEST_DEPS,
5482)
5483
5484xnnpack_unit_test(
5485 name = "f16_vmin_test",
5486 srcs = [
5487 "test/f16-vmin.cc",
5488 "test/vbinary-microkernel-tester.h",
5489 ] + MICROKERNEL_TEST_HDRS,
5490 deps = MICROKERNEL_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
5494 name = "f16_vminc_test",
5495 srcs = [
5496 "test/f16-vminc.cc",
5497 "test/vbinaryc-microkernel-tester.h",
5498 ] + MICROKERNEL_TEST_HDRS,
5499 deps = MICROKERNEL_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
5503 name = "f16_vmul_minmax_test",
5504 srcs = [
5505 "test/f16-vmul-minmax.cc",
5506 "test/vbinary-microkernel-tester.h",
5507 ] + MICROKERNEL_TEST_HDRS,
5508 deps = MICROKERNEL_TEST_DEPS,
5509)
5510
5511xnnpack_unit_test(
5512 name = "f16_vmulc_minmax_test",
5513 srcs = [
5514 "test/f16-vmulc-minmax.cc",
5515 "test/vbinaryc-microkernel-tester.h",
5516 ] + MICROKERNEL_TEST_HDRS,
5517 deps = MICROKERNEL_TEST_DEPS,
5518)
5519
5520xnnpack_unit_test(
5521 name = "f16_vsub_minmax_test",
5522 srcs = [
5523 "test/f16-vsub-minmax.cc",
5524 "test/vbinary-microkernel-tester.h",
5525 ] + MICROKERNEL_TEST_HDRS,
5526 deps = MICROKERNEL_TEST_DEPS,
5527)
5528
5529xnnpack_unit_test(
5530 name = "f16_vsubc_minmax_test",
5531 srcs = [
5532 "test/f16-vsubc-minmax.cc",
5533 "test/vbinaryc-microkernel-tester.h",
5534 ] + MICROKERNEL_TEST_HDRS,
5535 deps = MICROKERNEL_TEST_DEPS,
5536)
5537
5538xnnpack_unit_test(
5539 name = "f16_vrsubc_minmax_test",
5540 srcs = [
5541 "test/f16-vrsubc-minmax.cc",
5542 "test/vbinaryc-microkernel-tester.h",
5543 ] + MICROKERNEL_TEST_HDRS,
5544 deps = MICROKERNEL_TEST_DEPS,
5545)
5546
5547xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005548 name = "f32_vabs_test",
5549 srcs = [
5550 "test/f32-vabs.cc",
5551 "test/vunary-microkernel-tester.h",
5552 ] + MICROKERNEL_TEST_HDRS,
5553 deps = MICROKERNEL_TEST_DEPS,
5554)
5555
5556xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005557 name = "f32_vadd_test",
5558 srcs = [
5559 "test/f32-vadd.cc",
5560 "test/vbinary-microkernel-tester.h",
5561 ] + MICROKERNEL_TEST_HDRS,
5562 deps = MICROKERNEL_TEST_DEPS,
5563)
5564
5565xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005566 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005567 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005568 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005569 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005570 ] + MICROKERNEL_TEST_HDRS,
5571 deps = MICROKERNEL_TEST_DEPS,
5572)
5573
5574xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005575 name = "f32_vadd_relu_test",
5576 srcs = [
5577 "test/f32-vadd-relu.cc",
5578 "test/vbinary-microkernel-tester.h",
5579 ] + MICROKERNEL_TEST_HDRS,
5580 deps = MICROKERNEL_TEST_DEPS,
5581)
5582
5583xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005584 name = "f32_vaddc_test",
5585 srcs = [
5586 "test/f32-vaddc.cc",
5587 "test/vbinaryc-microkernel-tester.h",
5588 ] + MICROKERNEL_TEST_HDRS,
5589 deps = MICROKERNEL_TEST_DEPS,
5590)
5591
5592xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005593 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005594 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005595 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005596 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005597 ] + MICROKERNEL_TEST_HDRS,
5598 deps = MICROKERNEL_TEST_DEPS,
5599)
5600
5601xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005602 name = "f32_vaddc_relu_test",
5603 srcs = [
5604 "test/f32-vaddc-relu.cc",
5605 "test/vbinaryc-microkernel-tester.h",
5606 ] + MICROKERNEL_TEST_HDRS,
5607 deps = MICROKERNEL_TEST_DEPS,
5608)
5609
5610xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005611 name = "f32_vdiv_test",
5612 srcs = [
5613 "test/f32-vdiv.cc",
5614 "test/vbinary-microkernel-tester.h",
5615 ] + MICROKERNEL_TEST_HDRS,
5616 deps = MICROKERNEL_TEST_DEPS,
5617)
5618
5619xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005620 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005621 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005622 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005623 "test/vbinary-microkernel-tester.h",
5624 ] + MICROKERNEL_TEST_HDRS,
5625 deps = MICROKERNEL_TEST_DEPS,
5626)
5627
5628xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005629 name = "f32_vdiv_relu_test",
5630 srcs = [
5631 "test/f32-vdiv-relu.cc",
5632 "test/vbinary-microkernel-tester.h",
5633 ] + MICROKERNEL_TEST_HDRS,
5634 deps = MICROKERNEL_TEST_DEPS,
5635)
5636
5637xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005638 name = "f32_vdivc_test",
5639 srcs = [
5640 "test/f32-vdivc.cc",
5641 "test/vbinaryc-microkernel-tester.h",
5642 ] + MICROKERNEL_TEST_HDRS,
5643 deps = MICROKERNEL_TEST_DEPS,
5644)
5645
5646xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005647 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005648 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005649 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005650 "test/vbinaryc-microkernel-tester.h",
5651 ] + MICROKERNEL_TEST_HDRS,
5652 deps = MICROKERNEL_TEST_DEPS,
5653)
5654
5655xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005656 name = "f32_vdivc_relu_test",
5657 srcs = [
5658 "test/f32-vdivc-relu.cc",
5659 "test/vbinaryc-microkernel-tester.h",
5660 ] + MICROKERNEL_TEST_HDRS,
5661 deps = MICROKERNEL_TEST_DEPS,
5662)
5663
5664xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005665 name = "f32_vrdivc_test",
5666 srcs = [
5667 "test/f32-vrdivc.cc",
5668 "test/vbinaryc-microkernel-tester.h",
5669 ] + MICROKERNEL_TEST_HDRS,
5670 deps = MICROKERNEL_TEST_DEPS,
5671)
5672
5673xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005674 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005675 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005676 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005677 "test/vbinaryc-microkernel-tester.h",
5678 ] + MICROKERNEL_TEST_HDRS,
5679 deps = MICROKERNEL_TEST_DEPS,
5680)
5681
5682xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005683 name = "f32_vrdivc_relu_test",
5684 srcs = [
5685 "test/f32-vrdivc-relu.cc",
5686 "test/vbinaryc-microkernel-tester.h",
5687 ] + MICROKERNEL_TEST_HDRS,
5688 deps = MICROKERNEL_TEST_DEPS,
5689)
5690
5691xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005692 name = "f32_vmax_test",
5693 srcs = [
5694 "test/f32-vmax.cc",
5695 "test/vbinary-microkernel-tester.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 deps = MICROKERNEL_TEST_DEPS,
5698)
5699
5700xnnpack_unit_test(
5701 name = "f32_vmaxc_test",
5702 srcs = [
5703 "test/f32-vmaxc.cc",
5704 "test/vbinaryc-microkernel-tester.h",
5705 ] + MICROKERNEL_TEST_HDRS,
5706 deps = MICROKERNEL_TEST_DEPS,
5707)
5708
5709xnnpack_unit_test(
5710 name = "f32_vmin_test",
5711 srcs = [
5712 "test/f32-vmin.cc",
5713 "test/vbinary-microkernel-tester.h",
5714 ] + MICROKERNEL_TEST_HDRS,
5715 deps = MICROKERNEL_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
5719 name = "f32_vminc_test",
5720 srcs = [
5721 "test/f32-vminc.cc",
5722 "test/vbinaryc-microkernel-tester.h",
5723 ] + MICROKERNEL_TEST_HDRS,
5724 deps = MICROKERNEL_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005728 name = "f32_vmul_test",
5729 srcs = [
5730 "test/f32-vmul.cc",
5731 "test/vbinary-microkernel-tester.h",
5732 ] + MICROKERNEL_TEST_HDRS,
5733 deps = MICROKERNEL_TEST_DEPS,
5734)
5735
5736xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005737 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005738 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005739 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005740 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005741 ] + MICROKERNEL_TEST_HDRS,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005746 name = "f32_vmul_relu_test",
5747 srcs = [
5748 "test/f32-vmul-relu.cc",
5749 "test/vbinary-microkernel-tester.h",
5750 ] + MICROKERNEL_TEST_HDRS,
5751 deps = MICROKERNEL_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005755 name = "f32_vmulc_test",
5756 srcs = [
5757 "test/f32-vmulc.cc",
5758 "test/vbinaryc-microkernel-tester.h",
5759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005764 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005765 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005766 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005767 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005773 name = "f32_vmulc_relu_test",
5774 srcs = [
5775 "test/f32-vmulc-relu.cc",
5776 "test/vbinaryc-microkernel-tester.h",
5777 ] + MICROKERNEL_TEST_HDRS,
5778 deps = MICROKERNEL_TEST_DEPS,
5779)
5780
5781xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005782 name = "f16_vmulcaddc_minmax_test",
5783 srcs = [
5784 "test/f16-vmulcaddc-minmax.cc",
5785 "test/vmulcaddc-microkernel-tester.h",
5786 "src/xnnpack/AlignedAllocator.h",
5787 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005788 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005789)
5790
5791xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005792 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005793 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005794 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005795 "test/vmulcaddc-microkernel-tester.h",
5796 "src/xnnpack/AlignedAllocator.h",
5797 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005798 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005799)
5800
5801xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005802 name = "f32_vlrelu_test",
5803 srcs = [
5804 "test/f32-vlrelu.cc",
5805 "test/vunary-microkernel-tester.h",
5806 ] + MICROKERNEL_TEST_HDRS,
5807 deps = MICROKERNEL_TEST_DEPS,
5808)
5809
5810xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005811 name = "f32_vneg_test",
5812 srcs = [
5813 "test/f32-vneg.cc",
5814 "test/vunary-microkernel-tester.h",
5815 ] + MICROKERNEL_TEST_HDRS,
5816 deps = MICROKERNEL_TEST_DEPS,
5817)
5818
5819xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005820 name = "f32_vrndne_test",
5821 srcs = [
5822 "test/f32-vrndne.cc",
5823 "test/vunary-microkernel-tester.h",
5824 ] + MICROKERNEL_TEST_HDRS,
5825 deps = MICROKERNEL_TEST_DEPS,
5826)
5827
5828xnnpack_unit_test(
5829 name = "f32_vrndz_test",
5830 srcs = [
5831 "test/f32-vrndz.cc",
5832 "test/vunary-microkernel-tester.h",
5833 ] + MICROKERNEL_TEST_HDRS,
5834 deps = MICROKERNEL_TEST_DEPS,
5835)
5836
5837xnnpack_unit_test(
5838 name = "f32_vrndu_test",
5839 srcs = [
5840 "test/f32-vrndu.cc",
5841 "test/vunary-microkernel-tester.h",
5842 ] + MICROKERNEL_TEST_HDRS,
5843 deps = MICROKERNEL_TEST_DEPS,
5844)
5845
5846xnnpack_unit_test(
5847 name = "f32_vrndd_test",
5848 srcs = [
5849 "test/f32-vrndd.cc",
5850 "test/vunary-microkernel-tester.h",
5851 ] + MICROKERNEL_TEST_HDRS,
5852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
5855xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005856 name = "f32_vscale_test",
5857 srcs = [
5858 "test/f32-vscale.cc",
5859 "test/vscale-microkernel-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005865 name = "f32_vscaleexpminusmax_test",
5866 srcs = [
5867 "test/f32-vscaleexpminusmax.cc",
5868 "test/vscaleexpminusmax-microkernel-tester.h",
5869 ] + MICROKERNEL_TEST_HDRS,
5870 deps = MICROKERNEL_TEST_DEPS,
5871)
5872
5873xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005874 name = "f32_vscaleextexp_test",
5875 srcs = [
5876 "test/f32-vscaleextexp.cc",
5877 "test/vscaleextexp-microkernel-tester.h",
5878 ] + MICROKERNEL_TEST_HDRS,
5879 deps = MICROKERNEL_TEST_DEPS,
5880)
5881
5882xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005883 name = "f32_vsqr_test",
5884 srcs = [
5885 "test/f32-vsqr.cc",
5886 "test/vunary-microkernel-tester.h",
5887 ] + MICROKERNEL_TEST_HDRS,
5888 deps = MICROKERNEL_TEST_DEPS,
5889)
5890
5891xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005892 name = "f32_vsqrdiff_test",
5893 srcs = [
5894 "test/f32-vsqrdiff.cc",
5895 "test/vbinary-microkernel-tester.h",
5896 ] + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS,
5898)
5899
5900xnnpack_unit_test(
5901 name = "f32_vsqrdiffc_test",
5902 srcs = [
5903 "test/f32-vsqrdiffc.cc",
5904 "test/vbinaryc-microkernel-tester.h",
5905 ] + MICROKERNEL_TEST_HDRS,
5906 deps = MICROKERNEL_TEST_DEPS,
5907)
5908
5909xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005910 name = "f32_vsqrt_test",
5911 srcs = [
5912 "test/f32-vsqrt.cc",
5913 "test/vunary-microkernel-tester.h",
5914 ] + MICROKERNEL_TEST_HDRS,
5915 deps = MICROKERNEL_TEST_DEPS,
5916)
5917
5918xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005919 name = "f32_vsub_test",
5920 srcs = [
5921 "test/f32-vsub.cc",
5922 "test/vbinary-microkernel-tester.h",
5923 ] + MICROKERNEL_TEST_HDRS,
5924 deps = MICROKERNEL_TEST_DEPS,
5925)
5926
5927xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005928 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005929 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005930 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005931 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005932 ] + MICROKERNEL_TEST_HDRS,
5933 deps = MICROKERNEL_TEST_DEPS,
5934)
5935
5936xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005937 name = "f32_vsub_relu_test",
5938 srcs = [
5939 "test/f32-vsub-relu.cc",
5940 "test/vbinary-microkernel-tester.h",
5941 ] + MICROKERNEL_TEST_HDRS,
5942 deps = MICROKERNEL_TEST_DEPS,
5943)
5944
5945xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005946 name = "f32_vsubc_test",
5947 srcs = [
5948 "test/f32-vsubc.cc",
5949 "test/vbinaryc-microkernel-tester.h",
5950 ] + MICROKERNEL_TEST_HDRS,
5951 deps = MICROKERNEL_TEST_DEPS,
5952)
5953
5954xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005955 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005956 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005957 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005958 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005959 ] + MICROKERNEL_TEST_HDRS,
5960 deps = MICROKERNEL_TEST_DEPS,
5961)
5962
5963xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005964 name = "f32_vsubc_relu_test",
5965 srcs = [
5966 "test/f32-vsubc-relu.cc",
5967 "test/vbinaryc-microkernel-tester.h",
5968 ] + MICROKERNEL_TEST_HDRS,
5969 deps = MICROKERNEL_TEST_DEPS,
5970)
5971
5972xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005973 name = "f32_vrsubc_test",
5974 srcs = [
5975 "test/f32-vrsubc.cc",
5976 "test/vbinaryc-microkernel-tester.h",
5977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005982 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005983 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005984 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005985 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005986 ] + MICROKERNEL_TEST_HDRS,
5987 deps = MICROKERNEL_TEST_DEPS,
5988)
5989
5990xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005991 name = "f32_vrsubc_relu_test",
5992 srcs = [
5993 "test/f32-vrsubc-relu.cc",
5994 "test/vbinaryc-microkernel-tester.h",
5995 ] + MICROKERNEL_TEST_HDRS,
5996 deps = MICROKERNEL_TEST_DEPS,
5997)
5998
5999xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006000 name = "qs8_dwconv_minmax_test",
6001 srcs = [
6002 "test/qs8-dwconv-minmax.cc",
6003 "test/dwconv-microkernel-tester.h",
6004 "src/xnnpack/AlignedAllocator.h",
6005 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6006 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6007)
6008
6009xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006010 name = "qs8_gavgpool_minmax_test",
6011 srcs = [
6012 "test/qs8-gavgpool-minmax.cc",
6013 "test/gavgpool-microkernel-tester.h",
6014 "src/xnnpack/AlignedAllocator.h",
6015 ] + MICROKERNEL_TEST_HDRS,
6016 deps = MICROKERNEL_TEST_DEPS,
6017)
6018
6019xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006020 name = "qs8_gemm_minmax_test",
6021 srcs = [
6022 "test/qs8-gemm-minmax.cc",
6023 "test/gemm-microkernel-tester.h",
6024 "src/xnnpack/AlignedAllocator.h",
6025 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6026 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6027)
6028
6029xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006030 name = "qs8_igemm_minmax_test",
6031 srcs = [
6032 "test/qs8-igemm-minmax.cc",
6033 "test/gemm-microkernel-tester.h",
6034 "src/xnnpack/AlignedAllocator.h",
6035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6036 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6037)
6038
6039xnnpack_unit_test(
6040 name = "qs8_requantization_test",
6041 srcs = [
6042 "src/xnnpack/requantization-stubs.h",
6043 "test/qs8-requantization.cc",
6044 "test/requantization-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006050 name = "qs8_vadd_minmax_test",
6051 srcs = [
6052 "test/qs8-vadd-minmax.cc",
6053 "test/vadd-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006059 name = "qs8_vaddc_minmax_test",
6060 srcs = [
6061 "test/qs8-vaddc-minmax.cc",
6062 "test/vaddc-microkernel-tester.h",
6063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006068 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006069 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006070 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071 "test/avgpool-microkernel-tester.h",
6072 "src/xnnpack/AlignedAllocator.h",
6073 ] + MICROKERNEL_TEST_HDRS,
6074 deps = MICROKERNEL_TEST_DEPS,
6075)
6076
6077xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006078 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006080 "test/qu8-dwconv-minmax.cc",
6081 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082 "src/xnnpack/AlignedAllocator.h",
6083 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006084 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006085)
6086
6087xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006088 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006089 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006090 "test/qu8-igemm-minmax.cc",
6091 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006092 "src/xnnpack/AlignedAllocator.h",
6093 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006094 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095)
6096
6097xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006098 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006099 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006100 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101 "test/gavgpool-microkernel-tester.h",
6102 "src/xnnpack/AlignedAllocator.h",
6103 ] + MICROKERNEL_TEST_HDRS,
6104 deps = MICROKERNEL_TEST_DEPS,
6105)
6106
6107xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006108 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006109 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006110 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006111 "test/gemm-microkernel-tester.h",
6112 "src/xnnpack/AlignedAllocator.h",
6113 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006114 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006115)
6116
6117xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006118 name = "qu8_requantization_test",
6119 srcs = [
6120 "src/xnnpack/requantization-stubs.h",
6121 "test/qu8-requantization.cc",
6122 "test/requantization-tester.h",
6123 ] + MICROKERNEL_TEST_HDRS,
6124 deps = MICROKERNEL_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006128 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006129 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006130 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131 "test/vadd-microkernel-tester.h",
6132 ] + MICROKERNEL_TEST_HDRS,
6133 deps = MICROKERNEL_TEST_DEPS,
6134)
6135
6136xnnpack_unit_test(
6137 name = "u8_clamp_test",
6138 srcs = [
6139 "test/u8-clamp.cc",
6140 "test/clamp-microkernel-tester.h",
6141 ] + MICROKERNEL_TEST_HDRS,
6142 deps = MICROKERNEL_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
6146 name = "u8_lut32norm_test",
6147 srcs = [
6148 "test/u8-lut32norm.cc",
6149 "test/lut-norm-microkernel-tester.h",
6150 ] + MICROKERNEL_TEST_HDRS,
6151 deps = MICROKERNEL_TEST_DEPS,
6152)
6153
6154xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006155 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006156 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006157 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006158 "test/maxpool-microkernel-tester.h",
6159 ] + MICROKERNEL_TEST_HDRS,
6160 deps = MICROKERNEL_TEST_DEPS,
6161)
6162
6163xnnpack_unit_test(
6164 name = "u8_rmax_test",
6165 srcs = [
6166 "test/u8-rmax.cc",
6167 "test/rmax-microkernel-tester.h",
6168 ] + MICROKERNEL_TEST_HDRS,
6169 deps = MICROKERNEL_TEST_DEPS,
6170)
6171
6172xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006173 name = "x32_fill_test",
6174 srcs = [
6175 "test/x32-fill.cc",
6176 "test/fill-microkernel-tester.h",
6177 ] + MICROKERNEL_TEST_HDRS,
6178 deps = MICROKERNEL_TEST_DEPS,
6179)
6180
6181xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006182 name = "x32_packx_test",
6183 srcs = [
6184 "test/x32-packx.cc",
6185 "test/pack-microkernel-tester.h",
6186 "src/xnnpack/AlignedAllocator.h",
6187 ] + MICROKERNEL_TEST_HDRS,
6188 deps = MICROKERNEL_TEST_DEPS,
6189)
6190
6191xnnpack_unit_test(
6192 name = "x32_pad_test",
6193 srcs = [
6194 "test/x32-pad.cc",
6195 "test/pad-microkernel-tester.h",
6196 ] + MICROKERNEL_TEST_HDRS,
6197 deps = MICROKERNEL_TEST_DEPS,
6198)
6199
6200xnnpack_unit_test(
6201 name = "x32_unpool_test",
6202 srcs = [
6203 "test/x32-unpool.cc",
6204 "test/unpool-microkernel-tester.h",
6205 ] + MICROKERNEL_TEST_HDRS,
6206 deps = MICROKERNEL_TEST_DEPS,
6207)
6208
6209xnnpack_unit_test(
6210 name = "x32_zip_test",
6211 srcs = [
6212 "test/x32-zip.cc",
6213 "test/zip-microkernel-tester.h",
6214 ] + MICROKERNEL_TEST_HDRS,
6215 deps = MICROKERNEL_TEST_DEPS,
6216)
6217
6218xnnpack_unit_test(
6219 name = "x8_lut_test",
6220 srcs = [
6221 "test/x8-lut.cc",
6222 "test/lut-microkernel-tester.h",
6223 ] + MICROKERNEL_TEST_HDRS,
6224 deps = MICROKERNEL_TEST_DEPS,
6225)
6226
6227xnnpack_unit_test(
6228 name = "x8_zip_test",
6229 srcs = [
6230 "test/x8-zip.cc",
6231 "test/zip-microkernel-tester.h",
6232 ] + MICROKERNEL_TEST_HDRS,
6233 deps = MICROKERNEL_TEST_DEPS,
6234)
6235
Marat Dukhan20c3b922020-03-10 03:45:06 -07006236########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006237
6238xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006239 name = "operator_size_test",
6240 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006241 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242)
6243
Marat Dukhan20c3b922020-03-10 03:45:06 -07006244xnnpack_binary(
6245 name = "subgraph_size_test",
6246 srcs = ["test/subgraph-size.c"],
6247 deps = [":XNNPACK"],
6248)
6249
6250########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006251
6252xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006253 name = "abs_nc_test",
6254 srcs = [
6255 "test/abs-nc.cc",
6256 "test/abs-operator-tester.h",
6257 ],
6258 deps = OPERATOR_TEST_DEPS,
6259)
6260
6261xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006262 name = "add_nd_test",
6263 srcs = [
6264 "test/add-nd.cc",
6265 "test/binary-elementwise-operator-tester.h",
6266 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006267 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006268)
6269
6270xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006271 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006273 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274 "test/argmax-pooling-operator-tester.h",
6275 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006276 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277)
6278
6279xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006280 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006282 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283 "test/average-pooling-operator-tester.h",
6284 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006285 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286)
6287
6288xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006289 name = "bankers_rounding_nc_test",
6290 srcs = [
6291 "test/bankers-rounding-nc.cc",
6292 "test/bankers-rounding-operator-tester.h",
6293 ],
6294 deps = OPERATOR_TEST_DEPS,
6295)
6296
6297xnnpack_unit_test(
6298 name = "ceiling_nc_test",
6299 srcs = [
6300 "test/ceiling-nc.cc",
6301 "test/ceiling-operator-tester.h",
6302 ],
6303 deps = OPERATOR_TEST_DEPS,
6304)
6305
6306xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006307 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006309 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006310 "test/channel-shuffle-operator-tester.h",
6311 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006312 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006313)
6314
6315xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006316 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006318 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006319 "test/clamp-operator-tester.h",
6320 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006321 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006322)
6323
6324xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006325 name = "constant_pad_nd_test",
6326 srcs = [
6327 "test/constant-pad-nd.cc",
6328 "test/constant-pad-operator-tester.h",
6329 ],
6330 deps = OPERATOR_TEST_DEPS,
6331)
6332
6333xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006334 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006335 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006336 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006337 "test/convolution-operator-tester.h",
6338 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006339 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006340)
6341
6342xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006343 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006344 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006345 "test/convolution-nchw.cc",
6346 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006347 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006348 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006349)
6350
6351xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006352 name = "copy_nc_test",
6353 srcs = [
6354 "test/copy-nc.cc",
6355 "test/copy-operator-tester.h",
6356 ],
6357 deps = OPERATOR_TEST_DEPS,
6358)
6359
6360xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006361 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006363 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006364 "test/deconvolution-operator-tester.h",
6365 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006366 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006367)
6368
6369xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006370 name = "divide_nd_test",
6371 srcs = [
6372 "test/binary-elementwise-operator-tester.h",
6373 "test/divide-nd.cc",
6374 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006375 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006376)
6377
6378xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006379 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006381 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006382 "test/fully-connected-operator-tester.h",
6383 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006384 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006385)
6386
6387xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006388 name = "floor_nc_test",
6389 srcs = [
6390 "test/floor-nc.cc",
6391 "test/floor-operator-tester.h",
6392 ],
6393 deps = OPERATOR_TEST_DEPS,
6394)
6395
6396xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006397 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006399 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006400 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006401 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006402 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006403)
6404
6405xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006406 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006408 "test/global-average-pooling-ncw.cc",
6409 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006410 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006411 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412)
6413
6414xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006415 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006416 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006417 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006418 "test/hardswish-operator-tester.h",
6419 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006420 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006421)
6422
6423xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006424 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006425 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006426 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006427 "test/leaky-relu-operator-tester.h",
6428 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006429 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006430)
6431
6432xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006433 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006434 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006435 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006436 "test/max-pooling-operator-tester.h",
6437 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006438 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006439)
6440
6441xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006442 name = "maximum_nd_test",
6443 srcs = [
6444 "test/binary-elementwise-operator-tester.h",
6445 "test/maximum-nd.cc",
6446 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006447 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006448)
6449
6450xnnpack_unit_test(
6451 name = "minimum_nd_test",
6452 srcs = [
6453 "test/binary-elementwise-operator-tester.h",
6454 "test/minimum-nd.cc",
6455 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006456 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006457)
6458
6459xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006460 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006461 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006462 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006463 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006464 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006465 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006466)
6467
6468xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006469 name = "negate_nc_test",
6470 srcs = [
6471 "test/negate-nc.cc",
6472 "test/negate-operator-tester.h",
6473 ],
6474 deps = OPERATOR_TEST_DEPS,
6475)
6476
6477xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006478 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006479 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006480 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006481 "test/prelu-operator-tester.h",
6482 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006483 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006484)
6485
6486xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006487 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006488 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006489 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006490 "test/resize-bilinear-operator-tester.h",
6491 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006492 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006493)
6494
6495xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006496 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006497 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006498 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006499 "test/sigmoid-operator-tester.h",
6500 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006501 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502)
6503
6504xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006505 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006506 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006507 "test/softmax-nc.cc",
6508 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006509 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006510 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006511)
6512
6513xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006514 name = "square_nc_test",
6515 srcs = [
6516 "test/square-nc.cc",
6517 "test/square-operator-tester.h",
6518 ],
6519 deps = OPERATOR_TEST_DEPS,
6520)
6521
6522xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006523 name = "square_root_nc_test",
6524 srcs = [
6525 "test/square-root-nc.cc",
6526 "test/square-root-operator-tester.h",
6527 ],
6528 deps = OPERATOR_TEST_DEPS,
6529)
6530
6531xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006532 name = "squared_difference_nd_test",
6533 srcs = [
6534 "test/binary-elementwise-operator-tester.h",
6535 "test/squared-difference-nd.cc",
6536 ],
6537 deps = OPERATOR_TEST_DEPS,
6538)
6539
6540xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006541 name = "subtract_nd_test",
6542 srcs = [
6543 "test/binary-elementwise-operator-tester.h",
6544 "test/subtract-nd.cc",
6545 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006546 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006547)
6548
6549xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006550 name = "truncation_nc_test",
6551 srcs = [
6552 "test/truncation-nc.cc",
6553 "test/truncation-operator-tester.h",
6554 ],
6555 deps = OPERATOR_TEST_DEPS,
6556)
6557
6558xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006559 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006560 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006561 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006562 "test/unpooling-operator-tester.h",
6563 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006564 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006565)
6566
Chao Mei6ddfc602020-05-13 22:29:36 -07006567############################### Misc unit tests ###############################
6568
6569xnnpack_unit_test(
6570 name = "memory_planner_test",
6571 srcs = [
6572 "test/memory-planner-test.cc",
6573 ],
6574 deps = [
6575 ":XNNPACK",
6576 ":memory_planner",
6577 ],
6578)
6579
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006580xnnpack_unit_test(
6581 name = "subgraph_nchw_test",
6582 srcs = [
6583 "src/xnnpack/subgraph.h",
6584 "test/subgraph-nchw.cc",
6585 "test/subgraph-tester.h",
6586 ],
6587 deps = [
6588 ":XNNPACK",
6589 ],
6590)
6591
Marat Dukhan08c4a432019-10-03 09:29:21 -07006592############################# Build configurations #############################
6593
Marat Dukhanb8642352019-10-30 15:43:02 -07006594# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006595config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006596 name = "xnn_enable_assembly_explicit_true",
6597 define_values = {"xnn_enable_assembly": "true"},
6598)
6599
6600# Disables usage of assembly kernels.
6601config_setting(
6602 name = "xnn_enable_assembly_explicit_false",
6603 define_values = {"xnn_enable_assembly": "false"},
6604)
6605
Marat Dukhan9de90e02020-06-18 16:04:12 -07006606# Enables usage of sparse inference.
6607config_setting(
6608 name = "xnn_enable_sparse_explicit_true",
6609 define_values = {"xnn_enable_sparse": "true"},
6610)
6611
6612# Disables usage of sparse inference.
6613config_setting(
6614 name = "xnn_enable_sparse_explicit_false",
6615 define_values = {"xnn_enable_sparse": "false"},
6616)
6617
Marat Dukhan05702cf2020-03-26 15:41:33 -07006618# Disables usage of HMP-aware optimizations.
6619config_setting(
6620 name = "xnn_enable_hmp_explicit_false",
6621 define_values = {"xnn_enable_hmp": "false"},
6622)
6623
Chao Mei6ddfc602020-05-13 22:29:36 -07006624# Enable usage of optimized memory allocation
6625config_setting(
6626 name = "xnn_enable_memopt_explicit_true",
6627 define_values = {"xnnpack_enable_memopt": "true"},
6628)
6629
6630# Disable usage of optimized memory allocation
6631config_setting(
6632 name = "xnn_enable_memopt_explicit_false",
6633 define_values = {"xnnpack_enable_memopt": "false"},
6634)
6635
Marat Dukhanb8642352019-10-30 15:43:02 -07006636# Builds with -c dbg
6637config_setting(
6638 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006639 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006640 "compilation_mode": "dbg",
6641 },
6642)
6643
6644# Builds with -c opt
6645config_setting(
6646 name = "optimized_build",
6647 values = {
6648 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006649 },
6650)
6651
6652config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006653 name = "linux_k8",
6654 values = {"cpu": "k8"},
6655)
6656
6657config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006658 name = "linux_arm",
6659 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006660)
6661
6662config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006663 name = "linux_armeabi",
6664 values = {"cpu": "armeabi"},
6665)
6666
6667config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006668 name = "linux_armhf",
6669 values = {"cpu": "armhf"},
6670)
6671
6672config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006673 name = "linux_armv7a",
6674 values = {"cpu": "armv7a"},
6675)
6676
6677config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006678 name = "linux_aarch64",
6679 values = {"cpu": "aarch64"},
6680)
6681
6682config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006683 name = "android",
6684 values = {"crosstool_top": "//external:android/crosstool"},
6685)
6686
6687config_setting(
6688 name = "android_armv7",
6689 values = {
6690 "crosstool_top": "//external:android/crosstool",
6691 "cpu": "armeabi-v7a",
6692 },
6693)
6694
6695config_setting(
6696 name = "android_arm64",
6697 values = {
6698 "crosstool_top": "//external:android/crosstool",
6699 "cpu": "arm64-v8a",
6700 },
6701)
6702
6703config_setting(
6704 name = "android_x86",
6705 values = {
6706 "crosstool_top": "//external:android/crosstool",
6707 "cpu": "x86",
6708 },
6709)
6710
6711config_setting(
6712 name = "android_x86_64",
6713 values = {
6714 "crosstool_top": "//external:android/crosstool",
6715 "cpu": "x86_64",
6716 },
6717)
6718
6719config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006720 name = "windows_x86_64",
6721 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006722)
6723
6724config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006725 name = "windows_x86_64_clang",
6726 values = {
6727 "compiler": "clang-cl",
6728 "cpu": "x64_windows",
6729 },
6730)
6731
6732config_setting(
6733 name = "windows_x86_64_mingw",
6734 values = {
6735 "compiler": "mingw-gcc",
6736 "cpu": "x64_windows",
6737 },
6738)
6739
6740config_setting(
6741 name = "windows_x86_64_msys",
6742 values = {
6743 "compiler": "msys-gcc",
6744 "cpu": "x64_windows",
6745 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006746)
6747
6748config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006749 name = "macos_x86_64",
6750 values = {
6751 "apple_platform_type": "macos",
6752 "cpu": "darwin",
6753 },
6754)
6755
6756config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006757 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006758 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006759)
6760
6761config_setting(
6762 name = "emscripten_wasm",
6763 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006764 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006765 "cpu": "wasm",
6766 },
6767)
6768
6769config_setting(
6770 name = "emscripten_wasmsimd",
6771 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006772 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006773 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006774 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006775 },
6776)
6777
6778config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006779 name = "ios_armv7",
6780 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006781 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006782 "cpu": "ios_armv7",
6783 },
6784)
6785
6786config_setting(
6787 name = "ios_arm64",
6788 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006789 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006790 "cpu": "ios_arm64",
6791 },
6792)
6793
6794config_setting(
6795 name = "ios_arm64e",
6796 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006797 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006798 "cpu": "ios_arm64e",
6799 },
6800)
6801
6802config_setting(
6803 name = "ios_x86",
6804 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006805 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006806 "cpu": "ios_i386",
6807 },
6808)
6809
6810config_setting(
6811 name = "ios_x86_64",
6812 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006813 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006814 "cpu": "ios_x86_64",
6815 },
6816)
6817
6818config_setting(
6819 name = "watchos_armv7k",
6820 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006821 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006822 "cpu": "watchos_armv7k",
6823 },
6824)
6825
6826config_setting(
6827 name = "watchos_arm64_32",
6828 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006829 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006830 "cpu": "watchos_arm64_32",
6831 },
6832)
6833
6834config_setting(
6835 name = "watchos_x86",
6836 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006837 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006838 "cpu": "watchos_i386",
6839 },
6840)
6841
6842config_setting(
6843 name = "watchos_x86_64",
6844 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006845 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006846 "cpu": "watchos_x86_64",
6847 },
6848)
6849
6850config_setting(
6851 name = "tvos_arm64",
6852 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006853 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006854 "cpu": "tvos_arm64",
6855 },
6856)
6857
6858config_setting(
6859 name = "tvos_x86_64",
6860 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006861 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006862 "cpu": "tvos_x86_64",
6863 },
6864)