blob: 2f45293c484bfa16f882d518cca8af9edd95d1f7 [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 Dukhan82f0c322020-10-25 19:17:35 -07001191 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
1192 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1193 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1194 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
1195 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1196 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1197 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
1198 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001199 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001200 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1201 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001202 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001203 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1204 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001205 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001206 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1207 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1208 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1209 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1210 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001211 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1212 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001213 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1214 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001215 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1216 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001217 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1218 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1219 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1220 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1221 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1222 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1223 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1224 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1225 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1226 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1227 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1228 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1229 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1230 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1231 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1232 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001233 "src/f32-hswish/gen/hswish-neon-x4.c",
1234 "src/f32-hswish/gen/hswish-neon-x8.c",
1235 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001236 "src/f32-ibilinear/gen/neon-c4.c",
1237 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001238 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001239 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001240 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001241 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1242 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001243 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001244 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1245 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1246 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1247 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001248 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1249 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001250 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1251 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001252 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1253 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001254 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1255 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1256 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001257 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1258 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001259 "src/f32-prelu/gen/neon-1x4.c",
1260 "src/f32-prelu/gen/neon-1x8.c",
1261 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001262 "src/f32-prelu/gen/neon-2x4.c",
1263 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001264 "src/f32-prelu/gen/neon-2x16.c",
1265 "src/f32-prelu/gen/neon-4x4.c",
1266 "src/f32-prelu/gen/neon-4x8.c",
1267 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001268 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001269 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001270 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001271 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1272 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001273 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001274 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1275 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001276 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001277 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1278 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001279 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1280 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1281 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1282 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1283 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1284 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1285 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1286 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1287 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1288 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1289 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1290 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1291 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001292 "src/f32-relu/gen/neon-x4.c",
1293 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001294 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001295 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001296 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1297 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1298 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1299 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1300 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1301 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1302 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1303 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1304 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1305 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1306 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1307 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001308 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1309 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1310 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1311 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1312 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1313 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001314 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1315 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1316 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1317 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001318 "src/f32-vbinary/gen/vmax-neon-x4.c",
1319 "src/f32-vbinary/gen/vmax-neon-x8.c",
1320 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1321 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1322 "src/f32-vbinary/gen/vmin-neon-x4.c",
1323 "src/f32-vbinary/gen/vmin-neon-x8.c",
1324 "src/f32-vbinary/gen/vminc-neon-x4.c",
1325 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001326 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1327 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1328 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1329 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1330 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1331 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001332 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1333 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1334 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1335 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001336 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1337 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1338 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1339 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001340 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1341 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001342 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1343 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001344 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1345 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001346 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1347 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1348 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1349 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1350 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1351 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001352 "src/f32-vunary/gen/vabs-neon-x4.c",
1353 "src/f32-vunary/gen/vabs-neon-x8.c",
1354 "src/f32-vunary/gen/vneg-neon-x4.c",
1355 "src/f32-vunary/gen/vneg-neon-x8.c",
1356 "src/f32-vunary/gen/vsqr-neon-x4.c",
1357 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001358 "src/math/roundd-neon-addsub.c",
1359 "src/math/roundd-neon-cvt.c",
1360 "src/math/roundne-neon-addsub.c",
1361 "src/math/roundu-neon-addsub.c",
1362 "src/math/roundu-neon-cvt.c",
1363 "src/math/roundz-neon-addsub.c",
1364 "src/math/roundz-neon-cvt.c",
1365 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1366 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1367 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1368 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1369 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1370 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1371 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1372 "src/math/sqrt-neon-nr1rsqrts.c",
1373 "src/math/sqrt-neon-nr2rsqrts.c",
1374 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001375 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1376 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1377 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1378 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001379 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1380 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1381 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1382 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1384 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1385 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1386 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001387 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001388 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001389 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001390 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1391 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001392 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001393 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001394 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001395 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001396 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001397 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001398 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1399 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1400 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1401 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1402 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1403 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1404 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1405 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001406 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1407 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1408 "src/qu8-dwconv/up8x9-minmax-neon.c",
1409 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1410 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1411 "src/qu8-gemm/4x8-minmax-neon.c",
1412 "src/qu8-gemm/8x8-minmax-neon.c",
1413 "src/qu8-igemm/4x8-minmax-neon.c",
1414 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001415 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001416 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001417 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001418 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001419 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001420 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001421 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001422 "src/x8-zip/x2-neon.c",
1423 "src/x8-zip/x3-neon.c",
1424 "src/x8-zip/x4-neon.c",
1425 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001426 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001427 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001428 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001429 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001430 "src/x32-zip/x2-neon.c",
1431 "src/x32-zip/x3-neon.c",
1432 "src/x32-zip/x4-neon.c",
1433 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001434]
1435
1436NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1438 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1439 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1440 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1441 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1442 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1443 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1444 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1445 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1446 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1447 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1448 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1449 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1450 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1451 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1452 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1453 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1454 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1455 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1456 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1457 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1458 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1459 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1460 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1461 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1462 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1463 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1464 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1465 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1466 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001467 "src/f32-ibilinear/gen/neonfma-c4.c",
1468 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001469 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001470 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001471 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001472 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1473 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001474 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1475 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001476 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1477 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001478 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1479 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001480 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001481 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001482 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001483 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1484 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001485 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001486 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1487 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001488 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001489 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1490 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001491 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1492 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1493 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1494 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1495 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1496 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1497 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1498 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1499 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1500 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1501 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1502 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1503 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001504 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001510 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1511 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1512 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1513 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1514 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1515 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001516 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1517 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1518 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1519 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1520 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1521 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001522 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1523 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1524 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1525 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1526 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1527 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1529 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1530 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1531 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1532 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1533 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001534 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1535 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1536 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1537 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1538 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1539 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001540 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1541 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1542 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1543 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1544 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1545 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1546 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1547 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1548 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1549 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1550 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1551 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1552 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1553 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1554 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1555 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1556 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1557 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001558 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1559 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001560 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1561 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1562 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1563 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1564 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1565 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1566 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1567 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1568 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1569 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1570 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1571 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1572 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1573 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1574 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1575 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1576 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1577 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1578 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1579 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001580 "src/math/exp-neonfma-lut64-p2.c",
1581 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001582 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001583 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001584 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001585 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1586 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1587 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001588 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1589 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1590 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001591 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1592 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1593 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001594 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1595 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1596 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001597 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1598 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1599 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001600 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1601 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1602 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001603 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001604 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001605 "src/math/sqrt-neonfma-nr2fma.c",
1606 "src/math/sqrt-neonfma-nr2fma1adj.c",
1607 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001608]
1609
1610AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001611 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001612 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001613 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001614 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001615 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001616 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001617 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001618 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001619 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
1621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1624 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1625 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
1626 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1627 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1628 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
1629 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001630 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
1631 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1632 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1633 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
1634 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1635 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1636 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
1637 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001638 "src/f32-dwconv2d-chw/5x5p2-neonfma-3x4.c",
1639 "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001640 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1641 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1642 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1643 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1644 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1645 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1646 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1647 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1648 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1649 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1650 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1651 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1652 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1653 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1654 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1655 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1656 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1657 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1658 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1659 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001660 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1661 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1662 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1663 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1664 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1665 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1666 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1667 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1668 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1669 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1670 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1671 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1673 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1674 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1675 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1676 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1677 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001678 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1679 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1680 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1681 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1682 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1683 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1684 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1685 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1686 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1687 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001688 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1689 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1690 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1691 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1692 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1693 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1694 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1695 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1696 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1697 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1698 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1699 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1700 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1701 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001702 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1703 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001704 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001705 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001706 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001707 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001708 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001709 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001710]
1711
Marat Dukhan8853b822020-05-07 12:19:01 -07001712NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001713 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1714 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001715 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1716 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1717 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1718 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1719 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1720 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001721 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001722 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001723 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001724 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001725]
1726
Marat Dukhan08c4a432019-10-03 09:29:21 -07001727AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001728 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001729 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001730 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1731 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1732 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1733 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001734 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1735 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1736 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1737 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1738 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1739 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1740 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1741 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001742 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1743 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001744 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1745 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1746 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1747 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1748 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1749 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1750 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1751 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1752 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1753 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1754 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1755 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1756 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1757 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1758 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1759 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001760 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001761 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1762 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1763 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1764 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1765 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1766 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1767 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1768 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1769 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001770 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001771 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001772 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001773 "src/f16-relu/gen/neonfp16arith-x16.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001774 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001775 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001776 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001777 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001778 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001779 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001780 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001781 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1782 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1783 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1784 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1785 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1786 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1787 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1788 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1789 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1790 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1791 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1792 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1793 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1794 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1795 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1796 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1797 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1798 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1799 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1800 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1801 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1802 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1803 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1804 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1805 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1806 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1807 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1808 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1809 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1810 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1811 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001812]
1813
Benoit Jacoba9644732020-08-13 12:48:55 -07001814NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001815 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001816 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001817 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001818 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001819 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001820 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001821 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001822 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001823 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001824 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001825 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001826 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001827 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001828 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001829 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001831 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001832 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001833]
1834
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001836 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1837 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001838 "src/f32-clamp/gen/sse-x4.c",
1839 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001840 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1841 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001842 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1843 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1844 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1845 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001846 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1847 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001848 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1849 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1850 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1851 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001852 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1853 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001854 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1855 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1856 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1857 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1858 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1859 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1860 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1861 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1862 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1863 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001864 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
1865 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1866 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1867 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
1868 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1869 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1870 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
1871 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001872 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001873 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1874 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001875 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1876 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1877 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1878 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1879 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1880 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001881 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1882 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1883 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1884 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1885 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1886 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1887 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001888 "src/f32-hswish/gen/hswish-sse-x4.c",
1889 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001890 "src/f32-ibilinear/gen/sse-c4.c",
1891 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1893 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1894 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1895 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1896 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1897 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1898 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001899 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1900 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1901 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001902 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001903 "src/f32-prelu/gen/sse-2x4.c",
1904 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001905 "src/f32-relu/gen/sse-x4.c",
1906 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001907 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001908 "src/f32-spmm/gen/4x1-minmax-sse.c",
1909 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001910 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001911 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1912 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1913 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1914 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1915 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1916 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1917 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1918 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001919 "src/f32-vbinary/gen/vmax-sse-x4.c",
1920 "src/f32-vbinary/gen/vmax-sse-x8.c",
1921 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1922 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1923 "src/f32-vbinary/gen/vmin-sse-x4.c",
1924 "src/f32-vbinary/gen/vmin-sse-x8.c",
1925 "src/f32-vbinary/gen/vminc-sse-x4.c",
1926 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001927 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1928 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1929 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1930 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1931 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1932 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1933 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1934 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001935 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1936 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1937 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1938 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001939 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1940 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1941 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1942 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001943 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1944 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001945 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1946 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001947 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1948 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001949 "src/f32-vunary/gen/vabs-sse-x4.c",
1950 "src/f32-vunary/gen/vabs-sse-x8.c",
1951 "src/f32-vunary/gen/vneg-sse-x4.c",
1952 "src/f32-vunary/gen/vneg-sse-x8.c",
1953 "src/f32-vunary/gen/vsqr-sse-x4.c",
1954 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001955 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001957 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001958 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001959 "src/math/sqrt-sse-hh1mac.c",
1960 "src/math/sqrt-sse-nr1mac.c",
1961 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001962 "src/x32-fill/sse.c",
1963 "src/x32-packx/x4-sse.c",
1964 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001965]
1966
1967SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001968 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001969 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08001970 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001971 "src/f32-prelu/gen/sse2-2x4.c",
1972 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001973 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001974 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001975 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001976 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1977 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001978 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001979 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1980 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001981 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001982 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1983 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001984 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001985 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1986 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1987 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1988 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1989 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1990 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001991 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001992 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001993 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001994 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001995 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1996 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001997 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1998 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001999 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2000 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002001 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2002 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2003 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2004 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2005 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2006 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
2007 "src/math/exp-sse2-lut64-p2.c",
2008 "src/math/exp-sse2-p5.c",
2009 "src/math/expminus-sse2-p5.c",
2010 "src/math/roundd-sse2-cvt.c",
2011 "src/math/roundne-sse2-cvt.c",
2012 "src/math/roundu-sse2-cvt.c",
2013 "src/math/roundz-sse2-cvt.c",
2014 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2015 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2016 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2017 "src/math/sigmoid-sse2-rr2-p5-div.c",
2018 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2019 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002020 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2021 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2022 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002023 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2024 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2025 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002026 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2027 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2028 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002029 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002030 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002031 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002032 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002033 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002034 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002035 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2036 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002037 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002038 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2039 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002040 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002041 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2042 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2043 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002044 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002045 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002046 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002047 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002049 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002050 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002051 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002052 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2053 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002054 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002055 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002056 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002057 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2058 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2059 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2060 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002061 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2062 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2063 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2064 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002065 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2066 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002067 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2068 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2069 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2070 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2071 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002072 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002073 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002074 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002075 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002076 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002077 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002078 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002080 "src/x8-zip/x2-sse2.c",
2081 "src/x8-zip/x3-sse2.c",
2082 "src/x8-zip/x4-sse2.c",
2083 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002084 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 "src/x32-zip/x2-sse2.c",
2086 "src/x32-zip/x3-sse2.c",
2087 "src/x32-zip/x4-sse2.c",
2088 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002089]
2090
2091SSSE3_UKERNELS = [
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002092 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
2093 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2094 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2095 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2096 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2097 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
2098 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2099 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2100 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
2101 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002102 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2103 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2104 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002105 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2106 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2107 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002108 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2109 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2110 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002111 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002112 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002113 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002114 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002115 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002116 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002117 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2118 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002119 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002120 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2121 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002122 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002123 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2124 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2125 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002126 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002127 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002128 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002129 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002130 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002131 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002133 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002134 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2135 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002136 "src/qs8-requantization/precise-ssse3.c",
2137 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002138 "src/qu8-requantization/precise-ssse3.c",
2139 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140]
2141
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002142SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002143 "src/f32-prelu/gen/sse41-2x4.c",
2144 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002145 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2146 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2147 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2148 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2149 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2150 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002151 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002152 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002153 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002154 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002155 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2156 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002157 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2158 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002159 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2160 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2162 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2163 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2164 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2165 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2166 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2167 "src/math/roundd-sse41.c",
2168 "src/math/roundne-sse41.c",
2169 "src/math/roundu-sse41.c",
2170 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002171 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2172 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2173 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002174 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2175 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2176 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2178 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2179 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002180 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002181 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002182 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002183 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002184 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002185 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002186 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2187 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002188 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002189 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2190 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002191 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002192 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2193 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2194 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002195 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002196 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002197 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002198 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002199 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002200 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002202 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2204 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002205 "src/qs8-requantization/fp32-sse4.c",
2206 "src/qs8-requantization/precise-sse4.c",
2207 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002208 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2209 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2210 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2211 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002212 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2213 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2214 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2215 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002216 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2217 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2218 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2219 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002220 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2221 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2222 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2223 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002224 "src/qu8-requantization/precise-sse4.c",
2225 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002226]
2227
Marat Dukhan08c4a432019-10-03 09:29:21 -07002228AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002229 "src/f32-clamp/gen/avx-x8.c",
2230 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002231 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2232 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002233 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2234 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002235 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2236 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2238 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2239 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2240 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2241 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2242 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002243 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002244 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2245 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002246 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002247 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002248 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002249 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002250 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2251 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2252 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2253 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2254 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2255 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2256 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2257 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2258 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2259 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2260 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002261 "src/f32-hswish/gen/hswish-avx-x8.c",
2262 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002263 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002264 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2265 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002266 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002267 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002268 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002269 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2271 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002272 "src/f32-prelu/gen/avx-2x8.c",
2273 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002274 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002277 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2278 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2279 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2280 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2281 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2282 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2283 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2284 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2285 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2286 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2287 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2288 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2289 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2290 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2291 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2292 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2293 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2294 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2295 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2296 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002297 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2298 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2299 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2300 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2301 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2302 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2303 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2304 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002305 "src/f32-vbinary/gen/vmax-avx-x8.c",
2306 "src/f32-vbinary/gen/vmax-avx-x16.c",
2307 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2308 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2309 "src/f32-vbinary/gen/vmin-avx-x8.c",
2310 "src/f32-vbinary/gen/vmin-avx-x16.c",
2311 "src/f32-vbinary/gen/vminc-avx-x8.c",
2312 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002313 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2314 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2315 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2316 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2317 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2318 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2319 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2320 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002321 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2322 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2323 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2324 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002325 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2326 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2327 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2328 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002329 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2330 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002331 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2332 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002333 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2334 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2335 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2336 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2337 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2338 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2339 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002340 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2341 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002342 "src/f32-vunary/gen/vabs-avx-x8.c",
2343 "src/f32-vunary/gen/vabs-avx-x16.c",
2344 "src/f32-vunary/gen/vneg-avx-x8.c",
2345 "src/f32-vunary/gen/vneg-avx-x16.c",
2346 "src/f32-vunary/gen/vsqr-avx-x8.c",
2347 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002348 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002349 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002350 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002351 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2352 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353]
2354
Marat Dukhan1566fee2020-08-02 21:55:41 -07002355XOP_UKERNELS = [
2356 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002357 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002358 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002359 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002360 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002361 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002362 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2363 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002364 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002365 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2366 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002367 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002368 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2369 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2370 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002371 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002372 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002373 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002374 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002375 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002376 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002377 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002378 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002379 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2380 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002381 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2382 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2383 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2384 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2385 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2386 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2387 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2388 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002389]
2390
Marat Dukhanfda12b82019-11-21 12:27:59 -08002391FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002392 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2393 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002394 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2395 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002396 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2397 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002398 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2399 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2400 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2401 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2402 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2403 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002404 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002405 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2406 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2407 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2408 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002409 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002410 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2411 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002412 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002413 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2414 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002415 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2416 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2417 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002418 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2419 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2420 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2421 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2422 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2423 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2424 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2425 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2426 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2427 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2428 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2429 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2430 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2431 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002432 "src/f32-hswish/gen/hswish-fma3-x8.c",
2433 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002434 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2436 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2437 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2438 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002439 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002440 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2441 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002442 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002443 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2444 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002445 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2446 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2447 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002448 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2449 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2450 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2451 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2452 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2453 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2454 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2455 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002456 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002457 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002458 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002459]
2460
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002461AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002462 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2463 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002464 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002465 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002466 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002467 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2468 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002469 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002470 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2471 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2472 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002473 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002474 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2475 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002476 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002477 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002478 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002479 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2480 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002481 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002482 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2483 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2484 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002486 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2487 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002488 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002489 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002490 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002491 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2492 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002493 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002494 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2495 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2496 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002497 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002498 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2499 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2500 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2501 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2502 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2503 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2504 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2505 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2506 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2507 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2508 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2509 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2510 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2511 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2512 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2513 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2514 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2515 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2516 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2517 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2518 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2519 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2520 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2521 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2522 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2523 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2524 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2525 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2526 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2527 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002528 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2529 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2530 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2531 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2532 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2533 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2534 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2535 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2536 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2537 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2538 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2539 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2540 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2541 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2542 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2543 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2544 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2545 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2546 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2547 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2548 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2549 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2550 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2551 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002552 "src/math/exp-avx2-lut8-p3-perm.c",
2553 "src/math/exp-avx2-lut8-p4-perm.c",
2554 "src/math/exp-avx2-p5.c",
2555 "src/math/expminus-avx2-p5.c",
2556 "src/math/extexp-avx2-p5.c",
2557 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2558 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2559 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2560 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2561 "src/math/sigmoid-avx2-rr1-p5-div.c",
2562 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2563 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2564 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2565 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2566 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2567 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2568 "src/math/sigmoid-avx2-rr2-p5-div.c",
2569 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2570 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002571 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002573 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2574 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002575 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002576 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002577 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002578 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002579 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002580 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002581 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002582 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002583 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2584 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2585 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002586 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2587 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2588 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2589 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2590 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2591 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2592 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2593 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002594]
2595
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002597 "src/f32-clamp/gen/avx512f-x16.c",
2598 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002599 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2600 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002601 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2602 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002603 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2604 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002605 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2606 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2607 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2608 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2609 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2610 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002611 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2612 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2613 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2614 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2615 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2616 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002617 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2618 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2619 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2620 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2621 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2622 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002623 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2624 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002625 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2626 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2627 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2628 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2629 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2630 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002631 "src/f32-prelu/gen/avx512f-2x16.c",
2632 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002633 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2634 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002635 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002636 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002638 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2639 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002640 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002641 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2642 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2643 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002645 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2646 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002647 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002648 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002649 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002650 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2651 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002652 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002653 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2654 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2655 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002657 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2658 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002659 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002660 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002661 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002662 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2663 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002664 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002665 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2666 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2667 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002668 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002669 "src/f32-relu/gen/avx512f-x16.c",
2670 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002671 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002672 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2673 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2674 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2675 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2676 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2677 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2678 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2679 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2680 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2681 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2682 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2683 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2684 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2685 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2686 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2687 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002688 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2689 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2690 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2691 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2692 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2693 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2694 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2695 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2696 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2697 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2698 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2699 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2700 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2701 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2702 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2703 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002704 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2705 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2706 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2707 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2708 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2709 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2710 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2711 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2712 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2713 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2714 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2715 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2716 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2717 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2718 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2719 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002720 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2721 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2722 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2723 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2724 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2725 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2726 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2727 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002728 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2729 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2730 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2731 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2732 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2733 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2734 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2735 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002736 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2737 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2738 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2739 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2740 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2741 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2742 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2743 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002744 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2745 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2746 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2747 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002748 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2749 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2750 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2751 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002752 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2753 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002754 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2755 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2756 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2757 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2758 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2759 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2760 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2761 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002762 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002763 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2764 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2765 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2766 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2767 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2768 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2769 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2770 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2771 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2772 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2773 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2774 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2775 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2776 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2777 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2778 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2779 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2780 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2781 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2782 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2783 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2784 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2785 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2786 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002787 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2788 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2789 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2790 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2791 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2792 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2793 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2794 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002795 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2796 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2797 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2798 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2799 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2800 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002801 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2802 "src/math/exp-avx512f-lut16-p3-perm.c",
2803 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2804 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002805 "src/math/exp-avx512f-p5-scalef.c",
2806 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002807 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002808 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002809 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002810 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002811 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002812 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002813 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002814 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002815 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002816 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2817 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2818 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2819 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2820 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2821 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2822 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2823 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2824 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2825 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002826 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002827 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002828 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2829 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2830 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2831 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002832 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002833 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002834 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835]
2836
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002837AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002838 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2839 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002840 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2841 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2842 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2843 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2844 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2845 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2846 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2847 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2848]
2849
Frank Barchardbcedc082020-08-17 18:00:51 -07002850WASM32_ASM_UKERNELS = [
2851 "src/f32-relu/wasm_shr_x1.S",
2852 "src/f32-relu/wasm_shr_x2.S",
2853 "src/f32-relu/wasm_shr_x4.S",
2854]
2855
Marat Dukhan08c4a432019-10-03 09:29:21 -07002856AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002857 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002858 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002859 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2860 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002861 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002862 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002863 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002864 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002865 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2866 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002867 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2868 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2869 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2870 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871]
2872
2873AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002874 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002875 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002876 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002877 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002878 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002879 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002880 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002881 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2882 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2883 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2884 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2885 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2886 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2887 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002888 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2889 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002890 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2891 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2892 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002893 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2894 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002895 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2896 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2897 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2898 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002899 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002900 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2901 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002902 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2903 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2904 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2905 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002906 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002907 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002908 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002909 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002910 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2911 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2912 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2913 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2914 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2915 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2916 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2917 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2918 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2919 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2920 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2921 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2922 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2923 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2924 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2925 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2926 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2927 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2928 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2929 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2930 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2931 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002932 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002933 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002934 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2935 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002936 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2937 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2938 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2939 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2940 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2941 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002942 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2943 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2944 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2945 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002946 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2947 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002948 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002949 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002950 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951]
2952
Marat Dukhan1b354632020-03-23 12:50:22 -07002953INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954 "src/xnnpack/argmaxpool.h",
2955 "src/xnnpack/avgpool.h",
2956 "src/xnnpack/clamp.h",
2957 "src/xnnpack/common.h",
2958 "src/xnnpack/conv.h",
2959 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002960 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 "src/xnnpack/gavgpool.h",
2962 "src/xnnpack/gemm.h",
2963 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002964 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002965 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002966 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 "src/xnnpack/lut.h",
2968 "src/xnnpack/math.h",
2969 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002970 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 "src/xnnpack/packx.h",
2972 "src/xnnpack/pad.h",
2973 "src/xnnpack/params.h",
2974 "src/xnnpack/pavgpool.h",
2975 "src/xnnpack/ppmm.h",
2976 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002977 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002978 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002979 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002980 "src/xnnpack/rmax.h",
2981 "src/xnnpack/scalar-utils.h",
2982 "src/xnnpack/spmm.h",
2983 "src/xnnpack/unpool.h",
2984 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002985 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002987 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002988 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002989 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002990 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002992]
2993
2994INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 "include/xnnpack.h",
2996 "src/xnnpack/allocator.h",
2997 "src/xnnpack/compute.h",
2998 "src/xnnpack/im2col.h",
2999 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003000 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003001 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 "src/xnnpack/operator.h",
3003 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003004 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003006 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003007 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003008]
3009
Marat Dukhan1b354632020-03-23 12:50:22 -07003010ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003011 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003012]
3013
Marat Dukhan1b354632020-03-23 12:50:22 -07003014MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003015 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 "include/xnnpack.h",
3017]
3018
Marat Dukhan1b354632020-03-23 12:50:22 -07003019MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003020 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003021 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 "src/xnnpack/requantization.h",
3023 "include/xnnpack.h",
3024]
3025
3026OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029]
3030
3031WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003032 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003033 "src/xnnpack/operator.h",
3034 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003035]
3036
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003037LOGGING_COPTS = select({
3038 # No logging in optimized mode
3039 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3040 # Full logging in debug mode
3041 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3042 # Error-only logging in default (fastbuild) mode
3043 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3044})
3045
Marat Dukhan3b59de22020-06-03 20:15:19 -07003046LOGGING_SRCS = select({
3047 # No logging in optimized mode
3048 ":optimized_build": [],
3049 "//conditions:default": [
3050 "src/operator-strings.c",
3051 "src/subgraph-strings.c",
3052 ],
3053})
3054
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003055LOGGING_HDRS = [
3056 "src/xnnpack/log.h",
3057]
3058
Marat Dukhan08c4a432019-10-03 09:29:21 -07003059xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003060 name = "tables",
3061 srcs = TABLE_SRCS,
3062 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003063 gcc_copts = xnnpack_gcc_std_copts(),
3064 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003065)
3066
3067xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 name = "scalar_ukernels",
3069 srcs = SCALAR_UKERNELS,
3070 hdrs = INTERNAL_HDRS,
3071 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003072 gcc_copts = xnnpack_gcc_std_copts(),
3073 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003075 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 "@FP16",
3077 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003078 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003079 ],
3080)
3081
3082xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003083 name = "scalar_ukernels_test_mode",
3084 srcs = SCALAR_UKERNELS,
3085 hdrs = INTERNAL_HDRS,
3086 aarch32_copts = ["-marm"],
3087 copts = [
3088 "-UNDEBUG",
3089 "-DXNN_TEST_MODE=1",
3090 ],
3091 gcc_copts = xnnpack_gcc_std_copts(),
3092 msvc_copts = xnnpack_msvc_std_copts(),
3093 deps = [
3094 ":tables",
3095 "@FP16",
3096 "@FXdiv",
3097 "@pthreadpool",
3098 ],
3099)
3100
3101xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003102 name = "wasm_ukernels",
3103 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003104 gcc_copts = xnnpack_gcc_std_copts(),
3105 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003106 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003107 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003108 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003109 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003110 "@FP16",
3111 "@FXdiv",
3112 "@pthreadpool",
3113 ],
3114)
3115
3116xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003117 name = "wasm_ukernels_test_mode",
3118 hdrs = INTERNAL_HDRS,
3119 copts = [
3120 "-UNDEBUG",
3121 "-DXNN_TEST_MODE=1",
3122 ],
3123 gcc_copts = xnnpack_gcc_std_copts(),
3124 msvc_copts = xnnpack_msvc_std_copts(),
3125 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003126 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003127 deps = [
3128 ":tables",
3129 "@FP16",
3130 "@FXdiv",
3131 "@pthreadpool",
3132 ],
3133)
3134
3135xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003136 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 hdrs = INTERNAL_HDRS,
3138 aarch32_copts = [
3139 "-marm",
3140 "-mfpu=neon",
3141 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003142 gcc_copts = xnnpack_gcc_std_copts(),
3143 gcc_x86_copts = ["-msse2"],
3144 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 optimized_copts = [
3146 "-O3",
3147 "-ffast-math",
3148 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003149 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003151 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 "@FP16",
3153 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003154 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003155 ],
3156)
3157
3158xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003159 name = "psimd_fastmath_ukernels_test_mode",
3160 hdrs = INTERNAL_HDRS,
3161 aarch32_copts = [
3162 "-marm",
3163 "-mfpu=neon",
3164 ],
3165 copts = [
3166 "-UNDEBUG",
3167 "-DXNN_TEST_MODE=1",
3168 ],
3169 gcc_copts = xnnpack_gcc_std_copts(),
3170 gcc_x86_copts = ["-msse2"],
3171 msvc_copts = xnnpack_msvc_std_copts(),
3172 optimized_copts = [
3173 "-O3",
3174 "-ffast-math",
3175 ],
3176 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3177 deps = [
3178 ":tables",
3179 "@FP16",
3180 "@psimd",
3181 "@pthreadpool",
3182 ],
3183)
3184
3185xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003186 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003187 hdrs = INTERNAL_HDRS,
3188 aarch32_copts = [
3189 "-marm",
3190 "-mfpu=neon",
3191 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003192 gcc_copts = xnnpack_gcc_std_copts(),
3193 gcc_x86_copts = ["-msse2"],
3194 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003195 optimized_copts = [
3196 "-O3",
3197 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003198 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003199 deps = [
3200 ":tables",
3201 "@FP16",
3202 "@psimd",
3203 "@pthreadpool",
3204 ],
3205)
3206
3207xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003208 name = "psimd_accmath_ukernels_test_mode",
3209 hdrs = INTERNAL_HDRS,
3210 aarch32_copts = [
3211 "-marm",
3212 "-mfpu=neon",
3213 ],
3214 copts = [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ],
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 gcc_x86_copts = ["-msse2"],
3220 msvc_copts = xnnpack_msvc_std_copts(),
3221 optimized_copts = [
3222 "-O3",
3223 ],
3224 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3225 deps = [
3226 ":tables",
3227 "@FP16",
3228 "@psimd",
3229 "@pthreadpool",
3230 ],
3231)
3232
3233xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 name = "neon_ukernels",
3235 hdrs = INTERNAL_HDRS,
3236 aarch32_copts = [
3237 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003238 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 "-mfpu=neon",
3240 ],
3241 aarch32_srcs = NEON_UKERNELS,
3242 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003243 gcc_copts = xnnpack_gcc_std_copts(),
3244 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003245 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003246 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003247 "@FP16",
3248 "@pthreadpool",
3249 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250)
3251
3252xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003253 name = "neon_ukernels_test_mode",
3254 hdrs = INTERNAL_HDRS,
3255 aarch32_copts = [
3256 "-marm",
3257 "-march=armv7-a",
3258 "-mfpu=neon",
3259 ],
3260 aarch32_srcs = NEON_UKERNELS,
3261 aarch64_srcs = NEON_UKERNELS,
3262 copts = [
3263 "-UNDEBUG",
3264 "-DXNN_TEST_MODE=1",
3265 ],
3266 gcc_copts = xnnpack_gcc_std_copts(),
3267 msvc_copts = xnnpack_msvc_std_copts(),
3268 deps = [
3269 ":tables",
3270 "@FP16",
3271 "@pthreadpool",
3272 ],
3273)
3274
3275xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 name = "neonfma_ukernels",
3277 hdrs = INTERNAL_HDRS,
3278 aarch32_copts = [
3279 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003280 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281 "-mfpu=neon-vfpv4",
3282 ],
3283 aarch32_srcs = NEONFMA_UKERNELS,
3284 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003285 apple_aarch32_copts = [
3286 "-mcpu=swift",
3287 "-mtune=generic",
3288 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003289 gcc_copts = xnnpack_gcc_std_copts(),
3290 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003291 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003292 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003293 "@FP16",
3294 "@pthreadpool",
3295 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296)
3297
3298xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003299 name = "neonfma_ukernels_test_mode",
3300 hdrs = INTERNAL_HDRS,
3301 aarch32_copts = [
3302 "-marm",
3303 "-march=armv7-a",
3304 "-mfpu=neon-vfpv4",
3305 ],
3306 aarch32_srcs = NEONFMA_UKERNELS,
3307 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003308 apple_aarch32_copts = [
3309 "-mcpu=swift",
3310 "-mtune=generic",
3311 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003312 copts = [
3313 "-UNDEBUG",
3314 "-DXNN_TEST_MODE=1",
3315 ],
3316 gcc_copts = xnnpack_gcc_std_copts(),
3317 msvc_copts = xnnpack_msvc_std_copts(),
3318 deps = [
3319 ":tables",
3320 "@FP16",
3321 "@pthreadpool",
3322 ],
3323)
3324
3325xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003326 name = "neonv8_ukernels",
3327 hdrs = INTERNAL_HDRS,
3328 aarch32_copts = [
3329 "-marm",
3330 "-march=armv8-a",
3331 "-mfpu=neon-fp-armv8",
3332 ],
3333 aarch32_srcs = NEONV8_UKERNELS,
3334 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003335 apple_aarch32_copts = [
3336 "-mcpu=cyclone",
3337 "-mtune=generic",
3338 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003339 gcc_copts = xnnpack_gcc_std_copts(),
3340 msvc_copts = xnnpack_msvc_std_copts(),
3341 deps = [
3342 ":tables",
3343 "@FP16",
3344 "@pthreadpool",
3345 ],
3346)
3347
3348xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003349 name = "neonv8_ukernels_test_mode",
3350 hdrs = INTERNAL_HDRS,
3351 aarch32_copts = [
3352 "-marm",
3353 "-march=armv8-a",
3354 "-mfpu=neon-fp-armv8",
3355 ],
3356 aarch32_srcs = NEONV8_UKERNELS,
3357 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003358 apple_aarch32_copts = [
3359 "-mcpu=cyclone",
3360 "-mtune=generic",
3361 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003362 copts = [
3363 "-UNDEBUG",
3364 "-DXNN_TEST_MODE=1",
3365 ],
3366 gcc_copts = xnnpack_gcc_std_copts(),
3367 msvc_copts = xnnpack_msvc_std_copts(),
3368 deps = [
3369 ":tables",
3370 "@FP16",
3371 "@pthreadpool",
3372 ],
3373)
3374
3375xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376 name = "neonfp16arith_ukernels",
3377 hdrs = INTERNAL_HDRS,
3378 aarch64_copts = ["-march=armv8.2-a+fp16"],
3379 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003380 gcc_copts = xnnpack_gcc_std_copts(),
3381 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003382 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003383 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003384 "@FP16",
3385 "@pthreadpool",
3386 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387)
3388
3389xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003390 name = "neonfp16arith_ukernels_test_mode",
3391 hdrs = INTERNAL_HDRS,
3392 aarch64_copts = ["-march=armv8.2-a+fp16"],
3393 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3394 copts = [
3395 "-UNDEBUG",
3396 "-DXNN_TEST_MODE=1",
3397 ],
3398 gcc_copts = xnnpack_gcc_std_copts(),
3399 msvc_copts = xnnpack_msvc_std_copts(),
3400 deps = [
3401 ":tables",
3402 "@FP16",
3403 "@pthreadpool",
3404 ],
3405)
3406
3407xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003408 name = "neondot_ukernels",
3409 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003410 aarch32_copts = [
3411 "-marm",
3412 "-march=armv8.2-a+dotprod",
3413 "-mfpu=neon-fp-armv8",
3414 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003415 aarch32_srcs = NEONDOT_UKERNELS,
3416 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3417 aarch64_srcs = NEONDOT_UKERNELS,
3418 gcc_copts = xnnpack_gcc_std_copts(),
3419 msvc_copts = xnnpack_msvc_std_copts(),
3420 deps = [
3421 ":tables",
3422 "@FP16",
3423 "@pthreadpool",
3424 ],
3425)
3426
3427xnnpack_cc_library(
3428 name = "neondot_ukernels_test_mode",
3429 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003430 aarch32_copts = [
3431 "-marm",
3432 "-march=armv8.2-a+dotprod",
3433 "-mfpu=neon-fp-armv8",
3434 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003435 aarch32_srcs = NEONDOT_UKERNELS,
3436 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3437 aarch64_srcs = NEONDOT_UKERNELS,
3438 copts = [
3439 "-UNDEBUG",
3440 "-DXNN_TEST_MODE=1",
3441 ],
3442 gcc_copts = xnnpack_gcc_std_copts(),
3443 msvc_copts = xnnpack_msvc_std_copts(),
3444 deps = [
3445 ":tables",
3446 "@FP16",
3447 "@pthreadpool",
3448 ],
3449)
3450
3451xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452 name = "sse2_ukernels",
3453 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003454 gcc_copts = xnnpack_gcc_std_copts(),
3455 gcc_x86_copts = ["-msse2"],
3456 msvc_copts = xnnpack_msvc_std_copts(),
3457 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003458 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003459 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003460 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003461 "@FP16",
3462 "@pthreadpool",
3463 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464)
3465
3466xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 name = "sse2_ukernels_test_mode",
3468 hdrs = INTERNAL_HDRS,
3469 copts = [
3470 "-UNDEBUG",
3471 "-DXNN_TEST_MODE=1",
3472 ],
3473 gcc_copts = xnnpack_gcc_std_copts(),
3474 gcc_x86_copts = ["-msse2"],
3475 msvc_copts = xnnpack_msvc_std_copts(),
3476 msvc_x86_32_copts = ["/arch:SSE2"],
3477 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3478 deps = [
3479 ":tables",
3480 "@FP16",
3481 "@pthreadpool",
3482 ],
3483)
3484
3485xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003486 name = "ssse3_ukernels",
3487 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003488 gcc_copts = xnnpack_gcc_std_copts(),
3489 gcc_x86_copts = ["-mssse3"],
3490 msvc_copts = xnnpack_msvc_std_copts(),
3491 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003492 x86_srcs = SSSE3_UKERNELS,
3493 deps = [
3494 ":tables",
3495 "@FP16",
3496 "@pthreadpool",
3497 ],
3498)
3499
3500xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003501 name = "ssse3_ukernels_test_mode",
3502 hdrs = INTERNAL_HDRS,
3503 copts = [
3504 "-UNDEBUG",
3505 "-DXNN_TEST_MODE=1",
3506 ],
3507 gcc_copts = xnnpack_gcc_std_copts(),
3508 gcc_x86_copts = ["-mssse3"],
3509 msvc_copts = xnnpack_msvc_std_copts(),
3510 msvc_x86_32_copts = ["/arch:SSE2"],
3511 x86_srcs = SSSE3_UKERNELS,
3512 deps = [
3513 ":tables",
3514 "@FP16",
3515 "@pthreadpool",
3516 ],
3517)
3518
3519xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003520 name = "sse41_ukernels",
3521 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003522 gcc_copts = xnnpack_gcc_std_copts(),
3523 gcc_x86_copts = ["-msse4.1"],
3524 msvc_copts = xnnpack_msvc_std_copts(),
3525 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003526 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003527 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003528 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003529 "@FP16",
3530 "@pthreadpool",
3531 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003532)
3533
3534xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003535 name = "sse41_ukernels_test_mode",
3536 hdrs = INTERNAL_HDRS,
3537 copts = [
3538 "-UNDEBUG",
3539 "-DXNN_TEST_MODE=1",
3540 ],
3541 gcc_copts = xnnpack_gcc_std_copts(),
3542 gcc_x86_copts = ["-msse4.1"],
3543 msvc_copts = xnnpack_msvc_std_copts(),
3544 msvc_x86_32_copts = ["/arch:SSE2"],
3545 x86_srcs = SSE41_UKERNELS,
3546 deps = [
3547 ":tables",
3548 "@FP16",
3549 "@pthreadpool",
3550 ],
3551)
3552
3553xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 name = "avx_ukernels",
3555 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003556 gcc_copts = xnnpack_gcc_std_copts(),
3557 gcc_x86_copts = ["-mavx"],
3558 msvc_copts = xnnpack_msvc_std_copts(),
3559 msvc_x86_32_copts = ["/arch:AVX"],
3560 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003562 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003563 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003564 "@FP16",
3565 "@pthreadpool",
3566 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003567)
3568
3569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003570 name = "avx_ukernels_test_mode",
3571 hdrs = INTERNAL_HDRS,
3572 copts = [
3573 "-UNDEBUG",
3574 "-DXNN_TEST_MODE=1",
3575 ],
3576 gcc_copts = xnnpack_gcc_std_copts(),
3577 gcc_x86_copts = ["-mavx"],
3578 msvc_copts = xnnpack_msvc_std_copts(),
3579 msvc_x86_32_copts = ["/arch:AVX"],
3580 msvc_x86_64_copts = ["/arch:AVX"],
3581 x86_srcs = AVX_UKERNELS,
3582 deps = [
3583 ":tables",
3584 "@FP16",
3585 "@pthreadpool",
3586 ],
3587)
3588
3589xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003590 name = "xop_ukernels",
3591 hdrs = INTERNAL_HDRS,
3592 gcc_copts = xnnpack_gcc_std_copts(),
3593 gcc_x86_copts = ["-mxop"],
3594 msvc_copts = xnnpack_msvc_std_copts(),
3595 msvc_x86_32_copts = ["/arch:AVX"],
3596 msvc_x86_64_copts = ["/arch:AVX"],
3597 x86_srcs = XOP_UKERNELS,
3598 deps = [
3599 ":tables",
3600 "@FP16",
3601 "@pthreadpool",
3602 ],
3603)
3604
3605xnnpack_cc_library(
3606 name = "xop_ukernels_test_mode",
3607 hdrs = INTERNAL_HDRS,
3608 copts = [
3609 "-UNDEBUG",
3610 "-DXNN_TEST_MODE=1",
3611 ],
3612 gcc_copts = xnnpack_gcc_std_copts(),
3613 gcc_x86_copts = ["-mxop"],
3614 msvc_copts = xnnpack_msvc_std_copts(),
3615 msvc_x86_32_copts = ["/arch:AVX"],
3616 msvc_x86_64_copts = ["/arch:AVX"],
3617 x86_srcs = XOP_UKERNELS,
3618 deps = [
3619 ":tables",
3620 "@FP16",
3621 "@pthreadpool",
3622 ],
3623)
3624
3625xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003626 name = "fma3_ukernels",
3627 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003628 gcc_copts = xnnpack_gcc_std_copts(),
3629 gcc_x86_copts = ["-mfma"],
3630 msvc_copts = xnnpack_msvc_std_copts(),
3631 msvc_x86_32_copts = ["/arch:AVX"],
3632 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003633 x86_srcs = FMA3_UKERNELS,
3634 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003635 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003636 "@FP16",
3637 "@pthreadpool",
3638 ],
3639)
3640
3641xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003642 name = "fma3_ukernels_test_mode",
3643 hdrs = INTERNAL_HDRS,
3644 copts = [
3645 "-UNDEBUG",
3646 "-DXNN_TEST_MODE=1",
3647 ],
3648 gcc_copts = xnnpack_gcc_std_copts(),
3649 gcc_x86_copts = ["-mfma"],
3650 msvc_copts = xnnpack_msvc_std_copts(),
3651 msvc_x86_32_copts = ["/arch:AVX"],
3652 msvc_x86_64_copts = ["/arch:AVX"],
3653 x86_srcs = FMA3_UKERNELS,
3654 deps = [
3655 ":tables",
3656 "@FP16",
3657 "@pthreadpool",
3658 ],
3659)
3660
3661xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003662 name = "avx2_ukernels",
3663 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003664 gcc_copts = xnnpack_gcc_std_copts(),
3665 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003666 "-mfma",
3667 "-mavx2",
3668 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003669 msvc_copts = xnnpack_msvc_std_copts(),
3670 msvc_x86_32_copts = ["/arch:AVX2"],
3671 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003672 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003673 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003674 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003675 "@FP16",
3676 "@pthreadpool",
3677 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003678)
3679
3680xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003681 name = "avx2_ukernels_test_mode",
3682 hdrs = INTERNAL_HDRS,
3683 copts = [
3684 "-UNDEBUG",
3685 "-DXNN_TEST_MODE=1",
3686 ],
3687 gcc_copts = xnnpack_gcc_std_copts(),
3688 gcc_x86_copts = [
3689 "-mfma",
3690 "-mavx2",
3691 ],
3692 msvc_copts = xnnpack_msvc_std_copts(),
3693 msvc_x86_32_copts = ["/arch:AVX2"],
3694 msvc_x86_64_copts = ["/arch:AVX2"],
3695 x86_srcs = AVX2_UKERNELS,
3696 deps = [
3697 ":tables",
3698 "@FP16",
3699 "@pthreadpool",
3700 ],
3701)
3702
3703xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704 name = "avx512f_ukernels",
3705 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003706 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"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003713 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003714 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003715 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003716 "@FP16",
3717 "@pthreadpool",
3718 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719)
3720
3721xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003722 name = "avx512f_ukernels_test_mode",
3723 hdrs = INTERNAL_HDRS,
3724 copts = [
3725 "-UNDEBUG",
3726 "-DXNN_TEST_MODE=1",
3727 ],
3728 gcc_copts = xnnpack_gcc_std_copts(),
3729 gcc_x86_copts = ["-mavx512f"],
3730 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3731 msvc_copts = xnnpack_msvc_std_copts(),
3732 msvc_x86_32_copts = ["/arch:AVX512"],
3733 msvc_x86_64_copts = ["/arch:AVX512"],
3734 msys_copts = ["-fno-asynchronous-unwind-tables"],
3735 x86_srcs = AVX512F_UKERNELS,
3736 deps = [
3737 ":tables",
3738 "@FP16",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003744 name = "avx512skx_ukernels",
3745 hdrs = INTERNAL_HDRS,
3746 gcc_copts = xnnpack_gcc_std_copts(),
3747 gcc_x86_copts = [
3748 "-mavx512f",
3749 "-mavx512cd",
3750 "-mavx512bw",
3751 "-mavx512dq",
3752 "-mavx512vl",
3753 ],
3754 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3755 msvc_copts = xnnpack_msvc_std_copts(),
3756 msvc_x86_32_copts = ["/arch:AVX512"],
3757 msvc_x86_64_copts = ["/arch:AVX512"],
3758 msys_copts = ["-fno-asynchronous-unwind-tables"],
3759 x86_srcs = AVX512SKX_UKERNELS,
3760 deps = [
3761 ":tables",
3762 "@FP16",
3763 "@pthreadpool",
3764 ],
3765)
3766
3767xnnpack_cc_library(
3768 name = "avx512skx_ukernels_test_mode",
3769 hdrs = INTERNAL_HDRS,
3770 copts = [
3771 "-UNDEBUG",
3772 "-DXNN_TEST_MODE=1",
3773 ],
3774 gcc_copts = xnnpack_gcc_std_copts(),
3775 gcc_x86_copts = [
3776 "-mavx512f",
3777 "-mavx512cd",
3778 "-mavx512bw",
3779 "-mavx512dq",
3780 "-mavx512vl",
3781 ],
3782 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3783 msvc_copts = xnnpack_msvc_std_copts(),
3784 msvc_x86_32_copts = ["/arch:AVX512"],
3785 msvc_x86_64_copts = ["/arch:AVX512"],
3786 msys_copts = ["-fno-asynchronous-unwind-tables"],
3787 x86_srcs = AVX512SKX_UKERNELS,
3788 deps = [
3789 ":tables",
3790 "@FP16",
3791 "@pthreadpool",
3792 ],
3793)
3794
3795xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003796 name = "asm_ukernels",
3797 hdrs = ["src/xnnpack/assembly.h"],
3798 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003799 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003800 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003801 wasm_srcs = WASM32_ASM_UKERNELS,
3802 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003803)
3804
Marat Dukhan3b59de22020-06-03 20:15:19 -07003805xnnpack_cc_library(
3806 name = "logging_utils",
3807 srcs = LOGGING_SRCS,
3808 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3809 copts = LOGGING_COPTS + [
3810 "-Isrc",
3811 "-Iinclude",
3812 ] + select({
3813 ":debug_build": [],
3814 "//conditions:default": xnnpack_min_size_copts(),
3815 }),
3816 gcc_copts = xnnpack_gcc_std_copts(),
3817 msvc_copts = xnnpack_msvc_std_copts(),
3818 visibility = xnnpack_visibility(),
3819 deps = [
3820 "@FP16",
3821 "@clog",
3822 "@pthreadpool",
3823 ],
3824)
3825
Marat Dukhan08c4a432019-10-03 09:29:21 -07003826xnnpack_aggregate_library(
3827 name = "ukernels",
3828 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829 ":neon_ukernels",
3830 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003831 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003832 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003833 ":asm_ukernels",
3834 ],
3835 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 ":neon_ukernels",
3837 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003838 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003839 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003840 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003841 ":asm_ukernels",
3842 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003843 generic_deps = [
3844 ":scalar_ukernels",
3845 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003846 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003847 ":psimd_fastmath_ukernels",
3848 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003849 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003850 wasm_deps = [
3851 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003852 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003853 ],
3854 wasmsimd_deps = [
3855 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003856 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003857 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003860 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003861 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003863 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003864 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003865 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003867 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868 ],
3869)
3870
Marat Dukhan33fcf782020-05-24 14:27:15 -07003871xnnpack_aggregate_library(
3872 name = "ukernels_test_mode",
3873 aarch32_deps = [
3874 ":neon_ukernels_test_mode",
3875 ":neonfma_ukernels_test_mode",
3876 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003877 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003878 ":asm_ukernels",
3879 ],
3880 aarch64_deps = [
3881 ":neon_ukernels_test_mode",
3882 ":neonfma_ukernels_test_mode",
3883 ":neonv8_ukernels_test_mode",
3884 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003885 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003886 ":asm_ukernels",
3887 ],
3888 generic_deps = [
3889 ":scalar_ukernels_test_mode",
3890 ],
3891 psimd_deps = [
3892 ":psimd_fastmath_ukernels_test_mode",
3893 ":psimd_accmath_ukernels_test_mode",
3894 ],
3895 wasm_deps = [
3896 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003897 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003898 ],
3899 wasmsimd_deps = [
3900 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003901 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003902 ],
3903 x86_deps = [
3904 ":sse2_ukernels_test_mode",
3905 ":ssse3_ukernels_test_mode",
3906 ":sse41_ukernels_test_mode",
3907 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003908 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003909 ":fma3_ukernels_test_mode",
3910 ":avx2_ukernels_test_mode",
3911 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003912 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003913 ],
3914)
3915
Marat Dukhan86df3842020-06-17 14:03:31 -07003916# This target must be compatible with Android NDK r17 and compile with Clang 6.
3917# It is similar to :ukernels target, but excludes microkernels which require
3918# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003919xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003920 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003921 aarch32_deps = [
3922 ":neon_ukernels",
3923 ":neonfma_ukernels",
3924 ":neonv8_ukernels",
3925 ":asm_ukernels",
3926 ],
3927 aarch64_deps = [
3928 ":neon_ukernels",
3929 ":neonfma_ukernels",
3930 ":neonv8_ukernels",
3931 ":asm_ukernels",
3932 ],
3933 generic_deps = [
3934 ":scalar_ukernels",
3935 ],
3936 psimd_deps = [
3937 ":psimd_fastmath_ukernels",
3938 ":psimd_accmath_ukernels",
3939 ],
3940 wasm_deps = [
3941 ":wasm_ukernels",
3942 ],
3943 wasmsimd_deps = [
3944 ":wasm_ukernels",
3945 ],
3946 x86_deps = [
3947 ":sse2_ukernels",
3948 ":ssse3_ukernels",
3949 ":sse41_ukernels",
3950 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003951 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003952 ":fma3_ukernels",
3953 ":avx2_ukernels",
3954 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003955 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003956 ],
3957)
3958
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959xnnpack_cc_library(
3960 name = "im2col",
3961 srcs = ["src/im2col.c"],
3962 hdrs = [
3963 "src/xnnpack/common.h",
3964 "src/xnnpack/im2col.h",
3965 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003966 gcc_copts = xnnpack_gcc_std_copts(),
3967 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003968)
3969
3970xnnpack_cc_library(
3971 name = "indirection",
3972 srcs = ["src/indirection.c"],
3973 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003974 gcc_copts = xnnpack_gcc_std_copts(),
3975 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003976 deps = [
3977 "@FP16",
3978 "@FXdiv",
3979 "@pthreadpool",
3980 ],
3981)
3982
3983xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003984 name = "indirection_test_mode",
3985 srcs = ["src/indirection.c"],
3986 hdrs = INTERNAL_HDRS,
3987 copts = [
3988 "-UNDEBUG",
3989 "-DXNN_TEST_MODE=1",
3990 ],
3991 gcc_copts = xnnpack_gcc_std_copts(),
3992 msvc_copts = xnnpack_msvc_std_copts(),
3993 deps = [
3994 "@FP16",
3995 "@FXdiv",
3996 "@pthreadpool",
3997 ],
3998)
3999
4000xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004001 name = "packing",
4002 srcs = ["src/packing.c"],
4003 hdrs = INTERNAL_HDRS,
4004 gcc_copts = xnnpack_gcc_std_copts(),
4005 msvc_copts = xnnpack_msvc_std_copts(),
4006 deps = [
4007 "@FP16",
4008 "@FXdiv",
4009 "@pthreadpool",
4010 ],
4011)
4012
4013xnnpack_cc_library(
4014 name = "packing_test_mode",
4015 srcs = ["src/packing.c"],
4016 hdrs = INTERNAL_HDRS,
4017 copts = [
4018 "-UNDEBUG",
4019 "-DXNN_TEST_MODE=1",
4020 ],
4021 gcc_copts = xnnpack_gcc_std_copts(),
4022 msvc_copts = xnnpack_msvc_std_copts(),
4023 deps = [
4024 "@FP16",
4025 "@FXdiv",
4026 "@pthreadpool",
4027 ],
4028)
4029
4030xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004031 name = "operator_run",
4032 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004033 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004034 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004035 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4036 "//conditions:default": [],
4037 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004038 gcc_copts = xnnpack_gcc_std_copts(),
4039 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004041 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004042 "@FP16",
4043 "@FXdiv",
4044 "@clog",
4045 "@pthreadpool",
4046 ],
4047)
4048
Chao Mei6ddfc602020-05-13 22:29:36 -07004049xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004050 name = "operator_run_test_mode",
4051 srcs = ["src/operator-run.c"],
4052 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4053 copts = LOGGING_COPTS + [
4054 "-UNDEBUG",
4055 "-DXNN_TEST_MODE=1",
4056 ] + select({
4057 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4058 "//conditions:default": [],
4059 }),
4060 gcc_copts = xnnpack_gcc_std_copts(),
4061 msvc_copts = xnnpack_msvc_std_copts(),
4062 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004063 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004064 "@FP16",
4065 "@FXdiv",
4066 "@clog",
4067 "@pthreadpool",
4068 ],
4069)
4070
4071xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004072 name = "memory_planner",
4073 srcs = ["src/memory-planner.c"],
4074 hdrs = INTERNAL_HDRS,
4075 defines = select({
4076 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4077 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4078 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4079 }),
4080 gcc_copts = xnnpack_gcc_std_copts(),
4081 msvc_copts = xnnpack_msvc_std_copts(),
4082 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004083 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004084 "@pthreadpool",
4085 ],
4086)
4087
Marat Dukhan33fcf782020-05-24 14:27:15 -07004088xnnpack_cc_library(
4089 name = "memory_planner_test_mode",
4090 srcs = ["src/memory-planner.c"],
4091 hdrs = INTERNAL_HDRS,
4092 copts = [
4093 "-UNDEBUG",
4094 "-DXNN_TEST_MODE=1",
4095 ],
4096 defines = select({
4097 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4098 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4099 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4100 }),
4101 gcc_copts = xnnpack_gcc_std_copts(),
4102 msvc_copts = xnnpack_msvc_std_copts(),
4103 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004104 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004105 "@pthreadpool",
4106 ],
4107)
4108
Marat Dukhan08c4a432019-10-03 09:29:21 -07004109cc_library(
4110 name = "enable_assembly",
4111 defines = select({
4112 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4113 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004114 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004115 }),
4116)
4117
Marat Dukhan9de90e02020-06-18 16:04:12 -07004118cc_library(
4119 name = "enable_sparse",
4120 defines = select({
4121 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4122 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004123 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004124 }),
4125)
4126
Marat Dukhancf056b22019-10-07 10:26:29 -07004127xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004128 name = "operators",
4129 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004130 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004132 ],
4133 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004134 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004135 "-Isrc",
4136 "-Iinclude",
4137 ] + select({
4138 ":debug_build": [],
4139 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004140 }) + select({
4141 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4142 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004143 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004144 gcc_copts = xnnpack_gcc_std_copts(),
4145 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004146 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004148 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004149 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004150 "@FP16",
4151 "@FXdiv",
4152 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004154 ],
4155)
4156
Marat Dukhan10a38082020-04-17 03:58:35 -07004157xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004158 name = "operators_test_mode",
4159 srcs = OPERATOR_SRCS + [
4160 "src/memory.c",
4161 "src/operator-delete.c",
4162 ],
4163 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4164 copts = LOGGING_COPTS + [
4165 "-Isrc",
4166 "-Iinclude",
4167 "-UNDEBUG",
4168 "-DXNN_TEST_MODE=1",
4169 ] + select({
4170 ":debug_build": [],
4171 "//conditions:default": xnnpack_min_size_copts(),
4172 }) + select({
4173 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4174 "//conditions:default": [],
4175 }),
4176 gcc_copts = xnnpack_gcc_std_copts(),
4177 msvc_copts = xnnpack_msvc_std_copts(),
4178 deps = [
4179 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004180 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004181 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004182 "@FP16",
4183 "@FXdiv",
4184 "@clog",
4185 "@pthreadpool",
4186 ],
4187)
4188
4189xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004190 name = "XNNPACK",
4191 srcs = [
4192 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004193 "src/runtime.c",
4194 "src/subgraph.c",
4195 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004196 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004197 hdrs = ["include/xnnpack.h"],
4198 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004199 "-Isrc",
4200 "-Iinclude",
4201 ] + select({
4202 ":debug_build": [],
4203 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004204 }) + select({
4205 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4206 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004207 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004208 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004209 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004210 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004211 visibility = xnnpack_visibility(),
4212 deps = [
4213 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004214 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004215 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004216 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004217 ":operator_run",
4218 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004219 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004220 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004221 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004222 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004223 ] + select({
4224 ":emscripten": [],
4225 "//conditions:default": ["@cpuinfo"],
4226 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227)
4228
Marat Dukhan10a38082020-04-17 03:58:35 -07004229xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004230 name = "XNNPACK_test_mode",
4231 srcs = [
4232 "src/init.c",
4233 "src/runtime.c",
4234 "src/subgraph.c",
4235 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004236 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004237 hdrs = ["include/xnnpack.h"],
4238 copts = LOGGING_COPTS + [
4239 "-Isrc",
4240 "-Iinclude",
4241 "-UNDEBUG",
4242 "-DXNN_TEST_MODE=1",
4243 ] + select({
4244 ":debug_build": [],
4245 "//conditions:default": xnnpack_min_size_copts(),
4246 }) + select({
4247 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4248 "//conditions:default": [],
4249 }),
4250 gcc_copts = xnnpack_gcc_std_copts(),
4251 includes = ["include"],
4252 msvc_copts = xnnpack_msvc_std_copts(),
4253 visibility = xnnpack_visibility(),
4254 deps = [
4255 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004256 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004257 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004258 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004259 ":operator_run_test_mode",
4260 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004261 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004262 "@clog",
4263 "@FP16",
4264 "@pthreadpool",
4265 ] + select({
4266 ":emscripten": [],
4267 "//conditions:default": ["@cpuinfo"],
4268 }),
4269)
4270
4271xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004272 name = "xnnpack_f32",
4273 srcs = [
4274 "src/init.c",
4275 "src/runtime.c",
4276 "src/subgraph.c",
4277 "src/tensor.c",
4278 ] + SUBGRAPH_SRCS,
4279 hdrs = ["include/xnnpack.h"],
4280 copts = LOGGING_COPTS + [
4281 "-Isrc",
4282 "-Iinclude",
4283 ] + select({
4284 ":debug_build": [],
4285 "//conditions:default": xnnpack_min_size_copts(),
4286 }) + select({
4287 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4288 "//conditions:default": [],
4289 }),
4290 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004291 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004292 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004293 "XNN_NO_U8_OPERATORS",
4294 "XNN_NO_X8_OPERATORS",
4295 "XNN_NO_F16_OPERATORS",
4296 "XNN_NO_X16_OPERATORS",
4297 ],
4298 gcc_copts = xnnpack_gcc_std_copts(),
4299 includes = ["include"],
4300 msvc_copts = xnnpack_msvc_std_copts(),
4301 visibility = xnnpack_visibility(),
4302 deps = [
4303 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004304 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004305 ":logging_utils",
4306 ":memory_planner",
4307 ":operator_run",
4308 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004309 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004310 "@clog",
4311 "@FP16",
4312 "@pthreadpool",
4313 ] + select({
4314 ":emscripten": [],
4315 "//conditions:default": ["@cpuinfo"],
4316 }),
4317)
4318
4319xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004320 name = "xnnpack_operators_nhwc_f32",
4321 srcs = [
4322 "src/init.c",
4323 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004324 hdrs = ["include/xnnpack.h"],
4325 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004326 "-Isrc",
4327 "-Iinclude",
4328 ] + select({
4329 ":debug_build": [],
4330 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004331 }) + select({
4332 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4333 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004334 }),
4335 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004336 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004337 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004338 "XNN_NO_U8_OPERATORS",
4339 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004340 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004341 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004342 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004344 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 visibility = xnnpack_visibility(),
4346 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004347 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004348 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349 ":operator_run",
4350 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004351 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004352 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004354 ] + select({
4355 ":emscripten": [],
4356 "//conditions:default": ["@cpuinfo"],
4357 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004358)
4359
Marat Dukhancf056b22019-10-07 10:26:29 -07004360xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361 name = "bench_utils",
4362 srcs = ["bench/utils.cc"],
4363 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004364 deps = [
4365 "@com_google_benchmark//:benchmark",
4366 "@cpuinfo",
4367 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368)
4369
Frank Barchard7e955972019-10-11 10:34:25 -07004370######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371
4372xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004373 name = "qs8_gemm_bench",
4374 srcs = [
4375 "bench/gemm.h",
4376 "bench/qs8-gemm.cc",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004379 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4380 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004381)
4382
4383xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004384 name = "qs8_requantization_bench",
4385 srcs = [
4386 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004387 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004388 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004389 ] + MICROKERNEL_BENCHMARK_HDRS,
4390 deps = MICROKERNEL_BENCHMARK_DEPS,
4391)
4392
4393xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004394 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395 srcs = [
4396 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004397 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 "src/xnnpack/AlignedAllocator.h",
4399 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004400 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004401 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402)
4403
4404xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004405 name = "qu8_requantization_bench",
4406 srcs = [
4407 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004408 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004409 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004410 ] + MICROKERNEL_BENCHMARK_HDRS,
4411 deps = MICROKERNEL_BENCHMARK_DEPS,
4412)
4413
4414xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004415 name = "f16_igemm_bench",
4416 srcs = [
4417 "bench/f16-igemm.cc",
4418 "bench/conv.h",
4419 "bench/google/conv.h",
4420 "src/xnnpack/AlignedAllocator.h",
4421 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004422 deps = MICROKERNEL_BENCHMARK_DEPS + [
4423 ":indirection",
4424 ":packing",
4425 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004426)
4427
4428xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429 name = "f16_gemm_bench",
4430 srcs = [
4431 "bench/f16-gemm.cc",
4432 "bench/gemm.h",
4433 "src/xnnpack/AlignedAllocator.h",
4434 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004435 deps = MICROKERNEL_BENCHMARK_DEPS + [
4436 ":packing",
4437 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438)
4439
4440xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004441 name = "f16_spmm_bench",
4442 srcs = [
4443 "bench/f16-spmm.cc",
4444 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004445 "src/xnnpack/AlignedAllocator.h",
4446 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004447 deps = MICROKERNEL_BENCHMARK_DEPS,
4448)
4449
4450xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451 name = "f32_igemm_bench",
4452 srcs = [
4453 "bench/f32-igemm.cc",
4454 "bench/conv.h",
4455 "src/xnnpack/AlignedAllocator.h",
4456 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004457 deps = MICROKERNEL_BENCHMARK_DEPS + [
4458 ":indirection",
4459 ":packing",
4460 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461)
4462
4463xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004464 name = "f16_relu_bench",
4465 srcs = [
4466 "bench/f16-relu.cc",
4467 "src/xnnpack/AlignedAllocator.h",
4468 ] + MICROKERNEL_BENCHMARK_HDRS,
4469 deps = MICROKERNEL_BENCHMARK_DEPS,
4470)
4471
4472xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 name = "f32_conv_hwc_bench",
4474 srcs = [
4475 "bench/f32-conv-hwc.cc",
4476 "bench/dconv.h",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004479 deps = MICROKERNEL_BENCHMARK_DEPS + [
4480 ":packing",
4481 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482)
4483
4484xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004485 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004486 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004487 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004488 "bench/dconv.h",
4489 "src/xnnpack/AlignedAllocator.h",
4490 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004491 deps = MICROKERNEL_BENCHMARK_DEPS + [
4492 ":packing",
4493 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004494)
4495
4496xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004497 name = "f16_dwconv_bench",
4498 srcs = [
4499 "bench/f16-dwconv.cc",
4500 "bench/dwconv.h",
4501 "bench/google/dwconv.h",
4502 "src/xnnpack/AlignedAllocator.h",
4503 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004504 deps = MICROKERNEL_BENCHMARK_DEPS + [
4505 ":indirection",
4506 ":packing",
4507 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004508)
4509
4510xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511 name = "f32_dwconv_bench",
4512 srcs = [
4513 "bench/f32-dwconv.cc",
4514 "bench/dwconv.h",
4515 "src/xnnpack/AlignedAllocator.h",
4516 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004517 deps = MICROKERNEL_BENCHMARK_DEPS + [
4518 ":indirection",
4519 ":packing",
4520 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521)
4522
4523xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004524 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004526 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004527 "bench/dwconv.h",
4528 "src/xnnpack/AlignedAllocator.h",
4529 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004530 deps = MICROKERNEL_BENCHMARK_DEPS + [
4531 ":indirection",
4532 ":packing",
4533 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534)
4535
4536xnnpack_benchmark(
4537 name = "f32_gemm_bench",
4538 srcs = [
4539 "bench/f32-gemm.cc",
4540 "bench/gemm.h",
4541 "src/xnnpack/AlignedAllocator.h",
4542 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004543 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004544 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545)
4546
4547xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004548 name = "f32_hswish_bench",
4549 srcs = [
4550 "bench/f32-hswish.cc",
4551 "src/xnnpack/AlignedAllocator.h",
4552 ] + MICROKERNEL_BENCHMARK_HDRS,
4553 deps = MICROKERNEL_BENCHMARK_DEPS,
4554)
4555
4556xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004557 name = "f32_raddexpminusmax_bench",
4558 srcs = [
4559 "bench/f32-raddexpminusmax.cc",
4560 "src/xnnpack/AlignedAllocator.h",
4561 ] + MICROKERNEL_BENCHMARK_HDRS,
4562 deps = MICROKERNEL_BENCHMARK_DEPS,
4563)
4564
4565xnnpack_benchmark(
4566 name = "f32_raddextexp_bench",
4567 srcs = [
4568 "bench/f32-raddextexp.cc",
4569 "src/xnnpack/AlignedAllocator.h",
4570 ] + MICROKERNEL_BENCHMARK_HDRS,
4571 deps = MICROKERNEL_BENCHMARK_DEPS,
4572)
4573
4574xnnpack_benchmark(
4575 name = "f32_raddstoreexpminusmax_bench",
4576 srcs = [
4577 "bench/f32-raddstoreexpminusmax.cc",
4578 "src/xnnpack/AlignedAllocator.h",
4579 ] + MICROKERNEL_BENCHMARK_HDRS,
4580 deps = MICROKERNEL_BENCHMARK_DEPS,
4581)
4582
4583xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004584 name = "f32_relu_bench",
4585 srcs = [
4586 "bench/f32-relu.cc",
4587 "src/xnnpack/AlignedAllocator.h",
4588 ] + MICROKERNEL_BENCHMARK_HDRS,
4589 deps = MICROKERNEL_BENCHMARK_DEPS,
4590)
4591
4592xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 name = "f32_rmax_bench",
4594 srcs = [
4595 "bench/f32-rmax.cc",
4596 "src/xnnpack/AlignedAllocator.h",
4597 ] + MICROKERNEL_BENCHMARK_HDRS,
4598 deps = MICROKERNEL_BENCHMARK_DEPS,
4599)
4600
4601xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004602 name = "f32_sigmoid_bench",
4603 srcs = [
4604 "bench/f32-sigmoid.cc",
4605 "src/xnnpack/AlignedAllocator.h",
4606 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004607 deps = MICROKERNEL_BENCHMARK_DEPS,
4608)
4609
4610xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004611 name = "f32_spmm_bench",
4612 srcs = [
4613 "bench/f32-spmm.cc",
4614 "bench/gemm.h",
4615 "src/xnnpack/AlignedAllocator.h",
4616 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004617 deps = MICROKERNEL_BENCHMARK_DEPS,
4618)
4619
4620xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004621 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004622 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004623 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004624 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004625 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004626 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004627)
4628
4629xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004630 name = "f32_vscaleexpminusmax_bench",
4631 srcs = [
4632 "bench/f32-vscaleexpminusmax.cc",
4633 "src/xnnpack/AlignedAllocator.h",
4634 ] + MICROKERNEL_BENCHMARK_HDRS,
4635 deps = MICROKERNEL_BENCHMARK_DEPS,
4636)
4637
4638xnnpack_benchmark(
4639 name = "f32_vscaleextexp_bench",
4640 srcs = [
4641 "bench/f32-vscaleextexp.cc",
4642 "src/xnnpack/AlignedAllocator.h",
4643 ] + MICROKERNEL_BENCHMARK_HDRS,
4644 deps = MICROKERNEL_BENCHMARK_DEPS,
4645)
4646
4647xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004648 name = "f32_vsqrt_bench",
4649 srcs = [
4650 "bench/f32-vsqrt.cc",
4651 "src/xnnpack/AlignedAllocator.h",
4652 ] + MICROKERNEL_BENCHMARK_HDRS,
4653 deps = MICROKERNEL_BENCHMARK_DEPS,
4654)
4655
4656xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657 name = "f32_im2col_gemm_bench",
4658 srcs = [
4659 "bench/f32-im2col-gemm.cc",
4660 "bench/conv.h",
4661 "src/xnnpack/AlignedAllocator.h",
4662 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004663 deps = MICROKERNEL_BENCHMARK_DEPS + [
4664 ":im2col",
4665 ":packing",
4666 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667)
4668
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004669xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004670 name = "rounding_bench",
4671 srcs = [
4672 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004673 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004674 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004675 ] + MICROKERNEL_BENCHMARK_HDRS,
4676 deps = MICROKERNEL_BENCHMARK_DEPS,
4677)
4678
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679########################### Benchmarks for operators ###########################
4680
4681xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682 name = "average_pooling_bench",
4683 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004684 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004685 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004686 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004687)
4688
4689xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004690 name = "bankers_rounding_bench",
4691 srcs = ["bench/bankers-rounding.cc"],
4692 copts = xnnpack_optional_tflite_copts(),
4693 tags = ["nowin32"],
4694 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4695)
4696
4697xnnpack_benchmark(
4698 name = "ceiling_bench",
4699 srcs = ["bench/ceiling.cc"],
4700 copts = xnnpack_optional_tflite_copts(),
4701 tags = ["nowin32"],
4702 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4703)
4704
4705xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004706 name = "channel_shuffle_bench",
4707 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004708 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004709)
4710
4711xnnpack_benchmark(
4712 name = "convolution_bench",
4713 srcs = ["bench/convolution.cc"],
4714 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004715 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004716 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717)
4718
4719xnnpack_benchmark(
4720 name = "deconvolution_bench",
4721 srcs = ["bench/deconvolution.cc"],
4722 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004723 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004724 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725)
4726
4727xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004728 name = "floor_bench",
4729 srcs = ["bench/floor.cc"],
4730 copts = xnnpack_optional_tflite_copts(),
4731 tags = ["nowin32"],
4732 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4733)
4734
4735xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004736 name = "global_average_pooling_bench",
4737 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004738 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004739)
4740
4741xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004742 name = "hardswish_bench",
4743 srcs = ["bench/hardswish.cc"],
4744 copts = xnnpack_optional_tflite_copts(),
4745 tags = ["nowin32"],
4746 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4747)
4748
4749xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004750 name = "max_pooling_bench",
4751 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004752 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753)
4754
4755xnnpack_benchmark(
4756 name = "sigmoid_bench",
4757 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004758 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004759 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004760 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004761)
4762
4763xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004764 name = "prelu_bench",
4765 srcs = ["bench/prelu.cc"],
4766 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004767 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004768 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004769)
4770
4771xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004772 name = "softmax_bench",
4773 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004774 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004775 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004776 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777)
4778
Marat Dukhan87727142020-06-24 15:24:10 -07004779xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004780 name = "square_root_bench",
4781 srcs = ["bench/square-root.cc"],
4782 copts = xnnpack_optional_tflite_copts(),
4783 tags = ["nowin32"],
4784 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4785)
4786
4787xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004788 name = "truncation_bench",
4789 srcs = ["bench/truncation.cc"],
4790 deps = OPERATOR_BENCHMARK_DEPS,
4791)
4792
Marat Dukhanc068bb62019-10-04 13:24:39 -07004793############################# End-to-end benchmarks ############################
4794
4795cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004796 name = "fp32_mobilenet_v1",
4797 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004798 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004799 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004800 linkstatic = True,
4801 deps = [
4802 ":XNNPACK",
4803 "@pthreadpool",
4804 ],
4805)
4806
4807cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004808 name = "fp16_mobilenet_v1",
4809 srcs = ["models/fp16-mobilenet-v1.cc"],
4810 hdrs = ["models/models.h"],
4811 copts = xnnpack_std_cxxopts(),
4812 linkstatic = True,
4813 deps = [
4814 ":XNNPACK",
4815 "@FP16",
4816 "@pthreadpool",
4817 ],
4818)
4819
4820cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004821 name = "qs8_mobilenet_v1",
4822 srcs = ["models/qs8-mobilenet-v1.cc"],
4823 hdrs = ["models/models.h"],
4824 copts = xnnpack_std_cxxopts(),
4825 linkstatic = True,
4826 deps = [
4827 ":XNNPACK",
4828 "@pthreadpool",
4829 ],
4830)
4831
4832cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004833 name = "qs8_mobilenet_v2",
4834 srcs = ["models/qs8-mobilenet-v2.cc"],
4835 hdrs = ["models/models.h"],
4836 copts = xnnpack_std_cxxopts(),
4837 linkstatic = True,
4838 deps = [
4839 ":XNNPACK",
4840 "@pthreadpool",
4841 ],
4842)
4843
4844cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004845 name = "fp32_mobilenet_v2",
4846 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004847 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004848 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004849 linkstatic = True,
4850 deps = [
4851 ":XNNPACK",
4852 "@pthreadpool",
4853 ],
4854)
4855
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004856cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004857 name = "fp16_mobilenet_v2",
4858 srcs = ["models/fp16-mobilenet-v2.cc"],
4859 hdrs = ["models/models.h"],
4860 copts = xnnpack_std_cxxopts(),
4861 linkstatic = True,
4862 deps = [
4863 ":XNNPACK",
4864 "@FP16",
4865 "@pthreadpool",
4866 ],
4867)
4868
4869cc_library(
4870 name = "fp32_mobilenet_v3_large",
4871 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004872 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004873 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004874 linkstatic = True,
4875 deps = [
4876 ":XNNPACK",
4877 "@pthreadpool",
4878 ],
4879)
4880
4881cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004882 name = "fp16_mobilenet_v3_large",
4883 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4884 hdrs = ["models/models.h"],
4885 copts = xnnpack_std_cxxopts(),
4886 linkstatic = True,
4887 deps = [
4888 ":XNNPACK",
4889 "@FP16",
4890 "@pthreadpool",
4891 ],
4892)
4893
4894cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004895 name = "fp32_mobilenet_v3_small",
4896 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004897 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004898 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004899 linkstatic = True,
4900 deps = [
4901 ":XNNPACK",
4902 "@pthreadpool",
4903 ],
4904)
4905
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004906cc_library(
4907 name = "fp16_mobilenet_v3_small",
4908 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4909 hdrs = ["models/models.h"],
4910 copts = xnnpack_std_cxxopts(),
4911 linkstatic = True,
4912 deps = [
4913 ":XNNPACK",
4914 "@FP16",
4915 "@pthreadpool",
4916 ],
4917)
4918
Marat Dukhanc068bb62019-10-04 13:24:39 -07004919xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004920 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004921 srcs = [
4922 "bench/f32-dwconv-e2e.cc",
4923 "bench/end2end.h",
4924 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004925 deps = MICROKERNEL_BENCHMARK_DEPS + [
4926 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004927 ":fp32_mobilenet_v1",
4928 ":fp32_mobilenet_v2",
4929 ":fp32_mobilenet_v3_large",
4930 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004931 ],
4932)
4933
4934xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004935 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004936 srcs = [
4937 "bench/f32-gemm-e2e.cc",
4938 "bench/end2end.h",
4939 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004940 deps = MICROKERNEL_BENCHMARK_DEPS + [
4941 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004942 ":fp32_mobilenet_v1",
4943 ":fp32_mobilenet_v2",
4944 ":fp32_mobilenet_v3_large",
4945 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004946 ],
4947)
4948
4949xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004950 name = "end2end_bench",
4951 srcs = ["bench/end2end.cc"],
4952 deps = [
4953 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004954 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004955 ":fp16_mobilenet_v1",
4956 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004957 ":fp16_mobilenet_v3_large",
4958 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004959 ":fp32_mobilenet_v1",
4960 ":fp32_mobilenet_v2",
4961 ":fp32_mobilenet_v3_large",
4962 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004963 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004964 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004965 "@pthreadpool",
4966 ],
4967)
4968
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004969#################### Accuracy evaluation for math functions ####################
4970
4971xnnpack_benchmark(
4972 name = "f32_exp_eval",
4973 srcs = [
4974 "eval/f32-exp.cc",
4975 "src/xnnpack/AlignedAllocator.h",
4976 ] + ACCURACY_EVAL_HDRS,
4977 deps = ACCURACY_EVAL_DEPS,
4978)
4979
Marat Dukhan515c9772019-10-17 18:07:57 -07004980xnnpack_benchmark(
4981 name = "f32_expminus_eval",
4982 srcs = [
4983 "eval/f32-expminus.cc",
4984 "src/xnnpack/AlignedAllocator.h",
4985 ] + ACCURACY_EVAL_HDRS,
4986 deps = ACCURACY_EVAL_DEPS,
4987)
4988
Marat Dukhan98ba4412019-10-23 02:14:28 -07004989xnnpack_benchmark(
4990 name = "f32_extexp_eval",
4991 srcs = [
4992 "eval/f32-extexp.cc",
4993 "src/xnnpack/AlignedAllocator.h",
4994 ] + ACCURACY_EVAL_HDRS,
4995 deps = ACCURACY_EVAL_DEPS,
4996)
4997
Marat Dukhan8853b822020-05-07 12:19:01 -07004998xnnpack_unit_test(
4999 name = "f32_roundne_eval",
5000 srcs = [
5001 "eval/f32-roundne.cc",
5002 "src/xnnpack/AlignedAllocator.h",
5003 "src/xnnpack/math-stubs.h",
5004 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005005 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005006 deps = MICROKERNEL_TEST_DEPS,
5007)
5008
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005009xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005010 name = "f32_roundd_eval",
5011 srcs = [
5012 "eval/f32-roundd.cc",
5013 "src/xnnpack/AlignedAllocator.h",
5014 "src/xnnpack/math-stubs.h",
5015 ] + MICROKERNEL_TEST_HDRS,
5016 automatic = False,
5017 deps = MICROKERNEL_TEST_DEPS,
5018)
5019
5020xnnpack_unit_test(
5021 name = "f32_roundu_eval",
5022 srcs = [
5023 "eval/f32-roundu.cc",
5024 "src/xnnpack/AlignedAllocator.h",
5025 "src/xnnpack/math-stubs.h",
5026 ] + MICROKERNEL_TEST_HDRS,
5027 automatic = False,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005032 name = "f32_roundz_eval",
5033 srcs = [
5034 "eval/f32-roundz.cc",
5035 "src/xnnpack/AlignedAllocator.h",
5036 "src/xnnpack/math-stubs.h",
5037 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005038 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005039 deps = MICROKERNEL_TEST_DEPS,
5040)
5041
Marat Dukhan346a9e52019-11-15 09:06:30 -08005042xnnpack_benchmark(
5043 name = "f32_sigmoid_eval",
5044 srcs = [
5045 "eval/f32-sigmoid.cc",
5046 "src/xnnpack/AlignedAllocator.h",
5047 ] + ACCURACY_EVAL_HDRS,
5048 deps = ACCURACY_EVAL_DEPS,
5049)
5050
Marat Dukhan84000762020-06-29 18:38:43 -07005051xnnpack_benchmark(
5052 name = "f32_sqrt_eval",
5053 srcs = [
5054 "eval/f32-sqrt.cc",
5055 "src/xnnpack/AlignedAllocator.h",
5056 ] + ACCURACY_EVAL_HDRS,
5057 deps = ACCURACY_EVAL_DEPS + [
5058 ":bench_utils",
5059 ],
5060)
5061
Marat Dukhan08c4a432019-10-03 09:29:21 -07005062######################### Unit tests for micro-kernels #########################
5063
5064xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005065 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005066 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005067 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005068 "test/gemm-microkernel-tester.h",
5069 "src/xnnpack/AlignedAllocator.h",
5070 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005071 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072)
5073
5074xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005075 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005076 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005077 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005078 "test/spmm-microkernel-tester.h",
5079 "src/xnnpack/AlignedAllocator.h",
5080 ] + MICROKERNEL_TEST_HDRS,
5081 deps = MICROKERNEL_TEST_DEPS,
5082)
5083
5084xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 name = "f32_argmaxpool_test",
5086 srcs = [
5087 "test/f32-argmaxpool.cc",
5088 "test/argmaxpool-microkernel-tester.h",
5089 "src/xnnpack/AlignedAllocator.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005095 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005096 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005097 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005098 "test/avgpool-microkernel-tester.h",
5099 "src/xnnpack/AlignedAllocator.h",
5100 ] + MICROKERNEL_TEST_HDRS,
5101 deps = MICROKERNEL_TEST_DEPS,
5102)
5103
5104xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005105 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005106 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005107 "test/f32-ibilinear.cc",
5108 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005109 "src/xnnpack/AlignedAllocator.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005115 name = "f32_ibilinear_chw_test",
5116 srcs = [
5117 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005118 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005119 "src/xnnpack/AlignedAllocator.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 deps = MICROKERNEL_TEST_DEPS,
5122)
5123
5124xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005125 name = "f16_clamp_test",
5126 srcs = [
5127 "test/f16-clamp.cc",
5128 "test/clamp-microkernel-tester.h",
5129 ] + MICROKERNEL_TEST_HDRS,
5130 deps = MICROKERNEL_TEST_DEPS,
5131)
5132
5133xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 name = "f32_clamp_test",
5135 srcs = [
5136 "test/f32-clamp.cc",
5137 "test/clamp-microkernel-tester.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005143 name = "f32_igemm_test",
5144 srcs = [
5145 "test/f32-igemm.cc",
5146 "test/gemm-microkernel-tester.h",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005149 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005150)
5151
5152xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005153 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005155 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005156 "test/gemm-microkernel-tester.h",
5157 "src/xnnpack/AlignedAllocator.h",
5158 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005159 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005160)
5161
5162xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005163 name = "f32_igemm_minmax_test",
5164 srcs = [
5165 "test/f32-igemm-minmax.cc",
5166 "test/gemm-microkernel-tester.h",
5167 "src/xnnpack/AlignedAllocator.h",
5168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005169 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005170)
5171
5172xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005173 name = "f16_igemm_minmax_test",
5174 srcs = [
5175 "test/f16-igemm-minmax.cc",
5176 "test/gemm-microkernel-tester.h",
5177 "src/xnnpack/AlignedAllocator.h",
5178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005179 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005180)
5181
5182xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 name = "f32_conv_hwc_test",
5184 srcs = [
5185 "test/f32-conv-hwc.cc",
5186 "test/conv-hwc-microkernel-tester.h",
5187 "src/xnnpack/AlignedAllocator.h",
5188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005189 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190)
5191
5192xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005193 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005195 "test/f32-conv-hwc2chw.cc",
5196 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 "src/xnnpack/AlignedAllocator.h",
5198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005199 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200)
5201
5202xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005203 name = "f32_dwconv_test",
5204 srcs = [
5205 "test/f32-dwconv.cc",
5206 "test/dwconv-microkernel-tester.h",
5207 "src/xnnpack/AlignedAllocator.h",
5208 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005209 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005210)
5211
5212xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005213 name = "f16_dwconv_minmax_test",
5214 srcs = [
5215 "test/f16-dwconv-minmax.cc",
5216 "test/dwconv-microkernel-tester.h",
5217 "src/xnnpack/AlignedAllocator.h",
5218 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005219 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005220)
5221
5222xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005223 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005225 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226 "test/dwconv-microkernel-tester.h",
5227 "src/xnnpack/AlignedAllocator.h",
5228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005229 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230)
5231
5232xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005233 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005234 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005235 "test/f32-dwconv2d-chw.cc",
5236 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 "src/xnnpack/AlignedAllocator.h",
5238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005239 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240)
5241
5242xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005243 name = "f16_gavgpool_minmax_test",
5244 srcs = [
5245 "test/f16-gavgpool-minmax.cc",
5246 "test/gavgpool-microkernel-tester.h",
5247 "src/xnnpack/AlignedAllocator.h",
5248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005253 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005255 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256 "test/gavgpool-microkernel-tester.h",
5257 "src/xnnpack/AlignedAllocator.h",
5258 ] + MICROKERNEL_TEST_HDRS,
5259 deps = MICROKERNEL_TEST_DEPS,
5260)
5261
5262xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005263 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005265 "test/f32-gavgpool-cw.cc",
5266 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 "src/xnnpack/AlignedAllocator.h",
5268 ] + MICROKERNEL_TEST_HDRS,
5269 deps = MICROKERNEL_TEST_DEPS,
5270)
5271
5272xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005273 name = "f32_gemm_test",
5274 srcs = [
5275 "test/f32-gemm.cc",
5276 "test/gemm-microkernel-tester.h",
5277 "src/xnnpack/AlignedAllocator.h",
5278 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005279 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005280)
5281
5282xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005283 name = "f32_gemm_relu_test",
5284 srcs = [
5285 "test/f32-gemm-relu.cc",
5286 "test/gemm-microkernel-tester.h",
5287 "src/xnnpack/AlignedAllocator.h",
5288 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005289 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005290)
5291
5292xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005293 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005295 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "test/gemm-microkernel-tester.h",
5297 "src/xnnpack/AlignedAllocator.h",
5298 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005299 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300)
5301
5302xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005303 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005304 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005305 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306 "test/gemm-microkernel-tester.h",
5307 "src/xnnpack/AlignedAllocator.h",
5308 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005309 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310)
5311
5312xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005313 name = "f16_hswish_test",
5314 srcs = [
5315 "test/f16-hswish.cc",
5316 "test/hswish-microkernel-tester.h",
5317 ] + MICROKERNEL_TEST_HDRS,
5318 deps = MICROKERNEL_TEST_DEPS,
5319)
5320
5321xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 name = "f32_hswish_test",
5323 srcs = [
5324 "test/f32-hswish.cc",
5325 "test/hswish-microkernel-tester.h",
5326 ] + MICROKERNEL_TEST_HDRS,
5327 deps = MICROKERNEL_TEST_DEPS,
5328)
5329
5330xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005331 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005333 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 "test/maxpool-microkernel-tester.h",
5335 ] + MICROKERNEL_TEST_HDRS,
5336 deps = MICROKERNEL_TEST_DEPS,
5337)
5338
5339xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005340 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005342 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 "test/avgpool-microkernel-tester.h",
5344 "src/xnnpack/AlignedAllocator.h",
5345 ] + MICROKERNEL_TEST_HDRS,
5346 deps = MICROKERNEL_TEST_DEPS,
5347)
5348
5349xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005350 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005351 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005352 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353 "test/gemm-microkernel-tester.h",
5354 "src/xnnpack/AlignedAllocator.h",
5355 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005356 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357)
5358
5359xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005360 name = "f16_prelu_test",
5361 srcs = [
5362 "test/f16-prelu.cc",
5363 "test/prelu-microkernel-tester.h",
5364 "src/xnnpack/AlignedAllocator.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005370 name = "f32_prelu_test",
5371 srcs = [
5372 "test/f32-prelu.cc",
5373 "test/prelu-microkernel-tester.h",
5374 "src/xnnpack/AlignedAllocator.h",
5375 ] + MICROKERNEL_TEST_HDRS,
5376 deps = MICROKERNEL_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005380 name = "f32_raddexpminusmax_test",
5381 srcs = [
5382 "test/f32-raddexpminusmax.cc",
5383 "test/raddexpminusmax-microkernel-tester.h",
5384 ] + MICROKERNEL_TEST_HDRS,
5385 deps = MICROKERNEL_TEST_DEPS,
5386)
5387
5388xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005389 name = "f32_raddextexp_test",
5390 srcs = [
5391 "test/f32-raddextexp.cc",
5392 "test/raddextexp-microkernel-tester.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005398 name = "f32_raddstoreexpminusmax_test",
5399 srcs = [
5400 "test/f32-raddstoreexpminusmax.cc",
5401 "test/raddstoreexpminusmax-microkernel-tester.h",
5402 ] + MICROKERNEL_TEST_HDRS,
5403 deps = MICROKERNEL_TEST_DEPS,
5404)
5405
5406xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005407 name = "f32_relu_test",
5408 srcs = [
5409 "test/f32-relu.cc",
5410 "test/vunary-microkernel-tester.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416 name = "f32_rmax_test",
5417 srcs = [
5418 "test/f32-rmax.cc",
5419 "test/rmax-microkernel-tester.h",
5420 ] + MICROKERNEL_TEST_HDRS,
5421 deps = MICROKERNEL_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005425 name = "f32_sigmoid_test",
5426 srcs = [
5427 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005428 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005429 ] + MICROKERNEL_TEST_HDRS,
5430 deps = MICROKERNEL_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005434 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005436 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005437 "test/spmm-microkernel-tester.h",
5438 "src/xnnpack/AlignedAllocator.h",
5439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005444 name = "f16_vadd_minmax_test",
5445 srcs = [
5446 "test/f16-vadd-minmax.cc",
5447 "test/vbinary-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
5453 name = "f16_vaddc_minmax_test",
5454 srcs = [
5455 "test/f16-vaddc-minmax.cc",
5456 "test/vbinaryc-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
5462 name = "f16_vdiv_minmax_test",
5463 srcs = [
5464 "test/f16-vdiv-minmax.cc",
5465 "test/vbinary-microkernel-tester.h",
5466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
5471 name = "f16_vdivc_minmax_test",
5472 srcs = [
5473 "test/f16-vdivc-minmax.cc",
5474 "test/vbinaryc-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
5480 name = "f16_vrdivc_minmax_test",
5481 srcs = [
5482 "test/f16-vrdivc-minmax.cc",
5483 "test/vbinaryc-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
5489 name = "f16_vmax_test",
5490 srcs = [
5491 "test/f16-vmax.cc",
5492 "test/vbinary-microkernel-tester.h",
5493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
5498 name = "f16_vmaxc_test",
5499 srcs = [
5500 "test/f16-vmaxc.cc",
5501 "test/vbinaryc-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
5507 name = "f16_vmin_test",
5508 srcs = [
5509 "test/f16-vmin.cc",
5510 "test/vbinary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
5516 name = "f16_vminc_test",
5517 srcs = [
5518 "test/f16-vminc.cc",
5519 "test/vbinaryc-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
5525 name = "f16_vmul_minmax_test",
5526 srcs = [
5527 "test/f16-vmul-minmax.cc",
5528 "test/vbinary-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
5534 name = "f16_vmulc_minmax_test",
5535 srcs = [
5536 "test/f16-vmulc-minmax.cc",
5537 "test/vbinaryc-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
5543 name = "f16_vsub_minmax_test",
5544 srcs = [
5545 "test/f16-vsub-minmax.cc",
5546 "test/vbinary-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
5552 name = "f16_vsubc_minmax_test",
5553 srcs = [
5554 "test/f16-vsubc-minmax.cc",
5555 "test/vbinaryc-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
5561 name = "f16_vrsubc_minmax_test",
5562 srcs = [
5563 "test/f16-vrsubc-minmax.cc",
5564 "test/vbinaryc-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005570 name = "f32_vabs_test",
5571 srcs = [
5572 "test/f32-vabs.cc",
5573 "test/vunary-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005579 name = "f32_vadd_test",
5580 srcs = [
5581 "test/f32-vadd.cc",
5582 "test/vbinary-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005588 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005589 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005590 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005591 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005597 name = "f32_vadd_relu_test",
5598 srcs = [
5599 "test/f32-vadd-relu.cc",
5600 "test/vbinary-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005606 name = "f32_vaddc_test",
5607 srcs = [
5608 "test/f32-vaddc.cc",
5609 "test/vbinaryc-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005615 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005616 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005617 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005618 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005624 name = "f32_vaddc_relu_test",
5625 srcs = [
5626 "test/f32-vaddc-relu.cc",
5627 "test/vbinaryc-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005633 name = "f32_vdiv_test",
5634 srcs = [
5635 "test/f32-vdiv.cc",
5636 "test/vbinary-microkernel-tester.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005642 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005643 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005644 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005645 "test/vbinary-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005651 name = "f32_vdiv_relu_test",
5652 srcs = [
5653 "test/f32-vdiv-relu.cc",
5654 "test/vbinary-microkernel-tester.h",
5655 ] + MICROKERNEL_TEST_HDRS,
5656 deps = MICROKERNEL_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005660 name = "f32_vdivc_test",
5661 srcs = [
5662 "test/f32-vdivc.cc",
5663 "test/vbinaryc-microkernel-tester.h",
5664 ] + MICROKERNEL_TEST_HDRS,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005669 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005670 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005671 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005672 "test/vbinaryc-microkernel-tester.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005678 name = "f32_vdivc_relu_test",
5679 srcs = [
5680 "test/f32-vdivc-relu.cc",
5681 "test/vbinaryc-microkernel-tester.h",
5682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005687 name = "f32_vrdivc_test",
5688 srcs = [
5689 "test/f32-vrdivc.cc",
5690 "test/vbinaryc-microkernel-tester.h",
5691 ] + MICROKERNEL_TEST_HDRS,
5692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
5695xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005696 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005697 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005698 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005699 "test/vbinaryc-microkernel-tester.h",
5700 ] + MICROKERNEL_TEST_HDRS,
5701 deps = MICROKERNEL_TEST_DEPS,
5702)
5703
5704xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005705 name = "f32_vrdivc_relu_test",
5706 srcs = [
5707 "test/f32-vrdivc-relu.cc",
5708 "test/vbinaryc-microkernel-tester.h",
5709 ] + MICROKERNEL_TEST_HDRS,
5710 deps = MICROKERNEL_TEST_DEPS,
5711)
5712
5713xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005714 name = "f32_vmax_test",
5715 srcs = [
5716 "test/f32-vmax.cc",
5717 "test/vbinary-microkernel-tester.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 deps = MICROKERNEL_TEST_DEPS,
5720)
5721
5722xnnpack_unit_test(
5723 name = "f32_vmaxc_test",
5724 srcs = [
5725 "test/f32-vmaxc.cc",
5726 "test/vbinaryc-microkernel-tester.h",
5727 ] + MICROKERNEL_TEST_HDRS,
5728 deps = MICROKERNEL_TEST_DEPS,
5729)
5730
5731xnnpack_unit_test(
5732 name = "f32_vmin_test",
5733 srcs = [
5734 "test/f32-vmin.cc",
5735 "test/vbinary-microkernel-tester.h",
5736 ] + MICROKERNEL_TEST_HDRS,
5737 deps = MICROKERNEL_TEST_DEPS,
5738)
5739
5740xnnpack_unit_test(
5741 name = "f32_vminc_test",
5742 srcs = [
5743 "test/f32-vminc.cc",
5744 "test/vbinaryc-microkernel-tester.h",
5745 ] + MICROKERNEL_TEST_HDRS,
5746 deps = MICROKERNEL_TEST_DEPS,
5747)
5748
5749xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005750 name = "f32_vmul_test",
5751 srcs = [
5752 "test/f32-vmul.cc",
5753 "test/vbinary-microkernel-tester.h",
5754 ] + MICROKERNEL_TEST_HDRS,
5755 deps = MICROKERNEL_TEST_DEPS,
5756)
5757
5758xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005759 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005760 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005761 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005762 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005763 ] + MICROKERNEL_TEST_HDRS,
5764 deps = MICROKERNEL_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005768 name = "f32_vmul_relu_test",
5769 srcs = [
5770 "test/f32-vmul-relu.cc",
5771 "test/vbinary-microkernel-tester.h",
5772 ] + MICROKERNEL_TEST_HDRS,
5773 deps = MICROKERNEL_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005777 name = "f32_vmulc_test",
5778 srcs = [
5779 "test/f32-vmulc.cc",
5780 "test/vbinaryc-microkernel-tester.h",
5781 ] + MICROKERNEL_TEST_HDRS,
5782 deps = MICROKERNEL_TEST_DEPS,
5783)
5784
5785xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005786 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005787 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005788 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005789 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005790 ] + MICROKERNEL_TEST_HDRS,
5791 deps = MICROKERNEL_TEST_DEPS,
5792)
5793
5794xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005795 name = "f32_vmulc_relu_test",
5796 srcs = [
5797 "test/f32-vmulc-relu.cc",
5798 "test/vbinaryc-microkernel-tester.h",
5799 ] + MICROKERNEL_TEST_HDRS,
5800 deps = MICROKERNEL_TEST_DEPS,
5801)
5802
5803xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005804 name = "f16_vmulcaddc_minmax_test",
5805 srcs = [
5806 "test/f16-vmulcaddc-minmax.cc",
5807 "test/vmulcaddc-microkernel-tester.h",
5808 "src/xnnpack/AlignedAllocator.h",
5809 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005810 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005811)
5812
5813xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005814 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005816 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005817 "test/vmulcaddc-microkernel-tester.h",
5818 "src/xnnpack/AlignedAllocator.h",
5819 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005820 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005821)
5822
5823xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005824 name = "f32_vlrelu_test",
5825 srcs = [
5826 "test/f32-vlrelu.cc",
5827 "test/vunary-microkernel-tester.h",
5828 ] + MICROKERNEL_TEST_HDRS,
5829 deps = MICROKERNEL_TEST_DEPS,
5830)
5831
5832xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005833 name = "f32_vneg_test",
5834 srcs = [
5835 "test/f32-vneg.cc",
5836 "test/vunary-microkernel-tester.h",
5837 ] + MICROKERNEL_TEST_HDRS,
5838 deps = MICROKERNEL_TEST_DEPS,
5839)
5840
5841xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005842 name = "f32_vrndne_test",
5843 srcs = [
5844 "test/f32-vrndne.cc",
5845 "test/vunary-microkernel-tester.h",
5846 ] + MICROKERNEL_TEST_HDRS,
5847 deps = MICROKERNEL_TEST_DEPS,
5848)
5849
5850xnnpack_unit_test(
5851 name = "f32_vrndz_test",
5852 srcs = [
5853 "test/f32-vrndz.cc",
5854 "test/vunary-microkernel-tester.h",
5855 ] + MICROKERNEL_TEST_HDRS,
5856 deps = MICROKERNEL_TEST_DEPS,
5857)
5858
5859xnnpack_unit_test(
5860 name = "f32_vrndu_test",
5861 srcs = [
5862 "test/f32-vrndu.cc",
5863 "test/vunary-microkernel-tester.h",
5864 ] + MICROKERNEL_TEST_HDRS,
5865 deps = MICROKERNEL_TEST_DEPS,
5866)
5867
5868xnnpack_unit_test(
5869 name = "f32_vrndd_test",
5870 srcs = [
5871 "test/f32-vrndd.cc",
5872 "test/vunary-microkernel-tester.h",
5873 ] + MICROKERNEL_TEST_HDRS,
5874 deps = MICROKERNEL_TEST_DEPS,
5875)
5876
5877xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005878 name = "f32_vscale_test",
5879 srcs = [
5880 "test/f32-vscale.cc",
5881 "test/vscale-microkernel-tester.h",
5882 ] + MICROKERNEL_TEST_HDRS,
5883 deps = MICROKERNEL_TEST_DEPS,
5884)
5885
5886xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005887 name = "f32_vscaleexpminusmax_test",
5888 srcs = [
5889 "test/f32-vscaleexpminusmax.cc",
5890 "test/vscaleexpminusmax-microkernel-tester.h",
5891 ] + MICROKERNEL_TEST_HDRS,
5892 deps = MICROKERNEL_TEST_DEPS,
5893)
5894
5895xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005896 name = "f32_vscaleextexp_test",
5897 srcs = [
5898 "test/f32-vscaleextexp.cc",
5899 "test/vscaleextexp-microkernel-tester.h",
5900 ] + MICROKERNEL_TEST_HDRS,
5901 deps = MICROKERNEL_TEST_DEPS,
5902)
5903
5904xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005905 name = "f32_vsqr_test",
5906 srcs = [
5907 "test/f32-vsqr.cc",
5908 "test/vunary-microkernel-tester.h",
5909 ] + MICROKERNEL_TEST_HDRS,
5910 deps = MICROKERNEL_TEST_DEPS,
5911)
5912
5913xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005914 name = "f32_vsqrdiff_test",
5915 srcs = [
5916 "test/f32-vsqrdiff.cc",
5917 "test/vbinary-microkernel-tester.h",
5918 ] + MICROKERNEL_TEST_HDRS,
5919 deps = MICROKERNEL_TEST_DEPS,
5920)
5921
5922xnnpack_unit_test(
5923 name = "f32_vsqrdiffc_test",
5924 srcs = [
5925 "test/f32-vsqrdiffc.cc",
5926 "test/vbinaryc-microkernel-tester.h",
5927 ] + MICROKERNEL_TEST_HDRS,
5928 deps = MICROKERNEL_TEST_DEPS,
5929)
5930
5931xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005932 name = "f32_vsqrt_test",
5933 srcs = [
5934 "test/f32-vsqrt.cc",
5935 "test/vunary-microkernel-tester.h",
5936 ] + MICROKERNEL_TEST_HDRS,
5937 deps = MICROKERNEL_TEST_DEPS,
5938)
5939
5940xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005941 name = "f32_vsub_test",
5942 srcs = [
5943 "test/f32-vsub.cc",
5944 "test/vbinary-microkernel-tester.h",
5945 ] + MICROKERNEL_TEST_HDRS,
5946 deps = MICROKERNEL_TEST_DEPS,
5947)
5948
5949xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005950 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005951 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005952 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005953 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005954 ] + MICROKERNEL_TEST_HDRS,
5955 deps = MICROKERNEL_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005959 name = "f32_vsub_relu_test",
5960 srcs = [
5961 "test/f32-vsub-relu.cc",
5962 "test/vbinary-microkernel-tester.h",
5963 ] + MICROKERNEL_TEST_HDRS,
5964 deps = MICROKERNEL_TEST_DEPS,
5965)
5966
5967xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005968 name = "f32_vsubc_test",
5969 srcs = [
5970 "test/f32-vsubc.cc",
5971 "test/vbinaryc-microkernel-tester.h",
5972 ] + MICROKERNEL_TEST_HDRS,
5973 deps = MICROKERNEL_TEST_DEPS,
5974)
5975
5976xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005977 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005978 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005979 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005980 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005981 ] + MICROKERNEL_TEST_HDRS,
5982 deps = MICROKERNEL_TEST_DEPS,
5983)
5984
5985xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005986 name = "f32_vsubc_relu_test",
5987 srcs = [
5988 "test/f32-vsubc-relu.cc",
5989 "test/vbinaryc-microkernel-tester.h",
5990 ] + MICROKERNEL_TEST_HDRS,
5991 deps = MICROKERNEL_TEST_DEPS,
5992)
5993
5994xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005995 name = "f32_vrsubc_test",
5996 srcs = [
5997 "test/f32-vrsubc.cc",
5998 "test/vbinaryc-microkernel-tester.h",
5999 ] + MICROKERNEL_TEST_HDRS,
6000 deps = MICROKERNEL_TEST_DEPS,
6001)
6002
6003xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006004 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006005 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006006 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006007 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006008 ] + MICROKERNEL_TEST_HDRS,
6009 deps = MICROKERNEL_TEST_DEPS,
6010)
6011
6012xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006013 name = "f32_vrsubc_relu_test",
6014 srcs = [
6015 "test/f32-vrsubc-relu.cc",
6016 "test/vbinaryc-microkernel-tester.h",
6017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
6021xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006022 name = "qs8_dwconv_minmax_test",
6023 srcs = [
6024 "test/qs8-dwconv-minmax.cc",
6025 "test/dwconv-microkernel-tester.h",
6026 "src/xnnpack/AlignedAllocator.h",
6027 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6029)
6030
6031xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006032 name = "qs8_gavgpool_minmax_test",
6033 srcs = [
6034 "test/qs8-gavgpool-minmax.cc",
6035 "test/gavgpool-microkernel-tester.h",
6036 "src/xnnpack/AlignedAllocator.h",
6037 ] + MICROKERNEL_TEST_HDRS,
6038 deps = MICROKERNEL_TEST_DEPS,
6039)
6040
6041xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006042 name = "qs8_gemm_minmax_test",
6043 srcs = [
6044 "test/qs8-gemm-minmax.cc",
6045 "test/gemm-microkernel-tester.h",
6046 "src/xnnpack/AlignedAllocator.h",
6047 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6048 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6049)
6050
6051xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006052 name = "qs8_igemm_minmax_test",
6053 srcs = [
6054 "test/qs8-igemm-minmax.cc",
6055 "test/gemm-microkernel-tester.h",
6056 "src/xnnpack/AlignedAllocator.h",
6057 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6058 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6059)
6060
6061xnnpack_unit_test(
6062 name = "qs8_requantization_test",
6063 srcs = [
6064 "src/xnnpack/requantization-stubs.h",
6065 "test/qs8-requantization.cc",
6066 "test/requantization-tester.h",
6067 ] + MICROKERNEL_TEST_HDRS,
6068 deps = MICROKERNEL_TEST_DEPS,
6069)
6070
6071xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006072 name = "qs8_vadd_minmax_test",
6073 srcs = [
6074 "test/qs8-vadd-minmax.cc",
6075 "test/vadd-microkernel-tester.h",
6076 ] + MICROKERNEL_TEST_HDRS,
6077 deps = MICROKERNEL_TEST_DEPS,
6078)
6079
6080xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006081 name = "qs8_vaddc_minmax_test",
6082 srcs = [
6083 "test/qs8-vaddc-minmax.cc",
6084 "test/vaddc-microkernel-tester.h",
6085 ] + MICROKERNEL_TEST_HDRS,
6086 deps = MICROKERNEL_TEST_DEPS,
6087)
6088
6089xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006090 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006091 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006092 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093 "test/avgpool-microkernel-tester.h",
6094 "src/xnnpack/AlignedAllocator.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006100 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006102 "test/qu8-dwconv-minmax.cc",
6103 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006104 "src/xnnpack/AlignedAllocator.h",
6105 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006106 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006107)
6108
6109xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006110 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006111 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006112 "test/qu8-igemm-minmax.cc",
6113 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006114 "src/xnnpack/AlignedAllocator.h",
6115 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006116 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006117)
6118
6119xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006120 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006121 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006122 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006123 "test/gavgpool-microkernel-tester.h",
6124 "src/xnnpack/AlignedAllocator.h",
6125 ] + MICROKERNEL_TEST_HDRS,
6126 deps = MICROKERNEL_TEST_DEPS,
6127)
6128
6129xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006130 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006132 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006133 "test/gemm-microkernel-tester.h",
6134 "src/xnnpack/AlignedAllocator.h",
6135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006136 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006137)
6138
6139xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006140 name = "qu8_requantization_test",
6141 srcs = [
6142 "src/xnnpack/requantization-stubs.h",
6143 "test/qu8-requantization.cc",
6144 "test/requantization-tester.h",
6145 ] + MICROKERNEL_TEST_HDRS,
6146 deps = MICROKERNEL_TEST_DEPS,
6147)
6148
6149xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006150 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006151 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006152 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153 "test/vadd-microkernel-tester.h",
6154 ] + MICROKERNEL_TEST_HDRS,
6155 deps = MICROKERNEL_TEST_DEPS,
6156)
6157
6158xnnpack_unit_test(
6159 name = "u8_clamp_test",
6160 srcs = [
6161 "test/u8-clamp.cc",
6162 "test/clamp-microkernel-tester.h",
6163 ] + MICROKERNEL_TEST_HDRS,
6164 deps = MICROKERNEL_TEST_DEPS,
6165)
6166
6167xnnpack_unit_test(
6168 name = "u8_lut32norm_test",
6169 srcs = [
6170 "test/u8-lut32norm.cc",
6171 "test/lut-norm-microkernel-tester.h",
6172 ] + MICROKERNEL_TEST_HDRS,
6173 deps = MICROKERNEL_TEST_DEPS,
6174)
6175
6176xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006177 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006178 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006179 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006180 "test/maxpool-microkernel-tester.h",
6181 ] + MICROKERNEL_TEST_HDRS,
6182 deps = MICROKERNEL_TEST_DEPS,
6183)
6184
6185xnnpack_unit_test(
6186 name = "u8_rmax_test",
6187 srcs = [
6188 "test/u8-rmax.cc",
6189 "test/rmax-microkernel-tester.h",
6190 ] + MICROKERNEL_TEST_HDRS,
6191 deps = MICROKERNEL_TEST_DEPS,
6192)
6193
6194xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006195 name = "x32_fill_test",
6196 srcs = [
6197 "test/x32-fill.cc",
6198 "test/fill-microkernel-tester.h",
6199 ] + MICROKERNEL_TEST_HDRS,
6200 deps = MICROKERNEL_TEST_DEPS,
6201)
6202
6203xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204 name = "x32_packx_test",
6205 srcs = [
6206 "test/x32-packx.cc",
6207 "test/pack-microkernel-tester.h",
6208 "src/xnnpack/AlignedAllocator.h",
6209 ] + MICROKERNEL_TEST_HDRS,
6210 deps = MICROKERNEL_TEST_DEPS,
6211)
6212
6213xnnpack_unit_test(
6214 name = "x32_pad_test",
6215 srcs = [
6216 "test/x32-pad.cc",
6217 "test/pad-microkernel-tester.h",
6218 ] + MICROKERNEL_TEST_HDRS,
6219 deps = MICROKERNEL_TEST_DEPS,
6220)
6221
6222xnnpack_unit_test(
6223 name = "x32_unpool_test",
6224 srcs = [
6225 "test/x32-unpool.cc",
6226 "test/unpool-microkernel-tester.h",
6227 ] + MICROKERNEL_TEST_HDRS,
6228 deps = MICROKERNEL_TEST_DEPS,
6229)
6230
6231xnnpack_unit_test(
6232 name = "x32_zip_test",
6233 srcs = [
6234 "test/x32-zip.cc",
6235 "test/zip-microkernel-tester.h",
6236 ] + MICROKERNEL_TEST_HDRS,
6237 deps = MICROKERNEL_TEST_DEPS,
6238)
6239
6240xnnpack_unit_test(
6241 name = "x8_lut_test",
6242 srcs = [
6243 "test/x8-lut.cc",
6244 "test/lut-microkernel-tester.h",
6245 ] + MICROKERNEL_TEST_HDRS,
6246 deps = MICROKERNEL_TEST_DEPS,
6247)
6248
6249xnnpack_unit_test(
6250 name = "x8_zip_test",
6251 srcs = [
6252 "test/x8-zip.cc",
6253 "test/zip-microkernel-tester.h",
6254 ] + MICROKERNEL_TEST_HDRS,
6255 deps = MICROKERNEL_TEST_DEPS,
6256)
6257
Marat Dukhan20c3b922020-03-10 03:45:06 -07006258########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259
6260xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006261 name = "operator_size_test",
6262 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006263 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006264)
6265
Marat Dukhan20c3b922020-03-10 03:45:06 -07006266xnnpack_binary(
6267 name = "subgraph_size_test",
6268 srcs = ["test/subgraph-size.c"],
6269 deps = [":XNNPACK"],
6270)
6271
6272########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006273
6274xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006275 name = "abs_nc_test",
6276 srcs = [
6277 "test/abs-nc.cc",
6278 "test/abs-operator-tester.h",
6279 ],
6280 deps = OPERATOR_TEST_DEPS,
6281)
6282
6283xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006284 name = "add_nd_test",
6285 srcs = [
6286 "test/add-nd.cc",
6287 "test/binary-elementwise-operator-tester.h",
6288 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006289 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006290)
6291
6292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006293 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006295 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 "test/argmax-pooling-operator-tester.h",
6297 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299)
6300
6301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006302 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006304 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 "test/average-pooling-operator-tester.h",
6306 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308)
6309
6310xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006311 name = "bankers_rounding_nc_test",
6312 srcs = [
6313 "test/bankers-rounding-nc.cc",
6314 "test/bankers-rounding-operator-tester.h",
6315 ],
6316 deps = OPERATOR_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
6320 name = "ceiling_nc_test",
6321 srcs = [
6322 "test/ceiling-nc.cc",
6323 "test/ceiling-operator-tester.h",
6324 ],
6325 deps = OPERATOR_TEST_DEPS,
6326)
6327
6328xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006329 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006330 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006331 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006332 "test/channel-shuffle-operator-tester.h",
6333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006335)
6336
6337xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006338 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006339 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006340 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006341 "test/clamp-operator-tester.h",
6342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006344)
6345
6346xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006347 name = "constant_pad_nd_test",
6348 srcs = [
6349 "test/constant-pad-nd.cc",
6350 "test/constant-pad-operator-tester.h",
6351 ],
6352 deps = OPERATOR_TEST_DEPS,
6353)
6354
6355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006356 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 "test/convolution-operator-tester.h",
6360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362)
6363
6364xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006365 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006367 "test/convolution-nchw.cc",
6368 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006369 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371)
6372
6373xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006374 name = "copy_nc_test",
6375 srcs = [
6376 "test/copy-nc.cc",
6377 "test/copy-operator-tester.h",
6378 ],
6379 deps = OPERATOR_TEST_DEPS,
6380)
6381
6382xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006383 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006384 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006385 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006386 "test/deconvolution-operator-tester.h",
6387 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006388 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389)
6390
6391xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006392 name = "divide_nd_test",
6393 srcs = [
6394 "test/binary-elementwise-operator-tester.h",
6395 "test/divide-nd.cc",
6396 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006397 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006398)
6399
6400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006401 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006403 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006404 "test/fully-connected-operator-tester.h",
6405 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407)
6408
6409xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006410 name = "floor_nc_test",
6411 srcs = [
6412 "test/floor-nc.cc",
6413 "test/floor-operator-tester.h",
6414 ],
6415 deps = OPERATOR_TEST_DEPS,
6416)
6417
6418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006419 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006421 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006422 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006423 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006425)
6426
6427xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006428 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006429 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006430 "test/global-average-pooling-ncw.cc",
6431 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006434)
6435
6436xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006437 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006438 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006439 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006440 "test/hardswish-operator-tester.h",
6441 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006442 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006443)
6444
6445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006446 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006448 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006449 "test/leaky-relu-operator-tester.h",
6450 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006451 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452)
6453
6454xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006455 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006456 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006457 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006458 "test/max-pooling-operator-tester.h",
6459 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006460 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006461)
6462
6463xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006464 name = "maximum_nd_test",
6465 srcs = [
6466 "test/binary-elementwise-operator-tester.h",
6467 "test/maximum-nd.cc",
6468 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006470)
6471
6472xnnpack_unit_test(
6473 name = "minimum_nd_test",
6474 srcs = [
6475 "test/binary-elementwise-operator-tester.h",
6476 "test/minimum-nd.cc",
6477 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006479)
6480
6481xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006482 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006483 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006484 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006485 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006486 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006487 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006488)
6489
6490xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006491 name = "negate_nc_test",
6492 srcs = [
6493 "test/negate-nc.cc",
6494 "test/negate-operator-tester.h",
6495 ],
6496 deps = OPERATOR_TEST_DEPS,
6497)
6498
6499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006500 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006502 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503 "test/prelu-operator-tester.h",
6504 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006506)
6507
6508xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006509 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006510 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006511 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006512 "test/resize-bilinear-operator-tester.h",
6513 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006514 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006515)
6516
6517xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006518 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006519 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006520 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006521 "test/sigmoid-operator-tester.h",
6522 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006523 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006524)
6525
6526xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006527 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006528 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006529 "test/softmax-nc.cc",
6530 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006533)
6534
6535xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006536 name = "square_nc_test",
6537 srcs = [
6538 "test/square-nc.cc",
6539 "test/square-operator-tester.h",
6540 ],
6541 deps = OPERATOR_TEST_DEPS,
6542)
6543
6544xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006545 name = "square_root_nc_test",
6546 srcs = [
6547 "test/square-root-nc.cc",
6548 "test/square-root-operator-tester.h",
6549 ],
6550 deps = OPERATOR_TEST_DEPS,
6551)
6552
6553xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006554 name = "squared_difference_nd_test",
6555 srcs = [
6556 "test/binary-elementwise-operator-tester.h",
6557 "test/squared-difference-nd.cc",
6558 ],
6559 deps = OPERATOR_TEST_DEPS,
6560)
6561
6562xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006563 name = "subtract_nd_test",
6564 srcs = [
6565 "test/binary-elementwise-operator-tester.h",
6566 "test/subtract-nd.cc",
6567 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006568 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006569)
6570
6571xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006572 name = "truncation_nc_test",
6573 srcs = [
6574 "test/truncation-nc.cc",
6575 "test/truncation-operator-tester.h",
6576 ],
6577 deps = OPERATOR_TEST_DEPS,
6578)
6579
6580xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006581 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006582 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006583 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006584 "test/unpooling-operator-tester.h",
6585 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006586 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006587)
6588
Chao Mei6ddfc602020-05-13 22:29:36 -07006589############################### Misc unit tests ###############################
6590
6591xnnpack_unit_test(
6592 name = "memory_planner_test",
6593 srcs = [
6594 "test/memory-planner-test.cc",
6595 ],
6596 deps = [
6597 ":XNNPACK",
6598 ":memory_planner",
6599 ],
6600)
6601
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006602xnnpack_unit_test(
6603 name = "subgraph_nchw_test",
6604 srcs = [
6605 "src/xnnpack/subgraph.h",
6606 "test/subgraph-nchw.cc",
6607 "test/subgraph-tester.h",
6608 ],
6609 deps = [
6610 ":XNNPACK",
6611 ],
6612)
6613
Marat Dukhan08c4a432019-10-03 09:29:21 -07006614############################# Build configurations #############################
6615
Marat Dukhanb8642352019-10-30 15:43:02 -07006616# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006617config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006618 name = "xnn_enable_assembly_explicit_true",
6619 define_values = {"xnn_enable_assembly": "true"},
6620)
6621
6622# Disables usage of assembly kernels.
6623config_setting(
6624 name = "xnn_enable_assembly_explicit_false",
6625 define_values = {"xnn_enable_assembly": "false"},
6626)
6627
Marat Dukhan9de90e02020-06-18 16:04:12 -07006628# Enables usage of sparse inference.
6629config_setting(
6630 name = "xnn_enable_sparse_explicit_true",
6631 define_values = {"xnn_enable_sparse": "true"},
6632)
6633
6634# Disables usage of sparse inference.
6635config_setting(
6636 name = "xnn_enable_sparse_explicit_false",
6637 define_values = {"xnn_enable_sparse": "false"},
6638)
6639
Marat Dukhan05702cf2020-03-26 15:41:33 -07006640# Disables usage of HMP-aware optimizations.
6641config_setting(
6642 name = "xnn_enable_hmp_explicit_false",
6643 define_values = {"xnn_enable_hmp": "false"},
6644)
6645
Chao Mei6ddfc602020-05-13 22:29:36 -07006646# Enable usage of optimized memory allocation
6647config_setting(
6648 name = "xnn_enable_memopt_explicit_true",
6649 define_values = {"xnnpack_enable_memopt": "true"},
6650)
6651
6652# Disable usage of optimized memory allocation
6653config_setting(
6654 name = "xnn_enable_memopt_explicit_false",
6655 define_values = {"xnnpack_enable_memopt": "false"},
6656)
6657
Marat Dukhanb8642352019-10-30 15:43:02 -07006658# Builds with -c dbg
6659config_setting(
6660 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006661 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006662 "compilation_mode": "dbg",
6663 },
6664)
6665
6666# Builds with -c opt
6667config_setting(
6668 name = "optimized_build",
6669 values = {
6670 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006671 },
6672)
6673
6674config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006675 name = "linux_k8",
6676 values = {"cpu": "k8"},
6677)
6678
6679config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006680 name = "linux_arm",
6681 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006682)
6683
6684config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006685 name = "linux_armeabi",
6686 values = {"cpu": "armeabi"},
6687)
6688
6689config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006690 name = "linux_armhf",
6691 values = {"cpu": "armhf"},
6692)
6693
6694config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006695 name = "linux_armv7a",
6696 values = {"cpu": "armv7a"},
6697)
6698
6699config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006700 name = "linux_aarch64",
6701 values = {"cpu": "aarch64"},
6702)
6703
6704config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006705 name = "android",
6706 values = {"crosstool_top": "//external:android/crosstool"},
6707)
6708
6709config_setting(
6710 name = "android_armv7",
6711 values = {
6712 "crosstool_top": "//external:android/crosstool",
6713 "cpu": "armeabi-v7a",
6714 },
6715)
6716
6717config_setting(
6718 name = "android_arm64",
6719 values = {
6720 "crosstool_top": "//external:android/crosstool",
6721 "cpu": "arm64-v8a",
6722 },
6723)
6724
6725config_setting(
6726 name = "android_x86",
6727 values = {
6728 "crosstool_top": "//external:android/crosstool",
6729 "cpu": "x86",
6730 },
6731)
6732
6733config_setting(
6734 name = "android_x86_64",
6735 values = {
6736 "crosstool_top": "//external:android/crosstool",
6737 "cpu": "x86_64",
6738 },
6739)
6740
6741config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006742 name = "windows_x86_64",
6743 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006744)
6745
6746config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006747 name = "windows_x86_64_clang",
6748 values = {
6749 "compiler": "clang-cl",
6750 "cpu": "x64_windows",
6751 },
6752)
6753
6754config_setting(
6755 name = "windows_x86_64_mingw",
6756 values = {
6757 "compiler": "mingw-gcc",
6758 "cpu": "x64_windows",
6759 },
6760)
6761
6762config_setting(
6763 name = "windows_x86_64_msys",
6764 values = {
6765 "compiler": "msys-gcc",
6766 "cpu": "x64_windows",
6767 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006768)
6769
6770config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006771 name = "macos_x86_64",
6772 values = {
6773 "apple_platform_type": "macos",
6774 "cpu": "darwin",
6775 },
6776)
6777
6778config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006779 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006780 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006781)
6782
6783config_setting(
6784 name = "emscripten_wasm",
6785 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006786 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006787 "cpu": "wasm",
6788 },
6789)
6790
6791config_setting(
6792 name = "emscripten_wasmsimd",
6793 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006794 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006795 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006796 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006797 },
6798)
6799
6800config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006801 name = "ios_armv7",
6802 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006803 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006804 "cpu": "ios_armv7",
6805 },
6806)
6807
6808config_setting(
6809 name = "ios_arm64",
6810 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006811 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006812 "cpu": "ios_arm64",
6813 },
6814)
6815
6816config_setting(
6817 name = "ios_arm64e",
6818 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006819 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006820 "cpu": "ios_arm64e",
6821 },
6822)
6823
6824config_setting(
6825 name = "ios_x86",
6826 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006827 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006828 "cpu": "ios_i386",
6829 },
6830)
6831
6832config_setting(
6833 name = "ios_x86_64",
6834 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006835 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006836 "cpu": "ios_x86_64",
6837 },
6838)
6839
6840config_setting(
6841 name = "watchos_armv7k",
6842 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006843 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006844 "cpu": "watchos_armv7k",
6845 },
6846)
6847
6848config_setting(
6849 name = "watchos_arm64_32",
6850 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006851 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006852 "cpu": "watchos_arm64_32",
6853 },
6854)
6855
6856config_setting(
6857 name = "watchos_x86",
6858 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006859 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006860 "cpu": "watchos_i386",
6861 },
6862)
6863
6864config_setting(
6865 name = "watchos_x86_64",
6866 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006867 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006868 "cpu": "watchos_x86_64",
6869 },
6870)
6871
6872config_setting(
6873 name = "tvos_arm64",
6874 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006875 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006876 "cpu": "tvos_arm64",
6877 },
6878)
6879
6880config_setting(
6881 name = "tvos_x86_64",
6882 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006883 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006884 "cpu": "tvos_x86_64",
6885 },
6886)