blob: f5f5a93d854d766679f5a520dd197d77cc5f6752 [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 Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.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",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "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",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800219 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
225 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700228 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/1x1-minmax-scalar.c",
230 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar.c",
232 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar.c",
234 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar.c",
236 "src/f32-spmm/gen/8x2-minmax-scalar.c",
237 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700238 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700241 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vadd-scalar-x1.c",
245 "src/f32-vbinary/gen/vadd-scalar-x2.c",
246 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700250 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
251 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
252 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700253 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700259 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
260 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
261 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700262 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
263 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
264 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
272 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
273 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800274 "src/f32-vbinary/gen/vmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vmax-scalar-x4.c",
277 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
278 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
279 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
280 "src/f32-vbinary/gen/vmin-scalar-x1.c",
281 "src/f32-vbinary/gen/vmin-scalar-x2.c",
282 "src/f32-vbinary/gen/vmin-scalar-x4.c",
283 "src/f32-vbinary/gen/vminc-scalar-x1.c",
284 "src/f32-vbinary/gen/vminc-scalar-x2.c",
285 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700289 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
290 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
291 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vmul-scalar-x1.c",
293 "src/f32-vbinary/gen/vmul-scalar-x2.c",
294 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700298 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
299 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
300 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700301 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
302 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
303 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
311 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
312 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700313 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
314 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
315 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700316 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
317 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
318 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700319 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
320 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
321 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700322 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
323 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
324 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
325 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
326 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
327 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700334 "src/f32-vbinary/gen/vsub-scalar-x1.c",
335 "src/f32-vbinary/gen/vsub-scalar-x2.c",
336 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700337 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
338 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
339 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700340 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
341 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
342 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700343 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
344 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
345 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700346 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
347 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
348 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700349 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
350 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
351 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700352 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
353 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
354 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
355 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
356 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
357 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
358 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
359 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
360 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700361 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
362 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
363 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700364 "src/f32-vunary/gen/vabs-scalar-x1.c",
365 "src/f32-vunary/gen/vabs-scalar-x2.c",
366 "src/f32-vunary/gen/vabs-scalar-x4.c",
367 "src/f32-vunary/gen/vneg-scalar-x1.c",
368 "src/f32-vunary/gen/vneg-scalar-x2.c",
369 "src/f32-vunary/gen/vneg-scalar-x4.c",
370 "src/f32-vunary/gen/vsqr-scalar-x1.c",
371 "src/f32-vunary/gen/vsqr-scalar-x2.c",
372 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800373 "src/math/expminus-scalar-lut2048-p1.c",
374 "src/math/expminus-scalar-lut64-p2.c",
375 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700376 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700377 "src/math/roundne-scalar-nearbyint.c",
378 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700379 "src/math/roundd-scalar-addsub.c",
380 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700381 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700382 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700383 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700384 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700385 "src/math/roundz-scalar-addsub.c",
386 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700387 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800388 "src/math/sigmoid-scalar-lut2048-p1-div.c",
389 "src/math/sigmoid-scalar-lut64-p2-div.c",
390 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700391 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
392 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700393 "src/qs8-requantization/fp32-scalar-lrintf.c",
394 "src/qs8-requantization/fp32-scalar-magic.c",
395 "src/qs8-requantization/precise-scalar-signed64.c",
396 "src/qs8-requantization/precise-scalar-unsigned32.c",
397 "src/qs8-requantization/precise-scalar-unsigned64.c",
398 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700399 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
400 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
401 "src/qu8-dwconv/up1x9-minmax-scalar.c",
402 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
403 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700404 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
405 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700406 "src/qu8-gemm/2x2-minmax-scalar.c",
407 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700408 "src/qu8-requantization/fp32-scalar-lrintf.c",
409 "src/qu8-requantization/fp32-scalar-magic.c",
410 "src/qu8-requantization/precise-scalar-signed64.c",
411 "src/qu8-requantization/precise-scalar-unsigned32.c",
412 "src/qu8-requantization/precise-scalar-unsigned64.c",
413 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700414 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700415 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700416 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700417 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700419 "src/x32-fill/scalar-float.c",
420 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x2-scalar.c",
422 "src/x32-packx/x3-scalar.c",
423 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700424 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700425 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700426 "src/x32-unpool/scalar.c",
427 "src/x32-zip/x2-scalar.c",
428 "src/x32-zip/x3-scalar.c",
429 "src/x32-zip/x4-scalar.c",
430 "src/x32-zip/xm-scalar.c",
431 "src/x8-lut/scalar.c",
432 "src/x8-zip/x2-scalar.c",
433 "src/x8-zip/x3-scalar.c",
434 "src/x8-zip/x4-scalar.c",
435 "src/x8-zip/xm-scalar.c",
436]
437
Marat Dukhan436ebe62019-12-04 15:10:12 -0800438WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700439 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
440 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700441 "src/f32-clamp/gen/wasm-x1.c",
442 "src/f32-clamp/gen/wasm-x2.c",
443 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700444 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
445 "src/f32-dwconv/gen/up1x4-wasm.c",
446 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
447 "src/f32-dwconv/gen/up1x9-wasm.c",
448 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
449 "src/f32-dwconv/gen/up1x25-wasm.c",
450 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
451 "src/f32-dwconv/gen/up2x4-wasm.c",
452 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
453 "src/f32-dwconv/gen/up2x9-wasm.c",
454 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
455 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700456 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
457 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
458 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
459 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700460 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
461 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
463 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
464 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
465 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700466 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
467 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
469 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700470 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
471 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
472 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700473 "src/f32-gemm/gen/1x4-wasm.c",
474 "src/f32-gemm/gen/2x4-wasm.c",
475 "src/f32-gemm/gen/4x2-wasm.c",
476 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700477 "src/f32-gemm/gen/1x4-relu-wasm.c",
478 "src/f32-gemm/gen/2x4-relu-wasm.c",
479 "src/f32-gemm/gen/4x2-relu-wasm.c",
480 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700481 "src/f32-gemm/gen/1x4-minmax-wasm.c",
482 "src/f32-gemm/gen/2x4-minmax-wasm.c",
483 "src/f32-gemm/gen/4x2-minmax-wasm.c",
484 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700485 "src/f32-hswish/gen/hswish-wasm-x1.c",
486 "src/f32-hswish/gen/hswish-wasm-x2.c",
487 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700488 "src/f32-igemm/gen/1x4-wasm.c",
489 "src/f32-igemm/gen/2x4-wasm.c",
490 "src/f32-igemm/gen/4x2-wasm.c",
491 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700492 "src/f32-igemm/gen/1x4-relu-wasm.c",
493 "src/f32-igemm/gen/2x4-relu-wasm.c",
494 "src/f32-igemm/gen/4x2-relu-wasm.c",
495 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700496 "src/f32-igemm/gen/1x4-minmax-wasm.c",
497 "src/f32-igemm/gen/2x4-minmax-wasm.c",
498 "src/f32-igemm/gen/4x2-minmax-wasm.c",
499 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700500 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
501 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
502 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700503 "src/f32-prelu/gen/wasm-2x1.c",
504 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700505 "src/f32-relu/gen/wasm-x1.c",
506 "src/f32-relu/gen/wasm-x2.c",
507 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700508 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
509 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
510 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
511 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
512 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
513 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
514 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
515 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
516 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
517 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
518 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
519 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700520 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
521 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
522 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
523 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
524 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
525 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
526 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
527 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
528 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
529 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
530 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
531 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800532 "src/f32-vbinary/gen/vmax-wasm-x1.c",
533 "src/f32-vbinary/gen/vmax-wasm-x2.c",
534 "src/f32-vbinary/gen/vmax-wasm-x4.c",
535 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
536 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
537 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
538 "src/f32-vbinary/gen/vmin-wasm-x1.c",
539 "src/f32-vbinary/gen/vmin-wasm-x2.c",
540 "src/f32-vbinary/gen/vmin-wasm-x4.c",
541 "src/f32-vbinary/gen/vminc-wasm-x1.c",
542 "src/f32-vbinary/gen/vminc-wasm-x2.c",
543 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700544 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
545 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
546 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
547 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
548 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
549 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
550 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
553 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
554 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
555 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
556 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
557 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
558 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
559 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
560 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
561 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
565 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
566 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
567 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
568 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
569 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
570 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
571 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
572 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
573 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
574 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
577 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
578 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
579 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700580 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
581 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
582 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700583 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
584 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
585 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800586]
587
Marat Dukhan290055c2020-06-09 12:24:29 -0700588WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700589 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
590 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
591 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700592 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
593 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
594 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
595 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700596 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
597 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
598 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
599 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700600 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
601 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
602 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
603 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
604 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
605 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700606 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
607 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
608 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
609 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
610 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
611 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
612 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
613 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
614 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
615 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
616 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
617 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
618 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
619 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
620 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
621 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
622 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
623 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
624 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
625 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
626 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
627 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
628 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
629 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700630 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
631 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
632 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700633 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700634 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
635 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
636 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
637 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700638 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
639 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
640 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
641 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
642 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
643 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
644 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
645 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
646 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
647 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
648 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
649 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
650 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
651 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
652 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
653 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
654 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
655 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
656 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
657 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
658 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
659 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
660 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
661 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
662 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
663 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
664 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
665 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
666 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
667 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700668 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
669 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700670 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
671 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
672 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
673 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
674 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
675 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
676 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
677 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
678 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
679 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
680 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
681 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
682 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
683 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
684 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
685 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
686 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
687 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
688 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
689 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
690 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
691 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
692 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
693 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
694 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
695 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
696 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
697 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
698 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
699 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700700 "src/f32-ibilinear/gen/wasmsimd-c4.c",
701 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700702 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
703 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
704 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700705 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700706 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
707 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
708 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
709 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700710 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
711 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
712 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
713 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
714 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
715 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
716 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
717 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
718 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
719 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
720 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
721 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
722 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
723 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
724 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
725 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
726 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
727 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
728 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
729 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
730 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
731 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
732 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
733 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
734 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
735 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
736 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
737 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
738 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
739 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700740 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
741 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700742 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
743 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
744 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
745 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700746 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
747 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700748 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
749 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
750 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700751 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
752 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700753 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
754 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
755 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
756 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700757 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
758 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700759 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
760 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
761 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700762 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
763 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700764 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
765 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
766 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
767 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
768 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
769 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
770 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700771 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
772 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700773 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
774 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
775 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
776 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700777 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
778 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
779 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
780 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
781 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
782 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
783 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
784 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
785 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
786 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
787 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
788 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700789 "src/f32-relu/gen/wasmsimd-x4.c",
790 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700791 "src/f32-rmax/wasmsimd-arm.c",
792 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700793 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
794 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
795 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
796 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
797 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
798 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700799 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
800 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
801 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
802 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
803 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
804 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
806 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
807 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
808 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700809 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
810 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700811 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
812 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700813 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
814 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
815 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
816 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700817 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
818 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700821 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
822 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
823 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
824 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700825 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
826 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700827 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
828 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700829 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
830 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
831 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
832 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700833 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
834 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700835 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
836 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700837 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
838 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
839 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
840 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700841 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
842 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
843 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
844 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700845 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
846 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
847 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
848 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700849 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
850 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
851 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
852 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700853 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
854 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
855 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
856 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700857 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
858 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700859 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
860 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700861 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
862 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
863 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
864 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700865 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
866 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700867 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
868 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700869 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
870 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
871 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700873 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
874 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700875 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
876 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700877 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
878 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
879 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
880 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700881 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
882 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700883 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
884 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700885 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
886 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700887 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
888 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700889 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
890 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
891 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
892 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700893 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
894 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700895 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
896 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700897 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
898 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
899 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
900 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700901 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700903 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
904 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700905 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
906 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
907 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
908 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700909 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
910 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
911 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
912 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700913 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
914 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
915 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
916 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
917 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
918 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
919 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
920 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
921 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
922 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
923 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
924 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
925 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
926 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700927 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
928 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700929 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
930 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
931 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
932 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
933 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
934 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700935 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
936 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
937 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700938 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
939 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
940 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700941 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
942 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
943 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700944 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
945 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
946 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
947 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
948 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
949 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
950 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
951 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
952 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
953 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
954 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
955 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
956 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
957 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
958 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700959 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700960 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700961 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700962 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700963 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700964 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700965 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700966 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700967 "src/x32-zip/x2-wasmsimd.c",
968 "src/x32-zip/x3-wasmsimd.c",
969 "src/x32-zip/x4-wasmsimd.c",
970 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700971 "src/math/roundne-wasmsimd-addsub.c",
972 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700973 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700974 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700975 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700976 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700977 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700978 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
979 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700980]
981
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800982PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700983 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700984 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700985 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700986 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700987 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700988 "src/f32-spmm/gen/4x1-minmax-psimd.c",
989 "src/f32-spmm/gen/8x1-minmax-psimd.c",
990 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700991]
992
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800993PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700994 "src/qs8-requantization/precise-psimd.c",
995 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700996 "src/qu8-requantization/precise-psimd.c",
997 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800998]
999
Marat Dukhan08c4a432019-10-03 09:29:21 -07001000# ISA-specific micro-kernels
1001NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001002 "src/f32-argmaxpool/4x-neon-c4.c",
1003 "src/f32-argmaxpool/9p8x-neon-c4.c",
1004 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001005 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1006 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001007 "src/f32-clamp/gen/neon-x4.c",
1008 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001009 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1010 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001011 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1012 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001013 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1014 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001015 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1016 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001017 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1018 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1019 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1020 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001021 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1022 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1023 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1024 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001025 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1026 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1027 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1028 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001029 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001030 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1031 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001032 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1033 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1034 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1035 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1036 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1038 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1039 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1040 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1041 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1042 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1044 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1045 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1046 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1047 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1048 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1049 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1050 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1051 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1052 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1053 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1054 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1055 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1056 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1057 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1058 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1059 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1060 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1061 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1062 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001063 "src/f32-hswish/gen/hswish-neon-x4.c",
1064 "src/f32-hswish/gen/hswish-neon-x8.c",
1065 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001066 "src/f32-ibilinear/gen/neon-c4.c",
1067 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001068 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1069 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1070 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1071 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1072 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1073 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1074 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1075 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1076 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1077 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1078 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1079 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1080 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1081 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1082 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1083 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001084 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1085 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1086 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001087 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1088 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001089 "src/f32-prelu/gen/neon-1x4.c",
1090 "src/f32-prelu/gen/neon-1x8.c",
1091 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001092 "src/f32-prelu/gen/neon-2x4.c",
1093 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001094 "src/f32-prelu/gen/neon-2x16.c",
1095 "src/f32-prelu/gen/neon-4x4.c",
1096 "src/f32-prelu/gen/neon-4x8.c",
1097 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001098 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1099 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1100 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1101 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1103 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1104 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1105 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1106 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1107 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1108 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1109 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1110 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1111 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1112 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1113 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1114 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1115 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1116 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1117 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1118 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1119 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1120 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1121 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001122 "src/f32-relu/gen/neon-x4.c",
1123 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001125 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001126 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1127 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1128 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1129 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1130 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1131 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1132 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1133 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1134 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1135 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1136 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1137 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1138 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1139 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1140 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1141 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1142 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1143 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001144 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1145 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1146 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1147 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001148 "src/f32-vbinary/gen/vmax-neon-x4.c",
1149 "src/f32-vbinary/gen/vmax-neon-x8.c",
1150 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1151 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1152 "src/f32-vbinary/gen/vmin-neon-x4.c",
1153 "src/f32-vbinary/gen/vmin-neon-x8.c",
1154 "src/f32-vbinary/gen/vminc-neon-x4.c",
1155 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001156 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1157 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1158 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1159 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1160 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1161 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001162 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1163 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1164 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1165 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001166 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1167 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1168 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1169 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001170 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1171 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001172 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1173 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001174 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1175 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1176 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1177 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1178 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1179 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1180 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1181 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001182 "src/f32-vunary/gen/vabs-neon-x4.c",
1183 "src/f32-vunary/gen/vabs-neon-x8.c",
1184 "src/f32-vunary/gen/vneg-neon-x4.c",
1185 "src/f32-vunary/gen/vneg-neon-x8.c",
1186 "src/f32-vunary/gen/vsqr-neon-x4.c",
1187 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001188 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1189 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1190 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1191 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1192 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1193 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1194 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1195 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001196 "src/qs8-requantization/precise-neon.c",
1197 "src/qs8-requantization/fp32-neon.c",
1198 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001199 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1200 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1201 "src/qu8-dwconv/up8x9-minmax-neon.c",
1202 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1203 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1204 "src/qu8-gemm/4x8-minmax-neon.c",
1205 "src/qu8-gemm/8x8-minmax-neon.c",
1206 "src/qu8-igemm/4x8-minmax-neon.c",
1207 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001208 "src/qu8-requantization/precise-neon.c",
1209 "src/qu8-requantization/fp32-neon.c",
1210 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001211 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001212 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001213 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001214 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001215 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001216 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001217 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001218 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001219 "src/x32-zip/x2-neon.c",
1220 "src/x32-zip/x3-neon.c",
1221 "src/x32-zip/x4-neon.c",
1222 "src/x32-zip/xm-neon.c",
1223 "src/x8-zip/x2-neon.c",
1224 "src/x8-zip/x3-neon.c",
1225 "src/x8-zip/x4-neon.c",
1226 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001227 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001228 "src/math/roundd-neon-addsub.c",
1229 "src/math/roundd-neon-cvt.c",
1230 "src/math/roundu-neon-addsub.c",
1231 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001232 "src/math/roundz-neon-addsub.c",
1233 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001234 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001235 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1236 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1237 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1238 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1239 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1240 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001241 "src/math/sqrt-neon-nr1rsqrts.c",
1242 "src/math/sqrt-neon-nr2rsqrts.c",
1243 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001244]
1245
1246NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001247 "src/f32-ibilinear/gen/neonfma-c4.c",
1248 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001249 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1250 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1251 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1252 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1253 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1254 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1255 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1256 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1257 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001258 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1259 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1260 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1261 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001262 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1263 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1264 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1265 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001266 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1267 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1268 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1269 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001270 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1271 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1272 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1273 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1274 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1275 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1276 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1277 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1278 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1279 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1280 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1281 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1282 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1283 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1284 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1285 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1286 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1287 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001288 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1289 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001290 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1291 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1292 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1293 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1294 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1295 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1296 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1297 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1298 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1299 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1300 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1301 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1302 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1303 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1304 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1305 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1306 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1307 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1308 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1309 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1310 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1311 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1312 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1313 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001314 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1315 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1316 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1317 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1318 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1319 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1320 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1321 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1322 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1323 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001368 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1369 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001370 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1371 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1372 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1373 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1374 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1375 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1376 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1377 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1378 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1379 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1380 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1381 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1382 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1383 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1384 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1385 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1386 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1387 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1388 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1389 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001390 "src/math/exp-neonfma-lut64-p2.c",
1391 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001392 "src/math/expminus-neonfma-lut2048-p1.c",
1393 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001394 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001395 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1396 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1397 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1398 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1399 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1400 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1401 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1402 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1403 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1404 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1405 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1406 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1407 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1408 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1409 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1410 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1411 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1412 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001413 "src/math/sqrt-neonfma-nr1fma.c",
1414 "src/math/sqrt-neonfma-nr2fma.c",
1415 "src/math/sqrt-neonfma-nr3fma.c",
1416 "src/math/sqrt-neonfma-nr2fma1adj.c",
1417 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001418]
1419
1420AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001421 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1422 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1423 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1424 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1425 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1426 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001427 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1428 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1429 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1430 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1431 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1432 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1433 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1434 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1435 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1436 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1437 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1438 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1439 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1440 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1441 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1442 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1443 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1444 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1445 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1446 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001447 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1448 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001449 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1450 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001451 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1452 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001453 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1454 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001455 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1456 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1457 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1458 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1459 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001460 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001478 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1479 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1480 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1481 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1482 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1483 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1484 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1485 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1486 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1487 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1488 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1489 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1490 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1491 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1492 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1493 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1494 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1495 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001496 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1497 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001498 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1499 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1500 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1501 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1502 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1503 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001504]
1505
Marat Dukhan8853b822020-05-07 12:19:01 -07001506NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001507 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1508 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1509 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1510 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1511 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1512 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1513 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1514 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001515 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001516 "src/math/roundd-neonv8.c",
1517 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001518 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001519]
1520
Marat Dukhan08c4a432019-10-03 09:29:21 -07001521AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001522 "src/f16-clamp/gen/neonfp16arith-x16.c",
1523 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001524 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1525 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1526 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1527 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1528 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1529 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1530 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1531 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1532 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1533 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1534 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1535 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001536 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1537 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001538 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1539 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001540 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001541 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001542 "src/f16-relu/gen/neonfp16arith-x16.c",
1543 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001544 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1545 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1546 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1547 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1548 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1549 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1550 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1551 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1552 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1553 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1554 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1555 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1556 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1557 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1558 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1559 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1560 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1561 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1562 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1563 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1564 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1565 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1566 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1567 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1568 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1569 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1570 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1571 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001572 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1573 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001574 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1575 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1576 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1577 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1578 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1579 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1580 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1581 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001582 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1583 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1584 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1585 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001586 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1587 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1588 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1589 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1590 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1591 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1592 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1593 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1594 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1595 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1596 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1597 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001598 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1599 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1600 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1601 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1602 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1603 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1604 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1605 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606]
1607
1608SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001609 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1610 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001611 "src/f32-clamp/gen/sse-x4.c",
1612 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001613 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1614 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001615 "src/f32-dwconv-chw/3x3p1-sse.c",
1616 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001617 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1618 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1619 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1620 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1621 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1622 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1623 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1624 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1625 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1626 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1627 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1628 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001629 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001630 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1631 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001632 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1633 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1634 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1635 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1636 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1637 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1638 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1639 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1640 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1641 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1642 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1643 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1644 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001645 "src/f32-hswish/gen/hswish-sse-x4.c",
1646 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001647 "src/f32-ibilinear/gen/sse-c4.c",
1648 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001649 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1650 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1651 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1652 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1653 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1654 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1655 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001656 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1657 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1658 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001659 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001660 "src/f32-prelu/gen/sse-2x4.c",
1661 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001662 "src/f32-relu/gen/sse-x4.c",
1663 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001665 "src/f32-spmm/gen/4x1-minmax-sse.c",
1666 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001667 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001668 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1669 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1670 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1671 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1672 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1673 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1674 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1675 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001676 "src/f32-vbinary/gen/vmax-sse-x4.c",
1677 "src/f32-vbinary/gen/vmax-sse-x8.c",
1678 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1679 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1680 "src/f32-vbinary/gen/vmin-sse-x4.c",
1681 "src/f32-vbinary/gen/vmin-sse-x8.c",
1682 "src/f32-vbinary/gen/vminc-sse-x4.c",
1683 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001684 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1685 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1686 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1687 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1688 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1689 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1690 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1691 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001692 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1693 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1694 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1695 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001696 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1697 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1698 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1699 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001700 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1701 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001702 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1703 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001704 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1705 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001706 "src/f32-vunary/gen/vabs-sse-x4.c",
1707 "src/f32-vunary/gen/vabs-sse-x8.c",
1708 "src/f32-vunary/gen/vneg-sse-x4.c",
1709 "src/f32-vunary/gen/vneg-sse-x8.c",
1710 "src/f32-vunary/gen/vsqr-sse-x4.c",
1711 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001712 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001714 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001715 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001716 "src/math/roundd-sse-addsub.c",
1717 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001718 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001719 "src/math/sqrt-sse-hh1mac.c",
1720 "src/math/sqrt-sse-nr1mac.c",
1721 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722]
1723
1724SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001725 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1726 "src/f32-argmaxpool/4x-sse2-c4.c",
1727 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001728 "src/f32-prelu/gen/sse2-2x4.c",
1729 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001730 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1731 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1732 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1733 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1734 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1735 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1736 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1737 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1738 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1739 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1740 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1741 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001742 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001743 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001744 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001745 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001746 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1747 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001748 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1749 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001750 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1751 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1752 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1753 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1754 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1755 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1756 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1757 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001758 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1759 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1760 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001761 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1762 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1763 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001764 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1765 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1766 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001767 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1768 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001769 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1770 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001771 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1772 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001773 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1774 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001775 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001776 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1777 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001778 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001779 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1780 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1781 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001782 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1783 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001784 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1785 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001786 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1787 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001788 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001789 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1790 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001791 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001792 "src/qs8-requantization/precise-sse2.c",
1793 "src/qs8-requantization/fp32-sse2.c",
1794 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001795 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1796 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1797 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1798 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1799 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1800 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1801 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1802 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001803 "src/qu8-requantization/precise-sse2.c",
1804 "src/qu8-requantization/fp32-sse2.c",
1805 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001806 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001807 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001808 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001810 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811 "src/x32-zip/x2-sse2.c",
1812 "src/x32-zip/x3-sse2.c",
1813 "src/x32-zip/x4-sse2.c",
1814 "src/x32-zip/xm-sse2.c",
1815 "src/x8-zip/x2-sse2.c",
1816 "src/x8-zip/x3-sse2.c",
1817 "src/x8-zip/x4-sse2.c",
1818 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001819 "src/math/exp-sse2-p5.c",
1820 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001821 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001822 "src/math/roundd-sse2-cvt.c",
1823 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001824 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001825 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001826]
1827
1828SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001829 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1830 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1831 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001832 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1833 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1834 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001835 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1836 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1837 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001838 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1839 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001840 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1841 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001842 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1843 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001844 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1845 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001846 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001847 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1848 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001849 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001850 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1851 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1852 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001853 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1854 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001855 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1856 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001857 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1858 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001859 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001860 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1861 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001862 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001863 "src/qs8-requantization/precise-ssse3.c",
1864 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001865 "src/qu8-requantization/precise-ssse3.c",
1866 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867]
1868
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001869SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001870 "src/f32-prelu/gen/sse41-2x4.c",
1871 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001872 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001873 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001874 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001875 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001876 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1877 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001878 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1879 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001880 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1881 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1882 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1883 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1884 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1885 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1886 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1887 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001888 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1889 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1890 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001891 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1892 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1893 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001894 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1895 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1896 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001897 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1898 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001899 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1900 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001901 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1902 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001903 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1904 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001905 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001906 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1907 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001908 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001909 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1910 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1911 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001912 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1913 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001914 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1915 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001916 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1917 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001918 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001919 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1920 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001921 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001922 "src/qs8-requantization/fp32-sse4.c",
1923 "src/qs8-requantization/precise-sse4.c",
1924 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001925 "src/qu8-requantization/precise-sse4.c",
1926 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001927 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001928 "src/math/roundd-sse41.c",
1929 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001930 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001931]
1932
Marat Dukhan08c4a432019-10-03 09:29:21 -07001933AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001934 "src/f32-clamp/gen/avx-x8.c",
1935 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001936 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1937 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1938 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1939 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1940 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1941 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1942 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1943 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1944 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1945 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1946 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1947 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1948 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1949 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1950 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1951 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1952 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1953 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1954 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1955 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1956 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1957 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1958 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1959 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1960 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1961 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1962 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1963 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1964 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1965 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001966 "src/f32-hswish/gen/hswish-avx-x8.c",
1967 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001968 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1969 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1970 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1971 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1972 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1973 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1974 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1975 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1976 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001977 "src/f32-prelu/gen/avx-2x8.c",
1978 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001979 "src/f32-relu/gen/avx-x16.c",
1980 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001982 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1983 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1984 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1985 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1986 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1987 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1988 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1989 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001990 "src/f32-vbinary/gen/vmax-avx-x8.c",
1991 "src/f32-vbinary/gen/vmax-avx-x16.c",
1992 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1993 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1994 "src/f32-vbinary/gen/vmin-avx-x8.c",
1995 "src/f32-vbinary/gen/vmin-avx-x16.c",
1996 "src/f32-vbinary/gen/vminc-avx-x8.c",
1997 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001998 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1999 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2000 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2001 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2002 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2003 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2004 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2005 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002006 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2007 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2008 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2009 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002010 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2011 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2012 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2013 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002014 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2015 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002016 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002017 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2018 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2019 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2020 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2021 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2022 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2023 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2024 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002025 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2026 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002027 "src/f32-vunary/gen/vabs-avx-x8.c",
2028 "src/f32-vunary/gen/vabs-avx-x16.c",
2029 "src/f32-vunary/gen/vneg-avx-x8.c",
2030 "src/f32-vunary/gen/vneg-avx-x16.c",
2031 "src/f32-vunary/gen/vsqr-avx-x8.c",
2032 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002033]
2034
Marat Dukhan1566fee2020-08-02 21:55:41 -07002035XOP_UKERNELS = [
2036 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2037 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2038 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2039 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002040 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2041 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002042 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2043 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002044 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002045 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2046 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002047 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002048 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2049 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2050 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002051 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2052 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2053 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2054 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2055 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2056 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002057 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002058 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2059 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002060 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002061]
2062
Marat Dukhanfda12b82019-11-21 12:27:59 -08002063FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002064 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2065 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2066 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2067 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2068 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2069 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2070 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2071 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2072 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2073 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2074 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2075 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2076 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2077 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2078 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2079 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2080 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2081 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2082 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2083 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2084 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2085 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2086 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2087 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2088 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2089 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2090 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2091 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2092 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2093 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2094 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2095 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2096 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2097 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2098 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2099 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2100 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2101 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2102 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2103 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002104 "src/f32-hswish/gen/hswish-fma3-x8.c",
2105 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002106 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2107 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2108 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2109 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2110 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2111 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2112 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2113 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2114 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2115 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2116 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2117 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2118 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2119 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002120 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2121 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2122 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2123 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2124 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2125 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2126 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2127 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002128 "src/math/sqrt-fma3-nr1fma.c",
2129 "src/math/sqrt-fma3-nr2fma.c",
2130 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002131]
2132
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002133AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002134 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2135 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2136 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2137 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2138 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2139 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2140 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2141 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2142 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2143 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2144 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2145 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2146 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2147 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2148 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2149 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2150 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2151 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2152 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2153 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2154 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2155 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2156 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2157 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2158 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2159 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2160 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2162 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2163 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2164 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2165 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2166 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2167 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2168 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2169 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002170 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2171 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2172 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2173 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2174 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2175 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2176 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2177 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2178 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2179 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2180 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2181 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2192 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2193 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2194 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2195 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2196 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2197 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2198 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2199 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002200 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2201 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2202 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2203 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2204 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2205 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2206 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2207 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2208 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2209 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2210 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2211 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2212 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2213 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2214 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2215 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2216 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2217 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2218 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2219 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2220 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2221 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2222 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2223 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002224 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2225 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002226 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2227 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2228 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2229 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002230 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2231 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2232 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002233 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2234 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2235 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002236 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2237 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2238 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002239 "src/math/exp-avx2-p5.c",
2240 "src/math/exp-avx2-perm-p3.c",
2241 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002242 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002243 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002244 "src/math/sigmoid-avx2-rr2-p5-div.c",
2245 "src/math/sigmoid-avx2-rr1-p5-div.c",
2246 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2247 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2248 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2249 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002250]
2251
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002253 "src/f32-clamp/gen/avx512f-x16.c",
2254 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002255 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2256 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2257 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2258 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2259 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2260 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2261 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2262 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2263 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2264 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2265 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2266 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2267 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2268 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2269 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2270 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2271 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2272 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2273 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2274 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2275 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2276 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2277 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2278 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002279 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2280 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002281 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2282 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2283 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2284 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2285 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2286 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002287 "src/f32-prelu/gen/avx512f-2x16.c",
2288 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002289 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2290 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2291 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2292 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2293 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2294 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2295 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2296 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2297 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2298 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2299 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2300 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2301 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2302 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2303 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2304 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2305 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2306 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2307 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2308 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2309 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2310 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2311 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2312 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2313 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2314 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2315 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2316 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2317 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2318 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2319 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2320 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2321 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2322 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2323 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2324 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002325 "src/f32-relu/gen/avx512f-x16.c",
2326 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002328 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2329 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2330 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2331 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2332 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2333 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2334 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2335 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002336 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2337 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2338 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2339 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2340 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2341 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2342 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2343 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002344 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2345 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2346 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2347 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2348 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2349 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2350 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2351 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002352 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2353 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2354 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2355 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002356 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2357 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2358 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2359 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002360 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2361 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002362 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002363 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2364 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2365 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2366 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2367 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2368 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2369 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2370 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2371 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2372 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2373 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2374 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2375 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2376 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2377 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2378 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2379 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2380 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2381 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2382 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2383 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2384 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2385 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2386 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002387 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2388 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2389 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2390 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2391 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2392 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2393 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2394 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002395 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2396 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2397 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2398 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2399 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2400 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2401 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2402 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002403 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2404 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2405 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2406 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2407 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2408 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002409 "src/math/exp-avx512f-p5-scalef.c",
2410 "src/math/exp-avx512f-p5.c",
2411 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002412 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002413 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002414 "src/math/sqrt-avx512f-nr1fma.c",
2415 "src/math/sqrt-avx512f-nr2fma.c",
2416 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417]
2418
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002419AVX512SKX_UKERNELS = [
2420 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2421 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2422 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2423 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2424 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2425 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2426 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2427 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2428]
2429
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002431 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002432 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002433 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2434 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002435 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002436 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002437 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002438 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002439 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2440 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002441 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2442 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2443 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2444 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002445]
2446
2447AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002448 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002449 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2450 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002451 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002452 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2453 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002454 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2455 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2456 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002457 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002458 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2459 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2460 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002461 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002462 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2463 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002464 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002465 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2466 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2467 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2468 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2469 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2470 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2471 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2472 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2473 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2474 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2475 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2476 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2477 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2478 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2479 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2480 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2481 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2482 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002483 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2484 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002485 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002486 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2487 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2488 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2489 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2490 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2491 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2492 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2493 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2494 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2495 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2496 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2497 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2498 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2499 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2500 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2501 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2502 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2503 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002504 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2505 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2506 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2507 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2508 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2509 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2510 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2511 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2512 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2513 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2514 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2515 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2516 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2517 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2518 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2519 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2520 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2521 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002522]
2523
Marat Dukhan1b354632020-03-23 12:50:22 -07002524INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002525 "src/xnnpack/argmaxpool.h",
2526 "src/xnnpack/avgpool.h",
2527 "src/xnnpack/clamp.h",
2528 "src/xnnpack/common.h",
2529 "src/xnnpack/conv.h",
2530 "src/xnnpack/dwconv.h",
2531 "src/xnnpack/gavgpool.h",
2532 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002533 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002535 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002537 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538 "src/xnnpack/lut.h",
2539 "src/xnnpack/math.h",
2540 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002541 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 "src/xnnpack/packx.h",
2543 "src/xnnpack/pad.h",
2544 "src/xnnpack/params.h",
2545 "src/xnnpack/pavgpool.h",
2546 "src/xnnpack/ppmm.h",
2547 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002548 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002549 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002550 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002551 "src/xnnpack/rmax.h",
2552 "src/xnnpack/scalar-utils.h",
2553 "src/xnnpack/spmm.h",
2554 "src/xnnpack/unpool.h",
2555 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002556 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002558 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002559 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002560 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002561 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002563]
2564
2565INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002566 "include/xnnpack.h",
2567 "src/xnnpack/allocator.h",
2568 "src/xnnpack/compute.h",
2569 "src/xnnpack/im2col.h",
2570 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002571 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002572 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002573 "src/xnnpack/operator.h",
2574 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002575 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002577 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002578 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002579]
2580
Marat Dukhan1b354632020-03-23 12:50:22 -07002581ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002582 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583]
2584
Marat Dukhan1b354632020-03-23 12:50:22 -07002585MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002586 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002587 "include/xnnpack.h",
2588]
2589
Marat Dukhan1b354632020-03-23 12:50:22 -07002590MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002591 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002592 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593 "src/xnnpack/requantization.h",
2594 "include/xnnpack.h",
2595]
2596
2597OPERATOR_TEST_PARAMS_HDRS = [
2598 "src/xnnpack/params.h",
2599 "src/xnnpack/common.h",
2600]
2601
2602WEIGHTS_PACK_HDRS = [
2603 "src/xnnpack/pack.h",
2604 "src/xnnpack/operator.h",
2605 "src/xnnpack/compute.h",
2606]
2607
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002608LOGGING_COPTS = select({
2609 # No logging in optimized mode
2610 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2611 # Full logging in debug mode
2612 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2613 # Error-only logging in default (fastbuild) mode
2614 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2615})
2616
Marat Dukhan3b59de22020-06-03 20:15:19 -07002617LOGGING_SRCS = select({
2618 # No logging in optimized mode
2619 ":optimized_build": [],
2620 "//conditions:default": [
2621 "src/operator-strings.c",
2622 "src/subgraph-strings.c",
2623 ],
2624})
2625
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002626LOGGING_HDRS = [
2627 "src/xnnpack/log.h",
2628]
2629
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002631 name = "tables",
2632 srcs = TABLE_SRCS,
2633 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002634 gcc_copts = xnnpack_gcc_std_copts(),
2635 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002636)
2637
2638xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 name = "scalar_ukernels",
2640 srcs = SCALAR_UKERNELS,
2641 hdrs = INTERNAL_HDRS,
2642 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002643 gcc_copts = xnnpack_gcc_std_copts(),
2644 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002646 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002647 "@FP16",
2648 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002649 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002650 ],
2651)
2652
2653xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002654 name = "scalar_ukernels_test_mode",
2655 srcs = SCALAR_UKERNELS,
2656 hdrs = INTERNAL_HDRS,
2657 aarch32_copts = ["-marm"],
2658 copts = [
2659 "-UNDEBUG",
2660 "-DXNN_TEST_MODE=1",
2661 ],
2662 gcc_copts = xnnpack_gcc_std_copts(),
2663 msvc_copts = xnnpack_msvc_std_copts(),
2664 deps = [
2665 ":tables",
2666 "@FP16",
2667 "@FXdiv",
2668 "@pthreadpool",
2669 ],
2670)
2671
2672xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002673 name = "wasm_ukernels",
2674 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002675 gcc_copts = xnnpack_gcc_std_copts(),
2676 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002677 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002678 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002679 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002680 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002681 "@FP16",
2682 "@FXdiv",
2683 "@pthreadpool",
2684 ],
2685)
2686
2687xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002688 name = "wasm_ukernels_test_mode",
2689 hdrs = INTERNAL_HDRS,
2690 copts = [
2691 "-UNDEBUG",
2692 "-DXNN_TEST_MODE=1",
2693 ],
2694 gcc_copts = xnnpack_gcc_std_copts(),
2695 msvc_copts = xnnpack_msvc_std_copts(),
2696 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002697 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002698 deps = [
2699 ":tables",
2700 "@FP16",
2701 "@FXdiv",
2702 "@pthreadpool",
2703 ],
2704)
2705
2706xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002707 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002708 hdrs = INTERNAL_HDRS,
2709 aarch32_copts = [
2710 "-marm",
2711 "-mfpu=neon",
2712 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002713 gcc_copts = xnnpack_gcc_std_copts(),
2714 gcc_x86_copts = ["-msse2"],
2715 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002716 optimized_copts = [
2717 "-O3",
2718 "-ffast-math",
2719 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002720 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002721 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002722 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002723 "@FP16",
2724 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002725 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 ],
2727)
2728
2729xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002730 name = "psimd_fastmath_ukernels_test_mode",
2731 hdrs = INTERNAL_HDRS,
2732 aarch32_copts = [
2733 "-marm",
2734 "-mfpu=neon",
2735 ],
2736 copts = [
2737 "-UNDEBUG",
2738 "-DXNN_TEST_MODE=1",
2739 ],
2740 gcc_copts = xnnpack_gcc_std_copts(),
2741 gcc_x86_copts = ["-msse2"],
2742 msvc_copts = xnnpack_msvc_std_copts(),
2743 optimized_copts = [
2744 "-O3",
2745 "-ffast-math",
2746 ],
2747 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2748 deps = [
2749 ":tables",
2750 "@FP16",
2751 "@psimd",
2752 "@pthreadpool",
2753 ],
2754)
2755
2756xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002757 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002758 hdrs = INTERNAL_HDRS,
2759 aarch32_copts = [
2760 "-marm",
2761 "-mfpu=neon",
2762 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002763 gcc_copts = xnnpack_gcc_std_copts(),
2764 gcc_x86_copts = ["-msse2"],
2765 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002766 optimized_copts = [
2767 "-O3",
2768 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002769 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002770 deps = [
2771 ":tables",
2772 "@FP16",
2773 "@psimd",
2774 "@pthreadpool",
2775 ],
2776)
2777
2778xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002779 name = "psimd_accmath_ukernels_test_mode",
2780 hdrs = INTERNAL_HDRS,
2781 aarch32_copts = [
2782 "-marm",
2783 "-mfpu=neon",
2784 ],
2785 copts = [
2786 "-UNDEBUG",
2787 "-DXNN_TEST_MODE=1",
2788 ],
2789 gcc_copts = xnnpack_gcc_std_copts(),
2790 gcc_x86_copts = ["-msse2"],
2791 msvc_copts = xnnpack_msvc_std_copts(),
2792 optimized_copts = [
2793 "-O3",
2794 ],
2795 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2796 deps = [
2797 ":tables",
2798 "@FP16",
2799 "@psimd",
2800 "@pthreadpool",
2801 ],
2802)
2803
2804xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002805 name = "neon_ukernels",
2806 hdrs = INTERNAL_HDRS,
2807 aarch32_copts = [
2808 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002809 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 "-mfpu=neon",
2811 ],
2812 aarch32_srcs = NEON_UKERNELS,
2813 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002814 gcc_copts = xnnpack_gcc_std_copts(),
2815 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002816 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002817 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002818 "@FP16",
2819 "@pthreadpool",
2820 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821)
2822
2823xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002824 name = "neon_ukernels_test_mode",
2825 hdrs = INTERNAL_HDRS,
2826 aarch32_copts = [
2827 "-marm",
2828 "-march=armv7-a",
2829 "-mfpu=neon",
2830 ],
2831 aarch32_srcs = NEON_UKERNELS,
2832 aarch64_srcs = NEON_UKERNELS,
2833 copts = [
2834 "-UNDEBUG",
2835 "-DXNN_TEST_MODE=1",
2836 ],
2837 gcc_copts = xnnpack_gcc_std_copts(),
2838 msvc_copts = xnnpack_msvc_std_copts(),
2839 deps = [
2840 ":tables",
2841 "@FP16",
2842 "@pthreadpool",
2843 ],
2844)
2845
2846xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002847 name = "neonfma_ukernels",
2848 hdrs = INTERNAL_HDRS,
2849 aarch32_copts = [
2850 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002851 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002852 "-mfpu=neon-vfpv4",
2853 ],
2854 aarch32_srcs = NEONFMA_UKERNELS,
2855 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002856 apple_aarch32_copts = [
2857 "-mcpu=swift",
2858 "-mtune=generic",
2859 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002860 gcc_copts = xnnpack_gcc_std_copts(),
2861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002863 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002864 "@FP16",
2865 "@pthreadpool",
2866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867)
2868
2869xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002870 name = "neonfma_ukernels_test_mode",
2871 hdrs = INTERNAL_HDRS,
2872 aarch32_copts = [
2873 "-marm",
2874 "-march=armv7-a",
2875 "-mfpu=neon-vfpv4",
2876 ],
2877 aarch32_srcs = NEONFMA_UKERNELS,
2878 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002879 apple_aarch32_copts = [
2880 "-mcpu=swift",
2881 "-mtune=generic",
2882 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002883 copts = [
2884 "-UNDEBUG",
2885 "-DXNN_TEST_MODE=1",
2886 ],
2887 gcc_copts = xnnpack_gcc_std_copts(),
2888 msvc_copts = xnnpack_msvc_std_copts(),
2889 deps = [
2890 ":tables",
2891 "@FP16",
2892 "@pthreadpool",
2893 ],
2894)
2895
2896xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002897 name = "neonv8_ukernels",
2898 hdrs = INTERNAL_HDRS,
2899 aarch32_copts = [
2900 "-marm",
2901 "-march=armv8-a",
2902 "-mfpu=neon-fp-armv8",
2903 ],
2904 aarch32_srcs = NEONV8_UKERNELS,
2905 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002906 apple_aarch32_copts = [
2907 "-mcpu=cyclone",
2908 "-mtune=generic",
2909 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002910 gcc_copts = xnnpack_gcc_std_copts(),
2911 msvc_copts = xnnpack_msvc_std_copts(),
2912 deps = [
2913 ":tables",
2914 "@FP16",
2915 "@pthreadpool",
2916 ],
2917)
2918
2919xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002920 name = "neonv8_ukernels_test_mode",
2921 hdrs = INTERNAL_HDRS,
2922 aarch32_copts = [
2923 "-marm",
2924 "-march=armv8-a",
2925 "-mfpu=neon-fp-armv8",
2926 ],
2927 aarch32_srcs = NEONV8_UKERNELS,
2928 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002929 apple_aarch32_copts = [
2930 "-mcpu=cyclone",
2931 "-mtune=generic",
2932 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002933 copts = [
2934 "-UNDEBUG",
2935 "-DXNN_TEST_MODE=1",
2936 ],
2937 gcc_copts = xnnpack_gcc_std_copts(),
2938 msvc_copts = xnnpack_msvc_std_copts(),
2939 deps = [
2940 ":tables",
2941 "@FP16",
2942 "@pthreadpool",
2943 ],
2944)
2945
2946xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002947 name = "neonfp16arith_ukernels",
2948 hdrs = INTERNAL_HDRS,
2949 aarch64_copts = ["-march=armv8.2-a+fp16"],
2950 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002951 gcc_copts = xnnpack_gcc_std_copts(),
2952 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002953 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002954 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002955 "@FP16",
2956 "@pthreadpool",
2957 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958)
2959
2960xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002961 name = "neonfp16arith_ukernels_test_mode",
2962 hdrs = INTERNAL_HDRS,
2963 aarch64_copts = ["-march=armv8.2-a+fp16"],
2964 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2965 copts = [
2966 "-UNDEBUG",
2967 "-DXNN_TEST_MODE=1",
2968 ],
2969 gcc_copts = xnnpack_gcc_std_copts(),
2970 msvc_copts = xnnpack_msvc_std_copts(),
2971 deps = [
2972 ":tables",
2973 "@FP16",
2974 "@pthreadpool",
2975 ],
2976)
2977
2978xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 name = "sse2_ukernels",
2980 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002981 gcc_copts = xnnpack_gcc_std_copts(),
2982 gcc_x86_copts = ["-msse2"],
2983 msvc_copts = xnnpack_msvc_std_copts(),
2984 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002986 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002987 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002988 "@FP16",
2989 "@pthreadpool",
2990 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991)
2992
2993xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002994 name = "sse2_ukernels_test_mode",
2995 hdrs = INTERNAL_HDRS,
2996 copts = [
2997 "-UNDEBUG",
2998 "-DXNN_TEST_MODE=1",
2999 ],
3000 gcc_copts = xnnpack_gcc_std_copts(),
3001 gcc_x86_copts = ["-msse2"],
3002 msvc_copts = xnnpack_msvc_std_copts(),
3003 msvc_x86_32_copts = ["/arch:SSE2"],
3004 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3005 deps = [
3006 ":tables",
3007 "@FP16",
3008 "@pthreadpool",
3009 ],
3010)
3011
3012xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003013 name = "ssse3_ukernels",
3014 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003015 gcc_copts = xnnpack_gcc_std_copts(),
3016 gcc_x86_copts = ["-mssse3"],
3017 msvc_copts = xnnpack_msvc_std_copts(),
3018 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003019 x86_srcs = SSSE3_UKERNELS,
3020 deps = [
3021 ":tables",
3022 "@FP16",
3023 "@pthreadpool",
3024 ],
3025)
3026
3027xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003028 name = "ssse3_ukernels_test_mode",
3029 hdrs = INTERNAL_HDRS,
3030 copts = [
3031 "-UNDEBUG",
3032 "-DXNN_TEST_MODE=1",
3033 ],
3034 gcc_copts = xnnpack_gcc_std_copts(),
3035 gcc_x86_copts = ["-mssse3"],
3036 msvc_copts = xnnpack_msvc_std_copts(),
3037 msvc_x86_32_copts = ["/arch:SSE2"],
3038 x86_srcs = SSSE3_UKERNELS,
3039 deps = [
3040 ":tables",
3041 "@FP16",
3042 "@pthreadpool",
3043 ],
3044)
3045
3046xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003047 name = "sse41_ukernels",
3048 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003049 gcc_copts = xnnpack_gcc_std_copts(),
3050 gcc_x86_copts = ["-msse4.1"],
3051 msvc_copts = xnnpack_msvc_std_copts(),
3052 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003053 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003054 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003055 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003056 "@FP16",
3057 "@pthreadpool",
3058 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003059)
3060
3061xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003062 name = "sse41_ukernels_test_mode",
3063 hdrs = INTERNAL_HDRS,
3064 copts = [
3065 "-UNDEBUG",
3066 "-DXNN_TEST_MODE=1",
3067 ],
3068 gcc_copts = xnnpack_gcc_std_copts(),
3069 gcc_x86_copts = ["-msse4.1"],
3070 msvc_copts = xnnpack_msvc_std_copts(),
3071 msvc_x86_32_copts = ["/arch:SSE2"],
3072 x86_srcs = SSE41_UKERNELS,
3073 deps = [
3074 ":tables",
3075 "@FP16",
3076 "@pthreadpool",
3077 ],
3078)
3079
3080xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081 name = "avx_ukernels",
3082 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003083 gcc_copts = xnnpack_gcc_std_copts(),
3084 gcc_x86_copts = ["-mavx"],
3085 msvc_copts = xnnpack_msvc_std_copts(),
3086 msvc_x86_32_copts = ["/arch:AVX"],
3087 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003088 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003089 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003090 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003091 "@FP16",
3092 "@pthreadpool",
3093 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094)
3095
3096xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003097 name = "avx_ukernels_test_mode",
3098 hdrs = INTERNAL_HDRS,
3099 copts = [
3100 "-UNDEBUG",
3101 "-DXNN_TEST_MODE=1",
3102 ],
3103 gcc_copts = xnnpack_gcc_std_copts(),
3104 gcc_x86_copts = ["-mavx"],
3105 msvc_copts = xnnpack_msvc_std_copts(),
3106 msvc_x86_32_copts = ["/arch:AVX"],
3107 msvc_x86_64_copts = ["/arch:AVX"],
3108 x86_srcs = AVX_UKERNELS,
3109 deps = [
3110 ":tables",
3111 "@FP16",
3112 "@pthreadpool",
3113 ],
3114)
3115
3116xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003117 name = "xop_ukernels",
3118 hdrs = INTERNAL_HDRS,
3119 gcc_copts = xnnpack_gcc_std_copts(),
3120 gcc_x86_copts = ["-mxop"],
3121 msvc_copts = xnnpack_msvc_std_copts(),
3122 msvc_x86_32_copts = ["/arch:AVX"],
3123 msvc_x86_64_copts = ["/arch:AVX"],
3124 x86_srcs = XOP_UKERNELS,
3125 deps = [
3126 ":tables",
3127 "@FP16",
3128 "@pthreadpool",
3129 ],
3130)
3131
3132xnnpack_cc_library(
3133 name = "xop_ukernels_test_mode",
3134 hdrs = INTERNAL_HDRS,
3135 copts = [
3136 "-UNDEBUG",
3137 "-DXNN_TEST_MODE=1",
3138 ],
3139 gcc_copts = xnnpack_gcc_std_copts(),
3140 gcc_x86_copts = ["-mxop"],
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 msvc_x86_32_copts = ["/arch:AVX"],
3143 msvc_x86_64_copts = ["/arch:AVX"],
3144 x86_srcs = XOP_UKERNELS,
3145 deps = [
3146 ":tables",
3147 "@FP16",
3148 "@pthreadpool",
3149 ],
3150)
3151
3152xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003153 name = "fma3_ukernels",
3154 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003155 gcc_copts = xnnpack_gcc_std_copts(),
3156 gcc_x86_copts = ["-mfma"],
3157 msvc_copts = xnnpack_msvc_std_copts(),
3158 msvc_x86_32_copts = ["/arch:AVX"],
3159 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003160 x86_srcs = FMA3_UKERNELS,
3161 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003162 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003163 "@FP16",
3164 "@pthreadpool",
3165 ],
3166)
3167
3168xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003169 name = "fma3_ukernels_test_mode",
3170 hdrs = INTERNAL_HDRS,
3171 copts = [
3172 "-UNDEBUG",
3173 "-DXNN_TEST_MODE=1",
3174 ],
3175 gcc_copts = xnnpack_gcc_std_copts(),
3176 gcc_x86_copts = ["-mfma"],
3177 msvc_copts = xnnpack_msvc_std_copts(),
3178 msvc_x86_32_copts = ["/arch:AVX"],
3179 msvc_x86_64_copts = ["/arch:AVX"],
3180 x86_srcs = FMA3_UKERNELS,
3181 deps = [
3182 ":tables",
3183 "@FP16",
3184 "@pthreadpool",
3185 ],
3186)
3187
3188xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003189 name = "avx2_ukernels",
3190 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003191 gcc_copts = xnnpack_gcc_std_copts(),
3192 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003193 "-mfma",
3194 "-mavx2",
3195 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003196 msvc_copts = xnnpack_msvc_std_copts(),
3197 msvc_x86_32_copts = ["/arch:AVX2"],
3198 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003199 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003200 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003201 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003202 "@FP16",
3203 "@pthreadpool",
3204 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003205)
3206
3207xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003208 name = "avx2_ukernels_test_mode",
3209 hdrs = INTERNAL_HDRS,
3210 copts = [
3211 "-UNDEBUG",
3212 "-DXNN_TEST_MODE=1",
3213 ],
3214 gcc_copts = xnnpack_gcc_std_copts(),
3215 gcc_x86_copts = [
3216 "-mfma",
3217 "-mavx2",
3218 ],
3219 msvc_copts = xnnpack_msvc_std_copts(),
3220 msvc_x86_32_copts = ["/arch:AVX2"],
3221 msvc_x86_64_copts = ["/arch:AVX2"],
3222 x86_srcs = AVX2_UKERNELS,
3223 deps = [
3224 ":tables",
3225 "@FP16",
3226 "@pthreadpool",
3227 ],
3228)
3229
3230xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 name = "avx512f_ukernels",
3232 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003233 gcc_copts = xnnpack_gcc_std_copts(),
3234 gcc_x86_copts = ["-mavx512f"],
3235 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 msvc_x86_32_copts = ["/arch:AVX512"],
3238 msvc_x86_64_copts = ["/arch:AVX512"],
3239 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003241 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003242 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003243 "@FP16",
3244 "@pthreadpool",
3245 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246)
3247
3248xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003249 name = "avx512f_ukernels_test_mode",
3250 hdrs = INTERNAL_HDRS,
3251 copts = [
3252 "-UNDEBUG",
3253 "-DXNN_TEST_MODE=1",
3254 ],
3255 gcc_copts = xnnpack_gcc_std_copts(),
3256 gcc_x86_copts = ["-mavx512f"],
3257 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3258 msvc_copts = xnnpack_msvc_std_copts(),
3259 msvc_x86_32_copts = ["/arch:AVX512"],
3260 msvc_x86_64_copts = ["/arch:AVX512"],
3261 msys_copts = ["-fno-asynchronous-unwind-tables"],
3262 x86_srcs = AVX512F_UKERNELS,
3263 deps = [
3264 ":tables",
3265 "@FP16",
3266 "@pthreadpool",
3267 ],
3268)
3269
3270xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003271 name = "avx512skx_ukernels",
3272 hdrs = INTERNAL_HDRS,
3273 gcc_copts = xnnpack_gcc_std_copts(),
3274 gcc_x86_copts = [
3275 "-mavx512f",
3276 "-mavx512cd",
3277 "-mavx512bw",
3278 "-mavx512dq",
3279 "-mavx512vl",
3280 ],
3281 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3282 msvc_copts = xnnpack_msvc_std_copts(),
3283 msvc_x86_32_copts = ["/arch:AVX512"],
3284 msvc_x86_64_copts = ["/arch:AVX512"],
3285 msys_copts = ["-fno-asynchronous-unwind-tables"],
3286 x86_srcs = AVX512SKX_UKERNELS,
3287 deps = [
3288 ":tables",
3289 "@FP16",
3290 "@pthreadpool",
3291 ],
3292)
3293
3294xnnpack_cc_library(
3295 name = "avx512skx_ukernels_test_mode",
3296 hdrs = INTERNAL_HDRS,
3297 copts = [
3298 "-UNDEBUG",
3299 "-DXNN_TEST_MODE=1",
3300 ],
3301 gcc_copts = xnnpack_gcc_std_copts(),
3302 gcc_x86_copts = [
3303 "-mavx512f",
3304 "-mavx512cd",
3305 "-mavx512bw",
3306 "-mavx512dq",
3307 "-mavx512vl",
3308 ],
3309 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3310 msvc_copts = xnnpack_msvc_std_copts(),
3311 msvc_x86_32_copts = ["/arch:AVX512"],
3312 msvc_x86_64_copts = ["/arch:AVX512"],
3313 msys_copts = ["-fno-asynchronous-unwind-tables"],
3314 x86_srcs = AVX512SKX_UKERNELS,
3315 deps = [
3316 ":tables",
3317 "@FP16",
3318 "@pthreadpool",
3319 ],
3320)
3321
3322xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003323 name = "asm_ukernels",
3324 hdrs = ["src/xnnpack/assembly.h"],
3325 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003326 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003327 aarch64_srcs = AARCH64_ASM_UKERNELS,
3328)
3329
Marat Dukhan3b59de22020-06-03 20:15:19 -07003330xnnpack_cc_library(
3331 name = "logging_utils",
3332 srcs = LOGGING_SRCS,
3333 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3334 copts = LOGGING_COPTS + [
3335 "-Isrc",
3336 "-Iinclude",
3337 ] + select({
3338 ":debug_build": [],
3339 "//conditions:default": xnnpack_min_size_copts(),
3340 }),
3341 gcc_copts = xnnpack_gcc_std_copts(),
3342 msvc_copts = xnnpack_msvc_std_copts(),
3343 visibility = xnnpack_visibility(),
3344 deps = [
3345 "@FP16",
3346 "@clog",
3347 "@pthreadpool",
3348 ],
3349)
3350
Marat Dukhan08c4a432019-10-03 09:29:21 -07003351xnnpack_aggregate_library(
3352 name = "ukernels",
3353 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003354 ":neon_ukernels",
3355 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003356 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357 ":asm_ukernels",
3358 ],
3359 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 ":neon_ukernels",
3361 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003362 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 ":neonfp16arith_ukernels",
3364 ":asm_ukernels",
3365 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003366 generic_deps = [
3367 ":scalar_ukernels",
3368 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003369 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003370 ":psimd_fastmath_ukernels",
3371 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003373 wasm_deps = [
3374 ":wasm_ukernels",
3375 ],
3376 wasmsimd_deps = [
3377 ":wasm_ukernels",
3378 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003379 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003381 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003382 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003384 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003385 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003386 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003387 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003388 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003389 ],
3390)
3391
Marat Dukhan33fcf782020-05-24 14:27:15 -07003392xnnpack_aggregate_library(
3393 name = "ukernels_test_mode",
3394 aarch32_deps = [
3395 ":neon_ukernels_test_mode",
3396 ":neonfma_ukernels_test_mode",
3397 ":neonv8_ukernels_test_mode",
3398 ":asm_ukernels",
3399 ],
3400 aarch64_deps = [
3401 ":neon_ukernels_test_mode",
3402 ":neonfma_ukernels_test_mode",
3403 ":neonv8_ukernels_test_mode",
3404 ":neonfp16arith_ukernels_test_mode",
3405 ":asm_ukernels",
3406 ],
3407 generic_deps = [
3408 ":scalar_ukernels_test_mode",
3409 ],
3410 psimd_deps = [
3411 ":psimd_fastmath_ukernels_test_mode",
3412 ":psimd_accmath_ukernels_test_mode",
3413 ],
3414 wasm_deps = [
3415 ":wasm_ukernels_test_mode",
3416 ],
3417 wasmsimd_deps = [
3418 ":wasm_ukernels_test_mode",
3419 ],
3420 x86_deps = [
3421 ":sse2_ukernels_test_mode",
3422 ":ssse3_ukernels_test_mode",
3423 ":sse41_ukernels_test_mode",
3424 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003425 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003426 ":fma3_ukernels_test_mode",
3427 ":avx2_ukernels_test_mode",
3428 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003429 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003430 ],
3431)
3432
Marat Dukhan86df3842020-06-17 14:03:31 -07003433# This target must be compatible with Android NDK r17 and compile with Clang 6.
3434# It is similar to :ukernels target, but excludes microkernels which require
3435# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003436xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003437 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003438 aarch32_deps = [
3439 ":neon_ukernels",
3440 ":neonfma_ukernels",
3441 ":neonv8_ukernels",
3442 ":asm_ukernels",
3443 ],
3444 aarch64_deps = [
3445 ":neon_ukernels",
3446 ":neonfma_ukernels",
3447 ":neonv8_ukernels",
3448 ":asm_ukernels",
3449 ],
3450 generic_deps = [
3451 ":scalar_ukernels",
3452 ],
3453 psimd_deps = [
3454 ":psimd_fastmath_ukernels",
3455 ":psimd_accmath_ukernels",
3456 ],
3457 wasm_deps = [
3458 ":wasm_ukernels",
3459 ],
3460 wasmsimd_deps = [
3461 ":wasm_ukernels",
3462 ],
3463 x86_deps = [
3464 ":sse2_ukernels",
3465 ":ssse3_ukernels",
3466 ":sse41_ukernels",
3467 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003468 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003469 ":fma3_ukernels",
3470 ":avx2_ukernels",
3471 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003472 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003473 ],
3474)
3475
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476xnnpack_cc_library(
3477 name = "im2col",
3478 srcs = ["src/im2col.c"],
3479 hdrs = [
3480 "src/xnnpack/common.h",
3481 "src/xnnpack/im2col.h",
3482 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003483 gcc_copts = xnnpack_gcc_std_copts(),
3484 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485)
3486
3487xnnpack_cc_library(
3488 name = "indirection",
3489 srcs = ["src/indirection.c"],
3490 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003491 gcc_copts = xnnpack_gcc_std_copts(),
3492 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 deps = [
3494 "@FP16",
3495 "@FXdiv",
3496 "@pthreadpool",
3497 ],
3498)
3499
3500xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003501 name = "indirection_test_mode",
3502 srcs = ["src/indirection.c"],
3503 hdrs = INTERNAL_HDRS,
3504 copts = [
3505 "-UNDEBUG",
3506 "-DXNN_TEST_MODE=1",
3507 ],
3508 gcc_copts = xnnpack_gcc_std_copts(),
3509 msvc_copts = xnnpack_msvc_std_copts(),
3510 deps = [
3511 "@FP16",
3512 "@FXdiv",
3513 "@pthreadpool",
3514 ],
3515)
3516
3517xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003518 name = "packing",
3519 srcs = ["src/packing.c"],
3520 hdrs = INTERNAL_HDRS,
3521 gcc_copts = xnnpack_gcc_std_copts(),
3522 msvc_copts = xnnpack_msvc_std_copts(),
3523 deps = [
3524 "@FP16",
3525 "@FXdiv",
3526 "@pthreadpool",
3527 ],
3528)
3529
3530xnnpack_cc_library(
3531 name = "packing_test_mode",
3532 srcs = ["src/packing.c"],
3533 hdrs = INTERNAL_HDRS,
3534 copts = [
3535 "-UNDEBUG",
3536 "-DXNN_TEST_MODE=1",
3537 ],
3538 gcc_copts = xnnpack_gcc_std_copts(),
3539 msvc_copts = xnnpack_msvc_std_copts(),
3540 deps = [
3541 "@FP16",
3542 "@FXdiv",
3543 "@pthreadpool",
3544 ],
3545)
3546
3547xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003548 name = "operator_run",
3549 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003550 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003551 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003552 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3553 "//conditions:default": [],
3554 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003555 gcc_copts = xnnpack_gcc_std_copts(),
3556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003558 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559 "@FP16",
3560 "@FXdiv",
3561 "@clog",
3562 "@pthreadpool",
3563 ],
3564)
3565
Chao Mei6ddfc602020-05-13 22:29:36 -07003566xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003567 name = "operator_run_test_mode",
3568 srcs = ["src/operator-run.c"],
3569 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3570 copts = LOGGING_COPTS + [
3571 "-UNDEBUG",
3572 "-DXNN_TEST_MODE=1",
3573 ] + select({
3574 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3575 "//conditions:default": [],
3576 }),
3577 gcc_copts = xnnpack_gcc_std_copts(),
3578 msvc_copts = xnnpack_msvc_std_copts(),
3579 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003580 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003581 "@FP16",
3582 "@FXdiv",
3583 "@clog",
3584 "@pthreadpool",
3585 ],
3586)
3587
3588xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003589 name = "memory_planner",
3590 srcs = ["src/memory-planner.c"],
3591 hdrs = INTERNAL_HDRS,
3592 defines = select({
3593 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3594 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3595 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3596 }),
3597 gcc_copts = xnnpack_gcc_std_copts(),
3598 msvc_copts = xnnpack_msvc_std_copts(),
3599 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003600 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003601 "@pthreadpool",
3602 ],
3603)
3604
Marat Dukhan33fcf782020-05-24 14:27:15 -07003605xnnpack_cc_library(
3606 name = "memory_planner_test_mode",
3607 srcs = ["src/memory-planner.c"],
3608 hdrs = INTERNAL_HDRS,
3609 copts = [
3610 "-UNDEBUG",
3611 "-DXNN_TEST_MODE=1",
3612 ],
3613 defines = select({
3614 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3615 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3616 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3617 }),
3618 gcc_copts = xnnpack_gcc_std_copts(),
3619 msvc_copts = xnnpack_msvc_std_copts(),
3620 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003621 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003622 "@pthreadpool",
3623 ],
3624)
3625
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626cc_library(
3627 name = "enable_assembly",
3628 defines = select({
3629 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3630 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003631 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 }),
3633)
3634
Marat Dukhan9de90e02020-06-18 16:04:12 -07003635cc_library(
3636 name = "enable_sparse",
3637 defines = select({
3638 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3639 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003640 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003641 }),
3642)
3643
Marat Dukhancf056b22019-10-07 10:26:29 -07003644xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 name = "operators",
3646 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003647 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003649 ],
3650 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003651 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 "-Isrc",
3653 "-Iinclude",
3654 ] + select({
3655 ":debug_build": [],
3656 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003657 }) + select({
3658 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3659 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003661 gcc_copts = xnnpack_gcc_std_copts(),
3662 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003665 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003666 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003667 "@FP16",
3668 "@FXdiv",
3669 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003671 ],
3672)
3673
Marat Dukhan10a38082020-04-17 03:58:35 -07003674xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003675 name = "operators_test_mode",
3676 srcs = OPERATOR_SRCS + [
3677 "src/memory.c",
3678 "src/operator-delete.c",
3679 ],
3680 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3681 copts = LOGGING_COPTS + [
3682 "-Isrc",
3683 "-Iinclude",
3684 "-UNDEBUG",
3685 "-DXNN_TEST_MODE=1",
3686 ] + select({
3687 ":debug_build": [],
3688 "//conditions:default": xnnpack_min_size_copts(),
3689 }) + select({
3690 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3691 "//conditions:default": [],
3692 }),
3693 gcc_copts = xnnpack_gcc_std_copts(),
3694 msvc_copts = xnnpack_msvc_std_copts(),
3695 deps = [
3696 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003697 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003698 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003699 "@FP16",
3700 "@FXdiv",
3701 "@clog",
3702 "@pthreadpool",
3703 ],
3704)
3705
3706xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003707 name = "XNNPACK",
3708 srcs = [
3709 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003710 "src/runtime.c",
3711 "src/subgraph.c",
3712 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003713 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003714 hdrs = ["include/xnnpack.h"],
3715 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003716 "-Isrc",
3717 "-Iinclude",
3718 ] + select({
3719 ":debug_build": [],
3720 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003721 }) + select({
3722 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3723 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003724 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003725 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003726 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003728 visibility = xnnpack_visibility(),
3729 deps = [
3730 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003731 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003732 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003733 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003734 ":operator_run",
3735 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003736 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003737 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003738 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003739 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003740 ] + select({
3741 ":emscripten": [],
3742 "//conditions:default": ["@cpuinfo"],
3743 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003744)
3745
Marat Dukhan10a38082020-04-17 03:58:35 -07003746xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003747 name = "XNNPACK_test_mode",
3748 srcs = [
3749 "src/init.c",
3750 "src/runtime.c",
3751 "src/subgraph.c",
3752 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003753 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003754 hdrs = ["include/xnnpack.h"],
3755 copts = LOGGING_COPTS + [
3756 "-Isrc",
3757 "-Iinclude",
3758 "-UNDEBUG",
3759 "-DXNN_TEST_MODE=1",
3760 ] + select({
3761 ":debug_build": [],
3762 "//conditions:default": xnnpack_min_size_copts(),
3763 }) + select({
3764 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3765 "//conditions:default": [],
3766 }),
3767 gcc_copts = xnnpack_gcc_std_copts(),
3768 includes = ["include"],
3769 msvc_copts = xnnpack_msvc_std_copts(),
3770 visibility = xnnpack_visibility(),
3771 deps = [
3772 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003773 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003774 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003775 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003776 ":operator_run_test_mode",
3777 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003778 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003779 "@clog",
3780 "@FP16",
3781 "@pthreadpool",
3782 ] + select({
3783 ":emscripten": [],
3784 "//conditions:default": ["@cpuinfo"],
3785 }),
3786)
3787
3788xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003789 name = "xnnpack_f32",
3790 srcs = [
3791 "src/init.c",
3792 "src/runtime.c",
3793 "src/subgraph.c",
3794 "src/tensor.c",
3795 ] + SUBGRAPH_SRCS,
3796 hdrs = ["include/xnnpack.h"],
3797 copts = LOGGING_COPTS + [
3798 "-Isrc",
3799 "-Iinclude",
3800 ] + select({
3801 ":debug_build": [],
3802 "//conditions:default": xnnpack_min_size_copts(),
3803 }) + select({
3804 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3805 "//conditions:default": [],
3806 }),
3807 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003808 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003809 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003810 "XNN_NO_U8_OPERATORS",
3811 "XNN_NO_X8_OPERATORS",
3812 "XNN_NO_F16_OPERATORS",
3813 "XNN_NO_X16_OPERATORS",
3814 ],
3815 gcc_copts = xnnpack_gcc_std_copts(),
3816 includes = ["include"],
3817 msvc_copts = xnnpack_msvc_std_copts(),
3818 visibility = xnnpack_visibility(),
3819 deps = [
3820 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003821 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003822 ":logging_utils",
3823 ":memory_planner",
3824 ":operator_run",
3825 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003826 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003827 "@clog",
3828 "@FP16",
3829 "@pthreadpool",
3830 ] + select({
3831 ":emscripten": [],
3832 "//conditions:default": ["@cpuinfo"],
3833 }),
3834)
3835
3836xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003837 name = "xnnpack_operators_nhwc_f32",
3838 srcs = [
3839 "src/init.c",
3840 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003841 hdrs = ["include/xnnpack.h"],
3842 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003843 "-Isrc",
3844 "-Iinclude",
3845 ] + select({
3846 ":debug_build": [],
3847 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003848 }) + select({
3849 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3850 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003851 }),
3852 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003853 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003854 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003855 "XNN_NO_U8_OPERATORS",
3856 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003857 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003858 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003859 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862 visibility = xnnpack_visibility(),
3863 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003864 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003865 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866 ":operator_run",
3867 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003868 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003869 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003871 ] + select({
3872 ":emscripten": [],
3873 "//conditions:default": ["@cpuinfo"],
3874 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003875)
3876
Marat Dukhancf056b22019-10-07 10:26:29 -07003877xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003878 name = "bench_utils",
3879 srcs = ["bench/utils.cc"],
3880 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003881 deps = [
3882 "@com_google_benchmark//:benchmark",
3883 "@cpuinfo",
3884 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003885)
3886
Frank Barchard7e955972019-10-11 10:34:25 -07003887######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003888
3889xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003890 name = "qs8_gemm_bench",
3891 srcs = [
3892 "bench/gemm.h",
3893 "bench/qs8-gemm.cc",
3894 "src/xnnpack/AlignedAllocator.h",
3895 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3896 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3897)
3898
3899xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003900 name = "qs8_requantization_bench",
3901 srcs = [
3902 "bench/qs8-requantization.cc",
3903 "src/xnnpack/requantization-stubs.h",
3904 "src/xnnpack/AlignedAllocator.h",
3905 ] + MICROKERNEL_BENCHMARK_HDRS,
3906 deps = MICROKERNEL_BENCHMARK_DEPS,
3907)
3908
3909xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003910 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911 srcs = [
3912 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003913 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 "src/xnnpack/AlignedAllocator.h",
3915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003916 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003917 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003918)
3919
3920xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003921 name = "qu8_requantization_bench",
3922 srcs = [
3923 "bench/qu8-requantization.cc",
3924 "src/xnnpack/requantization-stubs.h",
3925 "src/xnnpack/AlignedAllocator.h",
3926 ] + MICROKERNEL_BENCHMARK_HDRS,
3927 deps = MICROKERNEL_BENCHMARK_DEPS,
3928)
3929
3930xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003931 name = "f16_igemm_bench",
3932 srcs = [
3933 "bench/f16-igemm.cc",
3934 "bench/conv.h",
3935 "bench/google/conv.h",
3936 "src/xnnpack/AlignedAllocator.h",
3937 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003938 deps = MICROKERNEL_BENCHMARK_DEPS + [
3939 ":indirection",
3940 ":packing",
3941 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003942)
3943
3944xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003945 name = "f16_gemm_bench",
3946 srcs = [
3947 "bench/f16-gemm.cc",
3948 "bench/gemm.h",
3949 "src/xnnpack/AlignedAllocator.h",
3950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003951 deps = MICROKERNEL_BENCHMARK_DEPS + [
3952 ":packing",
3953 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954)
3955
3956xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003957 name = "f16_spmm_bench",
3958 srcs = [
3959 "bench/f16-spmm.cc",
3960 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003961 "src/xnnpack/AlignedAllocator.h",
3962 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003963 deps = MICROKERNEL_BENCHMARK_DEPS,
3964)
3965
3966xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003967 name = "f32_igemm_bench",
3968 srcs = [
3969 "bench/f32-igemm.cc",
3970 "bench/conv.h",
3971 "src/xnnpack/AlignedAllocator.h",
3972 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003973 deps = MICROKERNEL_BENCHMARK_DEPS + [
3974 ":indirection",
3975 ":packing",
3976 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003977)
3978
3979xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003980 name = "f16_relu_bench",
3981 srcs = [
3982 "bench/f16-relu.cc",
3983 "src/xnnpack/AlignedAllocator.h",
3984 ] + MICROKERNEL_BENCHMARK_HDRS,
3985 deps = MICROKERNEL_BENCHMARK_DEPS,
3986)
3987
3988xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003989 name = "f32_conv_hwc_bench",
3990 srcs = [
3991 "bench/f32-conv-hwc.cc",
3992 "bench/dconv.h",
3993 "src/xnnpack/AlignedAllocator.h",
3994 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003995 deps = MICROKERNEL_BENCHMARK_DEPS + [
3996 ":packing",
3997 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003998)
3999
4000xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004001 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004002 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004003 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004004 "bench/dconv.h",
4005 "src/xnnpack/AlignedAllocator.h",
4006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004007 deps = MICROKERNEL_BENCHMARK_DEPS + [
4008 ":packing",
4009 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004010)
4011
4012xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004013 name = "f16_dwconv_bench",
4014 srcs = [
4015 "bench/f16-dwconv.cc",
4016 "bench/dwconv.h",
4017 "bench/google/dwconv.h",
4018 "src/xnnpack/AlignedAllocator.h",
4019 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004020 deps = MICROKERNEL_BENCHMARK_DEPS + [
4021 ":indirection",
4022 ":packing",
4023 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004024)
4025
4026xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004027 name = "f32_dwconv_bench",
4028 srcs = [
4029 "bench/f32-dwconv.cc",
4030 "bench/dwconv.h",
4031 "src/xnnpack/AlignedAllocator.h",
4032 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004033 deps = MICROKERNEL_BENCHMARK_DEPS + [
4034 ":indirection",
4035 ":packing",
4036 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004037)
4038
4039xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004040 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004041 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004042 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004043 "bench/dwconv.h",
4044 "src/xnnpack/AlignedAllocator.h",
4045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004046 deps = MICROKERNEL_BENCHMARK_DEPS + [
4047 ":indirection",
4048 ":packing",
4049 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004050)
4051
4052xnnpack_benchmark(
4053 name = "f32_gemm_bench",
4054 srcs = [
4055 "bench/f32-gemm.cc",
4056 "bench/gemm.h",
4057 "src/xnnpack/AlignedAllocator.h",
4058 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004059 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004060 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004061)
4062
4063xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004064 name = "f32_hswish_bench",
4065 srcs = [
4066 "bench/f32-hswish.cc",
4067 "src/xnnpack/AlignedAllocator.h",
4068 ] + MICROKERNEL_BENCHMARK_HDRS,
4069 deps = MICROKERNEL_BENCHMARK_DEPS,
4070)
4071
4072xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004073 name = "f32_raddexpminusmax_bench",
4074 srcs = [
4075 "bench/f32-raddexpminusmax.cc",
4076 "src/xnnpack/AlignedAllocator.h",
4077 ] + MICROKERNEL_BENCHMARK_HDRS,
4078 deps = MICROKERNEL_BENCHMARK_DEPS,
4079)
4080
4081xnnpack_benchmark(
4082 name = "f32_raddextexp_bench",
4083 srcs = [
4084 "bench/f32-raddextexp.cc",
4085 "src/xnnpack/AlignedAllocator.h",
4086 ] + MICROKERNEL_BENCHMARK_HDRS,
4087 deps = MICROKERNEL_BENCHMARK_DEPS,
4088)
4089
4090xnnpack_benchmark(
4091 name = "f32_raddstoreexpminusmax_bench",
4092 srcs = [
4093 "bench/f32-raddstoreexpminusmax.cc",
4094 "src/xnnpack/AlignedAllocator.h",
4095 ] + MICROKERNEL_BENCHMARK_HDRS,
4096 deps = MICROKERNEL_BENCHMARK_DEPS,
4097)
4098
4099xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004100 name = "f32_relu_bench",
4101 srcs = [
4102 "bench/f32-relu.cc",
4103 "src/xnnpack/AlignedAllocator.h",
4104 ] + MICROKERNEL_BENCHMARK_HDRS,
4105 deps = MICROKERNEL_BENCHMARK_DEPS,
4106)
4107
4108xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004109 name = "f32_rmax_bench",
4110 srcs = [
4111 "bench/f32-rmax.cc",
4112 "src/xnnpack/AlignedAllocator.h",
4113 ] + MICROKERNEL_BENCHMARK_HDRS,
4114 deps = MICROKERNEL_BENCHMARK_DEPS,
4115)
4116
4117xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004118 name = "f32_sigmoid_bench",
4119 srcs = [
4120 "bench/f32-sigmoid.cc",
4121 "src/xnnpack/AlignedAllocator.h",
4122 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004123 deps = MICROKERNEL_BENCHMARK_DEPS,
4124)
4125
4126xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 name = "f32_spmm_bench",
4128 srcs = [
4129 "bench/f32-spmm.cc",
4130 "bench/gemm.h",
4131 "src/xnnpack/AlignedAllocator.h",
4132 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004133 deps = MICROKERNEL_BENCHMARK_DEPS,
4134)
4135
4136xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004137 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004138 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004139 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004140 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004141 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004142 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004143)
4144
4145xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004146 name = "f32_vscaleexpminusmax_bench",
4147 srcs = [
4148 "bench/f32-vscaleexpminusmax.cc",
4149 "src/xnnpack/AlignedAllocator.h",
4150 ] + MICROKERNEL_BENCHMARK_HDRS,
4151 deps = MICROKERNEL_BENCHMARK_DEPS,
4152)
4153
4154xnnpack_benchmark(
4155 name = "f32_vscaleextexp_bench",
4156 srcs = [
4157 "bench/f32-vscaleextexp.cc",
4158 "src/xnnpack/AlignedAllocator.h",
4159 ] + MICROKERNEL_BENCHMARK_HDRS,
4160 deps = MICROKERNEL_BENCHMARK_DEPS,
4161)
4162
4163xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004164 name = "f32_vsqrt_bench",
4165 srcs = [
4166 "bench/f32-vsqrt.cc",
4167 "src/xnnpack/AlignedAllocator.h",
4168 ] + MICROKERNEL_BENCHMARK_HDRS,
4169 deps = MICROKERNEL_BENCHMARK_DEPS,
4170)
4171
4172xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004173 name = "f32_im2col_gemm_bench",
4174 srcs = [
4175 "bench/f32-im2col-gemm.cc",
4176 "bench/conv.h",
4177 "src/xnnpack/AlignedAllocator.h",
4178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004179 deps = MICROKERNEL_BENCHMARK_DEPS + [
4180 ":im2col",
4181 ":packing",
4182 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004183)
4184
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004185xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004186 name = "rounding_bench",
4187 srcs = [
4188 "bench/rounding.cc",
4189 "src/xnnpack/math-stubs.h",
4190 "src/xnnpack/AlignedAllocator.h",
4191 ] + MICROKERNEL_BENCHMARK_HDRS,
4192 deps = MICROKERNEL_BENCHMARK_DEPS,
4193)
4194
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195########################### Benchmarks for operators ###########################
4196
4197xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 name = "average_pooling_bench",
4199 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004200 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004201 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004202 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203)
4204
4205xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004206 name = "bankers_rounding_bench",
4207 srcs = ["bench/bankers-rounding.cc"],
4208 copts = xnnpack_optional_tflite_copts(),
4209 tags = ["nowin32"],
4210 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4211)
4212
4213xnnpack_benchmark(
4214 name = "ceiling_bench",
4215 srcs = ["bench/ceiling.cc"],
4216 copts = xnnpack_optional_tflite_copts(),
4217 tags = ["nowin32"],
4218 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4219)
4220
4221xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004222 name = "channel_shuffle_bench",
4223 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004224 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004225)
4226
4227xnnpack_benchmark(
4228 name = "convolution_bench",
4229 srcs = ["bench/convolution.cc"],
4230 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004231 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004232 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233)
4234
4235xnnpack_benchmark(
4236 name = "deconvolution_bench",
4237 srcs = ["bench/deconvolution.cc"],
4238 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004239 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004240 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004241)
4242
4243xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004244 name = "floor_bench",
4245 srcs = ["bench/floor.cc"],
4246 copts = xnnpack_optional_tflite_copts(),
4247 tags = ["nowin32"],
4248 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4249)
4250
4251xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004252 name = "global_average_pooling_bench",
4253 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004254 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255)
4256
4257xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004258 name = "hardswish_bench",
4259 srcs = ["bench/hardswish.cc"],
4260 copts = xnnpack_optional_tflite_copts(),
4261 tags = ["nowin32"],
4262 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4263)
4264
4265xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004266 name = "max_pooling_bench",
4267 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004268 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269)
4270
4271xnnpack_benchmark(
4272 name = "sigmoid_bench",
4273 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004274 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004275 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004276 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277)
4278
4279xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004280 name = "prelu_bench",
4281 srcs = ["bench/prelu.cc"],
4282 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004283 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004284 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004285)
4286
4287xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004288 name = "softmax_bench",
4289 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004290 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004291 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004292 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004293)
4294
Marat Dukhan87727142020-06-24 15:24:10 -07004295xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004296 name = "square_root_bench",
4297 srcs = ["bench/square-root.cc"],
4298 copts = xnnpack_optional_tflite_copts(),
4299 tags = ["nowin32"],
4300 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4301)
4302
4303xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004304 name = "truncation_bench",
4305 srcs = ["bench/truncation.cc"],
4306 deps = OPERATOR_BENCHMARK_DEPS,
4307)
4308
Marat Dukhanc068bb62019-10-04 13:24:39 -07004309############################# End-to-end benchmarks ############################
4310
4311cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004312 name = "fp32_mobilenet_v1",
4313 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004314 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004315 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004316 linkstatic = True,
4317 deps = [
4318 ":XNNPACK",
4319 "@pthreadpool",
4320 ],
4321)
4322
4323cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004324 name = "fp16_mobilenet_v1",
4325 srcs = ["models/fp16-mobilenet-v1.cc"],
4326 hdrs = ["models/models.h"],
4327 copts = xnnpack_std_cxxopts(),
4328 linkstatic = True,
4329 deps = [
4330 ":XNNPACK",
4331 "@FP16",
4332 "@pthreadpool",
4333 ],
4334)
4335
4336cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004337 name = "qs8_mobilenet_v1",
4338 srcs = ["models/qs8-mobilenet-v1.cc"],
4339 hdrs = ["models/models.h"],
4340 copts = xnnpack_std_cxxopts(),
4341 linkstatic = True,
4342 deps = [
4343 ":XNNPACK",
4344 "@pthreadpool",
4345 ],
4346)
4347
4348cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004349 name = "fp32_mobilenet_v2",
4350 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004351 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004352 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004353 linkstatic = True,
4354 deps = [
4355 ":XNNPACK",
4356 "@pthreadpool",
4357 ],
4358)
4359
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004360cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004361 name = "fp16_mobilenet_v2",
4362 srcs = ["models/fp16-mobilenet-v2.cc"],
4363 hdrs = ["models/models.h"],
4364 copts = xnnpack_std_cxxopts(),
4365 linkstatic = True,
4366 deps = [
4367 ":XNNPACK",
4368 "@FP16",
4369 "@pthreadpool",
4370 ],
4371)
4372
4373cc_library(
4374 name = "fp32_mobilenet_v3_large",
4375 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004376 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004377 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004378 linkstatic = True,
4379 deps = [
4380 ":XNNPACK",
4381 "@pthreadpool",
4382 ],
4383)
4384
4385cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004386 name = "fp32_mobilenet_v3_small",
4387 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004388 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004389 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004390 linkstatic = True,
4391 deps = [
4392 ":XNNPACK",
4393 "@pthreadpool",
4394 ],
4395)
4396
Marat Dukhanc068bb62019-10-04 13:24:39 -07004397xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004398 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004399 srcs = [
4400 "bench/f32-dwconv-e2e.cc",
4401 "bench/end2end.h",
4402 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004403 deps = MICROKERNEL_BENCHMARK_DEPS + [
4404 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004405 ":fp32_mobilenet_v1",
4406 ":fp32_mobilenet_v2",
4407 ":fp32_mobilenet_v3_large",
4408 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004409 ],
4410)
4411
4412xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004413 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004414 srcs = [
4415 "bench/f32-gemm-e2e.cc",
4416 "bench/end2end.h",
4417 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004418 deps = MICROKERNEL_BENCHMARK_DEPS + [
4419 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004420 ":fp32_mobilenet_v1",
4421 ":fp32_mobilenet_v2",
4422 ":fp32_mobilenet_v3_large",
4423 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004424 ],
4425)
4426
4427xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004428 name = "end2end_bench",
4429 srcs = ["bench/end2end.cc"],
4430 deps = [
4431 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004432 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004433 ":fp16_mobilenet_v1",
4434 ":fp16_mobilenet_v2",
4435 ":fp32_mobilenet_v1",
4436 ":fp32_mobilenet_v2",
4437 ":fp32_mobilenet_v3_large",
4438 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004439 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004440 "@pthreadpool",
4441 ],
4442)
4443
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004444#################### Accuracy evaluation for math functions ####################
4445
4446xnnpack_benchmark(
4447 name = "f32_exp_eval",
4448 srcs = [
4449 "eval/f32-exp.cc",
4450 "src/xnnpack/AlignedAllocator.h",
4451 ] + ACCURACY_EVAL_HDRS,
4452 deps = ACCURACY_EVAL_DEPS,
4453)
4454
Marat Dukhan515c9772019-10-17 18:07:57 -07004455xnnpack_benchmark(
4456 name = "f32_expminus_eval",
4457 srcs = [
4458 "eval/f32-expminus.cc",
4459 "src/xnnpack/AlignedAllocator.h",
4460 ] + ACCURACY_EVAL_HDRS,
4461 deps = ACCURACY_EVAL_DEPS,
4462)
4463
Marat Dukhan98ba4412019-10-23 02:14:28 -07004464xnnpack_benchmark(
4465 name = "f32_extexp_eval",
4466 srcs = [
4467 "eval/f32-extexp.cc",
4468 "src/xnnpack/AlignedAllocator.h",
4469 ] + ACCURACY_EVAL_HDRS,
4470 deps = ACCURACY_EVAL_DEPS,
4471)
4472
Marat Dukhan8853b822020-05-07 12:19:01 -07004473xnnpack_unit_test(
4474 name = "f32_roundne_eval",
4475 srcs = [
4476 "eval/f32-roundne.cc",
4477 "src/xnnpack/AlignedAllocator.h",
4478 "src/xnnpack/math-stubs.h",
4479 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004480 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004481 deps = MICROKERNEL_TEST_DEPS,
4482)
4483
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004484xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004485 name = "f32_roundd_eval",
4486 srcs = [
4487 "eval/f32-roundd.cc",
4488 "src/xnnpack/AlignedAllocator.h",
4489 "src/xnnpack/math-stubs.h",
4490 ] + MICROKERNEL_TEST_HDRS,
4491 automatic = False,
4492 deps = MICROKERNEL_TEST_DEPS,
4493)
4494
4495xnnpack_unit_test(
4496 name = "f32_roundu_eval",
4497 srcs = [
4498 "eval/f32-roundu.cc",
4499 "src/xnnpack/AlignedAllocator.h",
4500 "src/xnnpack/math-stubs.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 automatic = False,
4503 deps = MICROKERNEL_TEST_DEPS,
4504)
4505
4506xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004507 name = "f32_roundz_eval",
4508 srcs = [
4509 "eval/f32-roundz.cc",
4510 "src/xnnpack/AlignedAllocator.h",
4511 "src/xnnpack/math-stubs.h",
4512 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004513 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
Marat Dukhan346a9e52019-11-15 09:06:30 -08004517xnnpack_benchmark(
4518 name = "f32_sigmoid_eval",
4519 srcs = [
4520 "eval/f32-sigmoid.cc",
4521 "src/xnnpack/AlignedAllocator.h",
4522 ] + ACCURACY_EVAL_HDRS,
4523 deps = ACCURACY_EVAL_DEPS,
4524)
4525
Marat Dukhan84000762020-06-29 18:38:43 -07004526xnnpack_benchmark(
4527 name = "f32_sqrt_eval",
4528 srcs = [
4529 "eval/f32-sqrt.cc",
4530 "src/xnnpack/AlignedAllocator.h",
4531 ] + ACCURACY_EVAL_HDRS,
4532 deps = ACCURACY_EVAL_DEPS + [
4533 ":bench_utils",
4534 ],
4535)
4536
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537######################### Unit tests for micro-kernels #########################
4538
4539xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004540 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004542 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543 "test/gemm-microkernel-tester.h",
4544 "src/xnnpack/AlignedAllocator.h",
4545 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004546 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004547)
4548
4549xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004550 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004551 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004552 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004553 "test/spmm-microkernel-tester.h",
4554 "src/xnnpack/AlignedAllocator.h",
4555 ] + MICROKERNEL_TEST_HDRS,
4556 deps = MICROKERNEL_TEST_DEPS,
4557)
4558
4559xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560 name = "f32_argmaxpool_test",
4561 srcs = [
4562 "test/f32-argmaxpool.cc",
4563 "test/argmaxpool-microkernel-tester.h",
4564 "src/xnnpack/AlignedAllocator.h",
4565 ] + MICROKERNEL_TEST_HDRS,
4566 deps = MICROKERNEL_TEST_DEPS,
4567)
4568
4569xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004570 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004572 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004573 "test/avgpool-microkernel-tester.h",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_TEST_HDRS,
4576 deps = MICROKERNEL_TEST_DEPS,
4577)
4578
4579xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004580 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004581 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004582 "test/f32-ibilinear.cc",
4583 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004584 "src/xnnpack/AlignedAllocator.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004590 name = "f16_clamp_test",
4591 srcs = [
4592 "test/f16-clamp.cc",
4593 "test/clamp-microkernel-tester.h",
4594 ] + MICROKERNEL_TEST_HDRS,
4595 deps = MICROKERNEL_TEST_DEPS,
4596)
4597
4598xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004599 name = "f32_clamp_test",
4600 srcs = [
4601 "test/f32-clamp.cc",
4602 "test/clamp-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004608 name = "f32_igemm_test",
4609 srcs = [
4610 "test/f32-igemm.cc",
4611 "test/gemm-microkernel-tester.h",
4612 "src/xnnpack/AlignedAllocator.h",
4613 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004614 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004615)
4616
4617xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004618 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004620 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621 "test/gemm-microkernel-tester.h",
4622 "src/xnnpack/AlignedAllocator.h",
4623 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004624 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004625)
4626
4627xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004628 name = "f32_igemm_minmax_test",
4629 srcs = [
4630 "test/f32-igemm-minmax.cc",
4631 "test/gemm-microkernel-tester.h",
4632 "src/xnnpack/AlignedAllocator.h",
4633 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004634 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004635)
4636
4637xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004638 name = "f16_igemm_minmax_test",
4639 srcs = [
4640 "test/f16-igemm-minmax.cc",
4641 "test/gemm-microkernel-tester.h",
4642 "src/xnnpack/AlignedAllocator.h",
4643 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004644 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004645)
4646
4647xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648 name = "f32_conv_hwc_test",
4649 srcs = [
4650 "test/f32-conv-hwc.cc",
4651 "test/conv-hwc-microkernel-tester.h",
4652 "src/xnnpack/AlignedAllocator.h",
4653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004654 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004655)
4656
4657xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004658 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004660 "test/f32-conv-hwc2chw.cc",
4661 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004662 "src/xnnpack/AlignedAllocator.h",
4663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004664 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665)
4666
4667xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004668 name = "f32_dwconv_test",
4669 srcs = [
4670 "test/f32-dwconv.cc",
4671 "test/dwconv-microkernel-tester.h",
4672 "src/xnnpack/AlignedAllocator.h",
4673 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004674 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004675)
4676
4677xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004678 name = "f16_dwconv_minmax_test",
4679 srcs = [
4680 "test/f16-dwconv-minmax.cc",
4681 "test/dwconv-microkernel-tester.h",
4682 "src/xnnpack/AlignedAllocator.h",
4683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004684 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004685)
4686
4687xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004688 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004689 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004690 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004691 "test/dwconv-microkernel-tester.h",
4692 "src/xnnpack/AlignedAllocator.h",
4693 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004694 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004695)
4696
4697xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004698 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004699 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004700 "test/f32-dwconv-chw.cc",
4701 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004702 "src/xnnpack/AlignedAllocator.h",
4703 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004704 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004705)
4706
4707xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004708 name = "f16_gavgpool_minmax_test",
4709 srcs = [
4710 "test/f16-gavgpool-minmax.cc",
4711 "test/gavgpool-microkernel-tester.h",
4712 "src/xnnpack/AlignedAllocator.h",
4713 ] + MICROKERNEL_TEST_HDRS,
4714 deps = MICROKERNEL_TEST_DEPS,
4715)
4716
4717xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004718 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004720 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004721 "test/gavgpool-microkernel-tester.h",
4722 "src/xnnpack/AlignedAllocator.h",
4723 ] + MICROKERNEL_TEST_HDRS,
4724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
4727xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004728 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004729 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004730 "test/f32-gavgpool-cw.cc",
4731 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004732 "src/xnnpack/AlignedAllocator.h",
4733 ] + MICROKERNEL_TEST_HDRS,
4734 deps = MICROKERNEL_TEST_DEPS,
4735)
4736
4737xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004738 name = "f32_gemm_test",
4739 srcs = [
4740 "test/f32-gemm.cc",
4741 "test/gemm-microkernel-tester.h",
4742 "src/xnnpack/AlignedAllocator.h",
4743 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004744 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004745)
4746
4747xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004748 name = "f32_gemm_relu_test",
4749 srcs = [
4750 "test/f32-gemm-relu.cc",
4751 "test/gemm-microkernel-tester.h",
4752 "src/xnnpack/AlignedAllocator.h",
4753 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004754 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004755)
4756
4757xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004758 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004760 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004761 "test/gemm-microkernel-tester.h",
4762 "src/xnnpack/AlignedAllocator.h",
4763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765)
4766
4767xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004768 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004769 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004770 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004771 "test/gemm-microkernel-tester.h",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004774 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004775)
4776
4777xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004778 name = "f16_hswish_test",
4779 srcs = [
4780 "test/f16-hswish.cc",
4781 "test/hswish-microkernel-tester.h",
4782 ] + MICROKERNEL_TEST_HDRS,
4783 deps = MICROKERNEL_TEST_DEPS,
4784)
4785
4786xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787 name = "f32_hswish_test",
4788 srcs = [
4789 "test/f32-hswish.cc",
4790 "test/hswish-microkernel-tester.h",
4791 ] + MICROKERNEL_TEST_HDRS,
4792 deps = MICROKERNEL_TEST_DEPS,
4793)
4794
4795xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004796 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004797 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004798 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799 "test/maxpool-microkernel-tester.h",
4800 ] + MICROKERNEL_TEST_HDRS,
4801 deps = MICROKERNEL_TEST_DEPS,
4802)
4803
4804xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004805 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004806 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004807 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004808 "test/avgpool-microkernel-tester.h",
4809 "src/xnnpack/AlignedAllocator.h",
4810 ] + MICROKERNEL_TEST_HDRS,
4811 deps = MICROKERNEL_TEST_DEPS,
4812)
4813
4814xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004815 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004816 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004817 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004818 "test/gemm-microkernel-tester.h",
4819 "src/xnnpack/AlignedAllocator.h",
4820 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004821 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004822)
4823
4824xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004825 name = "f16_prelu_test",
4826 srcs = [
4827 "test/f16-prelu.cc",
4828 "test/prelu-microkernel-tester.h",
4829 "src/xnnpack/AlignedAllocator.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835 name = "f32_prelu_test",
4836 srcs = [
4837 "test/f32-prelu.cc",
4838 "test/prelu-microkernel-tester.h",
4839 "src/xnnpack/AlignedAllocator.h",
4840 ] + MICROKERNEL_TEST_HDRS,
4841 deps = MICROKERNEL_TEST_DEPS,
4842)
4843
4844xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004845 name = "f32_raddexpminusmax_test",
4846 srcs = [
4847 "test/f32-raddexpminusmax.cc",
4848 "test/raddexpminusmax-microkernel-tester.h",
4849 ] + MICROKERNEL_TEST_HDRS,
4850 deps = MICROKERNEL_TEST_DEPS,
4851)
4852
4853xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004854 name = "f32_raddextexp_test",
4855 srcs = [
4856 "test/f32-raddextexp.cc",
4857 "test/raddextexp-microkernel-tester.h",
4858 ] + MICROKERNEL_TEST_HDRS,
4859 deps = MICROKERNEL_TEST_DEPS,
4860)
4861
4862xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004863 name = "f32_raddstoreexpminusmax_test",
4864 srcs = [
4865 "test/f32-raddstoreexpminusmax.cc",
4866 "test/raddstoreexpminusmax-microkernel-tester.h",
4867 ] + MICROKERNEL_TEST_HDRS,
4868 deps = MICROKERNEL_TEST_DEPS,
4869)
4870
4871xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004872 name = "f32_relu_test",
4873 srcs = [
4874 "test/f32-relu.cc",
4875 "test/vunary-microkernel-tester.h",
4876 ] + MICROKERNEL_TEST_HDRS,
4877 deps = MICROKERNEL_TEST_DEPS,
4878)
4879
4880xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 name = "f32_rmax_test",
4882 srcs = [
4883 "test/f32-rmax.cc",
4884 "test/rmax-microkernel-tester.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004890 name = "f32_sigmoid_test",
4891 srcs = [
4892 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004893 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004894 ] + MICROKERNEL_TEST_HDRS,
4895 deps = MICROKERNEL_TEST_DEPS,
4896)
4897
4898xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004899 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004900 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004901 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004902 "test/spmm-microkernel-tester.h",
4903 "src/xnnpack/AlignedAllocator.h",
4904 ] + MICROKERNEL_TEST_HDRS,
4905 deps = MICROKERNEL_TEST_DEPS,
4906)
4907
4908xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004909 name = "f16_vadd_minmax_test",
4910 srcs = [
4911 "test/f16-vadd-minmax.cc",
4912 "test/vbinary-microkernel-tester.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
4918 name = "f16_vaddc_minmax_test",
4919 srcs = [
4920 "test/f16-vaddc-minmax.cc",
4921 "test/vbinaryc-microkernel-tester.h",
4922 ] + MICROKERNEL_TEST_HDRS,
4923 deps = MICROKERNEL_TEST_DEPS,
4924)
4925
4926xnnpack_unit_test(
4927 name = "f16_vdiv_minmax_test",
4928 srcs = [
4929 "test/f16-vdiv-minmax.cc",
4930 "test/vbinary-microkernel-tester.h",
4931 ] + MICROKERNEL_TEST_HDRS,
4932 deps = MICROKERNEL_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
4936 name = "f16_vdivc_minmax_test",
4937 srcs = [
4938 "test/f16-vdivc-minmax.cc",
4939 "test/vbinaryc-microkernel-tester.h",
4940 ] + MICROKERNEL_TEST_HDRS,
4941 deps = MICROKERNEL_TEST_DEPS,
4942)
4943
4944xnnpack_unit_test(
4945 name = "f16_vrdivc_minmax_test",
4946 srcs = [
4947 "test/f16-vrdivc-minmax.cc",
4948 "test/vbinaryc-microkernel-tester.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
4954 name = "f16_vmax_test",
4955 srcs = [
4956 "test/f16-vmax.cc",
4957 "test/vbinary-microkernel-tester.h",
4958 ] + MICROKERNEL_TEST_HDRS,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
4963 name = "f16_vmaxc_test",
4964 srcs = [
4965 "test/f16-vmaxc.cc",
4966 "test/vbinaryc-microkernel-tester.h",
4967 ] + MICROKERNEL_TEST_HDRS,
4968 deps = MICROKERNEL_TEST_DEPS,
4969)
4970
4971xnnpack_unit_test(
4972 name = "f16_vmin_test",
4973 srcs = [
4974 "test/f16-vmin.cc",
4975 "test/vbinary-microkernel-tester.h",
4976 ] + MICROKERNEL_TEST_HDRS,
4977 deps = MICROKERNEL_TEST_DEPS,
4978)
4979
4980xnnpack_unit_test(
4981 name = "f16_vminc_test",
4982 srcs = [
4983 "test/f16-vminc.cc",
4984 "test/vbinaryc-microkernel-tester.h",
4985 ] + MICROKERNEL_TEST_HDRS,
4986 deps = MICROKERNEL_TEST_DEPS,
4987)
4988
4989xnnpack_unit_test(
4990 name = "f16_vmul_minmax_test",
4991 srcs = [
4992 "test/f16-vmul-minmax.cc",
4993 "test/vbinary-microkernel-tester.h",
4994 ] + MICROKERNEL_TEST_HDRS,
4995 deps = MICROKERNEL_TEST_DEPS,
4996)
4997
4998xnnpack_unit_test(
4999 name = "f16_vmulc_minmax_test",
5000 srcs = [
5001 "test/f16-vmulc-minmax.cc",
5002 "test/vbinaryc-microkernel-tester.h",
5003 ] + MICROKERNEL_TEST_HDRS,
5004 deps = MICROKERNEL_TEST_DEPS,
5005)
5006
5007xnnpack_unit_test(
5008 name = "f16_vsub_minmax_test",
5009 srcs = [
5010 "test/f16-vsub-minmax.cc",
5011 "test/vbinary-microkernel-tester.h",
5012 ] + MICROKERNEL_TEST_HDRS,
5013 deps = MICROKERNEL_TEST_DEPS,
5014)
5015
5016xnnpack_unit_test(
5017 name = "f16_vsubc_minmax_test",
5018 srcs = [
5019 "test/f16-vsubc-minmax.cc",
5020 "test/vbinaryc-microkernel-tester.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
5026 name = "f16_vrsubc_minmax_test",
5027 srcs = [
5028 "test/f16-vrsubc-minmax.cc",
5029 "test/vbinaryc-microkernel-tester.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005035 name = "f32_vabs_test",
5036 srcs = [
5037 "test/f32-vabs.cc",
5038 "test/vunary-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005044 name = "f32_vadd_test",
5045 srcs = [
5046 "test/f32-vadd.cc",
5047 "test/vbinary-microkernel-tester.h",
5048 ] + MICROKERNEL_TEST_HDRS,
5049 deps = MICROKERNEL_TEST_DEPS,
5050)
5051
5052xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005053 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005055 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005056 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005057 ] + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS,
5059)
5060
5061xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005062 name = "f32_vadd_relu_test",
5063 srcs = [
5064 "test/f32-vadd-relu.cc",
5065 "test/vbinary-microkernel-tester.h",
5066 ] + MICROKERNEL_TEST_HDRS,
5067 deps = MICROKERNEL_TEST_DEPS,
5068)
5069
5070xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005071 name = "f32_vaddc_test",
5072 srcs = [
5073 "test/f32-vaddc.cc",
5074 "test/vbinaryc-microkernel-tester.h",
5075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005080 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005081 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005082 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005083 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005089 name = "f32_vaddc_relu_test",
5090 srcs = [
5091 "test/f32-vaddc-relu.cc",
5092 "test/vbinaryc-microkernel-tester.h",
5093 ] + MICROKERNEL_TEST_HDRS,
5094 deps = MICROKERNEL_TEST_DEPS,
5095)
5096
5097xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005098 name = "f32_vdiv_test",
5099 srcs = [
5100 "test/f32-vdiv.cc",
5101 "test/vbinary-microkernel-tester.h",
5102 ] + MICROKERNEL_TEST_HDRS,
5103 deps = MICROKERNEL_TEST_DEPS,
5104)
5105
5106xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005107 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005108 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005109 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005110 "test/vbinary-microkernel-tester.h",
5111 ] + MICROKERNEL_TEST_HDRS,
5112 deps = MICROKERNEL_TEST_DEPS,
5113)
5114
5115xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005116 name = "f32_vdiv_relu_test",
5117 srcs = [
5118 "test/f32-vdiv-relu.cc",
5119 "test/vbinary-microkernel-tester.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 deps = MICROKERNEL_TEST_DEPS,
5122)
5123
5124xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005125 name = "f32_vdivc_test",
5126 srcs = [
5127 "test/f32-vdivc.cc",
5128 "test/vbinaryc-microkernel-tester.h",
5129 ] + MICROKERNEL_TEST_HDRS,
5130 deps = MICROKERNEL_TEST_DEPS,
5131)
5132
5133xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005134 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005135 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005136 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005137 "test/vbinaryc-microkernel-tester.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005143 name = "f32_vdivc_relu_test",
5144 srcs = [
5145 "test/f32-vdivc-relu.cc",
5146 "test/vbinaryc-microkernel-tester.h",
5147 ] + MICROKERNEL_TEST_HDRS,
5148 deps = MICROKERNEL_TEST_DEPS,
5149)
5150
5151xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005152 name = "f32_vrdivc_test",
5153 srcs = [
5154 "test/f32-vrdivc.cc",
5155 "test/vbinaryc-microkernel-tester.h",
5156 ] + MICROKERNEL_TEST_HDRS,
5157 deps = MICROKERNEL_TEST_DEPS,
5158)
5159
5160xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005161 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005162 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005163 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005164 "test/vbinaryc-microkernel-tester.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005170 name = "f32_vrdivc_relu_test",
5171 srcs = [
5172 "test/f32-vrdivc-relu.cc",
5173 "test/vbinaryc-microkernel-tester.h",
5174 ] + MICROKERNEL_TEST_HDRS,
5175 deps = MICROKERNEL_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005179 name = "f32_vmax_test",
5180 srcs = [
5181 "test/f32-vmax.cc",
5182 "test/vbinary-microkernel-tester.h",
5183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
5188 name = "f32_vmaxc_test",
5189 srcs = [
5190 "test/f32-vmaxc.cc",
5191 "test/vbinaryc-microkernel-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
5197 name = "f32_vmin_test",
5198 srcs = [
5199 "test/f32-vmin.cc",
5200 "test/vbinary-microkernel-tester.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
5206 name = "f32_vminc_test",
5207 srcs = [
5208 "test/f32-vminc.cc",
5209 "test/vbinaryc-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
5214xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005215 name = "f32_vmul_test",
5216 srcs = [
5217 "test/f32-vmul.cc",
5218 "test/vbinary-microkernel-tester.h",
5219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005224 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005226 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005227 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005228 ] + MICROKERNEL_TEST_HDRS,
5229 deps = MICROKERNEL_TEST_DEPS,
5230)
5231
5232xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005233 name = "f32_vmul_relu_test",
5234 srcs = [
5235 "test/f32-vmul-relu.cc",
5236 "test/vbinary-microkernel-tester.h",
5237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005242 name = "f32_vmulc_test",
5243 srcs = [
5244 "test/f32-vmulc.cc",
5245 "test/vbinaryc-microkernel-tester.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
5250xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005251 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005252 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005253 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005254 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 ] + MICROKERNEL_TEST_HDRS,
5256 deps = MICROKERNEL_TEST_DEPS,
5257)
5258
5259xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005260 name = "f32_vmulc_relu_test",
5261 srcs = [
5262 "test/f32-vmulc-relu.cc",
5263 "test/vbinaryc-microkernel-tester.h",
5264 ] + MICROKERNEL_TEST_HDRS,
5265 deps = MICROKERNEL_TEST_DEPS,
5266)
5267
5268xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005269 name = "f16_vmulcaddc_minmax_test",
5270 srcs = [
5271 "test/f16-vmulcaddc-minmax.cc",
5272 "test/vmulcaddc-microkernel-tester.h",
5273 "src/xnnpack/AlignedAllocator.h",
5274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005275 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005276)
5277
5278xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005279 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005281 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 "test/vmulcaddc-microkernel-tester.h",
5283 "src/xnnpack/AlignedAllocator.h",
5284 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005285 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286)
5287
5288xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005289 name = "f32_vlrelu_test",
5290 srcs = [
5291 "test/f32-vlrelu.cc",
5292 "test/vunary-microkernel-tester.h",
5293 ] + MICROKERNEL_TEST_HDRS,
5294 deps = MICROKERNEL_TEST_DEPS,
5295)
5296
5297xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005298 name = "f32_vneg_test",
5299 srcs = [
5300 "test/f32-vneg.cc",
5301 "test/vunary-microkernel-tester.h",
5302 ] + MICROKERNEL_TEST_HDRS,
5303 deps = MICROKERNEL_TEST_DEPS,
5304)
5305
5306xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005307 name = "f32_vrndne_test",
5308 srcs = [
5309 "test/f32-vrndne.cc",
5310 "test/vunary-microkernel-tester.h",
5311 ] + MICROKERNEL_TEST_HDRS,
5312 deps = MICROKERNEL_TEST_DEPS,
5313)
5314
5315xnnpack_unit_test(
5316 name = "f32_vrndz_test",
5317 srcs = [
5318 "test/f32-vrndz.cc",
5319 "test/vunary-microkernel-tester.h",
5320 ] + MICROKERNEL_TEST_HDRS,
5321 deps = MICROKERNEL_TEST_DEPS,
5322)
5323
5324xnnpack_unit_test(
5325 name = "f32_vrndu_test",
5326 srcs = [
5327 "test/f32-vrndu.cc",
5328 "test/vunary-microkernel-tester.h",
5329 ] + MICROKERNEL_TEST_HDRS,
5330 deps = MICROKERNEL_TEST_DEPS,
5331)
5332
5333xnnpack_unit_test(
5334 name = "f32_vrndd_test",
5335 srcs = [
5336 "test/f32-vrndd.cc",
5337 "test/vunary-microkernel-tester.h",
5338 ] + MICROKERNEL_TEST_HDRS,
5339 deps = MICROKERNEL_TEST_DEPS,
5340)
5341
5342xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005343 name = "f32_vscale_test",
5344 srcs = [
5345 "test/f32-vscale.cc",
5346 "test/vscale-microkernel-tester.h",
5347 ] + MICROKERNEL_TEST_HDRS,
5348 deps = MICROKERNEL_TEST_DEPS,
5349)
5350
5351xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005352 name = "f32_vscaleexpminusmax_test",
5353 srcs = [
5354 "test/f32-vscaleexpminusmax.cc",
5355 "test/vscaleexpminusmax-microkernel-tester.h",
5356 ] + MICROKERNEL_TEST_HDRS,
5357 deps = MICROKERNEL_TEST_DEPS,
5358)
5359
5360xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005361 name = "f32_vscaleextexp_test",
5362 srcs = [
5363 "test/f32-vscaleextexp.cc",
5364 "test/vscaleextexp-microkernel-tester.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005370 name = "f32_vsqr_test",
5371 srcs = [
5372 "test/f32-vsqr.cc",
5373 "test/vunary-microkernel-tester.h",
5374 ] + MICROKERNEL_TEST_HDRS,
5375 deps = MICROKERNEL_TEST_DEPS,
5376)
5377
5378xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005379 name = "f32_vsqrdiff_test",
5380 srcs = [
5381 "test/f32-vsqrdiff.cc",
5382 "test/vbinary-microkernel-tester.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
5388 name = "f32_vsqrdiffc_test",
5389 srcs = [
5390 "test/f32-vsqrdiffc.cc",
5391 "test/vbinaryc-microkernel-tester.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 deps = MICROKERNEL_TEST_DEPS,
5394)
5395
5396xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005397 name = "f32_vsqrt_test",
5398 srcs = [
5399 "test/f32-vsqrt.cc",
5400 "test/vunary-microkernel-tester.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005406 name = "f32_vsub_test",
5407 srcs = [
5408 "test/f32-vsub.cc",
5409 "test/vbinary-microkernel-tester.h",
5410 ] + MICROKERNEL_TEST_HDRS,
5411 deps = MICROKERNEL_TEST_DEPS,
5412)
5413
5414xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005415 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005416 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005417 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005418 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005419 ] + MICROKERNEL_TEST_HDRS,
5420 deps = MICROKERNEL_TEST_DEPS,
5421)
5422
5423xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005424 name = "f32_vsub_relu_test",
5425 srcs = [
5426 "test/f32-vsub-relu.cc",
5427 "test/vbinary-microkernel-tester.h",
5428 ] + MICROKERNEL_TEST_HDRS,
5429 deps = MICROKERNEL_TEST_DEPS,
5430)
5431
5432xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005433 name = "f32_vsubc_test",
5434 srcs = [
5435 "test/f32-vsubc.cc",
5436 "test/vbinaryc-microkernel-tester.h",
5437 ] + MICROKERNEL_TEST_HDRS,
5438 deps = MICROKERNEL_TEST_DEPS,
5439)
5440
5441xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005442 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005443 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005444 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005445 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005446 ] + MICROKERNEL_TEST_HDRS,
5447 deps = MICROKERNEL_TEST_DEPS,
5448)
5449
5450xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005451 name = "f32_vsubc_relu_test",
5452 srcs = [
5453 "test/f32-vsubc-relu.cc",
5454 "test/vbinaryc-microkernel-tester.h",
5455 ] + MICROKERNEL_TEST_HDRS,
5456 deps = MICROKERNEL_TEST_DEPS,
5457)
5458
5459xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005460 name = "f32_vrsubc_test",
5461 srcs = [
5462 "test/f32-vrsubc.cc",
5463 "test/vbinaryc-microkernel-tester.h",
5464 ] + MICROKERNEL_TEST_HDRS,
5465 deps = MICROKERNEL_TEST_DEPS,
5466)
5467
5468xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005469 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005470 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005471 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005472 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005473 ] + MICROKERNEL_TEST_HDRS,
5474 deps = MICROKERNEL_TEST_DEPS,
5475)
5476
5477xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005478 name = "f32_vrsubc_relu_test",
5479 srcs = [
5480 "test/f32-vrsubc-relu.cc",
5481 "test/vbinaryc-microkernel-tester.h",
5482 ] + MICROKERNEL_TEST_HDRS,
5483 deps = MICROKERNEL_TEST_DEPS,
5484)
5485
5486xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005487 name = "qs8_dwconv_minmax_test",
5488 srcs = [
5489 "test/qs8-dwconv-minmax.cc",
5490 "test/dwconv-microkernel-tester.h",
5491 "src/xnnpack/AlignedAllocator.h",
5492 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5493 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5494)
5495
5496xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005497 name = "qs8_gavgpool_minmax_test",
5498 srcs = [
5499 "test/qs8-gavgpool-minmax.cc",
5500 "test/gavgpool-microkernel-tester.h",
5501 "src/xnnpack/AlignedAllocator.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005507 name = "qs8_gemm_minmax_test",
5508 srcs = [
5509 "test/qs8-gemm-minmax.cc",
5510 "test/gemm-microkernel-tester.h",
5511 "src/xnnpack/AlignedAllocator.h",
5512 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5513 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5514)
5515
5516xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005517 name = "qs8_igemm_minmax_test",
5518 srcs = [
5519 "test/qs8-igemm-minmax.cc",
5520 "test/gemm-microkernel-tester.h",
5521 "src/xnnpack/AlignedAllocator.h",
5522 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5523 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5524)
5525
5526xnnpack_unit_test(
5527 name = "qs8_requantization_test",
5528 srcs = [
5529 "src/xnnpack/requantization-stubs.h",
5530 "test/qs8-requantization.cc",
5531 "test/requantization-tester.h",
5532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005537 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005538 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005539 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 "test/avgpool-microkernel-tester.h",
5541 "src/xnnpack/AlignedAllocator.h",
5542 ] + MICROKERNEL_TEST_HDRS,
5543 deps = MICROKERNEL_TEST_DEPS,
5544)
5545
5546xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005547 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005548 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005549 "test/qu8-dwconv-minmax.cc",
5550 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005551 "src/xnnpack/AlignedAllocator.h",
5552 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005553 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005554)
5555
5556xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005557 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005559 "test/qu8-igemm-minmax.cc",
5560 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005561 "src/xnnpack/AlignedAllocator.h",
5562 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005563 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005564)
5565
5566xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005567 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005568 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005569 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005570 "test/gavgpool-microkernel-tester.h",
5571 "src/xnnpack/AlignedAllocator.h",
5572 ] + MICROKERNEL_TEST_HDRS,
5573 deps = MICROKERNEL_TEST_DEPS,
5574)
5575
5576xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005577 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005579 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005580 "test/gemm-microkernel-tester.h",
5581 "src/xnnpack/AlignedAllocator.h",
5582 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005583 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005584)
5585
5586xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005587 name = "qu8_requantization_test",
5588 srcs = [
5589 "src/xnnpack/requantization-stubs.h",
5590 "test/qu8-requantization.cc",
5591 "test/requantization-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005597 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005598 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005599 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 "test/vadd-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
5606 name = "u8_clamp_test",
5607 srcs = [
5608 "test/u8-clamp.cc",
5609 "test/clamp-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
5615 name = "u8_lut32norm_test",
5616 srcs = [
5617 "test/u8-lut32norm.cc",
5618 "test/lut-norm-microkernel-tester.h",
5619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005624 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005625 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005626 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005627 "test/maxpool-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
5633 name = "u8_rmax_test",
5634 srcs = [
5635 "test/u8-rmax.cc",
5636 "test/rmax-microkernel-tester.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005642 name = "x32_fill_test",
5643 srcs = [
5644 "test/x32-fill.cc",
5645 "test/fill-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005651 name = "x32_packx_test",
5652 srcs = [
5653 "test/x32-packx.cc",
5654 "test/pack-microkernel-tester.h",
5655 "src/xnnpack/AlignedAllocator.h",
5656 ] + MICROKERNEL_TEST_HDRS,
5657 deps = MICROKERNEL_TEST_DEPS,
5658)
5659
5660xnnpack_unit_test(
5661 name = "x32_pad_test",
5662 srcs = [
5663 "test/x32-pad.cc",
5664 "test/pad-microkernel-tester.h",
5665 ] + MICROKERNEL_TEST_HDRS,
5666 deps = MICROKERNEL_TEST_DEPS,
5667)
5668
5669xnnpack_unit_test(
5670 name = "x32_unpool_test",
5671 srcs = [
5672 "test/x32-unpool.cc",
5673 "test/unpool-microkernel-tester.h",
5674 ] + MICROKERNEL_TEST_HDRS,
5675 deps = MICROKERNEL_TEST_DEPS,
5676)
5677
5678xnnpack_unit_test(
5679 name = "x32_zip_test",
5680 srcs = [
5681 "test/x32-zip.cc",
5682 "test/zip-microkernel-tester.h",
5683 ] + MICROKERNEL_TEST_HDRS,
5684 deps = MICROKERNEL_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
5688 name = "x8_lut_test",
5689 srcs = [
5690 "test/x8-lut.cc",
5691 "test/lut-microkernel-tester.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
5697 name = "x8_zip_test",
5698 srcs = [
5699 "test/x8-zip.cc",
5700 "test/zip-microkernel-tester.h",
5701 ] + MICROKERNEL_TEST_HDRS,
5702 deps = MICROKERNEL_TEST_DEPS,
5703)
5704
Marat Dukhan20c3b922020-03-10 03:45:06 -07005705########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005706
5707xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005708 name = "operator_size_test",
5709 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005710 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005711)
5712
Marat Dukhan20c3b922020-03-10 03:45:06 -07005713xnnpack_binary(
5714 name = "subgraph_size_test",
5715 srcs = ["test/subgraph-size.c"],
5716 deps = [":XNNPACK"],
5717)
5718
5719########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005720
5721xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005722 name = "abs_nc_test",
5723 srcs = [
5724 "test/abs-nc.cc",
5725 "test/abs-operator-tester.h",
5726 ],
5727 deps = OPERATOR_TEST_DEPS,
5728)
5729
5730xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005731 name = "add_nd_test",
5732 srcs = [
5733 "test/add-nd.cc",
5734 "test/binary-elementwise-operator-tester.h",
5735 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005736 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005737)
5738
5739xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005740 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005741 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005742 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005743 "test/argmax-pooling-operator-tester.h",
5744 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005745 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005746)
5747
5748xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005749 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005750 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005751 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 "test/average-pooling-operator-tester.h",
5753 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005754 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005755)
5756
5757xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005758 name = "bankers_rounding_nc_test",
5759 srcs = [
5760 "test/bankers-rounding-nc.cc",
5761 "test/bankers-rounding-operator-tester.h",
5762 ],
5763 deps = OPERATOR_TEST_DEPS,
5764)
5765
5766xnnpack_unit_test(
5767 name = "ceiling_nc_test",
5768 srcs = [
5769 "test/ceiling-nc.cc",
5770 "test/ceiling-operator-tester.h",
5771 ],
5772 deps = OPERATOR_TEST_DEPS,
5773)
5774
5775xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005776 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005777 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005778 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005779 "test/channel-shuffle-operator-tester.h",
5780 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005781 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005782)
5783
5784xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005785 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005786 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005787 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005788 "test/clamp-operator-tester.h",
5789 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005790 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005791)
5792
5793xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005794 name = "constant_pad_nd_test",
5795 srcs = [
5796 "test/constant-pad-nd.cc",
5797 "test/constant-pad-operator-tester.h",
5798 ],
5799 deps = OPERATOR_TEST_DEPS,
5800)
5801
5802xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005803 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005804 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005805 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806 "test/convolution-operator-tester.h",
5807 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005808 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005809)
5810
5811xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005812 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005813 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005814 "test/convolution-nchw.cc",
5815 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005816 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005817 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005818)
5819
5820xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005821 name = "copy_nc_test",
5822 srcs = [
5823 "test/copy-nc.cc",
5824 "test/copy-operator-tester.h",
5825 ],
5826 deps = OPERATOR_TEST_DEPS,
5827)
5828
5829xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005830 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005831 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005832 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005833 "test/deconvolution-operator-tester.h",
5834 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005835 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836)
5837
5838xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005839 name = "divide_nd_test",
5840 srcs = [
5841 "test/binary-elementwise-operator-tester.h",
5842 "test/divide-nd.cc",
5843 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005844 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005845)
5846
5847xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005848 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005849 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005850 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005851 "test/fully-connected-operator-tester.h",
5852 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005853 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005854)
5855
5856xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005857 name = "floor_nc_test",
5858 srcs = [
5859 "test/floor-nc.cc",
5860 "test/floor-operator-tester.h",
5861 ],
5862 deps = OPERATOR_TEST_DEPS,
5863)
5864
5865xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005866 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005867 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005868 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005869 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005870 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005871 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005872)
5873
5874xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005875 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005876 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005877 "test/global-average-pooling-ncw.cc",
5878 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005879 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005880 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005881)
5882
5883xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005884 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005885 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005886 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005887 "test/hardswish-operator-tester.h",
5888 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005889 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005890)
5891
5892xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005893 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005894 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005895 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 "test/leaky-relu-operator-tester.h",
5897 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005898 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005899)
5900
5901xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005902 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005903 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005904 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005905 "test/max-pooling-operator-tester.h",
5906 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005907 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908)
5909
5910xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005911 name = "maximum_nd_test",
5912 srcs = [
5913 "test/binary-elementwise-operator-tester.h",
5914 "test/maximum-nd.cc",
5915 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005916 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005917)
5918
5919xnnpack_unit_test(
5920 name = "minimum_nd_test",
5921 srcs = [
5922 "test/binary-elementwise-operator-tester.h",
5923 "test/minimum-nd.cc",
5924 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005925 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005926)
5927
5928xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005929 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005930 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005931 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005932 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005933 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005934 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005935)
5936
5937xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005938 name = "negate_nc_test",
5939 srcs = [
5940 "test/negate-nc.cc",
5941 "test/negate-operator-tester.h",
5942 ],
5943 deps = OPERATOR_TEST_DEPS,
5944)
5945
5946xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005947 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005948 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005949 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 "test/prelu-operator-tester.h",
5951 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005952 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953)
5954
5955xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005956 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005957 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005958 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005959 "test/resize-bilinear-operator-tester.h",
5960 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005961 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005962)
5963
5964xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005965 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005967 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005968 "test/sigmoid-operator-tester.h",
5969 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005970 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005971)
5972
5973xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005974 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005975 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005976 "test/softmax-nc.cc",
5977 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005979 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980)
5981
5982xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005983 name = "square_nc_test",
5984 srcs = [
5985 "test/square-nc.cc",
5986 "test/square-operator-tester.h",
5987 ],
5988 deps = OPERATOR_TEST_DEPS,
5989)
5990
5991xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005992 name = "square_root_nc_test",
5993 srcs = [
5994 "test/square-root-nc.cc",
5995 "test/square-root-operator-tester.h",
5996 ],
5997 deps = OPERATOR_TEST_DEPS,
5998)
5999
6000xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006001 name = "squared_difference_nd_test",
6002 srcs = [
6003 "test/binary-elementwise-operator-tester.h",
6004 "test/squared-difference-nd.cc",
6005 ],
6006 deps = OPERATOR_TEST_DEPS,
6007)
6008
6009xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006010 name = "subtract_nd_test",
6011 srcs = [
6012 "test/binary-elementwise-operator-tester.h",
6013 "test/subtract-nd.cc",
6014 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006015 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006016)
6017
6018xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006019 name = "truncation_nc_test",
6020 srcs = [
6021 "test/truncation-nc.cc",
6022 "test/truncation-operator-tester.h",
6023 ],
6024 deps = OPERATOR_TEST_DEPS,
6025)
6026
6027xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006028 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006029 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006030 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006031 "test/unpooling-operator-tester.h",
6032 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006033 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034)
6035
Chao Mei6ddfc602020-05-13 22:29:36 -07006036############################### Misc unit tests ###############################
6037
6038xnnpack_unit_test(
6039 name = "memory_planner_test",
6040 srcs = [
6041 "test/memory-planner-test.cc",
6042 ],
6043 deps = [
6044 ":XNNPACK",
6045 ":memory_planner",
6046 ],
6047)
6048
Marat Dukhan08c4a432019-10-03 09:29:21 -07006049############################# Build configurations #############################
6050
Marat Dukhanb8642352019-10-30 15:43:02 -07006051# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006053 name = "xnn_enable_assembly_explicit_true",
6054 define_values = {"xnn_enable_assembly": "true"},
6055)
6056
6057# Disables usage of assembly kernels.
6058config_setting(
6059 name = "xnn_enable_assembly_explicit_false",
6060 define_values = {"xnn_enable_assembly": "false"},
6061)
6062
Marat Dukhan9de90e02020-06-18 16:04:12 -07006063# Enables usage of sparse inference.
6064config_setting(
6065 name = "xnn_enable_sparse_explicit_true",
6066 define_values = {"xnn_enable_sparse": "true"},
6067)
6068
6069# Disables usage of sparse inference.
6070config_setting(
6071 name = "xnn_enable_sparse_explicit_false",
6072 define_values = {"xnn_enable_sparse": "false"},
6073)
6074
Marat Dukhan05702cf2020-03-26 15:41:33 -07006075# Disables usage of HMP-aware optimizations.
6076config_setting(
6077 name = "xnn_enable_hmp_explicit_false",
6078 define_values = {"xnn_enable_hmp": "false"},
6079)
6080
Chao Mei6ddfc602020-05-13 22:29:36 -07006081# Enable usage of optimized memory allocation
6082config_setting(
6083 name = "xnn_enable_memopt_explicit_true",
6084 define_values = {"xnnpack_enable_memopt": "true"},
6085)
6086
6087# Disable usage of optimized memory allocation
6088config_setting(
6089 name = "xnn_enable_memopt_explicit_false",
6090 define_values = {"xnnpack_enable_memopt": "false"},
6091)
6092
Marat Dukhanb8642352019-10-30 15:43:02 -07006093# Builds with -c dbg
6094config_setting(
6095 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006096 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006097 "compilation_mode": "dbg",
6098 },
6099)
6100
6101# Builds with -c opt
6102config_setting(
6103 name = "optimized_build",
6104 values = {
6105 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006106 },
6107)
6108
6109config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006110 name = "linux_k8",
6111 values = {"cpu": "k8"},
6112)
6113
6114config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006115 name = "linux_arm",
6116 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006117)
6118
6119config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006120 name = "linux_armeabi",
6121 values = {"cpu": "armeabi"},
6122)
6123
6124config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006125 name = "linux_armhf",
6126 values = {"cpu": "armhf"},
6127)
6128
6129config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006130 name = "linux_armv7a",
6131 values = {"cpu": "armv7a"},
6132)
6133
6134config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006135 name = "linux_aarch64",
6136 values = {"cpu": "aarch64"},
6137)
6138
6139config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006140 name = "android",
6141 values = {"crosstool_top": "//external:android/crosstool"},
6142)
6143
6144config_setting(
6145 name = "android_armv7",
6146 values = {
6147 "crosstool_top": "//external:android/crosstool",
6148 "cpu": "armeabi-v7a",
6149 },
6150)
6151
6152config_setting(
6153 name = "android_arm64",
6154 values = {
6155 "crosstool_top": "//external:android/crosstool",
6156 "cpu": "arm64-v8a",
6157 },
6158)
6159
6160config_setting(
6161 name = "android_x86",
6162 values = {
6163 "crosstool_top": "//external:android/crosstool",
6164 "cpu": "x86",
6165 },
6166)
6167
6168config_setting(
6169 name = "android_x86_64",
6170 values = {
6171 "crosstool_top": "//external:android/crosstool",
6172 "cpu": "x86_64",
6173 },
6174)
6175
6176config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006177 name = "windows_x86_64",
6178 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006179)
6180
6181config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006182 name = "windows_x86_64_clang",
6183 values = {
6184 "compiler": "clang-cl",
6185 "cpu": "x64_windows",
6186 },
6187)
6188
6189config_setting(
6190 name = "windows_x86_64_mingw",
6191 values = {
6192 "compiler": "mingw-gcc",
6193 "cpu": "x64_windows",
6194 },
6195)
6196
6197config_setting(
6198 name = "windows_x86_64_msys",
6199 values = {
6200 "compiler": "msys-gcc",
6201 "cpu": "x64_windows",
6202 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006203)
6204
6205config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006206 name = "macos_x86_64",
6207 values = {
6208 "apple_platform_type": "macos",
6209 "cpu": "darwin",
6210 },
6211)
6212
6213config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006215 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216)
6217
6218config_setting(
6219 name = "emscripten_wasm",
6220 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006221 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222 "cpu": "wasm",
6223 },
6224)
6225
6226config_setting(
6227 name = "emscripten_wasmsimd",
6228 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006229 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006231 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006232 },
6233)
6234
6235config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006236 name = "ios_armv7",
6237 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006238 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006239 "cpu": "ios_armv7",
6240 },
6241)
6242
6243config_setting(
6244 name = "ios_arm64",
6245 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006246 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006247 "cpu": "ios_arm64",
6248 },
6249)
6250
6251config_setting(
6252 name = "ios_arm64e",
6253 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006254 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006255 "cpu": "ios_arm64e",
6256 },
6257)
6258
6259config_setting(
6260 name = "ios_x86",
6261 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006262 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006263 "cpu": "ios_i386",
6264 },
6265)
6266
6267config_setting(
6268 name = "ios_x86_64",
6269 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006270 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006271 "cpu": "ios_x86_64",
6272 },
6273)
6274
6275config_setting(
6276 name = "watchos_armv7k",
6277 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006278 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006279 "cpu": "watchos_armv7k",
6280 },
6281)
6282
6283config_setting(
6284 name = "watchos_arm64_32",
6285 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006286 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006287 "cpu": "watchos_arm64_32",
6288 },
6289)
6290
6291config_setting(
6292 name = "watchos_x86",
6293 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006294 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006295 "cpu": "watchos_i386",
6296 },
6297)
6298
6299config_setting(
6300 name = "watchos_x86_64",
6301 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006302 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006303 "cpu": "watchos_x86_64",
6304 },
6305)
6306
6307config_setting(
6308 name = "tvos_arm64",
6309 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006310 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006311 "cpu": "tvos_arm64",
6312 },
6313)
6314
6315config_setting(
6316 name = "tvos_x86_64",
6317 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006318 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006319 "cpu": "tvos_x86_64",
6320 },
6321)