blob: 75d07371c33a43037bb6ef81a18bdc6ed3763e10 [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",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700425 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001074 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1075 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1076 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1078 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1079 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001082 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001083 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001084 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001085 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001086 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001087 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001088 "src/x32-zip/x2-wasmsimd.c",
1089 "src/x32-zip/x3-wasmsimd.c",
1090 "src/x32-zip/x4-wasmsimd.c",
1091 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001092 "src/math/roundne-wasmsimd-addsub.c",
1093 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001094 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001095 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001099 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1100 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001101]
1102
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001103PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001104 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001105 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001106 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001107 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001108 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001109 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1110 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1111 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/precise-psimd.c",
1116 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001117 "src/qu8-requantization/precise-psimd.c",
1118 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121# ISA-specific micro-kernels
1122NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001123 "src/f32-argmaxpool/4x-neon-c4.c",
1124 "src/f32-argmaxpool/9p8x-neon-c4.c",
1125 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001126 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1127 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001128 "src/f32-clamp/gen/neon-x4.c",
1129 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001130 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1131 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001134 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1135 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1144 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1152 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1154 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1160 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1161 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1162 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1164 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1165 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1166 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1167 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1170 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1171 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1173 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1181 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1182 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001184 "src/f32-hswish/gen/hswish-neon-x4.c",
1185 "src/f32-hswish/gen/hswish-neon-x8.c",
1186 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001187 "src/f32-ibilinear/gen/neon-c4.c",
1188 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1190 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1191 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1193 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1195 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1196 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1197 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1198 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1200 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1201 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1202 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1203 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001205 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1206 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1207 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001210 "src/f32-prelu/gen/neon-1x4.c",
1211 "src/f32-prelu/gen/neon-1x8.c",
1212 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001213 "src/f32-prelu/gen/neon-2x4.c",
1214 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001215 "src/f32-prelu/gen/neon-2x16.c",
1216 "src/f32-prelu/gen/neon-4x4.c",
1217 "src/f32-prelu/gen/neon-4x8.c",
1218 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001243 "src/f32-relu/gen/neon-x4.c",
1244 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001246 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001247 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1248 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1267 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001269 "src/f32-vbinary/gen/vmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1272 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1273 "src/f32-vbinary/gen/vmin-neon-x4.c",
1274 "src/f32-vbinary/gen/vmin-neon-x8.c",
1275 "src/f32-vbinary/gen/vminc-neon-x4.c",
1276 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001283 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1285 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001291 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1292 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001293 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1294 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001295 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1296 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1297 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001303 "src/f32-vunary/gen/vabs-neon-x4.c",
1304 "src/f32-vunary/gen/vabs-neon-x8.c",
1305 "src/f32-vunary/gen/vneg-neon-x4.c",
1306 "src/f32-vunary/gen/vneg-neon-x8.c",
1307 "src/f32-vunary/gen/vsqr-neon-x4.c",
1308 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001309 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1310 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1311 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001313 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1314 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1315 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1317 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1318 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001321 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1322 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1323 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1325 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1326 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001329 "src/qs8-requantization/precise-neon.c",
1330 "src/qs8-requantization/fp32-neon.c",
1331 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001332 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1333 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1334 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1336 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1337 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001340 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1341 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1342 "src/qu8-dwconv/up8x9-minmax-neon.c",
1343 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1344 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1345 "src/qu8-gemm/4x8-minmax-neon.c",
1346 "src/qu8-gemm/8x8-minmax-neon.c",
1347 "src/qu8-igemm/4x8-minmax-neon.c",
1348 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001349 "src/qu8-requantization/precise-neon.c",
1350 "src/qu8-requantization/fp32-neon.c",
1351 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001352 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001355 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001356 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001358 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001359 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360 "src/x32-zip/x2-neon.c",
1361 "src/x32-zip/x3-neon.c",
1362 "src/x32-zip/x4-neon.c",
1363 "src/x32-zip/xm-neon.c",
1364 "src/x8-zip/x2-neon.c",
1365 "src/x8-zip/x3-neon.c",
1366 "src/x8-zip/x4-neon.c",
1367 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001368 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-neon-addsub.c",
1370 "src/math/roundd-neon-cvt.c",
1371 "src/math/roundu-neon-addsub.c",
1372 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neon-addsub.c",
1374 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001375 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001376 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1377 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1378 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1379 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1380 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001382 "src/math/sqrt-neon-nr1rsqrts.c",
1383 "src/math/sqrt-neon-nr2rsqrts.c",
1384 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001385]
1386
1387NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001388 "src/f32-ibilinear/gen/neonfma-c4.c",
1389 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001390 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1391 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1392 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1394 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1395 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1396 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1397 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1405 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1412 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1413 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1416 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1417 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1418 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1424 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1425 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1430 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001509 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1510 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001531 "src/math/exp-neonfma-lut64-p2.c",
1532 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001533 "src/math/expminus-neonfma-lut2048-p1.c",
1534 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001535 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001536 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1537 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1539 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1542 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1543 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1545 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1546 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1548 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1551 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1552 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001554 "src/math/sqrt-neonfma-nr1fma.c",
1555 "src/math/sqrt-neonfma-nr2fma.c",
1556 "src/math/sqrt-neonfma-nr3fma.c",
1557 "src/math/sqrt-neonfma-nr2fma1adj.c",
1558 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001559]
1560
1561AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1569 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1573 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1575 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1579 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1581 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1583 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1585 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1587 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1589 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001592 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1593 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001596 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1598 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1599 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001601 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001619 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1620 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1621 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1622 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1623 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1625 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1626 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1627 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1628 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1630 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1631 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1632 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1633 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1635 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1636 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001637 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1638 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1640 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1641 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1643 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1644 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645]
1646
Marat Dukhan8853b822020-05-07 12:19:01 -07001647NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001648 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1649 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1650 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001656 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001657 "src/math/roundd-neonv8.c",
1658 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001659 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001660]
1661
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001663 "src/f16-clamp/gen/neonfp16arith-x16.c",
1664 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001665 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001677 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1678 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001679 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1680 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001681 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001682 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001683 "src/f16-relu/gen/neonfp16arith-x16.c",
1684 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001685 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001713 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1714 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001715 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001723 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001727 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001739 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1740 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1741 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Benoit Jacoba9644732020-08-13 12:48:55 -07001749NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001750 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001751 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001753 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1754 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001755 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001756 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001757 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001758 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001759 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001760 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
1761 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001762 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1763 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001764 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001765 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001766 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001767 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001768]
1769
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001771 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1772 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001773 "src/f32-clamp/gen/sse-x4.c",
1774 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001775 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1776 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001777 "src/f32-dwconv-chw/3x3p1-sse.c",
1778 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001779 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1780 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1781 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1783 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1784 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1785 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1786 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1787 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1788 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1789 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1790 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001791 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001792 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1793 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001794 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1795 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1796 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1797 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1798 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1799 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1800 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1801 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1802 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1803 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1804 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1805 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1806 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001807 "src/f32-hswish/gen/hswish-sse-x4.c",
1808 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001809 "src/f32-ibilinear/gen/sse-c4.c",
1810 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001811 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1812 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1813 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1814 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1815 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1816 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1817 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001818 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1819 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1820 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001821 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001822 "src/f32-prelu/gen/sse-2x4.c",
1823 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001824 "src/f32-relu/gen/sse-x4.c",
1825 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001827 "src/f32-spmm/gen/4x1-minmax-sse.c",
1828 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001829 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001830 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1832 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1834 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1835 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1836 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1837 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001838 "src/f32-vbinary/gen/vmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1841 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1842 "src/f32-vbinary/gen/vmin-sse-x4.c",
1843 "src/f32-vbinary/gen/vmin-sse-x8.c",
1844 "src/f32-vbinary/gen/vminc-sse-x4.c",
1845 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001846 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1848 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1849 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1850 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1851 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1852 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001854 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1855 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1856 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1857 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001858 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1859 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1860 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1861 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001862 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1863 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001864 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1865 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001866 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1867 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001868 "src/f32-vunary/gen/vabs-sse-x4.c",
1869 "src/f32-vunary/gen/vabs-sse-x8.c",
1870 "src/f32-vunary/gen/vneg-sse-x4.c",
1871 "src/f32-vunary/gen/vneg-sse-x8.c",
1872 "src/f32-vunary/gen/vsqr-sse-x4.c",
1873 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001874 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001875 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001876 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001877 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001878 "src/math/roundd-sse-addsub.c",
1879 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001880 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001881 "src/math/sqrt-sse-hh1mac.c",
1882 "src/math/sqrt-sse-nr1mac.c",
1883 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001884]
1885
1886SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001887 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1888 "src/f32-argmaxpool/4x-sse2-c4.c",
1889 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001890 "src/f32-prelu/gen/sse2-2x4.c",
1891 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1896 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1897 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1898 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1899 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1901 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1902 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001904 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1905 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1906 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1907 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1908 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1909 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001910 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001911 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001912 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001913 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001914 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1915 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001916 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1917 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001918 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1919 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1920 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1921 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1922 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1923 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1924 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1925 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001926 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1927 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1928 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001929 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1930 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1931 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001932 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1933 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1934 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001935 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1936 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001937 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1938 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001939 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1940 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001941 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1942 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001943 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001944 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1945 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001946 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001947 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1948 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1949 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001950 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1951 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001952 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1953 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001954 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1955 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001956 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001957 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1958 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001959 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001960 "src/qs8-requantization/precise-sse2.c",
1961 "src/qs8-requantization/fp32-sse2.c",
1962 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001963 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1964 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1965 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1966 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001967 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1968 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1969 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1970 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001971 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1972 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1973 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1974 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1975 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1976 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1977 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1978 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001979 "src/qu8-requantization/precise-sse2.c",
1980 "src/qu8-requantization/fp32-sse2.c",
1981 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001982 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001983 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001984 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001985 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001986 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987 "src/x32-zip/x2-sse2.c",
1988 "src/x32-zip/x3-sse2.c",
1989 "src/x32-zip/x4-sse2.c",
1990 "src/x32-zip/xm-sse2.c",
1991 "src/x8-zip/x2-sse2.c",
1992 "src/x8-zip/x3-sse2.c",
1993 "src/x8-zip/x4-sse2.c",
1994 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001995 "src/math/exp-sse2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07001996 "src/math/exp-sse2-lut64-p2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001997 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001998 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001999 "src/math/roundd-sse2-cvt.c",
2000 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002001 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07002002 "src/math/sigmoid-sse2-rr2-p5-div.c",
2003 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002004]
2005
2006SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002007 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2008 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2009 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002010 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2011 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2012 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002013 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2014 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2015 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002016 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2017 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002018 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2019 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002020 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2021 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002022 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2023 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002024 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002025 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2026 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002027 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002028 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2029 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2030 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002031 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2032 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002033 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2034 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002035 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2036 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002037 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002038 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2039 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002040 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002041 "src/qs8-requantization/precise-ssse3.c",
2042 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002043 "src/qu8-requantization/precise-ssse3.c",
2044 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002045]
2046
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002047SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002048 "src/f32-prelu/gen/sse41-2x4.c",
2049 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002050 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2051 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2052 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2053 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2054 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2055 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002056 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002057 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002058 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002059 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002060 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2061 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002062 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2063 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002064 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2065 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2066 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2067 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2068 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2069 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2070 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2071 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002072 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2073 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2074 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002075 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2076 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2077 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002078 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2079 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2080 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002081 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2082 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002083 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2084 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002085 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2086 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002087 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2088 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002089 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002090 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2091 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002092 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002093 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2094 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2095 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002096 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2097 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002098 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2099 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002100 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2101 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002102 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002103 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2104 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002105 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002106 "src/qs8-requantization/fp32-sse4.c",
2107 "src/qs8-requantization/precise-sse4.c",
2108 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002109 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2110 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2111 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2112 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002113 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2114 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2115 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2116 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002117 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2118 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2119 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2120 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002121 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2122 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2123 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2124 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002125 "src/qu8-requantization/precise-sse4.c",
2126 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002127 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002128 "src/math/roundd-sse41.c",
2129 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002130 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002131]
2132
Marat Dukhan08c4a432019-10-03 09:29:21 -07002133AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002134 "src/f32-clamp/gen/avx-x8.c",
2135 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002136 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2137 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2138 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2139 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2140 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2141 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2142 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2143 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2144 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2145 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2146 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2147 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2148 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2149 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2150 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2151 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2152 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2153 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2154 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2157 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2158 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2159 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2160 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2161 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2162 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2163 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2164 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2165 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002166 "src/f32-hswish/gen/hswish-avx-x8.c",
2167 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002168 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2169 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2170 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2171 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2172 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2173 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2174 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2175 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2176 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002177 "src/f32-prelu/gen/avx-2x8.c",
2178 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002179 "src/f32-relu/gen/avx-x16.c",
2180 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002181 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002182 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2183 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2184 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2185 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2186 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2187 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2188 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2189 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002190 "src/f32-vbinary/gen/vmax-avx-x8.c",
2191 "src/f32-vbinary/gen/vmax-avx-x16.c",
2192 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2193 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2194 "src/f32-vbinary/gen/vmin-avx-x8.c",
2195 "src/f32-vbinary/gen/vmin-avx-x16.c",
2196 "src/f32-vbinary/gen/vminc-avx-x8.c",
2197 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002198 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2199 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2200 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2201 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2202 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2203 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2204 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2205 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002206 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2207 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2208 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2209 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002210 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2211 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2212 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2213 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002214 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2215 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002216 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002217 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2218 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2219 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2220 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2221 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2222 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2223 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2224 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002225 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2226 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002227 "src/f32-vunary/gen/vabs-avx-x8.c",
2228 "src/f32-vunary/gen/vabs-avx-x16.c",
2229 "src/f32-vunary/gen/vneg-avx-x8.c",
2230 "src/f32-vunary/gen/vneg-avx-x16.c",
2231 "src/f32-vunary/gen/vsqr-avx-x8.c",
2232 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002233 "src/math/exp-avx-p5.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002234 "src/math/sigmoid-avx-rr2-p5-div.c",
2235 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236]
2237
Marat Dukhan1566fee2020-08-02 21:55:41 -07002238XOP_UKERNELS = [
2239 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2240 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2241 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2242 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002243 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2244 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002245 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2246 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002247 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002248 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2249 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002250 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002251 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2252 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2253 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002254 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2255 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2256 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2257 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2258 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2259 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002260 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002261 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2262 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002263 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002264 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2265 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2266 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2267 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2268 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2269 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2270 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2271 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002272]
2273
Marat Dukhanfda12b82019-11-21 12:27:59 -08002274FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002275 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2276 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2277 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2278 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2279 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2280 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2281 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2282 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2283 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2284 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2285 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2286 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2287 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2288 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2289 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2291 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2292 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2293 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2294 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2295 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2296 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2297 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2298 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2299 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2300 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2301 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2302 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2303 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2304 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2305 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2306 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2307 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2308 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2309 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2310 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2311 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2312 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2313 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2314 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002315 "src/f32-hswish/gen/hswish-fma3-x8.c",
2316 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002317 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2318 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2319 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2320 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2321 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2322 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2323 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2324 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2325 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2326 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2327 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2328 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2329 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2330 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002331 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2332 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2333 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2334 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2335 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2336 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2337 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2338 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002339 "src/math/sqrt-fma3-nr1fma.c",
2340 "src/math/sqrt-fma3-nr2fma.c",
2341 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002342]
2343
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002344AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002345 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2346 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2347 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2348 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2349 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2350 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2351 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2352 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2353 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2354 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2355 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2356 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2357 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2358 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2359 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2360 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2361 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2362 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2363 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2364 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2365 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2366 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2367 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2368 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2369 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2375 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2376 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2377 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2378 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2379 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2380 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002381 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2387 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2388 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2389 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2390 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2405 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2406 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2407 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2408 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2409 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2410 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002411 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2417 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2418 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2419 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2420 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2421 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2422 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2423 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2424 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2425 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2426 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2427 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2428 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2429 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2430 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2431 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2432 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2433 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2434 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002435 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2436 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002437 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2438 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2439 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2440 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002441 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2442 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2443 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002444 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2445 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2446 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002447 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2448 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2449 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002450 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2451 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2452 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2453 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2454 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2455 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2456 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2457 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002458 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002459 "src/math/exp-avx2-lut8-p3-perm.c",
2460 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002461 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002462 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002463 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002464 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002465 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002466 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002467 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002468 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002469]
2470
Marat Dukhan08c4a432019-10-03 09:29:21 -07002471AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002472 "src/f32-clamp/gen/avx512f-x16.c",
2473 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002474 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2475 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2476 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2477 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2478 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2479 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2480 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2481 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2482 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2483 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2484 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2485 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2486 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2487 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2488 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2489 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2490 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2491 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2492 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2493 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2494 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2495 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2496 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2497 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002498 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2499 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002500 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2501 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2502 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2503 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2504 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2505 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002506 "src/f32-prelu/gen/avx512f-2x16.c",
2507 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002508 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2509 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2510 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2511 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2512 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2513 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2514 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2515 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2516 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2517 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2518 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2519 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2520 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2521 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2522 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2523 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2524 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2525 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2526 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2527 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2528 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2529 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2530 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2531 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2532 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2533 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2534 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2535 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2536 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2537 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2538 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2539 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2540 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2541 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2542 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2543 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002544 "src/f32-relu/gen/avx512f-x16.c",
2545 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002547 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2548 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2549 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2550 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2551 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2552 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2553 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2554 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2555 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2556 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2557 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2558 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2559 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2560 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2561 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2562 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2563 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2564 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2565 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2566 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2567 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2568 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2569 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2570 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2571 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2572 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2573 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2574 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2575 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2576 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2577 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2578 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2579 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2580 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2581 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2582 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2583 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2584 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2585 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2586 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2587 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2588 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2589 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2590 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2591 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2592 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2593 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2594 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002595 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2596 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2597 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2598 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2599 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2600 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2601 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2602 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002603 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2604 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2605 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2606 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2607 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2608 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2609 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2610 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002611 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2612 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2613 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2614 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2615 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2616 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2617 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2618 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002619 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2620 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2621 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2622 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002623 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2624 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2625 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2626 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002627 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2628 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002629 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002630 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2631 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2632 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2633 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2634 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2635 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2636 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2637 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2638 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2639 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2640 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2641 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2642 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2643 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2644 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2645 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2646 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2647 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2648 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2649 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2650 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2651 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2652 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2653 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002654 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2655 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2656 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2657 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2658 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2659 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2660 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2661 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002662 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2663 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2664 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2665 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2666 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2667 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2668 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2669 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002670 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2671 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2672 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2673 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2674 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2675 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002676 "src/math/exp-avx512f-p5-scalef.c",
2677 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002678 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002679 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002680 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002681 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002682 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002683 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002684 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002685 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002686 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002687 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002688 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002689 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2690 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2691 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2692 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2693 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2694 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2695 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2696 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2697 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2698 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2699 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2700 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002701 "src/math/sqrt-avx512f-nr1fma.c",
2702 "src/math/sqrt-avx512f-nr2fma.c",
2703 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002704]
2705
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002706AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002707 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2708 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002709 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2710 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2711 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2712 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2713 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2714 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2715 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2716 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2717]
2718
Frank Barchardbcedc082020-08-17 18:00:51 -07002719WASM32_ASM_UKERNELS = [
2720 "src/f32-relu/wasm_shr_x1.S",
2721 "src/f32-relu/wasm_shr_x2.S",
2722 "src/f32-relu/wasm_shr_x4.S",
2723]
2724
Marat Dukhan08c4a432019-10-03 09:29:21 -07002725AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002726 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002727 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002728 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2729 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002730 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002731 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002732 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002733 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002734 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2735 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002736 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2737 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2738 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2739 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740]
2741
2742AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002743 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002744 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2745 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002746 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002747 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2748 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002749 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2750 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2751 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002752 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002753 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2754 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2755 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002756 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002757 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2758 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002759 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002760 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2761 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2762 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2763 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2764 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2765 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2766 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2767 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2768 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2769 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2770 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2771 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2772 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2773 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2774 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2775 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2776 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2777 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002778 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2779 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002780 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002781 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2782 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2783 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2784 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2785 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2786 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2787 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2788 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2789 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2790 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2791 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2792 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2793 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2794 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2795 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2796 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2797 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2798 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2800 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2801 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2802 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2803 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2804 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2805 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2806 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2807 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2808 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2809 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2810 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2811 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2812 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2813 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2814 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2815 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2816 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002817]
2818
Marat Dukhan1b354632020-03-23 12:50:22 -07002819INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 "src/xnnpack/argmaxpool.h",
2821 "src/xnnpack/avgpool.h",
2822 "src/xnnpack/clamp.h",
2823 "src/xnnpack/common.h",
2824 "src/xnnpack/conv.h",
2825 "src/xnnpack/dwconv.h",
2826 "src/xnnpack/gavgpool.h",
2827 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002828 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002829 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002830 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002832 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002833 "src/xnnpack/lut.h",
2834 "src/xnnpack/math.h",
2835 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002836 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 "src/xnnpack/packx.h",
2838 "src/xnnpack/pad.h",
2839 "src/xnnpack/params.h",
2840 "src/xnnpack/pavgpool.h",
2841 "src/xnnpack/ppmm.h",
2842 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002843 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002844 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002845 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002846 "src/xnnpack/rmax.h",
2847 "src/xnnpack/scalar-utils.h",
2848 "src/xnnpack/spmm.h",
2849 "src/xnnpack/unpool.h",
2850 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002851 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002852 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002853 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002854 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002855 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002856 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002857 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002858]
2859
2860INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861 "include/xnnpack.h",
2862 "src/xnnpack/allocator.h",
2863 "src/xnnpack/compute.h",
2864 "src/xnnpack/im2col.h",
2865 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002866 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002867 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002868 "src/xnnpack/operator.h",
2869 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002870 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002872 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002873 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002874]
2875
Marat Dukhan1b354632020-03-23 12:50:22 -07002876ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002877 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878]
2879
Marat Dukhan1b354632020-03-23 12:50:22 -07002880MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002881 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 "include/xnnpack.h",
2883]
2884
Marat Dukhan1b354632020-03-23 12:50:22 -07002885MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002886 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002887 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002888 "src/xnnpack/requantization.h",
2889 "include/xnnpack.h",
2890]
2891
2892OPERATOR_TEST_PARAMS_HDRS = [
2893 "src/xnnpack/params.h",
2894 "src/xnnpack/common.h",
2895]
2896
2897WEIGHTS_PACK_HDRS = [
2898 "src/xnnpack/pack.h",
2899 "src/xnnpack/operator.h",
2900 "src/xnnpack/compute.h",
2901]
2902
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002903LOGGING_COPTS = select({
2904 # No logging in optimized mode
2905 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2906 # Full logging in debug mode
2907 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2908 # Error-only logging in default (fastbuild) mode
2909 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2910})
2911
Marat Dukhan3b59de22020-06-03 20:15:19 -07002912LOGGING_SRCS = select({
2913 # No logging in optimized mode
2914 ":optimized_build": [],
2915 "//conditions:default": [
2916 "src/operator-strings.c",
2917 "src/subgraph-strings.c",
2918 ],
2919})
2920
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002921LOGGING_HDRS = [
2922 "src/xnnpack/log.h",
2923]
2924
Marat Dukhan08c4a432019-10-03 09:29:21 -07002925xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002926 name = "tables",
2927 srcs = TABLE_SRCS,
2928 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002929 gcc_copts = xnnpack_gcc_std_copts(),
2930 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002931)
2932
2933xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 name = "scalar_ukernels",
2935 srcs = SCALAR_UKERNELS,
2936 hdrs = INTERNAL_HDRS,
2937 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002938 gcc_copts = xnnpack_gcc_std_copts(),
2939 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002941 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 "@FP16",
2943 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002944 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 ],
2946)
2947
2948xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002949 name = "scalar_ukernels_test_mode",
2950 srcs = SCALAR_UKERNELS,
2951 hdrs = INTERNAL_HDRS,
2952 aarch32_copts = ["-marm"],
2953 copts = [
2954 "-UNDEBUG",
2955 "-DXNN_TEST_MODE=1",
2956 ],
2957 gcc_copts = xnnpack_gcc_std_copts(),
2958 msvc_copts = xnnpack_msvc_std_copts(),
2959 deps = [
2960 ":tables",
2961 "@FP16",
2962 "@FXdiv",
2963 "@pthreadpool",
2964 ],
2965)
2966
2967xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002968 name = "wasm_ukernels",
2969 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002970 gcc_copts = xnnpack_gcc_std_copts(),
2971 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002972 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002973 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002974 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002975 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002976 "@FP16",
2977 "@FXdiv",
2978 "@pthreadpool",
2979 ],
2980)
2981
2982xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002983 name = "wasm_ukernels_test_mode",
2984 hdrs = INTERNAL_HDRS,
2985 copts = [
2986 "-UNDEBUG",
2987 "-DXNN_TEST_MODE=1",
2988 ],
2989 gcc_copts = xnnpack_gcc_std_copts(),
2990 msvc_copts = xnnpack_msvc_std_copts(),
2991 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002992 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002993 deps = [
2994 ":tables",
2995 "@FP16",
2996 "@FXdiv",
2997 "@pthreadpool",
2998 ],
2999)
3000
3001xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003002 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003003 hdrs = INTERNAL_HDRS,
3004 aarch32_copts = [
3005 "-marm",
3006 "-mfpu=neon",
3007 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003008 gcc_copts = xnnpack_gcc_std_copts(),
3009 gcc_x86_copts = ["-msse2"],
3010 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 optimized_copts = [
3012 "-O3",
3013 "-ffast-math",
3014 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003015 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003017 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 "@FP16",
3019 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003020 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 ],
3022)
3023
3024xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003025 name = "psimd_fastmath_ukernels_test_mode",
3026 hdrs = INTERNAL_HDRS,
3027 aarch32_copts = [
3028 "-marm",
3029 "-mfpu=neon",
3030 ],
3031 copts = [
3032 "-UNDEBUG",
3033 "-DXNN_TEST_MODE=1",
3034 ],
3035 gcc_copts = xnnpack_gcc_std_copts(),
3036 gcc_x86_copts = ["-msse2"],
3037 msvc_copts = xnnpack_msvc_std_copts(),
3038 optimized_copts = [
3039 "-O3",
3040 "-ffast-math",
3041 ],
3042 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3043 deps = [
3044 ":tables",
3045 "@FP16",
3046 "@psimd",
3047 "@pthreadpool",
3048 ],
3049)
3050
3051xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003052 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003053 hdrs = INTERNAL_HDRS,
3054 aarch32_copts = [
3055 "-marm",
3056 "-mfpu=neon",
3057 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003058 gcc_copts = xnnpack_gcc_std_copts(),
3059 gcc_x86_copts = ["-msse2"],
3060 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003061 optimized_copts = [
3062 "-O3",
3063 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003064 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003065 deps = [
3066 ":tables",
3067 "@FP16",
3068 "@psimd",
3069 "@pthreadpool",
3070 ],
3071)
3072
3073xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003074 name = "psimd_accmath_ukernels_test_mode",
3075 hdrs = INTERNAL_HDRS,
3076 aarch32_copts = [
3077 "-marm",
3078 "-mfpu=neon",
3079 ],
3080 copts = [
3081 "-UNDEBUG",
3082 "-DXNN_TEST_MODE=1",
3083 ],
3084 gcc_copts = xnnpack_gcc_std_copts(),
3085 gcc_x86_copts = ["-msse2"],
3086 msvc_copts = xnnpack_msvc_std_copts(),
3087 optimized_copts = [
3088 "-O3",
3089 ],
3090 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3091 deps = [
3092 ":tables",
3093 "@FP16",
3094 "@psimd",
3095 "@pthreadpool",
3096 ],
3097)
3098
3099xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003100 name = "neon_ukernels",
3101 hdrs = INTERNAL_HDRS,
3102 aarch32_copts = [
3103 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003104 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 "-mfpu=neon",
3106 ],
3107 aarch32_srcs = NEON_UKERNELS,
3108 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003109 gcc_copts = xnnpack_gcc_std_copts(),
3110 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003111 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003112 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003113 "@FP16",
3114 "@pthreadpool",
3115 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003116)
3117
3118xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003119 name = "neon_ukernels_test_mode",
3120 hdrs = INTERNAL_HDRS,
3121 aarch32_copts = [
3122 "-marm",
3123 "-march=armv7-a",
3124 "-mfpu=neon",
3125 ],
3126 aarch32_srcs = NEON_UKERNELS,
3127 aarch64_srcs = NEON_UKERNELS,
3128 copts = [
3129 "-UNDEBUG",
3130 "-DXNN_TEST_MODE=1",
3131 ],
3132 gcc_copts = xnnpack_gcc_std_copts(),
3133 msvc_copts = xnnpack_msvc_std_copts(),
3134 deps = [
3135 ":tables",
3136 "@FP16",
3137 "@pthreadpool",
3138 ],
3139)
3140
3141xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 name = "neonfma_ukernels",
3143 hdrs = INTERNAL_HDRS,
3144 aarch32_copts = [
3145 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003146 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147 "-mfpu=neon-vfpv4",
3148 ],
3149 aarch32_srcs = NEONFMA_UKERNELS,
3150 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003151 apple_aarch32_copts = [
3152 "-mcpu=swift",
3153 "-mtune=generic",
3154 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003155 gcc_copts = xnnpack_gcc_std_copts(),
3156 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003157 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003158 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003159 "@FP16",
3160 "@pthreadpool",
3161 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162)
3163
3164xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003165 name = "neonfma_ukernels_test_mode",
3166 hdrs = INTERNAL_HDRS,
3167 aarch32_copts = [
3168 "-marm",
3169 "-march=armv7-a",
3170 "-mfpu=neon-vfpv4",
3171 ],
3172 aarch32_srcs = NEONFMA_UKERNELS,
3173 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003174 apple_aarch32_copts = [
3175 "-mcpu=swift",
3176 "-mtune=generic",
3177 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003178 copts = [
3179 "-UNDEBUG",
3180 "-DXNN_TEST_MODE=1",
3181 ],
3182 gcc_copts = xnnpack_gcc_std_copts(),
3183 msvc_copts = xnnpack_msvc_std_copts(),
3184 deps = [
3185 ":tables",
3186 "@FP16",
3187 "@pthreadpool",
3188 ],
3189)
3190
3191xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003192 name = "neonv8_ukernels",
3193 hdrs = INTERNAL_HDRS,
3194 aarch32_copts = [
3195 "-marm",
3196 "-march=armv8-a",
3197 "-mfpu=neon-fp-armv8",
3198 ],
3199 aarch32_srcs = NEONV8_UKERNELS,
3200 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003201 apple_aarch32_copts = [
3202 "-mcpu=cyclone",
3203 "-mtune=generic",
3204 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003205 gcc_copts = xnnpack_gcc_std_copts(),
3206 msvc_copts = xnnpack_msvc_std_copts(),
3207 deps = [
3208 ":tables",
3209 "@FP16",
3210 "@pthreadpool",
3211 ],
3212)
3213
3214xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003215 name = "neonv8_ukernels_test_mode",
3216 hdrs = INTERNAL_HDRS,
3217 aarch32_copts = [
3218 "-marm",
3219 "-march=armv8-a",
3220 "-mfpu=neon-fp-armv8",
3221 ],
3222 aarch32_srcs = NEONV8_UKERNELS,
3223 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003224 apple_aarch32_copts = [
3225 "-mcpu=cyclone",
3226 "-mtune=generic",
3227 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003228 copts = [
3229 "-UNDEBUG",
3230 "-DXNN_TEST_MODE=1",
3231 ],
3232 gcc_copts = xnnpack_gcc_std_copts(),
3233 msvc_copts = xnnpack_msvc_std_copts(),
3234 deps = [
3235 ":tables",
3236 "@FP16",
3237 "@pthreadpool",
3238 ],
3239)
3240
3241xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242 name = "neonfp16arith_ukernels",
3243 hdrs = INTERNAL_HDRS,
3244 aarch64_copts = ["-march=armv8.2-a+fp16"],
3245 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003246 gcc_copts = xnnpack_gcc_std_copts(),
3247 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003248 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003249 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003250 "@FP16",
3251 "@pthreadpool",
3252 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253)
3254
3255xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003256 name = "neonfp16arith_ukernels_test_mode",
3257 hdrs = INTERNAL_HDRS,
3258 aarch64_copts = ["-march=armv8.2-a+fp16"],
3259 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3260 copts = [
3261 "-UNDEBUG",
3262 "-DXNN_TEST_MODE=1",
3263 ],
3264 gcc_copts = xnnpack_gcc_std_copts(),
3265 msvc_copts = xnnpack_msvc_std_copts(),
3266 deps = [
3267 ":tables",
3268 "@FP16",
3269 "@pthreadpool",
3270 ],
3271)
3272
3273xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003274 name = "neondot_ukernels",
3275 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003276 aarch32_copts = [
3277 "-marm",
3278 "-march=armv8.2-a+dotprod",
3279 "-mfpu=neon-fp-armv8",
3280 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003281 aarch32_srcs = NEONDOT_UKERNELS,
3282 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3283 aarch64_srcs = NEONDOT_UKERNELS,
3284 gcc_copts = xnnpack_gcc_std_copts(),
3285 msvc_copts = xnnpack_msvc_std_copts(),
3286 deps = [
3287 ":tables",
3288 "@FP16",
3289 "@pthreadpool",
3290 ],
3291)
3292
3293xnnpack_cc_library(
3294 name = "neondot_ukernels_test_mode",
3295 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003296 aarch32_copts = [
3297 "-marm",
3298 "-march=armv8.2-a+dotprod",
3299 "-mfpu=neon-fp-armv8",
3300 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003301 aarch32_srcs = NEONDOT_UKERNELS,
3302 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3303 aarch64_srcs = NEONDOT_UKERNELS,
3304 copts = [
3305 "-UNDEBUG",
3306 "-DXNN_TEST_MODE=1",
3307 ],
3308 gcc_copts = xnnpack_gcc_std_copts(),
3309 msvc_copts = xnnpack_msvc_std_copts(),
3310 deps = [
3311 ":tables",
3312 "@FP16",
3313 "@pthreadpool",
3314 ],
3315)
3316
3317xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003318 name = "sse2_ukernels",
3319 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003320 gcc_copts = xnnpack_gcc_std_copts(),
3321 gcc_x86_copts = ["-msse2"],
3322 msvc_copts = xnnpack_msvc_std_copts(),
3323 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003324 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003325 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003326 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003327 "@FP16",
3328 "@pthreadpool",
3329 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003330)
3331
3332xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003333 name = "sse2_ukernels_test_mode",
3334 hdrs = INTERNAL_HDRS,
3335 copts = [
3336 "-UNDEBUG",
3337 "-DXNN_TEST_MODE=1",
3338 ],
3339 gcc_copts = xnnpack_gcc_std_copts(),
3340 gcc_x86_copts = ["-msse2"],
3341 msvc_copts = xnnpack_msvc_std_copts(),
3342 msvc_x86_32_copts = ["/arch:SSE2"],
3343 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3344 deps = [
3345 ":tables",
3346 "@FP16",
3347 "@pthreadpool",
3348 ],
3349)
3350
3351xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003352 name = "ssse3_ukernels",
3353 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003354 gcc_copts = xnnpack_gcc_std_copts(),
3355 gcc_x86_copts = ["-mssse3"],
3356 msvc_copts = xnnpack_msvc_std_copts(),
3357 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003358 x86_srcs = SSSE3_UKERNELS,
3359 deps = [
3360 ":tables",
3361 "@FP16",
3362 "@pthreadpool",
3363 ],
3364)
3365
3366xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003367 name = "ssse3_ukernels_test_mode",
3368 hdrs = INTERNAL_HDRS,
3369 copts = [
3370 "-UNDEBUG",
3371 "-DXNN_TEST_MODE=1",
3372 ],
3373 gcc_copts = xnnpack_gcc_std_copts(),
3374 gcc_x86_copts = ["-mssse3"],
3375 msvc_copts = xnnpack_msvc_std_copts(),
3376 msvc_x86_32_copts = ["/arch:SSE2"],
3377 x86_srcs = SSSE3_UKERNELS,
3378 deps = [
3379 ":tables",
3380 "@FP16",
3381 "@pthreadpool",
3382 ],
3383)
3384
3385xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003386 name = "sse41_ukernels",
3387 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003388 gcc_copts = xnnpack_gcc_std_copts(),
3389 gcc_x86_copts = ["-msse4.1"],
3390 msvc_copts = xnnpack_msvc_std_copts(),
3391 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003392 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003393 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003394 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003395 "@FP16",
3396 "@pthreadpool",
3397 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003398)
3399
3400xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003401 name = "sse41_ukernels_test_mode",
3402 hdrs = INTERNAL_HDRS,
3403 copts = [
3404 "-UNDEBUG",
3405 "-DXNN_TEST_MODE=1",
3406 ],
3407 gcc_copts = xnnpack_gcc_std_copts(),
3408 gcc_x86_copts = ["-msse4.1"],
3409 msvc_copts = xnnpack_msvc_std_copts(),
3410 msvc_x86_32_copts = ["/arch:SSE2"],
3411 x86_srcs = SSE41_UKERNELS,
3412 deps = [
3413 ":tables",
3414 "@FP16",
3415 "@pthreadpool",
3416 ],
3417)
3418
3419xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420 name = "avx_ukernels",
3421 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003422 gcc_copts = xnnpack_gcc_std_copts(),
3423 gcc_x86_copts = ["-mavx"],
3424 msvc_copts = xnnpack_msvc_std_copts(),
3425 msvc_x86_32_copts = ["/arch:AVX"],
3426 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003428 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003429 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003430 "@FP16",
3431 "@pthreadpool",
3432 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433)
3434
3435xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003436 name = "avx_ukernels_test_mode",
3437 hdrs = INTERNAL_HDRS,
3438 copts = [
3439 "-UNDEBUG",
3440 "-DXNN_TEST_MODE=1",
3441 ],
3442 gcc_copts = xnnpack_gcc_std_copts(),
3443 gcc_x86_copts = ["-mavx"],
3444 msvc_copts = xnnpack_msvc_std_copts(),
3445 msvc_x86_32_copts = ["/arch:AVX"],
3446 msvc_x86_64_copts = ["/arch:AVX"],
3447 x86_srcs = AVX_UKERNELS,
3448 deps = [
3449 ":tables",
3450 "@FP16",
3451 "@pthreadpool",
3452 ],
3453)
3454
3455xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003456 name = "xop_ukernels",
3457 hdrs = INTERNAL_HDRS,
3458 gcc_copts = xnnpack_gcc_std_copts(),
3459 gcc_x86_copts = ["-mxop"],
3460 msvc_copts = xnnpack_msvc_std_copts(),
3461 msvc_x86_32_copts = ["/arch:AVX"],
3462 msvc_x86_64_copts = ["/arch:AVX"],
3463 x86_srcs = XOP_UKERNELS,
3464 deps = [
3465 ":tables",
3466 "@FP16",
3467 "@pthreadpool",
3468 ],
3469)
3470
3471xnnpack_cc_library(
3472 name = "xop_ukernels_test_mode",
3473 hdrs = INTERNAL_HDRS,
3474 copts = [
3475 "-UNDEBUG",
3476 "-DXNN_TEST_MODE=1",
3477 ],
3478 gcc_copts = xnnpack_gcc_std_copts(),
3479 gcc_x86_copts = ["-mxop"],
3480 msvc_copts = xnnpack_msvc_std_copts(),
3481 msvc_x86_32_copts = ["/arch:AVX"],
3482 msvc_x86_64_copts = ["/arch:AVX"],
3483 x86_srcs = XOP_UKERNELS,
3484 deps = [
3485 ":tables",
3486 "@FP16",
3487 "@pthreadpool",
3488 ],
3489)
3490
3491xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003492 name = "fma3_ukernels",
3493 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003494 gcc_copts = xnnpack_gcc_std_copts(),
3495 gcc_x86_copts = ["-mfma"],
3496 msvc_copts = xnnpack_msvc_std_copts(),
3497 msvc_x86_32_copts = ["/arch:AVX"],
3498 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003499 x86_srcs = FMA3_UKERNELS,
3500 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003501 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003502 "@FP16",
3503 "@pthreadpool",
3504 ],
3505)
3506
3507xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003508 name = "fma3_ukernels_test_mode",
3509 hdrs = INTERNAL_HDRS,
3510 copts = [
3511 "-UNDEBUG",
3512 "-DXNN_TEST_MODE=1",
3513 ],
3514 gcc_copts = xnnpack_gcc_std_copts(),
3515 gcc_x86_copts = ["-mfma"],
3516 msvc_copts = xnnpack_msvc_std_copts(),
3517 msvc_x86_32_copts = ["/arch:AVX"],
3518 msvc_x86_64_copts = ["/arch:AVX"],
3519 x86_srcs = FMA3_UKERNELS,
3520 deps = [
3521 ":tables",
3522 "@FP16",
3523 "@pthreadpool",
3524 ],
3525)
3526
3527xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003528 name = "avx2_ukernels",
3529 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003530 gcc_copts = xnnpack_gcc_std_copts(),
3531 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003532 "-mfma",
3533 "-mavx2",
3534 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003535 msvc_copts = xnnpack_msvc_std_copts(),
3536 msvc_x86_32_copts = ["/arch:AVX2"],
3537 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003538 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003539 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003540 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003541 "@FP16",
3542 "@pthreadpool",
3543 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003544)
3545
3546xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003547 name = "avx2_ukernels_test_mode",
3548 hdrs = INTERNAL_HDRS,
3549 copts = [
3550 "-UNDEBUG",
3551 "-DXNN_TEST_MODE=1",
3552 ],
3553 gcc_copts = xnnpack_gcc_std_copts(),
3554 gcc_x86_copts = [
3555 "-mfma",
3556 "-mavx2",
3557 ],
3558 msvc_copts = xnnpack_msvc_std_copts(),
3559 msvc_x86_32_copts = ["/arch:AVX2"],
3560 msvc_x86_64_copts = ["/arch:AVX2"],
3561 x86_srcs = AVX2_UKERNELS,
3562 deps = [
3563 ":tables",
3564 "@FP16",
3565 "@pthreadpool",
3566 ],
3567)
3568
3569xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 name = "avx512f_ukernels",
3571 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003572 gcc_copts = xnnpack_gcc_std_copts(),
3573 gcc_x86_copts = ["-mavx512f"],
3574 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3575 msvc_copts = xnnpack_msvc_std_copts(),
3576 msvc_x86_32_copts = ["/arch:AVX512"],
3577 msvc_x86_64_copts = ["/arch:AVX512"],
3578 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003580 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003581 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003582 "@FP16",
3583 "@pthreadpool",
3584 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585)
3586
3587xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003588 name = "avx512f_ukernels_test_mode",
3589 hdrs = INTERNAL_HDRS,
3590 copts = [
3591 "-UNDEBUG",
3592 "-DXNN_TEST_MODE=1",
3593 ],
3594 gcc_copts = xnnpack_gcc_std_copts(),
3595 gcc_x86_copts = ["-mavx512f"],
3596 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3597 msvc_copts = xnnpack_msvc_std_copts(),
3598 msvc_x86_32_copts = ["/arch:AVX512"],
3599 msvc_x86_64_copts = ["/arch:AVX512"],
3600 msys_copts = ["-fno-asynchronous-unwind-tables"],
3601 x86_srcs = AVX512F_UKERNELS,
3602 deps = [
3603 ":tables",
3604 "@FP16",
3605 "@pthreadpool",
3606 ],
3607)
3608
3609xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003610 name = "avx512skx_ukernels",
3611 hdrs = INTERNAL_HDRS,
3612 gcc_copts = xnnpack_gcc_std_copts(),
3613 gcc_x86_copts = [
3614 "-mavx512f",
3615 "-mavx512cd",
3616 "-mavx512bw",
3617 "-mavx512dq",
3618 "-mavx512vl",
3619 ],
3620 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3621 msvc_copts = xnnpack_msvc_std_copts(),
3622 msvc_x86_32_copts = ["/arch:AVX512"],
3623 msvc_x86_64_copts = ["/arch:AVX512"],
3624 msys_copts = ["-fno-asynchronous-unwind-tables"],
3625 x86_srcs = AVX512SKX_UKERNELS,
3626 deps = [
3627 ":tables",
3628 "@FP16",
3629 "@pthreadpool",
3630 ],
3631)
3632
3633xnnpack_cc_library(
3634 name = "avx512skx_ukernels_test_mode",
3635 hdrs = INTERNAL_HDRS,
3636 copts = [
3637 "-UNDEBUG",
3638 "-DXNN_TEST_MODE=1",
3639 ],
3640 gcc_copts = xnnpack_gcc_std_copts(),
3641 gcc_x86_copts = [
3642 "-mavx512f",
3643 "-mavx512cd",
3644 "-mavx512bw",
3645 "-mavx512dq",
3646 "-mavx512vl",
3647 ],
3648 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3649 msvc_copts = xnnpack_msvc_std_copts(),
3650 msvc_x86_32_copts = ["/arch:AVX512"],
3651 msvc_x86_64_copts = ["/arch:AVX512"],
3652 msys_copts = ["-fno-asynchronous-unwind-tables"],
3653 x86_srcs = AVX512SKX_UKERNELS,
3654 deps = [
3655 ":tables",
3656 "@FP16",
3657 "@pthreadpool",
3658 ],
3659)
3660
3661xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 name = "asm_ukernels",
3663 hdrs = ["src/xnnpack/assembly.h"],
3664 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003665 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003667 wasm_srcs = WASM32_ASM_UKERNELS,
3668 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669)
3670
Marat Dukhan3b59de22020-06-03 20:15:19 -07003671xnnpack_cc_library(
3672 name = "logging_utils",
3673 srcs = LOGGING_SRCS,
3674 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3675 copts = LOGGING_COPTS + [
3676 "-Isrc",
3677 "-Iinclude",
3678 ] + select({
3679 ":debug_build": [],
3680 "//conditions:default": xnnpack_min_size_copts(),
3681 }),
3682 gcc_copts = xnnpack_gcc_std_copts(),
3683 msvc_copts = xnnpack_msvc_std_copts(),
3684 visibility = xnnpack_visibility(),
3685 deps = [
3686 "@FP16",
3687 "@clog",
3688 "@pthreadpool",
3689 ],
3690)
3691
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692xnnpack_aggregate_library(
3693 name = "ukernels",
3694 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 ":neon_ukernels",
3696 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003697 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003698 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 ":asm_ukernels",
3700 ],
3701 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702 ":neon_ukernels",
3703 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003704 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003706 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 ":asm_ukernels",
3708 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003709 generic_deps = [
3710 ":scalar_ukernels",
3711 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003712 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003713 ":psimd_fastmath_ukernels",
3714 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003716 wasm_deps = [
3717 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003718 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003719 ],
3720 wasmsimd_deps = [
3721 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003722 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003723 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003726 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003727 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003729 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003730 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003731 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003732 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003733 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003734 ],
3735)
3736
Marat Dukhan33fcf782020-05-24 14:27:15 -07003737xnnpack_aggregate_library(
3738 name = "ukernels_test_mode",
3739 aarch32_deps = [
3740 ":neon_ukernels_test_mode",
3741 ":neonfma_ukernels_test_mode",
3742 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003743 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003744 ":asm_ukernels",
3745 ],
3746 aarch64_deps = [
3747 ":neon_ukernels_test_mode",
3748 ":neonfma_ukernels_test_mode",
3749 ":neonv8_ukernels_test_mode",
3750 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003751 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003752 ":asm_ukernels",
3753 ],
3754 generic_deps = [
3755 ":scalar_ukernels_test_mode",
3756 ],
3757 psimd_deps = [
3758 ":psimd_fastmath_ukernels_test_mode",
3759 ":psimd_accmath_ukernels_test_mode",
3760 ],
3761 wasm_deps = [
3762 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003763 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003764 ],
3765 wasmsimd_deps = [
3766 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003767 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003768 ],
3769 x86_deps = [
3770 ":sse2_ukernels_test_mode",
3771 ":ssse3_ukernels_test_mode",
3772 ":sse41_ukernels_test_mode",
3773 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003774 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003775 ":fma3_ukernels_test_mode",
3776 ":avx2_ukernels_test_mode",
3777 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003778 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003779 ],
3780)
3781
Marat Dukhan86df3842020-06-17 14:03:31 -07003782# This target must be compatible with Android NDK r17 and compile with Clang 6.
3783# It is similar to :ukernels target, but excludes microkernels which require
3784# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003785xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003786 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003787 aarch32_deps = [
3788 ":neon_ukernels",
3789 ":neonfma_ukernels",
3790 ":neonv8_ukernels",
3791 ":asm_ukernels",
3792 ],
3793 aarch64_deps = [
3794 ":neon_ukernels",
3795 ":neonfma_ukernels",
3796 ":neonv8_ukernels",
3797 ":asm_ukernels",
3798 ],
3799 generic_deps = [
3800 ":scalar_ukernels",
3801 ],
3802 psimd_deps = [
3803 ":psimd_fastmath_ukernels",
3804 ":psimd_accmath_ukernels",
3805 ],
3806 wasm_deps = [
3807 ":wasm_ukernels",
3808 ],
3809 wasmsimd_deps = [
3810 ":wasm_ukernels",
3811 ],
3812 x86_deps = [
3813 ":sse2_ukernels",
3814 ":ssse3_ukernels",
3815 ":sse41_ukernels",
3816 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003817 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003818 ":fma3_ukernels",
3819 ":avx2_ukernels",
3820 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003821 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003822 ],
3823)
3824
Marat Dukhan08c4a432019-10-03 09:29:21 -07003825xnnpack_cc_library(
3826 name = "im2col",
3827 srcs = ["src/im2col.c"],
3828 hdrs = [
3829 "src/xnnpack/common.h",
3830 "src/xnnpack/im2col.h",
3831 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003832 gcc_copts = xnnpack_gcc_std_copts(),
3833 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834)
3835
3836xnnpack_cc_library(
3837 name = "indirection",
3838 srcs = ["src/indirection.c"],
3839 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003840 gcc_copts = xnnpack_gcc_std_copts(),
3841 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003842 deps = [
3843 "@FP16",
3844 "@FXdiv",
3845 "@pthreadpool",
3846 ],
3847)
3848
3849xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003850 name = "indirection_test_mode",
3851 srcs = ["src/indirection.c"],
3852 hdrs = INTERNAL_HDRS,
3853 copts = [
3854 "-UNDEBUG",
3855 "-DXNN_TEST_MODE=1",
3856 ],
3857 gcc_copts = xnnpack_gcc_std_copts(),
3858 msvc_copts = xnnpack_msvc_std_copts(),
3859 deps = [
3860 "@FP16",
3861 "@FXdiv",
3862 "@pthreadpool",
3863 ],
3864)
3865
3866xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003867 name = "packing",
3868 srcs = ["src/packing.c"],
3869 hdrs = INTERNAL_HDRS,
3870 gcc_copts = xnnpack_gcc_std_copts(),
3871 msvc_copts = xnnpack_msvc_std_copts(),
3872 deps = [
3873 "@FP16",
3874 "@FXdiv",
3875 "@pthreadpool",
3876 ],
3877)
3878
3879xnnpack_cc_library(
3880 name = "packing_test_mode",
3881 srcs = ["src/packing.c"],
3882 hdrs = INTERNAL_HDRS,
3883 copts = [
3884 "-UNDEBUG",
3885 "-DXNN_TEST_MODE=1",
3886 ],
3887 gcc_copts = xnnpack_gcc_std_copts(),
3888 msvc_copts = xnnpack_msvc_std_copts(),
3889 deps = [
3890 "@FP16",
3891 "@FXdiv",
3892 "@pthreadpool",
3893 ],
3894)
3895
3896xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897 name = "operator_run",
3898 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003899 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003900 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003901 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3902 "//conditions:default": [],
3903 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003904 gcc_copts = xnnpack_gcc_std_copts(),
3905 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003907 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908 "@FP16",
3909 "@FXdiv",
3910 "@clog",
3911 "@pthreadpool",
3912 ],
3913)
3914
Chao Mei6ddfc602020-05-13 22:29:36 -07003915xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003916 name = "operator_run_test_mode",
3917 srcs = ["src/operator-run.c"],
3918 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3919 copts = LOGGING_COPTS + [
3920 "-UNDEBUG",
3921 "-DXNN_TEST_MODE=1",
3922 ] + select({
3923 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3924 "//conditions:default": [],
3925 }),
3926 gcc_copts = xnnpack_gcc_std_copts(),
3927 msvc_copts = xnnpack_msvc_std_copts(),
3928 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003929 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003930 "@FP16",
3931 "@FXdiv",
3932 "@clog",
3933 "@pthreadpool",
3934 ],
3935)
3936
3937xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003938 name = "memory_planner",
3939 srcs = ["src/memory-planner.c"],
3940 hdrs = INTERNAL_HDRS,
3941 defines = select({
3942 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3943 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3944 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3945 }),
3946 gcc_copts = xnnpack_gcc_std_copts(),
3947 msvc_copts = xnnpack_msvc_std_copts(),
3948 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003949 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003950 "@pthreadpool",
3951 ],
3952)
3953
Marat Dukhan33fcf782020-05-24 14:27:15 -07003954xnnpack_cc_library(
3955 name = "memory_planner_test_mode",
3956 srcs = ["src/memory-planner.c"],
3957 hdrs = INTERNAL_HDRS,
3958 copts = [
3959 "-UNDEBUG",
3960 "-DXNN_TEST_MODE=1",
3961 ],
3962 defines = select({
3963 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3964 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3965 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3966 }),
3967 gcc_copts = xnnpack_gcc_std_copts(),
3968 msvc_copts = xnnpack_msvc_std_copts(),
3969 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003970 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003971 "@pthreadpool",
3972 ],
3973)
3974
Marat Dukhan08c4a432019-10-03 09:29:21 -07003975cc_library(
3976 name = "enable_assembly",
3977 defines = select({
3978 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3979 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003980 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 }),
3982)
3983
Marat Dukhan9de90e02020-06-18 16:04:12 -07003984cc_library(
3985 name = "enable_sparse",
3986 defines = select({
3987 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3988 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003989 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003990 }),
3991)
3992
Marat Dukhancf056b22019-10-07 10:26:29 -07003993xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003994 name = "operators",
3995 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003996 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003998 ],
3999 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004000 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001 "-Isrc",
4002 "-Iinclude",
4003 ] + select({
4004 ":debug_build": [],
4005 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004006 }) + select({
4007 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4008 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004010 gcc_copts = xnnpack_gcc_std_copts(),
4011 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004012 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004013 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004014 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004015 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016 "@FP16",
4017 "@FXdiv",
4018 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004019 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004020 ],
4021)
4022
Marat Dukhan10a38082020-04-17 03:58:35 -07004023xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004024 name = "operators_test_mode",
4025 srcs = OPERATOR_SRCS + [
4026 "src/memory.c",
4027 "src/operator-delete.c",
4028 ],
4029 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4030 copts = LOGGING_COPTS + [
4031 "-Isrc",
4032 "-Iinclude",
4033 "-UNDEBUG",
4034 "-DXNN_TEST_MODE=1",
4035 ] + select({
4036 ":debug_build": [],
4037 "//conditions:default": xnnpack_min_size_copts(),
4038 }) + select({
4039 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4040 "//conditions:default": [],
4041 }),
4042 gcc_copts = xnnpack_gcc_std_copts(),
4043 msvc_copts = xnnpack_msvc_std_copts(),
4044 deps = [
4045 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004046 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004047 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004048 "@FP16",
4049 "@FXdiv",
4050 "@clog",
4051 "@pthreadpool",
4052 ],
4053)
4054
4055xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004056 name = "XNNPACK",
4057 srcs = [
4058 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004059 "src/runtime.c",
4060 "src/subgraph.c",
4061 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004062 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004063 hdrs = ["include/xnnpack.h"],
4064 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004065 "-Isrc",
4066 "-Iinclude",
4067 ] + select({
4068 ":debug_build": [],
4069 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004070 }) + select({
4071 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4072 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004073 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004074 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004075 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004076 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004077 visibility = xnnpack_visibility(),
4078 deps = [
4079 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004080 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004081 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004082 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004083 ":operator_run",
4084 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004085 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004086 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004087 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004088 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004089 ] + select({
4090 ":emscripten": [],
4091 "//conditions:default": ["@cpuinfo"],
4092 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093)
4094
Marat Dukhan10a38082020-04-17 03:58:35 -07004095xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004096 name = "XNNPACK_test_mode",
4097 srcs = [
4098 "src/init.c",
4099 "src/runtime.c",
4100 "src/subgraph.c",
4101 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004102 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004103 hdrs = ["include/xnnpack.h"],
4104 copts = LOGGING_COPTS + [
4105 "-Isrc",
4106 "-Iinclude",
4107 "-UNDEBUG",
4108 "-DXNN_TEST_MODE=1",
4109 ] + select({
4110 ":debug_build": [],
4111 "//conditions:default": xnnpack_min_size_copts(),
4112 }) + select({
4113 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4114 "//conditions:default": [],
4115 }),
4116 gcc_copts = xnnpack_gcc_std_copts(),
4117 includes = ["include"],
4118 msvc_copts = xnnpack_msvc_std_copts(),
4119 visibility = xnnpack_visibility(),
4120 deps = [
4121 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004122 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004123 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004124 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004125 ":operator_run_test_mode",
4126 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004127 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004128 "@clog",
4129 "@FP16",
4130 "@pthreadpool",
4131 ] + select({
4132 ":emscripten": [],
4133 "//conditions:default": ["@cpuinfo"],
4134 }),
4135)
4136
4137xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004138 name = "xnnpack_f32",
4139 srcs = [
4140 "src/init.c",
4141 "src/runtime.c",
4142 "src/subgraph.c",
4143 "src/tensor.c",
4144 ] + SUBGRAPH_SRCS,
4145 hdrs = ["include/xnnpack.h"],
4146 copts = LOGGING_COPTS + [
4147 "-Isrc",
4148 "-Iinclude",
4149 ] + select({
4150 ":debug_build": [],
4151 "//conditions:default": xnnpack_min_size_copts(),
4152 }) + select({
4153 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4154 "//conditions:default": [],
4155 }),
4156 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004157 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004158 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004159 "XNN_NO_U8_OPERATORS",
4160 "XNN_NO_X8_OPERATORS",
4161 "XNN_NO_F16_OPERATORS",
4162 "XNN_NO_X16_OPERATORS",
4163 ],
4164 gcc_copts = xnnpack_gcc_std_copts(),
4165 includes = ["include"],
4166 msvc_copts = xnnpack_msvc_std_copts(),
4167 visibility = xnnpack_visibility(),
4168 deps = [
4169 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004170 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004171 ":logging_utils",
4172 ":memory_planner",
4173 ":operator_run",
4174 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004175 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004176 "@clog",
4177 "@FP16",
4178 "@pthreadpool",
4179 ] + select({
4180 ":emscripten": [],
4181 "//conditions:default": ["@cpuinfo"],
4182 }),
4183)
4184
4185xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004186 name = "xnnpack_operators_nhwc_f32",
4187 srcs = [
4188 "src/init.c",
4189 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004190 hdrs = ["include/xnnpack.h"],
4191 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004192 "-Isrc",
4193 "-Iinclude",
4194 ] + select({
4195 ":debug_build": [],
4196 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004197 }) + select({
4198 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4199 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004200 }),
4201 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004202 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004203 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004204 "XNN_NO_U8_OPERATORS",
4205 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004206 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004207 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004208 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004210 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211 visibility = xnnpack_visibility(),
4212 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004213 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004214 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 ":operator_run",
4216 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004217 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004218 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004220 ] + select({
4221 ":emscripten": [],
4222 "//conditions:default": ["@cpuinfo"],
4223 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224)
4225
Marat Dukhancf056b22019-10-07 10:26:29 -07004226xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227 name = "bench_utils",
4228 srcs = ["bench/utils.cc"],
4229 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004230 deps = [
4231 "@com_google_benchmark//:benchmark",
4232 "@cpuinfo",
4233 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004234)
4235
Frank Barchard7e955972019-10-11 10:34:25 -07004236######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237
4238xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004239 name = "qs8_gemm_bench",
4240 srcs = [
4241 "bench/gemm.h",
4242 "bench/qs8-gemm.cc",
4243 "src/xnnpack/AlignedAllocator.h",
4244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4245 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4246)
4247
4248xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004249 name = "qs8_requantization_bench",
4250 srcs = [
4251 "bench/qs8-requantization.cc",
4252 "src/xnnpack/requantization-stubs.h",
4253 "src/xnnpack/AlignedAllocator.h",
4254 ] + MICROKERNEL_BENCHMARK_HDRS,
4255 deps = MICROKERNEL_BENCHMARK_DEPS,
4256)
4257
4258xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004259 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 srcs = [
4261 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004262 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 "src/xnnpack/AlignedAllocator.h",
4264 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004265 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004266 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267)
4268
4269xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004270 name = "qu8_requantization_bench",
4271 srcs = [
4272 "bench/qu8-requantization.cc",
4273 "src/xnnpack/requantization-stubs.h",
4274 "src/xnnpack/AlignedAllocator.h",
4275 ] + MICROKERNEL_BENCHMARK_HDRS,
4276 deps = MICROKERNEL_BENCHMARK_DEPS,
4277)
4278
4279xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004280 name = "f16_igemm_bench",
4281 srcs = [
4282 "bench/f16-igemm.cc",
4283 "bench/conv.h",
4284 "bench/google/conv.h",
4285 "src/xnnpack/AlignedAllocator.h",
4286 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004287 deps = MICROKERNEL_BENCHMARK_DEPS + [
4288 ":indirection",
4289 ":packing",
4290 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004291)
4292
4293xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004294 name = "f16_gemm_bench",
4295 srcs = [
4296 "bench/f16-gemm.cc",
4297 "bench/gemm.h",
4298 "src/xnnpack/AlignedAllocator.h",
4299 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004300 deps = MICROKERNEL_BENCHMARK_DEPS + [
4301 ":packing",
4302 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303)
4304
4305xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004306 name = "f16_spmm_bench",
4307 srcs = [
4308 "bench/f16-spmm.cc",
4309 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004310 "src/xnnpack/AlignedAllocator.h",
4311 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004312 deps = MICROKERNEL_BENCHMARK_DEPS,
4313)
4314
4315xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316 name = "f32_igemm_bench",
4317 srcs = [
4318 "bench/f32-igemm.cc",
4319 "bench/conv.h",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004322 deps = MICROKERNEL_BENCHMARK_DEPS + [
4323 ":indirection",
4324 ":packing",
4325 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326)
4327
4328xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004329 name = "f16_relu_bench",
4330 srcs = [
4331 "bench/f16-relu.cc",
4332 "src/xnnpack/AlignedAllocator.h",
4333 ] + MICROKERNEL_BENCHMARK_HDRS,
4334 deps = MICROKERNEL_BENCHMARK_DEPS,
4335)
4336
4337xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338 name = "f32_conv_hwc_bench",
4339 srcs = [
4340 "bench/f32-conv-hwc.cc",
4341 "bench/dconv.h",
4342 "src/xnnpack/AlignedAllocator.h",
4343 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004344 deps = MICROKERNEL_BENCHMARK_DEPS + [
4345 ":packing",
4346 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004347)
4348
4349xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004350 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004351 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004352 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004353 "bench/dconv.h",
4354 "src/xnnpack/AlignedAllocator.h",
4355 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004356 deps = MICROKERNEL_BENCHMARK_DEPS + [
4357 ":packing",
4358 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004359)
4360
4361xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004362 name = "f16_dwconv_bench",
4363 srcs = [
4364 "bench/f16-dwconv.cc",
4365 "bench/dwconv.h",
4366 "bench/google/dwconv.h",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004369 deps = MICROKERNEL_BENCHMARK_DEPS + [
4370 ":indirection",
4371 ":packing",
4372 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004373)
4374
4375xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 name = "f32_dwconv_bench",
4377 srcs = [
4378 "bench/f32-dwconv.cc",
4379 "bench/dwconv.h",
4380 "src/xnnpack/AlignedAllocator.h",
4381 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004382 deps = MICROKERNEL_BENCHMARK_DEPS + [
4383 ":indirection",
4384 ":packing",
4385 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386)
4387
4388xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004389 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004391 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004392 "bench/dwconv.h",
4393 "src/xnnpack/AlignedAllocator.h",
4394 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004395 deps = MICROKERNEL_BENCHMARK_DEPS + [
4396 ":indirection",
4397 ":packing",
4398 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399)
4400
4401xnnpack_benchmark(
4402 name = "f32_gemm_bench",
4403 srcs = [
4404 "bench/f32-gemm.cc",
4405 "bench/gemm.h",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004408 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004409 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004410)
4411
4412xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004413 name = "f32_hswish_bench",
4414 srcs = [
4415 "bench/f32-hswish.cc",
4416 "src/xnnpack/AlignedAllocator.h",
4417 ] + MICROKERNEL_BENCHMARK_HDRS,
4418 deps = MICROKERNEL_BENCHMARK_DEPS,
4419)
4420
4421xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004422 name = "f32_raddexpminusmax_bench",
4423 srcs = [
4424 "bench/f32-raddexpminusmax.cc",
4425 "src/xnnpack/AlignedAllocator.h",
4426 ] + MICROKERNEL_BENCHMARK_HDRS,
4427 deps = MICROKERNEL_BENCHMARK_DEPS,
4428)
4429
4430xnnpack_benchmark(
4431 name = "f32_raddextexp_bench",
4432 srcs = [
4433 "bench/f32-raddextexp.cc",
4434 "src/xnnpack/AlignedAllocator.h",
4435 ] + MICROKERNEL_BENCHMARK_HDRS,
4436 deps = MICROKERNEL_BENCHMARK_DEPS,
4437)
4438
4439xnnpack_benchmark(
4440 name = "f32_raddstoreexpminusmax_bench",
4441 srcs = [
4442 "bench/f32-raddstoreexpminusmax.cc",
4443 "src/xnnpack/AlignedAllocator.h",
4444 ] + MICROKERNEL_BENCHMARK_HDRS,
4445 deps = MICROKERNEL_BENCHMARK_DEPS,
4446)
4447
4448xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004449 name = "f32_relu_bench",
4450 srcs = [
4451 "bench/f32-relu.cc",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + MICROKERNEL_BENCHMARK_HDRS,
4454 deps = MICROKERNEL_BENCHMARK_DEPS,
4455)
4456
4457xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 name = "f32_rmax_bench",
4459 srcs = [
4460 "bench/f32-rmax.cc",
4461 "src/xnnpack/AlignedAllocator.h",
4462 ] + MICROKERNEL_BENCHMARK_HDRS,
4463 deps = MICROKERNEL_BENCHMARK_DEPS,
4464)
4465
4466xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004467 name = "f32_sigmoid_bench",
4468 srcs = [
4469 "bench/f32-sigmoid.cc",
4470 "src/xnnpack/AlignedAllocator.h",
4471 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004472 deps = MICROKERNEL_BENCHMARK_DEPS,
4473)
4474
4475xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476 name = "f32_spmm_bench",
4477 srcs = [
4478 "bench/f32-spmm.cc",
4479 "bench/gemm.h",
4480 "src/xnnpack/AlignedAllocator.h",
4481 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482 deps = MICROKERNEL_BENCHMARK_DEPS,
4483)
4484
4485xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004486 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004487 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004488 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004489 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004490 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004491 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004492)
4493
4494xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004495 name = "f32_vscaleexpminusmax_bench",
4496 srcs = [
4497 "bench/f32-vscaleexpminusmax.cc",
4498 "src/xnnpack/AlignedAllocator.h",
4499 ] + MICROKERNEL_BENCHMARK_HDRS,
4500 deps = MICROKERNEL_BENCHMARK_DEPS,
4501)
4502
4503xnnpack_benchmark(
4504 name = "f32_vscaleextexp_bench",
4505 srcs = [
4506 "bench/f32-vscaleextexp.cc",
4507 "src/xnnpack/AlignedAllocator.h",
4508 ] + MICROKERNEL_BENCHMARK_HDRS,
4509 deps = MICROKERNEL_BENCHMARK_DEPS,
4510)
4511
4512xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004513 name = "f32_vsqrt_bench",
4514 srcs = [
4515 "bench/f32-vsqrt.cc",
4516 "src/xnnpack/AlignedAllocator.h",
4517 ] + MICROKERNEL_BENCHMARK_HDRS,
4518 deps = MICROKERNEL_BENCHMARK_DEPS,
4519)
4520
4521xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522 name = "f32_im2col_gemm_bench",
4523 srcs = [
4524 "bench/f32-im2col-gemm.cc",
4525 "bench/conv.h",
4526 "src/xnnpack/AlignedAllocator.h",
4527 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004528 deps = MICROKERNEL_BENCHMARK_DEPS + [
4529 ":im2col",
4530 ":packing",
4531 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532)
4533
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004534xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004535 name = "rounding_bench",
4536 srcs = [
4537 "bench/rounding.cc",
4538 "src/xnnpack/math-stubs.h",
4539 "src/xnnpack/AlignedAllocator.h",
4540 ] + MICROKERNEL_BENCHMARK_HDRS,
4541 deps = MICROKERNEL_BENCHMARK_DEPS,
4542)
4543
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544########################### Benchmarks for operators ###########################
4545
4546xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004547 name = "average_pooling_bench",
4548 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004549 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004550 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004551 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552)
4553
4554xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004555 name = "bankers_rounding_bench",
4556 srcs = ["bench/bankers-rounding.cc"],
4557 copts = xnnpack_optional_tflite_copts(),
4558 tags = ["nowin32"],
4559 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4560)
4561
4562xnnpack_benchmark(
4563 name = "ceiling_bench",
4564 srcs = ["bench/ceiling.cc"],
4565 copts = xnnpack_optional_tflite_copts(),
4566 tags = ["nowin32"],
4567 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4568)
4569
4570xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004571 name = "channel_shuffle_bench",
4572 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004573 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574)
4575
4576xnnpack_benchmark(
4577 name = "convolution_bench",
4578 srcs = ["bench/convolution.cc"],
4579 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004580 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004581 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004582)
4583
4584xnnpack_benchmark(
4585 name = "deconvolution_bench",
4586 srcs = ["bench/deconvolution.cc"],
4587 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004588 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004589 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004590)
4591
4592xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004593 name = "floor_bench",
4594 srcs = ["bench/floor.cc"],
4595 copts = xnnpack_optional_tflite_copts(),
4596 tags = ["nowin32"],
4597 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4598)
4599
4600xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004601 name = "global_average_pooling_bench",
4602 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004603 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604)
4605
4606xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004607 name = "hardswish_bench",
4608 srcs = ["bench/hardswish.cc"],
4609 copts = xnnpack_optional_tflite_copts(),
4610 tags = ["nowin32"],
4611 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4612)
4613
4614xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004615 name = "max_pooling_bench",
4616 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004617 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004618)
4619
4620xnnpack_benchmark(
4621 name = "sigmoid_bench",
4622 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004623 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004624 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004625 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626)
4627
4628xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004629 name = "prelu_bench",
4630 srcs = ["bench/prelu.cc"],
4631 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004632 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004633 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004634)
4635
4636xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004637 name = "softmax_bench",
4638 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004639 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004640 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004641 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004642)
4643
Marat Dukhan87727142020-06-24 15:24:10 -07004644xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004645 name = "square_root_bench",
4646 srcs = ["bench/square-root.cc"],
4647 copts = xnnpack_optional_tflite_copts(),
4648 tags = ["nowin32"],
4649 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4650)
4651
4652xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004653 name = "truncation_bench",
4654 srcs = ["bench/truncation.cc"],
4655 deps = OPERATOR_BENCHMARK_DEPS,
4656)
4657
Marat Dukhanc068bb62019-10-04 13:24:39 -07004658############################# End-to-end benchmarks ############################
4659
4660cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004661 name = "fp32_mobilenet_v1",
4662 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004663 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004664 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004665 linkstatic = True,
4666 deps = [
4667 ":XNNPACK",
4668 "@pthreadpool",
4669 ],
4670)
4671
4672cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004673 name = "fp16_mobilenet_v1",
4674 srcs = ["models/fp16-mobilenet-v1.cc"],
4675 hdrs = ["models/models.h"],
4676 copts = xnnpack_std_cxxopts(),
4677 linkstatic = True,
4678 deps = [
4679 ":XNNPACK",
4680 "@FP16",
4681 "@pthreadpool",
4682 ],
4683)
4684
4685cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004686 name = "qs8_mobilenet_v1",
4687 srcs = ["models/qs8-mobilenet-v1.cc"],
4688 hdrs = ["models/models.h"],
4689 copts = xnnpack_std_cxxopts(),
4690 linkstatic = True,
4691 deps = [
4692 ":XNNPACK",
4693 "@pthreadpool",
4694 ],
4695)
4696
4697cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004698 name = "qs8_mobilenet_v2",
4699 srcs = ["models/qs8-mobilenet-v2.cc"],
4700 hdrs = ["models/models.h"],
4701 copts = xnnpack_std_cxxopts(),
4702 linkstatic = True,
4703 deps = [
4704 ":XNNPACK",
4705 "@pthreadpool",
4706 ],
4707)
4708
4709cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004710 name = "fp32_mobilenet_v2",
4711 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004712 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004713 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004714 linkstatic = True,
4715 deps = [
4716 ":XNNPACK",
4717 "@pthreadpool",
4718 ],
4719)
4720
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004721cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004722 name = "fp16_mobilenet_v2",
4723 srcs = ["models/fp16-mobilenet-v2.cc"],
4724 hdrs = ["models/models.h"],
4725 copts = xnnpack_std_cxxopts(),
4726 linkstatic = True,
4727 deps = [
4728 ":XNNPACK",
4729 "@FP16",
4730 "@pthreadpool",
4731 ],
4732)
4733
4734cc_library(
4735 name = "fp32_mobilenet_v3_large",
4736 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004737 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004738 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004739 linkstatic = True,
4740 deps = [
4741 ":XNNPACK",
4742 "@pthreadpool",
4743 ],
4744)
4745
4746cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004747 name = "fp16_mobilenet_v3_large",
4748 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4749 hdrs = ["models/models.h"],
4750 copts = xnnpack_std_cxxopts(),
4751 linkstatic = True,
4752 deps = [
4753 ":XNNPACK",
4754 "@FP16",
4755 "@pthreadpool",
4756 ],
4757)
4758
4759cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004760 name = "fp32_mobilenet_v3_small",
4761 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004762 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004763 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004764 linkstatic = True,
4765 deps = [
4766 ":XNNPACK",
4767 "@pthreadpool",
4768 ],
4769)
4770
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004771cc_library(
4772 name = "fp16_mobilenet_v3_small",
4773 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4774 hdrs = ["models/models.h"],
4775 copts = xnnpack_std_cxxopts(),
4776 linkstatic = True,
4777 deps = [
4778 ":XNNPACK",
4779 "@FP16",
4780 "@pthreadpool",
4781 ],
4782)
4783
Marat Dukhanc068bb62019-10-04 13:24:39 -07004784xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004785 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004786 srcs = [
4787 "bench/f32-dwconv-e2e.cc",
4788 "bench/end2end.h",
4789 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004790 deps = MICROKERNEL_BENCHMARK_DEPS + [
4791 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004792 ":fp32_mobilenet_v1",
4793 ":fp32_mobilenet_v2",
4794 ":fp32_mobilenet_v3_large",
4795 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004796 ],
4797)
4798
4799xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004800 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004801 srcs = [
4802 "bench/f32-gemm-e2e.cc",
4803 "bench/end2end.h",
4804 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004805 deps = MICROKERNEL_BENCHMARK_DEPS + [
4806 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004807 ":fp32_mobilenet_v1",
4808 ":fp32_mobilenet_v2",
4809 ":fp32_mobilenet_v3_large",
4810 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004811 ],
4812)
4813
4814xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004815 name = "end2end_bench",
4816 srcs = ["bench/end2end.cc"],
4817 deps = [
4818 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004819 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004820 ":fp16_mobilenet_v1",
4821 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004822 ":fp16_mobilenet_v3_large",
4823 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004824 ":fp32_mobilenet_v1",
4825 ":fp32_mobilenet_v2",
4826 ":fp32_mobilenet_v3_large",
4827 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004828 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004829 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004830 "@pthreadpool",
4831 ],
4832)
4833
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004834#################### Accuracy evaluation for math functions ####################
4835
4836xnnpack_benchmark(
4837 name = "f32_exp_eval",
4838 srcs = [
4839 "eval/f32-exp.cc",
4840 "src/xnnpack/AlignedAllocator.h",
4841 ] + ACCURACY_EVAL_HDRS,
4842 deps = ACCURACY_EVAL_DEPS,
4843)
4844
Marat Dukhan515c9772019-10-17 18:07:57 -07004845xnnpack_benchmark(
4846 name = "f32_expminus_eval",
4847 srcs = [
4848 "eval/f32-expminus.cc",
4849 "src/xnnpack/AlignedAllocator.h",
4850 ] + ACCURACY_EVAL_HDRS,
4851 deps = ACCURACY_EVAL_DEPS,
4852)
4853
Marat Dukhan98ba4412019-10-23 02:14:28 -07004854xnnpack_benchmark(
4855 name = "f32_extexp_eval",
4856 srcs = [
4857 "eval/f32-extexp.cc",
4858 "src/xnnpack/AlignedAllocator.h",
4859 ] + ACCURACY_EVAL_HDRS,
4860 deps = ACCURACY_EVAL_DEPS,
4861)
4862
Marat Dukhan8853b822020-05-07 12:19:01 -07004863xnnpack_unit_test(
4864 name = "f32_roundne_eval",
4865 srcs = [
4866 "eval/f32-roundne.cc",
4867 "src/xnnpack/AlignedAllocator.h",
4868 "src/xnnpack/math-stubs.h",
4869 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004870 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004871 deps = MICROKERNEL_TEST_DEPS,
4872)
4873
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004874xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004875 name = "f32_roundd_eval",
4876 srcs = [
4877 "eval/f32-roundd.cc",
4878 "src/xnnpack/AlignedAllocator.h",
4879 "src/xnnpack/math-stubs.h",
4880 ] + MICROKERNEL_TEST_HDRS,
4881 automatic = False,
4882 deps = MICROKERNEL_TEST_DEPS,
4883)
4884
4885xnnpack_unit_test(
4886 name = "f32_roundu_eval",
4887 srcs = [
4888 "eval/f32-roundu.cc",
4889 "src/xnnpack/AlignedAllocator.h",
4890 "src/xnnpack/math-stubs.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 automatic = False,
4893 deps = MICROKERNEL_TEST_DEPS,
4894)
4895
4896xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004897 name = "f32_roundz_eval",
4898 srcs = [
4899 "eval/f32-roundz.cc",
4900 "src/xnnpack/AlignedAllocator.h",
4901 "src/xnnpack/math-stubs.h",
4902 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004903 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
Marat Dukhan346a9e52019-11-15 09:06:30 -08004907xnnpack_benchmark(
4908 name = "f32_sigmoid_eval",
4909 srcs = [
4910 "eval/f32-sigmoid.cc",
4911 "src/xnnpack/AlignedAllocator.h",
4912 ] + ACCURACY_EVAL_HDRS,
4913 deps = ACCURACY_EVAL_DEPS,
4914)
4915
Marat Dukhan84000762020-06-29 18:38:43 -07004916xnnpack_benchmark(
4917 name = "f32_sqrt_eval",
4918 srcs = [
4919 "eval/f32-sqrt.cc",
4920 "src/xnnpack/AlignedAllocator.h",
4921 ] + ACCURACY_EVAL_HDRS,
4922 deps = ACCURACY_EVAL_DEPS + [
4923 ":bench_utils",
4924 ],
4925)
4926
Marat Dukhan08c4a432019-10-03 09:29:21 -07004927######################### Unit tests for micro-kernels #########################
4928
4929xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004930 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004932 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004933 "test/gemm-microkernel-tester.h",
4934 "src/xnnpack/AlignedAllocator.h",
4935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004936 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937)
4938
4939xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004940 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004941 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004942 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004943 "test/spmm-microkernel-tester.h",
4944 "src/xnnpack/AlignedAllocator.h",
4945 ] + MICROKERNEL_TEST_HDRS,
4946 deps = MICROKERNEL_TEST_DEPS,
4947)
4948
4949xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950 name = "f32_argmaxpool_test",
4951 srcs = [
4952 "test/f32-argmaxpool.cc",
4953 "test/argmaxpool-microkernel-tester.h",
4954 "src/xnnpack/AlignedAllocator.h",
4955 ] + MICROKERNEL_TEST_HDRS,
4956 deps = MICROKERNEL_TEST_DEPS,
4957)
4958
4959xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004960 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004961 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004962 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004963 "test/avgpool-microkernel-tester.h",
4964 "src/xnnpack/AlignedAllocator.h",
4965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004970 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004971 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004972 "test/f32-ibilinear.cc",
4973 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004974 "src/xnnpack/AlignedAllocator.h",
4975 ] + MICROKERNEL_TEST_HDRS,
4976 deps = MICROKERNEL_TEST_DEPS,
4977)
4978
4979xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004980 name = "f16_clamp_test",
4981 srcs = [
4982 "test/f16-clamp.cc",
4983 "test/clamp-microkernel-tester.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 deps = MICROKERNEL_TEST_DEPS,
4986)
4987
4988xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 name = "f32_clamp_test",
4990 srcs = [
4991 "test/f32-clamp.cc",
4992 "test/clamp-microkernel-tester.h",
4993 ] + MICROKERNEL_TEST_HDRS,
4994 deps = MICROKERNEL_TEST_DEPS,
4995)
4996
4997xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004998 name = "f32_igemm_test",
4999 srcs = [
5000 "test/f32-igemm.cc",
5001 "test/gemm-microkernel-tester.h",
5002 "src/xnnpack/AlignedAllocator.h",
5003 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005004 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005005)
5006
5007xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005008 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005010 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "test/gemm-microkernel-tester.h",
5012 "src/xnnpack/AlignedAllocator.h",
5013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005014 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015)
5016
5017xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005018 name = "f32_igemm_minmax_test",
5019 srcs = [
5020 "test/f32-igemm-minmax.cc",
5021 "test/gemm-microkernel-tester.h",
5022 "src/xnnpack/AlignedAllocator.h",
5023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005024 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005025)
5026
5027xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005028 name = "f16_igemm_minmax_test",
5029 srcs = [
5030 "test/f16-igemm-minmax.cc",
5031 "test/gemm-microkernel-tester.h",
5032 "src/xnnpack/AlignedAllocator.h",
5033 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005034 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005035)
5036
5037xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 name = "f32_conv_hwc_test",
5039 srcs = [
5040 "test/f32-conv-hwc.cc",
5041 "test/conv-hwc-microkernel-tester.h",
5042 "src/xnnpack/AlignedAllocator.h",
5043 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005044 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045)
5046
5047xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005048 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005050 "test/f32-conv-hwc2chw.cc",
5051 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005052 "src/xnnpack/AlignedAllocator.h",
5053 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005054 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005055)
5056
5057xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005058 name = "f32_dwconv_test",
5059 srcs = [
5060 "test/f32-dwconv.cc",
5061 "test/dwconv-microkernel-tester.h",
5062 "src/xnnpack/AlignedAllocator.h",
5063 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005064 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005065)
5066
5067xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005068 name = "f16_dwconv_minmax_test",
5069 srcs = [
5070 "test/f16-dwconv-minmax.cc",
5071 "test/dwconv-microkernel-tester.h",
5072 "src/xnnpack/AlignedAllocator.h",
5073 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005074 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005075)
5076
5077xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005078 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005079 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005080 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005081 "test/dwconv-microkernel-tester.h",
5082 "src/xnnpack/AlignedAllocator.h",
5083 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005084 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085)
5086
5087xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005088 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005090 "test/f32-dwconv-chw.cc",
5091 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 "src/xnnpack/AlignedAllocator.h",
5093 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005094 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095)
5096
5097xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005098 name = "f16_gavgpool_minmax_test",
5099 srcs = [
5100 "test/f16-gavgpool-minmax.cc",
5101 "test/gavgpool-microkernel-tester.h",
5102 "src/xnnpack/AlignedAllocator.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005108 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005109 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005110 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005111 "test/gavgpool-microkernel-tester.h",
5112 "src/xnnpack/AlignedAllocator.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005118 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005119 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005120 "test/f32-gavgpool-cw.cc",
5121 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005122 "src/xnnpack/AlignedAllocator.h",
5123 ] + MICROKERNEL_TEST_HDRS,
5124 deps = MICROKERNEL_TEST_DEPS,
5125)
5126
5127xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005128 name = "f32_gemm_test",
5129 srcs = [
5130 "test/f32-gemm.cc",
5131 "test/gemm-microkernel-tester.h",
5132 "src/xnnpack/AlignedAllocator.h",
5133 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005134 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005135)
5136
5137xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005138 name = "f32_gemm_relu_test",
5139 srcs = [
5140 "test/f32-gemm-relu.cc",
5141 "test/gemm-microkernel-tester.h",
5142 "src/xnnpack/AlignedAllocator.h",
5143 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005144 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005145)
5146
5147xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005148 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005150 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151 "test/gemm-microkernel-tester.h",
5152 "src/xnnpack/AlignedAllocator.h",
5153 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005154 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155)
5156
5157xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005158 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005160 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161 "test/gemm-microkernel-tester.h",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005164 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165)
5166
5167xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005168 name = "f16_hswish_test",
5169 srcs = [
5170 "test/f16-hswish.cc",
5171 "test/hswish-microkernel-tester.h",
5172 ] + MICROKERNEL_TEST_HDRS,
5173 deps = MICROKERNEL_TEST_DEPS,
5174)
5175
5176xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 name = "f32_hswish_test",
5178 srcs = [
5179 "test/f32-hswish.cc",
5180 "test/hswish-microkernel-tester.h",
5181 ] + MICROKERNEL_TEST_HDRS,
5182 deps = MICROKERNEL_TEST_DEPS,
5183)
5184
5185xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005186 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005188 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005189 "test/maxpool-microkernel-tester.h",
5190 ] + MICROKERNEL_TEST_HDRS,
5191 deps = MICROKERNEL_TEST_DEPS,
5192)
5193
5194xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005195 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005196 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005197 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005198 "test/avgpool-microkernel-tester.h",
5199 "src/xnnpack/AlignedAllocator.h",
5200 ] + MICROKERNEL_TEST_HDRS,
5201 deps = MICROKERNEL_TEST_DEPS,
5202)
5203
5204xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005205 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005207 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208 "test/gemm-microkernel-tester.h",
5209 "src/xnnpack/AlignedAllocator.h",
5210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005211 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212)
5213
5214xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005215 name = "f16_prelu_test",
5216 srcs = [
5217 "test/f16-prelu.cc",
5218 "test/prelu-microkernel-tester.h",
5219 "src/xnnpack/AlignedAllocator.h",
5220 ] + MICROKERNEL_TEST_HDRS,
5221 deps = MICROKERNEL_TEST_DEPS,
5222)
5223
5224xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 name = "f32_prelu_test",
5226 srcs = [
5227 "test/f32-prelu.cc",
5228 "test/prelu-microkernel-tester.h",
5229 "src/xnnpack/AlignedAllocator.h",
5230 ] + MICROKERNEL_TEST_HDRS,
5231 deps = MICROKERNEL_TEST_DEPS,
5232)
5233
5234xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005235 name = "f32_raddexpminusmax_test",
5236 srcs = [
5237 "test/f32-raddexpminusmax.cc",
5238 "test/raddexpminusmax-microkernel-tester.h",
5239 ] + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS,
5241)
5242
5243xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005244 name = "f32_raddextexp_test",
5245 srcs = [
5246 "test/f32-raddextexp.cc",
5247 "test/raddextexp-microkernel-tester.h",
5248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005253 name = "f32_raddstoreexpminusmax_test",
5254 srcs = [
5255 "test/f32-raddstoreexpminusmax.cc",
5256 "test/raddstoreexpminusmax-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005262 name = "f32_relu_test",
5263 srcs = [
5264 "test/f32-relu.cc",
5265 "test/vunary-microkernel-tester.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005271 name = "f32_rmax_test",
5272 srcs = [
5273 "test/f32-rmax.cc",
5274 "test/rmax-microkernel-tester.h",
5275 ] + MICROKERNEL_TEST_HDRS,
5276 deps = MICROKERNEL_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005280 name = "f32_sigmoid_test",
5281 srcs = [
5282 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005283 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005284 ] + MICROKERNEL_TEST_HDRS,
5285 deps = MICROKERNEL_TEST_DEPS,
5286)
5287
5288xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005289 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005290 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005291 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 "test/spmm-microkernel-tester.h",
5293 "src/xnnpack/AlignedAllocator.h",
5294 ] + MICROKERNEL_TEST_HDRS,
5295 deps = MICROKERNEL_TEST_DEPS,
5296)
5297
5298xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005299 name = "f16_vadd_minmax_test",
5300 srcs = [
5301 "test/f16-vadd-minmax.cc",
5302 "test/vbinary-microkernel-tester.h",
5303 ] + MICROKERNEL_TEST_HDRS,
5304 deps = MICROKERNEL_TEST_DEPS,
5305)
5306
5307xnnpack_unit_test(
5308 name = "f16_vaddc_minmax_test",
5309 srcs = [
5310 "test/f16-vaddc-minmax.cc",
5311 "test/vbinaryc-microkernel-tester.h",
5312 ] + MICROKERNEL_TEST_HDRS,
5313 deps = MICROKERNEL_TEST_DEPS,
5314)
5315
5316xnnpack_unit_test(
5317 name = "f16_vdiv_minmax_test",
5318 srcs = [
5319 "test/f16-vdiv-minmax.cc",
5320 "test/vbinary-microkernel-tester.h",
5321 ] + MICROKERNEL_TEST_HDRS,
5322 deps = MICROKERNEL_TEST_DEPS,
5323)
5324
5325xnnpack_unit_test(
5326 name = "f16_vdivc_minmax_test",
5327 srcs = [
5328 "test/f16-vdivc-minmax.cc",
5329 "test/vbinaryc-microkernel-tester.h",
5330 ] + MICROKERNEL_TEST_HDRS,
5331 deps = MICROKERNEL_TEST_DEPS,
5332)
5333
5334xnnpack_unit_test(
5335 name = "f16_vrdivc_minmax_test",
5336 srcs = [
5337 "test/f16-vrdivc-minmax.cc",
5338 "test/vbinaryc-microkernel-tester.h",
5339 ] + MICROKERNEL_TEST_HDRS,
5340 deps = MICROKERNEL_TEST_DEPS,
5341)
5342
5343xnnpack_unit_test(
5344 name = "f16_vmax_test",
5345 srcs = [
5346 "test/f16-vmax.cc",
5347 "test/vbinary-microkernel-tester.h",
5348 ] + MICROKERNEL_TEST_HDRS,
5349 deps = MICROKERNEL_TEST_DEPS,
5350)
5351
5352xnnpack_unit_test(
5353 name = "f16_vmaxc_test",
5354 srcs = [
5355 "test/f16-vmaxc.cc",
5356 "test/vbinaryc-microkernel-tester.h",
5357 ] + MICROKERNEL_TEST_HDRS,
5358 deps = MICROKERNEL_TEST_DEPS,
5359)
5360
5361xnnpack_unit_test(
5362 name = "f16_vmin_test",
5363 srcs = [
5364 "test/f16-vmin.cc",
5365 "test/vbinary-microkernel-tester.h",
5366 ] + MICROKERNEL_TEST_HDRS,
5367 deps = MICROKERNEL_TEST_DEPS,
5368)
5369
5370xnnpack_unit_test(
5371 name = "f16_vminc_test",
5372 srcs = [
5373 "test/f16-vminc.cc",
5374 "test/vbinaryc-microkernel-tester.h",
5375 ] + MICROKERNEL_TEST_HDRS,
5376 deps = MICROKERNEL_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
5380 name = "f16_vmul_minmax_test",
5381 srcs = [
5382 "test/f16-vmul-minmax.cc",
5383 "test/vbinary-microkernel-tester.h",
5384 ] + MICROKERNEL_TEST_HDRS,
5385 deps = MICROKERNEL_TEST_DEPS,
5386)
5387
5388xnnpack_unit_test(
5389 name = "f16_vmulc_minmax_test",
5390 srcs = [
5391 "test/f16-vmulc-minmax.cc",
5392 "test/vbinaryc-microkernel-tester.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
5398 name = "f16_vsub_minmax_test",
5399 srcs = [
5400 "test/f16-vsub-minmax.cc",
5401 "test/vbinary-microkernel-tester.h",
5402 ] + MICROKERNEL_TEST_HDRS,
5403 deps = MICROKERNEL_TEST_DEPS,
5404)
5405
5406xnnpack_unit_test(
5407 name = "f16_vsubc_minmax_test",
5408 srcs = [
5409 "test/f16-vsubc-minmax.cc",
5410 "test/vbinaryc-microkernel-tester.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
5416 name = "f16_vrsubc_minmax_test",
5417 srcs = [
5418 "test/f16-vrsubc-minmax.cc",
5419 "test/vbinaryc-microkernel-tester.h",
5420 ] + MICROKERNEL_TEST_HDRS,
5421 deps = MICROKERNEL_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005425 name = "f32_vabs_test",
5426 srcs = [
5427 "test/f32-vabs.cc",
5428 "test/vunary-microkernel-tester.h",
5429 ] + MICROKERNEL_TEST_HDRS,
5430 deps = MICROKERNEL_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005434 name = "f32_vadd_test",
5435 srcs = [
5436 "test/f32-vadd.cc",
5437 "test/vbinary-microkernel-tester.h",
5438 ] + MICROKERNEL_TEST_HDRS,
5439 deps = MICROKERNEL_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005443 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005444 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005445 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005446 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005447 ] + MICROKERNEL_TEST_HDRS,
5448 deps = MICROKERNEL_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005452 name = "f32_vadd_relu_test",
5453 srcs = [
5454 "test/f32-vadd-relu.cc",
5455 "test/vbinary-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005461 name = "f32_vaddc_test",
5462 srcs = [
5463 "test/f32-vaddc.cc",
5464 "test/vbinaryc-microkernel-tester.h",
5465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005470 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005471 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005472 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005473 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005479 name = "f32_vaddc_relu_test",
5480 srcs = [
5481 "test/f32-vaddc-relu.cc",
5482 "test/vbinaryc-microkernel-tester.h",
5483 ] + MICROKERNEL_TEST_HDRS,
5484 deps = MICROKERNEL_TEST_DEPS,
5485)
5486
5487xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005488 name = "f32_vdiv_test",
5489 srcs = [
5490 "test/f32-vdiv.cc",
5491 "test/vbinary-microkernel-tester.h",
5492 ] + MICROKERNEL_TEST_HDRS,
5493 deps = MICROKERNEL_TEST_DEPS,
5494)
5495
5496xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005497 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005498 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005499 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005500 "test/vbinary-microkernel-tester.h",
5501 ] + MICROKERNEL_TEST_HDRS,
5502 deps = MICROKERNEL_TEST_DEPS,
5503)
5504
5505xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005506 name = "f32_vdiv_relu_test",
5507 srcs = [
5508 "test/f32-vdiv-relu.cc",
5509 "test/vbinary-microkernel-tester.h",
5510 ] + MICROKERNEL_TEST_HDRS,
5511 deps = MICROKERNEL_TEST_DEPS,
5512)
5513
5514xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005515 name = "f32_vdivc_test",
5516 srcs = [
5517 "test/f32-vdivc.cc",
5518 "test/vbinaryc-microkernel-tester.h",
5519 ] + MICROKERNEL_TEST_HDRS,
5520 deps = MICROKERNEL_TEST_DEPS,
5521)
5522
5523xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005524 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005525 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005526 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005527 "test/vbinaryc-microkernel-tester.h",
5528 ] + MICROKERNEL_TEST_HDRS,
5529 deps = MICROKERNEL_TEST_DEPS,
5530)
5531
5532xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005533 name = "f32_vdivc_relu_test",
5534 srcs = [
5535 "test/f32-vdivc-relu.cc",
5536 "test/vbinaryc-microkernel-tester.h",
5537 ] + MICROKERNEL_TEST_HDRS,
5538 deps = MICROKERNEL_TEST_DEPS,
5539)
5540
5541xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005542 name = "f32_vrdivc_test",
5543 srcs = [
5544 "test/f32-vrdivc.cc",
5545 "test/vbinaryc-microkernel-tester.h",
5546 ] + MICROKERNEL_TEST_HDRS,
5547 deps = MICROKERNEL_TEST_DEPS,
5548)
5549
5550xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005551 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005552 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005553 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005554 "test/vbinaryc-microkernel-tester.h",
5555 ] + MICROKERNEL_TEST_HDRS,
5556 deps = MICROKERNEL_TEST_DEPS,
5557)
5558
5559xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005560 name = "f32_vrdivc_relu_test",
5561 srcs = [
5562 "test/f32-vrdivc-relu.cc",
5563 "test/vbinaryc-microkernel-tester.h",
5564 ] + MICROKERNEL_TEST_HDRS,
5565 deps = MICROKERNEL_TEST_DEPS,
5566)
5567
5568xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005569 name = "f32_vmax_test",
5570 srcs = [
5571 "test/f32-vmax.cc",
5572 "test/vbinary-microkernel-tester.h",
5573 ] + MICROKERNEL_TEST_HDRS,
5574 deps = MICROKERNEL_TEST_DEPS,
5575)
5576
5577xnnpack_unit_test(
5578 name = "f32_vmaxc_test",
5579 srcs = [
5580 "test/f32-vmaxc.cc",
5581 "test/vbinaryc-microkernel-tester.h",
5582 ] + MICROKERNEL_TEST_HDRS,
5583 deps = MICROKERNEL_TEST_DEPS,
5584)
5585
5586xnnpack_unit_test(
5587 name = "f32_vmin_test",
5588 srcs = [
5589 "test/f32-vmin.cc",
5590 "test/vbinary-microkernel-tester.h",
5591 ] + MICROKERNEL_TEST_HDRS,
5592 deps = MICROKERNEL_TEST_DEPS,
5593)
5594
5595xnnpack_unit_test(
5596 name = "f32_vminc_test",
5597 srcs = [
5598 "test/f32-vminc.cc",
5599 "test/vbinaryc-microkernel-tester.h",
5600 ] + MICROKERNEL_TEST_HDRS,
5601 deps = MICROKERNEL_TEST_DEPS,
5602)
5603
5604xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005605 name = "f32_vmul_test",
5606 srcs = [
5607 "test/f32-vmul.cc",
5608 "test/vbinary-microkernel-tester.h",
5609 ] + MICROKERNEL_TEST_HDRS,
5610 deps = MICROKERNEL_TEST_DEPS,
5611)
5612
5613xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005614 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005615 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005616 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005617 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005618 ] + MICROKERNEL_TEST_HDRS,
5619 deps = MICROKERNEL_TEST_DEPS,
5620)
5621
5622xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005623 name = "f32_vmul_relu_test",
5624 srcs = [
5625 "test/f32-vmul-relu.cc",
5626 "test/vbinary-microkernel-tester.h",
5627 ] + MICROKERNEL_TEST_HDRS,
5628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
5631xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005632 name = "f32_vmulc_test",
5633 srcs = [
5634 "test/f32-vmulc.cc",
5635 "test/vbinaryc-microkernel-tester.h",
5636 ] + MICROKERNEL_TEST_HDRS,
5637 deps = MICROKERNEL_TEST_DEPS,
5638)
5639
5640xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005641 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005642 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005643 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005644 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005645 ] + MICROKERNEL_TEST_HDRS,
5646 deps = MICROKERNEL_TEST_DEPS,
5647)
5648
5649xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005650 name = "f32_vmulc_relu_test",
5651 srcs = [
5652 "test/f32-vmulc-relu.cc",
5653 "test/vbinaryc-microkernel-tester.h",
5654 ] + MICROKERNEL_TEST_HDRS,
5655 deps = MICROKERNEL_TEST_DEPS,
5656)
5657
5658xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005659 name = "f16_vmulcaddc_minmax_test",
5660 srcs = [
5661 "test/f16-vmulcaddc-minmax.cc",
5662 "test/vmulcaddc-microkernel-tester.h",
5663 "src/xnnpack/AlignedAllocator.h",
5664 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005665 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005666)
5667
5668xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005669 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005671 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672 "test/vmulcaddc-microkernel-tester.h",
5673 "src/xnnpack/AlignedAllocator.h",
5674 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005675 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005676)
5677
5678xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005679 name = "f32_vlrelu_test",
5680 srcs = [
5681 "test/f32-vlrelu.cc",
5682 "test/vunary-microkernel-tester.h",
5683 ] + MICROKERNEL_TEST_HDRS,
5684 deps = MICROKERNEL_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005688 name = "f32_vneg_test",
5689 srcs = [
5690 "test/f32-vneg.cc",
5691 "test/vunary-microkernel-tester.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005697 name = "f32_vrndne_test",
5698 srcs = [
5699 "test/f32-vrndne.cc",
5700 "test/vunary-microkernel-tester.h",
5701 ] + MICROKERNEL_TEST_HDRS,
5702 deps = MICROKERNEL_TEST_DEPS,
5703)
5704
5705xnnpack_unit_test(
5706 name = "f32_vrndz_test",
5707 srcs = [
5708 "test/f32-vrndz.cc",
5709 "test/vunary-microkernel-tester.h",
5710 ] + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS,
5712)
5713
5714xnnpack_unit_test(
5715 name = "f32_vrndu_test",
5716 srcs = [
5717 "test/f32-vrndu.cc",
5718 "test/vunary-microkernel-tester.h",
5719 ] + MICROKERNEL_TEST_HDRS,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
5724 name = "f32_vrndd_test",
5725 srcs = [
5726 "test/f32-vrndd.cc",
5727 "test/vunary-microkernel-tester.h",
5728 ] + MICROKERNEL_TEST_HDRS,
5729 deps = MICROKERNEL_TEST_DEPS,
5730)
5731
5732xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005733 name = "f32_vscale_test",
5734 srcs = [
5735 "test/f32-vscale.cc",
5736 "test/vscale-microkernel-tester.h",
5737 ] + MICROKERNEL_TEST_HDRS,
5738 deps = MICROKERNEL_TEST_DEPS,
5739)
5740
5741xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005742 name = "f32_vscaleexpminusmax_test",
5743 srcs = [
5744 "test/f32-vscaleexpminusmax.cc",
5745 "test/vscaleexpminusmax-microkernel-tester.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005751 name = "f32_vscaleextexp_test",
5752 srcs = [
5753 "test/f32-vscaleextexp.cc",
5754 "test/vscaleextexp-microkernel-tester.h",
5755 ] + MICROKERNEL_TEST_HDRS,
5756 deps = MICROKERNEL_TEST_DEPS,
5757)
5758
5759xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005760 name = "f32_vsqr_test",
5761 srcs = [
5762 "test/f32-vsqr.cc",
5763 "test/vunary-microkernel-tester.h",
5764 ] + MICROKERNEL_TEST_HDRS,
5765 deps = MICROKERNEL_TEST_DEPS,
5766)
5767
5768xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005769 name = "f32_vsqrdiff_test",
5770 srcs = [
5771 "test/f32-vsqrdiff.cc",
5772 "test/vbinary-microkernel-tester.h",
5773 ] + MICROKERNEL_TEST_HDRS,
5774 deps = MICROKERNEL_TEST_DEPS,
5775)
5776
5777xnnpack_unit_test(
5778 name = "f32_vsqrdiffc_test",
5779 srcs = [
5780 "test/f32-vsqrdiffc.cc",
5781 "test/vbinaryc-microkernel-tester.h",
5782 ] + MICROKERNEL_TEST_HDRS,
5783 deps = MICROKERNEL_TEST_DEPS,
5784)
5785
5786xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005787 name = "f32_vsqrt_test",
5788 srcs = [
5789 "test/f32-vsqrt.cc",
5790 "test/vunary-microkernel-tester.h",
5791 ] + MICROKERNEL_TEST_HDRS,
5792 deps = MICROKERNEL_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005796 name = "f32_vsub_test",
5797 srcs = [
5798 "test/f32-vsub.cc",
5799 "test/vbinary-microkernel-tester.h",
5800 ] + MICROKERNEL_TEST_HDRS,
5801 deps = MICROKERNEL_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005805 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005806 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005807 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005808 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005809 ] + MICROKERNEL_TEST_HDRS,
5810 deps = MICROKERNEL_TEST_DEPS,
5811)
5812
5813xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005814 name = "f32_vsub_relu_test",
5815 srcs = [
5816 "test/f32-vsub-relu.cc",
5817 "test/vbinary-microkernel-tester.h",
5818 ] + MICROKERNEL_TEST_HDRS,
5819 deps = MICROKERNEL_TEST_DEPS,
5820)
5821
5822xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005823 name = "f32_vsubc_test",
5824 srcs = [
5825 "test/f32-vsubc.cc",
5826 "test/vbinaryc-microkernel-tester.h",
5827 ] + MICROKERNEL_TEST_HDRS,
5828 deps = MICROKERNEL_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005832 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005833 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005834 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005835 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005836 ] + MICROKERNEL_TEST_HDRS,
5837 deps = MICROKERNEL_TEST_DEPS,
5838)
5839
5840xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005841 name = "f32_vsubc_relu_test",
5842 srcs = [
5843 "test/f32-vsubc-relu.cc",
5844 "test/vbinaryc-microkernel-tester.h",
5845 ] + MICROKERNEL_TEST_HDRS,
5846 deps = MICROKERNEL_TEST_DEPS,
5847)
5848
5849xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005850 name = "f32_vrsubc_test",
5851 srcs = [
5852 "test/f32-vrsubc.cc",
5853 "test/vbinaryc-microkernel-tester.h",
5854 ] + MICROKERNEL_TEST_HDRS,
5855 deps = MICROKERNEL_TEST_DEPS,
5856)
5857
5858xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005859 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005860 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005861 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005862 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005863 ] + MICROKERNEL_TEST_HDRS,
5864 deps = MICROKERNEL_TEST_DEPS,
5865)
5866
5867xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005868 name = "f32_vrsubc_relu_test",
5869 srcs = [
5870 "test/f32-vrsubc-relu.cc",
5871 "test/vbinaryc-microkernel-tester.h",
5872 ] + MICROKERNEL_TEST_HDRS,
5873 deps = MICROKERNEL_TEST_DEPS,
5874)
5875
5876xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005877 name = "qs8_dwconv_minmax_test",
5878 srcs = [
5879 "test/qs8-dwconv-minmax.cc",
5880 "test/dwconv-microkernel-tester.h",
5881 "src/xnnpack/AlignedAllocator.h",
5882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5883 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5884)
5885
5886xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005887 name = "qs8_gavgpool_minmax_test",
5888 srcs = [
5889 "test/qs8-gavgpool-minmax.cc",
5890 "test/gavgpool-microkernel-tester.h",
5891 "src/xnnpack/AlignedAllocator.h",
5892 ] + MICROKERNEL_TEST_HDRS,
5893 deps = MICROKERNEL_TEST_DEPS,
5894)
5895
5896xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005897 name = "qs8_gemm_minmax_test",
5898 srcs = [
5899 "test/qs8-gemm-minmax.cc",
5900 "test/gemm-microkernel-tester.h",
5901 "src/xnnpack/AlignedAllocator.h",
5902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5903 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5904)
5905
5906xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005907 name = "qs8_igemm_minmax_test",
5908 srcs = [
5909 "test/qs8-igemm-minmax.cc",
5910 "test/gemm-microkernel-tester.h",
5911 "src/xnnpack/AlignedAllocator.h",
5912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5914)
5915
5916xnnpack_unit_test(
5917 name = "qs8_requantization_test",
5918 srcs = [
5919 "src/xnnpack/requantization-stubs.h",
5920 "test/qs8-requantization.cc",
5921 "test/requantization-tester.h",
5922 ] + MICROKERNEL_TEST_HDRS,
5923 deps = MICROKERNEL_TEST_DEPS,
5924)
5925
5926xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005927 name = "qs8_vadd_minmax_test",
5928 srcs = [
5929 "test/qs8-vadd-minmax.cc",
5930 "test/vadd-microkernel-tester.h",
5931 ] + MICROKERNEL_TEST_HDRS,
5932 deps = MICROKERNEL_TEST_DEPS,
5933)
5934
5935xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005936 name = "qs8_vaddc_minmax_test",
5937 srcs = [
5938 "test/qs8-vaddc-minmax.cc",
5939 "test/vaddc-microkernel-tester.h",
5940 ] + MICROKERNEL_TEST_HDRS,
5941 deps = MICROKERNEL_TEST_DEPS,
5942)
5943
5944xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005945 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005946 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005947 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005948 "test/avgpool-microkernel-tester.h",
5949 "src/xnnpack/AlignedAllocator.h",
5950 ] + MICROKERNEL_TEST_HDRS,
5951 deps = MICROKERNEL_TEST_DEPS,
5952)
5953
5954xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005955 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005956 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005957 "test/qu8-dwconv-minmax.cc",
5958 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005959 "src/xnnpack/AlignedAllocator.h",
5960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005961 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005962)
5963
5964xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005965 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005967 "test/qu8-igemm-minmax.cc",
5968 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005969 "src/xnnpack/AlignedAllocator.h",
5970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005971 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972)
5973
5974xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005975 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005976 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005977 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 "test/gavgpool-microkernel-tester.h",
5979 "src/xnnpack/AlignedAllocator.h",
5980 ] + MICROKERNEL_TEST_HDRS,
5981 deps = MICROKERNEL_TEST_DEPS,
5982)
5983
5984xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005985 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005986 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005987 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 "test/gemm-microkernel-tester.h",
5989 "src/xnnpack/AlignedAllocator.h",
5990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005991 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005992)
5993
5994xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005995 name = "qu8_requantization_test",
5996 srcs = [
5997 "src/xnnpack/requantization-stubs.h",
5998 "test/qu8-requantization.cc",
5999 "test/requantization-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006005 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006006 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006007 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006008 "test/vadd-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
6014 name = "u8_clamp_test",
6015 srcs = [
6016 "test/u8-clamp.cc",
6017 "test/clamp-microkernel-tester.h",
6018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
6023 name = "u8_lut32norm_test",
6024 srcs = [
6025 "test/u8-lut32norm.cc",
6026 "test/lut-norm-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006032 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006033 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006034 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006035 "test/maxpool-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
6041 name = "u8_rmax_test",
6042 srcs = [
6043 "test/u8-rmax.cc",
6044 "test/rmax-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006050 name = "x32_fill_test",
6051 srcs = [
6052 "test/x32-fill.cc",
6053 "test/fill-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006059 name = "x32_packx_test",
6060 srcs = [
6061 "test/x32-packx.cc",
6062 "test/pack-microkernel-tester.h",
6063 "src/xnnpack/AlignedAllocator.h",
6064 ] + MICROKERNEL_TEST_HDRS,
6065 deps = MICROKERNEL_TEST_DEPS,
6066)
6067
6068xnnpack_unit_test(
6069 name = "x32_pad_test",
6070 srcs = [
6071 "test/x32-pad.cc",
6072 "test/pad-microkernel-tester.h",
6073 ] + MICROKERNEL_TEST_HDRS,
6074 deps = MICROKERNEL_TEST_DEPS,
6075)
6076
6077xnnpack_unit_test(
6078 name = "x32_unpool_test",
6079 srcs = [
6080 "test/x32-unpool.cc",
6081 "test/unpool-microkernel-tester.h",
6082 ] + MICROKERNEL_TEST_HDRS,
6083 deps = MICROKERNEL_TEST_DEPS,
6084)
6085
6086xnnpack_unit_test(
6087 name = "x32_zip_test",
6088 srcs = [
6089 "test/x32-zip.cc",
6090 "test/zip-microkernel-tester.h",
6091 ] + MICROKERNEL_TEST_HDRS,
6092 deps = MICROKERNEL_TEST_DEPS,
6093)
6094
6095xnnpack_unit_test(
6096 name = "x8_lut_test",
6097 srcs = [
6098 "test/x8-lut.cc",
6099 "test/lut-microkernel-tester.h",
6100 ] + MICROKERNEL_TEST_HDRS,
6101 deps = MICROKERNEL_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
6105 name = "x8_zip_test",
6106 srcs = [
6107 "test/x8-zip.cc",
6108 "test/zip-microkernel-tester.h",
6109 ] + MICROKERNEL_TEST_HDRS,
6110 deps = MICROKERNEL_TEST_DEPS,
6111)
6112
Marat Dukhan20c3b922020-03-10 03:45:06 -07006113########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006114
6115xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006116 name = "operator_size_test",
6117 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006118 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006119)
6120
Marat Dukhan20c3b922020-03-10 03:45:06 -07006121xnnpack_binary(
6122 name = "subgraph_size_test",
6123 srcs = ["test/subgraph-size.c"],
6124 deps = [":XNNPACK"],
6125)
6126
6127########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006128
6129xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006130 name = "abs_nc_test",
6131 srcs = [
6132 "test/abs-nc.cc",
6133 "test/abs-operator-tester.h",
6134 ],
6135 deps = OPERATOR_TEST_DEPS,
6136)
6137
6138xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006139 name = "add_nd_test",
6140 srcs = [
6141 "test/add-nd.cc",
6142 "test/binary-elementwise-operator-tester.h",
6143 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006144 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006145)
6146
6147xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006148 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006150 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006151 "test/argmax-pooling-operator-tester.h",
6152 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006153 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154)
6155
6156xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006157 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006158 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006159 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006160 "test/average-pooling-operator-tester.h",
6161 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006162 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006163)
6164
6165xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006166 name = "bankers_rounding_nc_test",
6167 srcs = [
6168 "test/bankers-rounding-nc.cc",
6169 "test/bankers-rounding-operator-tester.h",
6170 ],
6171 deps = OPERATOR_TEST_DEPS,
6172)
6173
6174xnnpack_unit_test(
6175 name = "ceiling_nc_test",
6176 srcs = [
6177 "test/ceiling-nc.cc",
6178 "test/ceiling-operator-tester.h",
6179 ],
6180 deps = OPERATOR_TEST_DEPS,
6181)
6182
6183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006184 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006186 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006187 "test/channel-shuffle-operator-tester.h",
6188 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006190)
6191
6192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006193 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006195 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 "test/clamp-operator-tester.h",
6197 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199)
6200
6201xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006202 name = "constant_pad_nd_test",
6203 srcs = [
6204 "test/constant-pad-nd.cc",
6205 "test/constant-pad-operator-tester.h",
6206 ],
6207 deps = OPERATOR_TEST_DEPS,
6208)
6209
6210xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006211 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006212 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006213 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214 "test/convolution-operator-tester.h",
6215 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006216 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006217)
6218
6219xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006220 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006221 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006222 "test/convolution-nchw.cc",
6223 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006225 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226)
6227
6228xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006229 name = "copy_nc_test",
6230 srcs = [
6231 "test/copy-nc.cc",
6232 "test/copy-operator-tester.h",
6233 ],
6234 deps = OPERATOR_TEST_DEPS,
6235)
6236
6237xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006238 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006240 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006241 "test/deconvolution-operator-tester.h",
6242 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006243 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244)
6245
6246xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006247 name = "divide_nd_test",
6248 srcs = [
6249 "test/binary-elementwise-operator-tester.h",
6250 "test/divide-nd.cc",
6251 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006252 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006253)
6254
6255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006256 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006257 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006258 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259 "test/fully-connected-operator-tester.h",
6260 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006261 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262)
6263
6264xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006265 name = "floor_nc_test",
6266 srcs = [
6267 "test/floor-nc.cc",
6268 "test/floor-operator-tester.h",
6269 ],
6270 deps = OPERATOR_TEST_DEPS,
6271)
6272
6273xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006274 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006275 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006276 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006278 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006279 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006280)
6281
6282xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006283 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006285 "test/global-average-pooling-ncw.cc",
6286 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006288 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006289)
6290
6291xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006292 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006294 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006295 "test/hardswish-operator-tester.h",
6296 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006297 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006298)
6299
6300xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006301 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006302 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006303 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006304 "test/leaky-relu-operator-tester.h",
6305 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006306 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006307)
6308
6309xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006310 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006311 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006312 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006313 "test/max-pooling-operator-tester.h",
6314 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006315 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316)
6317
6318xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006319 name = "maximum_nd_test",
6320 srcs = [
6321 "test/binary-elementwise-operator-tester.h",
6322 "test/maximum-nd.cc",
6323 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006324 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006325)
6326
6327xnnpack_unit_test(
6328 name = "minimum_nd_test",
6329 srcs = [
6330 "test/binary-elementwise-operator-tester.h",
6331 "test/minimum-nd.cc",
6332 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006333 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006334)
6335
6336xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006337 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006338 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006339 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006340 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006341 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006342 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006343)
6344
6345xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006346 name = "negate_nc_test",
6347 srcs = [
6348 "test/negate-nc.cc",
6349 "test/negate-operator-tester.h",
6350 ],
6351 deps = OPERATOR_TEST_DEPS,
6352)
6353
6354xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006355 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006356 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006357 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006358 "test/prelu-operator-tester.h",
6359 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006360 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006361)
6362
6363xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006364 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006365 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006366 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006367 "test/resize-bilinear-operator-tester.h",
6368 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006369 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006370)
6371
6372xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006373 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006374 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006375 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006376 "test/sigmoid-operator-tester.h",
6377 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006378 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006379)
6380
6381xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006382 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006383 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006384 "test/softmax-nc.cc",
6385 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006386 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006387 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006388)
6389
6390xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006391 name = "square_nc_test",
6392 srcs = [
6393 "test/square-nc.cc",
6394 "test/square-operator-tester.h",
6395 ],
6396 deps = OPERATOR_TEST_DEPS,
6397)
6398
6399xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006400 name = "square_root_nc_test",
6401 srcs = [
6402 "test/square-root-nc.cc",
6403 "test/square-root-operator-tester.h",
6404 ],
6405 deps = OPERATOR_TEST_DEPS,
6406)
6407
6408xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006409 name = "squared_difference_nd_test",
6410 srcs = [
6411 "test/binary-elementwise-operator-tester.h",
6412 "test/squared-difference-nd.cc",
6413 ],
6414 deps = OPERATOR_TEST_DEPS,
6415)
6416
6417xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006418 name = "subtract_nd_test",
6419 srcs = [
6420 "test/binary-elementwise-operator-tester.h",
6421 "test/subtract-nd.cc",
6422 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006423 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006424)
6425
6426xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006427 name = "truncation_nc_test",
6428 srcs = [
6429 "test/truncation-nc.cc",
6430 "test/truncation-operator-tester.h",
6431 ],
6432 deps = OPERATOR_TEST_DEPS,
6433)
6434
6435xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006436 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006437 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006438 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006439 "test/unpooling-operator-tester.h",
6440 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006441 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442)
6443
Chao Mei6ddfc602020-05-13 22:29:36 -07006444############################### Misc unit tests ###############################
6445
6446xnnpack_unit_test(
6447 name = "memory_planner_test",
6448 srcs = [
6449 "test/memory-planner-test.cc",
6450 ],
6451 deps = [
6452 ":XNNPACK",
6453 ":memory_planner",
6454 ],
6455)
6456
Marat Dukhan08c4a432019-10-03 09:29:21 -07006457############################# Build configurations #############################
6458
Marat Dukhanb8642352019-10-30 15:43:02 -07006459# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006460config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006461 name = "xnn_enable_assembly_explicit_true",
6462 define_values = {"xnn_enable_assembly": "true"},
6463)
6464
6465# Disables usage of assembly kernels.
6466config_setting(
6467 name = "xnn_enable_assembly_explicit_false",
6468 define_values = {"xnn_enable_assembly": "false"},
6469)
6470
Marat Dukhan9de90e02020-06-18 16:04:12 -07006471# Enables usage of sparse inference.
6472config_setting(
6473 name = "xnn_enable_sparse_explicit_true",
6474 define_values = {"xnn_enable_sparse": "true"},
6475)
6476
6477# Disables usage of sparse inference.
6478config_setting(
6479 name = "xnn_enable_sparse_explicit_false",
6480 define_values = {"xnn_enable_sparse": "false"},
6481)
6482
Marat Dukhan05702cf2020-03-26 15:41:33 -07006483# Disables usage of HMP-aware optimizations.
6484config_setting(
6485 name = "xnn_enable_hmp_explicit_false",
6486 define_values = {"xnn_enable_hmp": "false"},
6487)
6488
Chao Mei6ddfc602020-05-13 22:29:36 -07006489# Enable usage of optimized memory allocation
6490config_setting(
6491 name = "xnn_enable_memopt_explicit_true",
6492 define_values = {"xnnpack_enable_memopt": "true"},
6493)
6494
6495# Disable usage of optimized memory allocation
6496config_setting(
6497 name = "xnn_enable_memopt_explicit_false",
6498 define_values = {"xnnpack_enable_memopt": "false"},
6499)
6500
Marat Dukhanb8642352019-10-30 15:43:02 -07006501# Builds with -c dbg
6502config_setting(
6503 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006504 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006505 "compilation_mode": "dbg",
6506 },
6507)
6508
6509# Builds with -c opt
6510config_setting(
6511 name = "optimized_build",
6512 values = {
6513 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006514 },
6515)
6516
6517config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006518 name = "linux_k8",
6519 values = {"cpu": "k8"},
6520)
6521
6522config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006523 name = "linux_arm",
6524 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006525)
6526
6527config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006528 name = "linux_armeabi",
6529 values = {"cpu": "armeabi"},
6530)
6531
6532config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006533 name = "linux_armhf",
6534 values = {"cpu": "armhf"},
6535)
6536
6537config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006538 name = "linux_armv7a",
6539 values = {"cpu": "armv7a"},
6540)
6541
6542config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006543 name = "linux_aarch64",
6544 values = {"cpu": "aarch64"},
6545)
6546
6547config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006548 name = "android",
6549 values = {"crosstool_top": "//external:android/crosstool"},
6550)
6551
6552config_setting(
6553 name = "android_armv7",
6554 values = {
6555 "crosstool_top": "//external:android/crosstool",
6556 "cpu": "armeabi-v7a",
6557 },
6558)
6559
6560config_setting(
6561 name = "android_arm64",
6562 values = {
6563 "crosstool_top": "//external:android/crosstool",
6564 "cpu": "arm64-v8a",
6565 },
6566)
6567
6568config_setting(
6569 name = "android_x86",
6570 values = {
6571 "crosstool_top": "//external:android/crosstool",
6572 "cpu": "x86",
6573 },
6574)
6575
6576config_setting(
6577 name = "android_x86_64",
6578 values = {
6579 "crosstool_top": "//external:android/crosstool",
6580 "cpu": "x86_64",
6581 },
6582)
6583
6584config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006585 name = "windows_x86_64",
6586 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006587)
6588
6589config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006590 name = "windows_x86_64_clang",
6591 values = {
6592 "compiler": "clang-cl",
6593 "cpu": "x64_windows",
6594 },
6595)
6596
6597config_setting(
6598 name = "windows_x86_64_mingw",
6599 values = {
6600 "compiler": "mingw-gcc",
6601 "cpu": "x64_windows",
6602 },
6603)
6604
6605config_setting(
6606 name = "windows_x86_64_msys",
6607 values = {
6608 "compiler": "msys-gcc",
6609 "cpu": "x64_windows",
6610 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006611)
6612
6613config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006614 name = "macos_x86_64",
6615 values = {
6616 "apple_platform_type": "macos",
6617 "cpu": "darwin",
6618 },
6619)
6620
6621config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006622 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006623 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006624)
6625
6626config_setting(
6627 name = "emscripten_wasm",
6628 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006629 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006630 "cpu": "wasm",
6631 },
6632)
6633
6634config_setting(
6635 name = "emscripten_wasmsimd",
6636 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006637 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006638 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006639 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006640 },
6641)
6642
6643config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006644 name = "ios_armv7",
6645 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006646 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006647 "cpu": "ios_armv7",
6648 },
6649)
6650
6651config_setting(
6652 name = "ios_arm64",
6653 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006654 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006655 "cpu": "ios_arm64",
6656 },
6657)
6658
6659config_setting(
6660 name = "ios_arm64e",
6661 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006662 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006663 "cpu": "ios_arm64e",
6664 },
6665)
6666
6667config_setting(
6668 name = "ios_x86",
6669 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006670 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006671 "cpu": "ios_i386",
6672 },
6673)
6674
6675config_setting(
6676 name = "ios_x86_64",
6677 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006678 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006679 "cpu": "ios_x86_64",
6680 },
6681)
6682
6683config_setting(
6684 name = "watchos_armv7k",
6685 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006686 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006687 "cpu": "watchos_armv7k",
6688 },
6689)
6690
6691config_setting(
6692 name = "watchos_arm64_32",
6693 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006694 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006695 "cpu": "watchos_arm64_32",
6696 },
6697)
6698
6699config_setting(
6700 name = "watchos_x86",
6701 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006702 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006703 "cpu": "watchos_i386",
6704 },
6705)
6706
6707config_setting(
6708 name = "watchos_x86_64",
6709 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006710 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006711 "cpu": "watchos_x86_64",
6712 },
6713)
6714
6715config_setting(
6716 name = "tvos_arm64",
6717 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006718 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006719 "cpu": "tvos_arm64",
6720 },
6721)
6722
6723config_setting(
6724 name = "tvos_x86_64",
6725 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006726 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006727 "cpu": "tvos_x86_64",
6728 },
6729)