blob: 524aa519d61576f09da4c314da1c7a8582ef961b [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 Dukhan5739f702019-12-22 19:45:09 -0800425 "src/math/sigmoid-scalar-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-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 Dukhan5b69f8b2020-07-24 15:26:48 -07001074 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001075 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001076 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001077 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001078 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001079 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001080 "src/x32-zip/x2-wasmsimd.c",
1081 "src/x32-zip/x3-wasmsimd.c",
1082 "src/x32-zip/x4-wasmsimd.c",
1083 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001084 "src/math/roundne-wasmsimd-addsub.c",
1085 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001086 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001087 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001088 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001089 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001090 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001091 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1092 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001093]
1094
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001095PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001096 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001097 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001098 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001099 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001100 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001101 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1102 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1103 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001104]
1105
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001106PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001107 "src/qs8-requantization/precise-psimd.c",
1108 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001109 "src/qu8-requantization/precise-psimd.c",
1110 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001111]
1112
Marat Dukhan08c4a432019-10-03 09:29:21 -07001113# ISA-specific micro-kernels
1114NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001115 "src/f32-argmaxpool/4x-neon-c4.c",
1116 "src/f32-argmaxpool/9p8x-neon-c4.c",
1117 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001118 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1119 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001120 "src/f32-clamp/gen/neon-x4.c",
1121 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001122 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1123 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001124 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1125 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001126 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1127 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001128 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1129 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001130 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1131 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1132 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1133 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001134 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1135 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1136 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1137 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001142 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001143 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1144 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001145 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1146 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1147 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1148 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1149 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1150 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1151 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1152 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1153 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1154 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1155 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1157 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1158 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1159 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1160 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1164 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1167 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1168 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1169 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1170 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1171 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1172 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1173 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1174 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1175 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001176 "src/f32-hswish/gen/hswish-neon-x4.c",
1177 "src/f32-hswish/gen/hswish-neon-x8.c",
1178 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001179 "src/f32-ibilinear/gen/neon-c4.c",
1180 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001181 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1182 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1183 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1184 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1185 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1186 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1187 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1188 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1189 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1190 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1191 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1192 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1193 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1194 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1195 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1196 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001197 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1198 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1199 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001200 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1201 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001202 "src/f32-prelu/gen/neon-1x4.c",
1203 "src/f32-prelu/gen/neon-1x8.c",
1204 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001205 "src/f32-prelu/gen/neon-2x4.c",
1206 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001207 "src/f32-prelu/gen/neon-2x16.c",
1208 "src/f32-prelu/gen/neon-4x4.c",
1209 "src/f32-prelu/gen/neon-4x8.c",
1210 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001211 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1212 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1213 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1214 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1215 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1216 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1217 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1218 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001235 "src/f32-relu/gen/neon-x4.c",
1236 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001237 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001238 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001239 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1240 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1241 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1242 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1243 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1244 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1245 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1246 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1247 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1248 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1249 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1250 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1251 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1252 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001257 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1258 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1259 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1260 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001261 "src/f32-vbinary/gen/vmax-neon-x4.c",
1262 "src/f32-vbinary/gen/vmax-neon-x8.c",
1263 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1264 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1265 "src/f32-vbinary/gen/vmin-neon-x4.c",
1266 "src/f32-vbinary/gen/vmin-neon-x8.c",
1267 "src/f32-vbinary/gen/vminc-neon-x4.c",
1268 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001269 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1272 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1273 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1274 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001275 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1276 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1277 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1278 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001279 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001283 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1284 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001285 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1286 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001287 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1288 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1289 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1290 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1291 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1292 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1293 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1294 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001295 "src/f32-vunary/gen/vabs-neon-x4.c",
1296 "src/f32-vunary/gen/vabs-neon-x8.c",
1297 "src/f32-vunary/gen/vneg-neon-x4.c",
1298 "src/f32-vunary/gen/vneg-neon-x8.c",
1299 "src/f32-vunary/gen/vsqr-neon-x4.c",
1300 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001301 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1302 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1303 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1304 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001305 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1306 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1307 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1308 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1309 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1310 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1311 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1312 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001313 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1314 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1315 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1316 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1317 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1318 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1319 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1320 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001321 "src/qs8-requantization/precise-neon.c",
1322 "src/qs8-requantization/fp32-neon.c",
1323 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001324 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1325 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1326 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1327 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1328 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1329 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1330 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1331 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001332 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1333 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1334 "src/qu8-dwconv/up8x9-minmax-neon.c",
1335 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1336 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1337 "src/qu8-gemm/4x8-minmax-neon.c",
1338 "src/qu8-gemm/8x8-minmax-neon.c",
1339 "src/qu8-igemm/4x8-minmax-neon.c",
1340 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001341 "src/qu8-requantization/precise-neon.c",
1342 "src/qu8-requantization/fp32-neon.c",
1343 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001344 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001345 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001346 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001347 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001348 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001349 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001350 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001351 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001352 "src/x32-zip/x2-neon.c",
1353 "src/x32-zip/x3-neon.c",
1354 "src/x32-zip/x4-neon.c",
1355 "src/x32-zip/xm-neon.c",
1356 "src/x8-zip/x2-neon.c",
1357 "src/x8-zip/x3-neon.c",
1358 "src/x8-zip/x4-neon.c",
1359 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001360 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001361 "src/math/roundd-neon-addsub.c",
1362 "src/math/roundd-neon-cvt.c",
1363 "src/math/roundu-neon-addsub.c",
1364 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001365 "src/math/roundz-neon-addsub.c",
1366 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001367 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001368 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1369 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1370 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1371 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1372 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1373 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001374 "src/math/sqrt-neon-nr1rsqrts.c",
1375 "src/math/sqrt-neon-nr2rsqrts.c",
1376 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377]
1378
1379NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001380 "src/f32-ibilinear/gen/neonfma-c4.c",
1381 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001382 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1383 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1384 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1385 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1386 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1387 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1388 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1389 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1390 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001391 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1392 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1393 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1394 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001395 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1396 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1397 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1398 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1404 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1405 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1406 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1407 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1408 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1409 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1410 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1411 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1412 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1413 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1414 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1416 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1417 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1418 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1419 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001421 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1422 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001423 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1424 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1425 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1426 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1427 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1428 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1429 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1430 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001447 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1448 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1449 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1450 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1451 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1452 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1453 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1454 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001501 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1502 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001503 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1504 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1505 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1506 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1507 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1508 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1509 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1510 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1513 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1514 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1515 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1516 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1517 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1518 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1519 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1520 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001523 "src/math/exp-neonfma-lut64-p2.c",
1524 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001525 "src/math/expminus-neonfma-lut2048-p1.c",
1526 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001527 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001528 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1529 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1530 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1531 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1532 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1533 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1534 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1535 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1536 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1537 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1538 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1539 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1540 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1541 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1542 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1543 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1544 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1545 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001546 "src/math/sqrt-neonfma-nr1fma.c",
1547 "src/math/sqrt-neonfma-nr2fma.c",
1548 "src/math/sqrt-neonfma-nr3fma.c",
1549 "src/math/sqrt-neonfma-nr2fma1adj.c",
1550 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001551]
1552
1553AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001554 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1556 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1558 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1559 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001560 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1561 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1562 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1563 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1564 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1565 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1566 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1567 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1568 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1569 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1571 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1573 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1575 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1576 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1577 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1578 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1579 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001580 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1581 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001582 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1583 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001584 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1585 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001586 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1587 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001588 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1589 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1590 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1591 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1592 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001593 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1594 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1595 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1596 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1600 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001611 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1612 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1613 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1614 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1615 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1616 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1617 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1618 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1619 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1620 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1621 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1622 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1623 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1624 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1625 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1626 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1627 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1628 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001629 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1630 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001631 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1632 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1633 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1634 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1635 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1636 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001637]
1638
Marat Dukhan8853b822020-05-07 12:19:01 -07001639NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001640 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1641 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1642 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1643 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1644 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1645 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1646 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1647 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001648 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001649 "src/math/roundd-neonv8.c",
1650 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001651 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001652]
1653
Marat Dukhan08c4a432019-10-03 09:29:21 -07001654AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001655 "src/f16-clamp/gen/neonfp16arith-x16.c",
1656 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001657 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1658 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1659 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1660 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1661 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1662 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1663 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1664 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1665 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001669 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1670 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001671 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1672 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001673 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001674 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001675 "src/f16-relu/gen/neonfp16arith-x16.c",
1676 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001677 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1678 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1679 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1680 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1681 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1682 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1683 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1684 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1685 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001705 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1706 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001707 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1708 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1709 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1710 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1711 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1712 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1713 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1714 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001715 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001719 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1723 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1724 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1725 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1726 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1727 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001731 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1732 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1733 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1734 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1735 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1736 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1737 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1738 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001739]
1740
Benoit Jacoba9644732020-08-13 12:48:55 -07001741NEONDOT_UKERNELS = [
1742 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1743 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1744 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001745 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1746 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1747 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001748]
1749
Marat Dukhan08c4a432019-10-03 09:29:21 -07001750SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001751 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1752 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001753 "src/f32-clamp/gen/sse-x4.c",
1754 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001755 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1756 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001757 "src/f32-dwconv-chw/3x3p1-sse.c",
1758 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001759 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1760 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1761 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1762 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1763 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1764 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1765 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1766 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1767 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1768 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1769 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1770 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001771 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001772 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1773 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001774 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1775 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1776 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1777 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1778 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1779 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1780 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1781 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1782 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1783 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1784 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1785 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1786 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001787 "src/f32-hswish/gen/hswish-sse-x4.c",
1788 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001789 "src/f32-ibilinear/gen/sse-c4.c",
1790 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001791 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1792 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1793 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1794 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1795 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1796 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1797 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001798 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1799 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1800 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001801 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001802 "src/f32-prelu/gen/sse-2x4.c",
1803 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001804 "src/f32-relu/gen/sse-x4.c",
1805 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001806 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001807 "src/f32-spmm/gen/4x1-minmax-sse.c",
1808 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001809 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001810 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1811 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1812 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1813 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1814 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1815 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1816 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1817 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001818 "src/f32-vbinary/gen/vmax-sse-x4.c",
1819 "src/f32-vbinary/gen/vmax-sse-x8.c",
1820 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1821 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1822 "src/f32-vbinary/gen/vmin-sse-x4.c",
1823 "src/f32-vbinary/gen/vmin-sse-x8.c",
1824 "src/f32-vbinary/gen/vminc-sse-x4.c",
1825 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001826 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1829 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1830 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1832 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001834 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1835 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1836 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1837 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001842 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1843 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001844 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1845 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001846 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1847 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001848 "src/f32-vunary/gen/vabs-sse-x4.c",
1849 "src/f32-vunary/gen/vabs-sse-x8.c",
1850 "src/f32-vunary/gen/vneg-sse-x4.c",
1851 "src/f32-vunary/gen/vneg-sse-x8.c",
1852 "src/f32-vunary/gen/vsqr-sse-x4.c",
1853 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001854 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001856 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001857 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001858 "src/math/roundd-sse-addsub.c",
1859 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001860 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001861 "src/math/sqrt-sse-hh1mac.c",
1862 "src/math/sqrt-sse-nr1mac.c",
1863 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864]
1865
1866SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001867 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1868 "src/f32-argmaxpool/4x-sse2-c4.c",
1869 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001870 "src/f32-prelu/gen/sse2-2x4.c",
1871 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001872 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1873 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1874 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1875 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1876 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1877 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1878 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1879 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1880 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1881 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1882 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1883 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001884 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001885 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001886 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001887 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001888 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1889 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001890 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1891 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001892 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1893 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1894 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1895 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1896 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1897 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1898 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1899 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001900 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1901 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1902 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001903 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1904 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1905 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001906 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1907 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1908 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001909 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1910 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001911 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1912 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001913 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1914 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001915 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1916 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001917 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001918 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1919 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001920 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001921 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1922 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1923 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001924 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1925 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001926 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1927 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001928 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1929 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001930 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001931 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1932 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001933 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001934 "src/qs8-requantization/precise-sse2.c",
1935 "src/qs8-requantization/fp32-sse2.c",
1936 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001937 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1938 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1939 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1940 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001941 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1942 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1943 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1944 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001945 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1946 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1947 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1948 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1949 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1950 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1951 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1952 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001953 "src/qu8-requantization/precise-sse2.c",
1954 "src/qu8-requantization/fp32-sse2.c",
1955 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001956 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001957 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001958 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001960 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001961 "src/x32-zip/x2-sse2.c",
1962 "src/x32-zip/x3-sse2.c",
1963 "src/x32-zip/x4-sse2.c",
1964 "src/x32-zip/xm-sse2.c",
1965 "src/x8-zip/x2-sse2.c",
1966 "src/x8-zip/x3-sse2.c",
1967 "src/x8-zip/x4-sse2.c",
1968 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001969 "src/math/exp-sse2-p5.c",
1970 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001971 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001972 "src/math/roundd-sse2-cvt.c",
1973 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001974 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001975 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001976]
1977
1978SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001979 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1980 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1981 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001982 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1983 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1984 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001985 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1986 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1987 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001988 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1989 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001990 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1991 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001992 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1993 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001994 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1995 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001996 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001997 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1998 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001999 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002000 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2001 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2002 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002003 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2004 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002005 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2006 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002007 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2008 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002009 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002010 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2011 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002012 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002013 "src/qs8-requantization/precise-ssse3.c",
2014 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002015 "src/qu8-requantization/precise-ssse3.c",
2016 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002017]
2018
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002019SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002020 "src/f32-prelu/gen/sse41-2x4.c",
2021 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002022 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002023 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002024 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002025 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002026 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2027 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002028 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2029 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002030 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2031 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2032 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2033 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2034 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2035 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2036 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2037 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002038 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2039 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2040 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002041 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2042 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2043 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002044 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2045 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2046 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002047 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2048 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002049 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2050 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002051 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2052 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002053 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2054 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002055 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002056 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2057 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002058 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002059 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2060 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2061 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002062 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2063 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002064 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2065 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002066 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2067 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002068 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002069 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2070 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002071 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002072 "src/qs8-requantization/fp32-sse4.c",
2073 "src/qs8-requantization/precise-sse4.c",
2074 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002075 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2076 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2077 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2078 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002079 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2080 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2081 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2082 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002083 "src/qu8-requantization/precise-sse4.c",
2084 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002085 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002086 "src/math/roundd-sse41.c",
2087 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002088 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002089]
2090
Marat Dukhan08c4a432019-10-03 09:29:21 -07002091AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002092 "src/f32-clamp/gen/avx-x8.c",
2093 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002094 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2095 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2096 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2097 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2098 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2099 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2100 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2101 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2102 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2103 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2104 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2105 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2106 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2107 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2108 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2109 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2110 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2111 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2112 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2113 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2114 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2115 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2116 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2117 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2118 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2119 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2120 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2121 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2122 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2123 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002124 "src/f32-hswish/gen/hswish-avx-x8.c",
2125 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002126 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2127 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2128 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2129 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2130 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2131 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2132 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2133 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2134 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002135 "src/f32-prelu/gen/avx-2x8.c",
2136 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002137 "src/f32-relu/gen/avx-x16.c",
2138 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002140 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2141 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2142 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2143 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2144 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2145 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2146 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2147 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002148 "src/f32-vbinary/gen/vmax-avx-x8.c",
2149 "src/f32-vbinary/gen/vmax-avx-x16.c",
2150 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2151 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2152 "src/f32-vbinary/gen/vmin-avx-x8.c",
2153 "src/f32-vbinary/gen/vmin-avx-x16.c",
2154 "src/f32-vbinary/gen/vminc-avx-x8.c",
2155 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002156 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2157 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2158 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2159 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2160 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2161 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2162 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2163 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002164 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2165 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2166 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2167 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002168 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2169 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2170 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2171 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002172 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2173 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002174 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002175 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2176 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2177 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2178 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2179 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2180 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2181 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2182 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002183 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2184 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002185 "src/f32-vunary/gen/vabs-avx-x8.c",
2186 "src/f32-vunary/gen/vabs-avx-x16.c",
2187 "src/f32-vunary/gen/vneg-avx-x8.c",
2188 "src/f32-vunary/gen/vneg-avx-x16.c",
2189 "src/f32-vunary/gen/vsqr-avx-x8.c",
2190 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002191]
2192
Marat Dukhan1566fee2020-08-02 21:55:41 -07002193XOP_UKERNELS = [
2194 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2195 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2196 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2197 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002198 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2199 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002200 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2201 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002202 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002203 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2204 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002205 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002206 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2207 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2208 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002209 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2210 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2211 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2212 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2213 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2214 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002215 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002216 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2217 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002218 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002219]
2220
Marat Dukhanfda12b82019-11-21 12:27:59 -08002221FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002222 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2223 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2224 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2225 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2226 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2227 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2228 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2229 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2230 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2231 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2232 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2233 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2234 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2235 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2236 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2237 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2238 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2239 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2240 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2241 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2242 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2243 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2244 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2245 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2246 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2247 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2248 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2249 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2250 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2251 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2252 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2253 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2254 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2255 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2256 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2257 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2258 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2259 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2260 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2261 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002262 "src/f32-hswish/gen/hswish-fma3-x8.c",
2263 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002264 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2265 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2266 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2267 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2268 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2269 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2270 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2271 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2272 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2273 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2274 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2275 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2276 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2277 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002278 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2279 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2280 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2281 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2282 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2283 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2284 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2285 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002286 "src/math/sqrt-fma3-nr1fma.c",
2287 "src/math/sqrt-fma3-nr2fma.c",
2288 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002289]
2290
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002291AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002292 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2293 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2294 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2295 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2296 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2297 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2298 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2299 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2300 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2301 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2302 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2303 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2304 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2305 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2306 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2307 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2308 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2309 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2310 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2311 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2312 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2313 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2314 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2315 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2316 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2317 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2318 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2319 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2320 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2321 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2322 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2323 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2324 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2325 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2326 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2327 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002328 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2329 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2330 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2331 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2332 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2333 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2334 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2335 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2336 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2337 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2338 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2339 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2340 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2341 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2342 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2343 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2344 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2345 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2350 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2351 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2352 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2353 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2354 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2355 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2356 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2357 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002358 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2359 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2360 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2361 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2362 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2363 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2364 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2365 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2366 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2367 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2368 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2369 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2370 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2371 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2372 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2373 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2374 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2375 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2376 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2377 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2378 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2379 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2380 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2381 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002382 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2383 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002384 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2385 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2386 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2387 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002388 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2389 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2390 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002391 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2392 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2393 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002394 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2395 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2396 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002397 "src/math/exp-avx2-p5.c",
2398 "src/math/exp-avx2-perm-p3.c",
2399 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002400 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002401 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002402 "src/math/sigmoid-avx2-rr2-p5-div.c",
2403 "src/math/sigmoid-avx2-rr1-p5-div.c",
2404 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2405 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2406 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2407 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002408]
2409
Marat Dukhan08c4a432019-10-03 09:29:21 -07002410AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002411 "src/f32-clamp/gen/avx512f-x16.c",
2412 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002413 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2414 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2415 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2416 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2417 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2418 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2419 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2420 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2421 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2422 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2423 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2424 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2425 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2426 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2427 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2428 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2429 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2430 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2431 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2432 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2433 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2434 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2435 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2436 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002437 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2438 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002439 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2440 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2441 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2442 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2443 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2444 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002445 "src/f32-prelu/gen/avx512f-2x16.c",
2446 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002447 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2448 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2449 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2450 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2451 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2452 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2453 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2454 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2455 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2456 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2457 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2458 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2459 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2460 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2461 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2462 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2463 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2464 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2465 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2466 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2467 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2468 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2469 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2470 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2471 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2472 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2473 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2474 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2475 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2476 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2477 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2478 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2479 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2480 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2481 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2482 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002483 "src/f32-relu/gen/avx512f-x16.c",
2484 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002485 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002486 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2487 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2488 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2489 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2490 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2491 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2492 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2493 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002494 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2495 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2496 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2497 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2498 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2499 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2500 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2501 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002502 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2503 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2504 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2505 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2506 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2507 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2508 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2509 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002510 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2511 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2512 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2513 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002514 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2515 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2516 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2517 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002518 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2519 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002520 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002521 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2522 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2523 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2524 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2525 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2526 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2527 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2528 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2529 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2530 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2531 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2532 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2533 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2534 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2535 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2536 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2537 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2538 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2539 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2540 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2541 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2542 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2543 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2544 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002545 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2546 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2547 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2548 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2549 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2550 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2551 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2552 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002553 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2554 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2555 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2556 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2557 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2558 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2559 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2560 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002561 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2562 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2563 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2564 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2565 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2566 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002567 "src/math/exp-avx512f-p5-scalef.c",
2568 "src/math/exp-avx512f-p5.c",
2569 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002570 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002571 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002572 "src/math/sqrt-avx512f-nr1fma.c",
2573 "src/math/sqrt-avx512f-nr2fma.c",
2574 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002575]
2576
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002577AVX512SKX_UKERNELS = [
2578 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2579 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2580 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2581 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2582 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2583 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2584 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2585 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2586]
2587
Frank Barchardbcedc082020-08-17 18:00:51 -07002588WASM32_ASM_UKERNELS = [
2589 "src/f32-relu/wasm_shr_x1.S",
2590 "src/f32-relu/wasm_shr_x2.S",
2591 "src/f32-relu/wasm_shr_x4.S",
2592]
2593
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002595 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002596 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002597 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2598 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002599 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002600 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002601 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002602 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002603 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2604 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002605 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2606 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2607 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2608 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609]
2610
2611AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002612 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002613 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2614 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002615 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002616 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2617 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002618 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2619 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2620 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002621 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002622 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2623 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2624 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002625 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002626 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2627 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002628 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002629 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2630 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2631 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2632 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2633 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2634 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2635 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2636 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2637 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2638 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2639 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2640 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2641 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2642 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2643 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2644 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2645 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2646 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002647 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2648 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002649 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002650 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2651 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2652 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2653 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2654 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2655 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2656 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2657 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2658 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2659 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2660 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2661 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2662 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2663 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2664 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2665 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2666 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2667 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002668 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2669 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2670 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2671 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2672 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2673 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2674 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2675 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2676 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2677 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2678 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2679 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2680 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2681 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2682 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2683 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2684 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2685 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686]
2687
Marat Dukhan1b354632020-03-23 12:50:22 -07002688INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002689 "src/xnnpack/argmaxpool.h",
2690 "src/xnnpack/avgpool.h",
2691 "src/xnnpack/clamp.h",
2692 "src/xnnpack/common.h",
2693 "src/xnnpack/conv.h",
2694 "src/xnnpack/dwconv.h",
2695 "src/xnnpack/gavgpool.h",
2696 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002697 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002699 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002700 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002701 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002702 "src/xnnpack/lut.h",
2703 "src/xnnpack/math.h",
2704 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002705 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 "src/xnnpack/packx.h",
2707 "src/xnnpack/pad.h",
2708 "src/xnnpack/params.h",
2709 "src/xnnpack/pavgpool.h",
2710 "src/xnnpack/ppmm.h",
2711 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002712 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002713 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002714 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002715 "src/xnnpack/rmax.h",
2716 "src/xnnpack/scalar-utils.h",
2717 "src/xnnpack/spmm.h",
2718 "src/xnnpack/unpool.h",
2719 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002720 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002721 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002722 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002723 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002724 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002725 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002727]
2728
2729INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002730 "include/xnnpack.h",
2731 "src/xnnpack/allocator.h",
2732 "src/xnnpack/compute.h",
2733 "src/xnnpack/im2col.h",
2734 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002735 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002736 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 "src/xnnpack/operator.h",
2738 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002739 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002741 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002742 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002743]
2744
Marat Dukhan1b354632020-03-23 12:50:22 -07002745ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002746 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747]
2748
Marat Dukhan1b354632020-03-23 12:50:22 -07002749MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002750 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 "include/xnnpack.h",
2752]
2753
Marat Dukhan1b354632020-03-23 12:50:22 -07002754MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002756 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757 "src/xnnpack/requantization.h",
2758 "include/xnnpack.h",
2759]
2760
2761OPERATOR_TEST_PARAMS_HDRS = [
2762 "src/xnnpack/params.h",
2763 "src/xnnpack/common.h",
2764]
2765
2766WEIGHTS_PACK_HDRS = [
2767 "src/xnnpack/pack.h",
2768 "src/xnnpack/operator.h",
2769 "src/xnnpack/compute.h",
2770]
2771
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002772LOGGING_COPTS = select({
2773 # No logging in optimized mode
2774 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2775 # Full logging in debug mode
2776 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2777 # Error-only logging in default (fastbuild) mode
2778 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2779})
2780
Marat Dukhan3b59de22020-06-03 20:15:19 -07002781LOGGING_SRCS = select({
2782 # No logging in optimized mode
2783 ":optimized_build": [],
2784 "//conditions:default": [
2785 "src/operator-strings.c",
2786 "src/subgraph-strings.c",
2787 ],
2788})
2789
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002790LOGGING_HDRS = [
2791 "src/xnnpack/log.h",
2792]
2793
Marat Dukhan08c4a432019-10-03 09:29:21 -07002794xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002795 name = "tables",
2796 srcs = TABLE_SRCS,
2797 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002798 gcc_copts = xnnpack_gcc_std_copts(),
2799 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002800)
2801
2802xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002803 name = "scalar_ukernels",
2804 srcs = SCALAR_UKERNELS,
2805 hdrs = INTERNAL_HDRS,
2806 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002807 gcc_copts = xnnpack_gcc_std_copts(),
2808 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002809 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002810 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002811 "@FP16",
2812 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002813 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 ],
2815)
2816
2817xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002818 name = "scalar_ukernels_test_mode",
2819 srcs = SCALAR_UKERNELS,
2820 hdrs = INTERNAL_HDRS,
2821 aarch32_copts = ["-marm"],
2822 copts = [
2823 "-UNDEBUG",
2824 "-DXNN_TEST_MODE=1",
2825 ],
2826 gcc_copts = xnnpack_gcc_std_copts(),
2827 msvc_copts = xnnpack_msvc_std_copts(),
2828 deps = [
2829 ":tables",
2830 "@FP16",
2831 "@FXdiv",
2832 "@pthreadpool",
2833 ],
2834)
2835
2836xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002837 name = "wasm_ukernels",
2838 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002839 gcc_copts = xnnpack_gcc_std_copts(),
2840 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002841 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002842 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002843 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002844 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002845 "@FP16",
2846 "@FXdiv",
2847 "@pthreadpool",
2848 ],
2849)
2850
2851xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002852 name = "wasm_ukernels_test_mode",
2853 hdrs = INTERNAL_HDRS,
2854 copts = [
2855 "-UNDEBUG",
2856 "-DXNN_TEST_MODE=1",
2857 ],
2858 gcc_copts = xnnpack_gcc_std_copts(),
2859 msvc_copts = xnnpack_msvc_std_copts(),
2860 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002861 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002862 deps = [
2863 ":tables",
2864 "@FP16",
2865 "@FXdiv",
2866 "@pthreadpool",
2867 ],
2868)
2869
2870xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002871 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 hdrs = INTERNAL_HDRS,
2873 aarch32_copts = [
2874 "-marm",
2875 "-mfpu=neon",
2876 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002877 gcc_copts = xnnpack_gcc_std_copts(),
2878 gcc_x86_copts = ["-msse2"],
2879 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 optimized_copts = [
2881 "-O3",
2882 "-ffast-math",
2883 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002884 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002886 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 "@FP16",
2888 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002889 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 ],
2891)
2892
2893xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002894 name = "psimd_fastmath_ukernels_test_mode",
2895 hdrs = INTERNAL_HDRS,
2896 aarch32_copts = [
2897 "-marm",
2898 "-mfpu=neon",
2899 ],
2900 copts = [
2901 "-UNDEBUG",
2902 "-DXNN_TEST_MODE=1",
2903 ],
2904 gcc_copts = xnnpack_gcc_std_copts(),
2905 gcc_x86_copts = ["-msse2"],
2906 msvc_copts = xnnpack_msvc_std_copts(),
2907 optimized_copts = [
2908 "-O3",
2909 "-ffast-math",
2910 ],
2911 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2912 deps = [
2913 ":tables",
2914 "@FP16",
2915 "@psimd",
2916 "@pthreadpool",
2917 ],
2918)
2919
2920xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002921 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002922 hdrs = INTERNAL_HDRS,
2923 aarch32_copts = [
2924 "-marm",
2925 "-mfpu=neon",
2926 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002927 gcc_copts = xnnpack_gcc_std_copts(),
2928 gcc_x86_copts = ["-msse2"],
2929 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002930 optimized_copts = [
2931 "-O3",
2932 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002933 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002934 deps = [
2935 ":tables",
2936 "@FP16",
2937 "@psimd",
2938 "@pthreadpool",
2939 ],
2940)
2941
2942xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002943 name = "psimd_accmath_ukernels_test_mode",
2944 hdrs = INTERNAL_HDRS,
2945 aarch32_copts = [
2946 "-marm",
2947 "-mfpu=neon",
2948 ],
2949 copts = [
2950 "-UNDEBUG",
2951 "-DXNN_TEST_MODE=1",
2952 ],
2953 gcc_copts = xnnpack_gcc_std_copts(),
2954 gcc_x86_copts = ["-msse2"],
2955 msvc_copts = xnnpack_msvc_std_copts(),
2956 optimized_copts = [
2957 "-O3",
2958 ],
2959 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2960 deps = [
2961 ":tables",
2962 "@FP16",
2963 "@psimd",
2964 "@pthreadpool",
2965 ],
2966)
2967
2968xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969 name = "neon_ukernels",
2970 hdrs = INTERNAL_HDRS,
2971 aarch32_copts = [
2972 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002973 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002974 "-mfpu=neon",
2975 ],
2976 aarch32_srcs = NEON_UKERNELS,
2977 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002978 gcc_copts = xnnpack_gcc_std_copts(),
2979 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002980 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002981 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002982 "@FP16",
2983 "@pthreadpool",
2984 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985)
2986
2987xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002988 name = "neon_ukernels_test_mode",
2989 hdrs = INTERNAL_HDRS,
2990 aarch32_copts = [
2991 "-marm",
2992 "-march=armv7-a",
2993 "-mfpu=neon",
2994 ],
2995 aarch32_srcs = NEON_UKERNELS,
2996 aarch64_srcs = NEON_UKERNELS,
2997 copts = [
2998 "-UNDEBUG",
2999 "-DXNN_TEST_MODE=1",
3000 ],
3001 gcc_copts = xnnpack_gcc_std_copts(),
3002 msvc_copts = xnnpack_msvc_std_copts(),
3003 deps = [
3004 ":tables",
3005 "@FP16",
3006 "@pthreadpool",
3007 ],
3008)
3009
3010xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 name = "neonfma_ukernels",
3012 hdrs = INTERNAL_HDRS,
3013 aarch32_copts = [
3014 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003015 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 "-mfpu=neon-vfpv4",
3017 ],
3018 aarch32_srcs = NEONFMA_UKERNELS,
3019 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003020 apple_aarch32_copts = [
3021 "-mcpu=swift",
3022 "-mtune=generic",
3023 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003024 gcc_copts = xnnpack_gcc_std_copts(),
3025 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003026 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003027 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003028 "@FP16",
3029 "@pthreadpool",
3030 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031)
3032
3033xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003034 name = "neonfma_ukernels_test_mode",
3035 hdrs = INTERNAL_HDRS,
3036 aarch32_copts = [
3037 "-marm",
3038 "-march=armv7-a",
3039 "-mfpu=neon-vfpv4",
3040 ],
3041 aarch32_srcs = NEONFMA_UKERNELS,
3042 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003043 apple_aarch32_copts = [
3044 "-mcpu=swift",
3045 "-mtune=generic",
3046 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003047 copts = [
3048 "-UNDEBUG",
3049 "-DXNN_TEST_MODE=1",
3050 ],
3051 gcc_copts = xnnpack_gcc_std_copts(),
3052 msvc_copts = xnnpack_msvc_std_copts(),
3053 deps = [
3054 ":tables",
3055 "@FP16",
3056 "@pthreadpool",
3057 ],
3058)
3059
3060xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003061 name = "neonv8_ukernels",
3062 hdrs = INTERNAL_HDRS,
3063 aarch32_copts = [
3064 "-marm",
3065 "-march=armv8-a",
3066 "-mfpu=neon-fp-armv8",
3067 ],
3068 aarch32_srcs = NEONV8_UKERNELS,
3069 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003070 apple_aarch32_copts = [
3071 "-mcpu=cyclone",
3072 "-mtune=generic",
3073 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003074 gcc_copts = xnnpack_gcc_std_copts(),
3075 msvc_copts = xnnpack_msvc_std_copts(),
3076 deps = [
3077 ":tables",
3078 "@FP16",
3079 "@pthreadpool",
3080 ],
3081)
3082
3083xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003084 name = "neonv8_ukernels_test_mode",
3085 hdrs = INTERNAL_HDRS,
3086 aarch32_copts = [
3087 "-marm",
3088 "-march=armv8-a",
3089 "-mfpu=neon-fp-armv8",
3090 ],
3091 aarch32_srcs = NEONV8_UKERNELS,
3092 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003093 apple_aarch32_copts = [
3094 "-mcpu=cyclone",
3095 "-mtune=generic",
3096 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003097 copts = [
3098 "-UNDEBUG",
3099 "-DXNN_TEST_MODE=1",
3100 ],
3101 gcc_copts = xnnpack_gcc_std_copts(),
3102 msvc_copts = xnnpack_msvc_std_copts(),
3103 deps = [
3104 ":tables",
3105 "@FP16",
3106 "@pthreadpool",
3107 ],
3108)
3109
3110xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 name = "neonfp16arith_ukernels",
3112 hdrs = INTERNAL_HDRS,
3113 aarch64_copts = ["-march=armv8.2-a+fp16"],
3114 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003115 gcc_copts = xnnpack_gcc_std_copts(),
3116 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003117 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003118 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003119 "@FP16",
3120 "@pthreadpool",
3121 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122)
3123
3124xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003125 name = "neonfp16arith_ukernels_test_mode",
3126 hdrs = INTERNAL_HDRS,
3127 aarch64_copts = ["-march=armv8.2-a+fp16"],
3128 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3129 copts = [
3130 "-UNDEBUG",
3131 "-DXNN_TEST_MODE=1",
3132 ],
3133 gcc_copts = xnnpack_gcc_std_copts(),
3134 msvc_copts = xnnpack_msvc_std_copts(),
3135 deps = [
3136 ":tables",
3137 "@FP16",
3138 "@pthreadpool",
3139 ],
3140)
3141
3142xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003143 name = "neondot_ukernels",
3144 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003145 aarch32_copts = [
3146 "-marm",
3147 "-march=armv8.2-a+dotprod",
3148 "-mfpu=neon-fp-armv8",
3149 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003150 aarch32_srcs = NEONDOT_UKERNELS,
3151 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3152 aarch64_srcs = NEONDOT_UKERNELS,
3153 gcc_copts = xnnpack_gcc_std_copts(),
3154 msvc_copts = xnnpack_msvc_std_copts(),
3155 deps = [
3156 ":tables",
3157 "@FP16",
3158 "@pthreadpool",
3159 ],
3160)
3161
3162xnnpack_cc_library(
3163 name = "neondot_ukernels_test_mode",
3164 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003165 aarch32_copts = [
3166 "-marm",
3167 "-march=armv8.2-a+dotprod",
3168 "-mfpu=neon-fp-armv8",
3169 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003170 aarch32_srcs = NEONDOT_UKERNELS,
3171 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3172 aarch64_srcs = NEONDOT_UKERNELS,
3173 copts = [
3174 "-UNDEBUG",
3175 "-DXNN_TEST_MODE=1",
3176 ],
3177 gcc_copts = xnnpack_gcc_std_copts(),
3178 msvc_copts = xnnpack_msvc_std_copts(),
3179 deps = [
3180 ":tables",
3181 "@FP16",
3182 "@pthreadpool",
3183 ],
3184)
3185
3186xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 name = "sse2_ukernels",
3188 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003189 gcc_copts = xnnpack_gcc_std_copts(),
3190 gcc_x86_copts = ["-msse2"],
3191 msvc_copts = xnnpack_msvc_std_copts(),
3192 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003194 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003195 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003196 "@FP16",
3197 "@pthreadpool",
3198 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199)
3200
3201xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003202 name = "sse2_ukernels_test_mode",
3203 hdrs = INTERNAL_HDRS,
3204 copts = [
3205 "-UNDEBUG",
3206 "-DXNN_TEST_MODE=1",
3207 ],
3208 gcc_copts = xnnpack_gcc_std_copts(),
3209 gcc_x86_copts = ["-msse2"],
3210 msvc_copts = xnnpack_msvc_std_copts(),
3211 msvc_x86_32_copts = ["/arch:SSE2"],
3212 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3213 deps = [
3214 ":tables",
3215 "@FP16",
3216 "@pthreadpool",
3217 ],
3218)
3219
3220xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003221 name = "ssse3_ukernels",
3222 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003223 gcc_copts = xnnpack_gcc_std_copts(),
3224 gcc_x86_copts = ["-mssse3"],
3225 msvc_copts = xnnpack_msvc_std_copts(),
3226 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003227 x86_srcs = SSSE3_UKERNELS,
3228 deps = [
3229 ":tables",
3230 "@FP16",
3231 "@pthreadpool",
3232 ],
3233)
3234
3235xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003236 name = "ssse3_ukernels_test_mode",
3237 hdrs = INTERNAL_HDRS,
3238 copts = [
3239 "-UNDEBUG",
3240 "-DXNN_TEST_MODE=1",
3241 ],
3242 gcc_copts = xnnpack_gcc_std_copts(),
3243 gcc_x86_copts = ["-mssse3"],
3244 msvc_copts = xnnpack_msvc_std_copts(),
3245 msvc_x86_32_copts = ["/arch:SSE2"],
3246 x86_srcs = SSSE3_UKERNELS,
3247 deps = [
3248 ":tables",
3249 "@FP16",
3250 "@pthreadpool",
3251 ],
3252)
3253
3254xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003255 name = "sse41_ukernels",
3256 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003257 gcc_copts = xnnpack_gcc_std_copts(),
3258 gcc_x86_copts = ["-msse4.1"],
3259 msvc_copts = xnnpack_msvc_std_copts(),
3260 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003261 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003262 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003263 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003264 "@FP16",
3265 "@pthreadpool",
3266 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003267)
3268
3269xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003270 name = "sse41_ukernels_test_mode",
3271 hdrs = INTERNAL_HDRS,
3272 copts = [
3273 "-UNDEBUG",
3274 "-DXNN_TEST_MODE=1",
3275 ],
3276 gcc_copts = xnnpack_gcc_std_copts(),
3277 gcc_x86_copts = ["-msse4.1"],
3278 msvc_copts = xnnpack_msvc_std_copts(),
3279 msvc_x86_32_copts = ["/arch:SSE2"],
3280 x86_srcs = SSE41_UKERNELS,
3281 deps = [
3282 ":tables",
3283 "@FP16",
3284 "@pthreadpool",
3285 ],
3286)
3287
3288xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003289 name = "avx_ukernels",
3290 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003291 gcc_copts = xnnpack_gcc_std_copts(),
3292 gcc_x86_copts = ["-mavx"],
3293 msvc_copts = xnnpack_msvc_std_copts(),
3294 msvc_x86_32_copts = ["/arch:AVX"],
3295 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003297 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003298 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003299 "@FP16",
3300 "@pthreadpool",
3301 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302)
3303
3304xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003305 name = "avx_ukernels_test_mode",
3306 hdrs = INTERNAL_HDRS,
3307 copts = [
3308 "-UNDEBUG",
3309 "-DXNN_TEST_MODE=1",
3310 ],
3311 gcc_copts = xnnpack_gcc_std_copts(),
3312 gcc_x86_copts = ["-mavx"],
3313 msvc_copts = xnnpack_msvc_std_copts(),
3314 msvc_x86_32_copts = ["/arch:AVX"],
3315 msvc_x86_64_copts = ["/arch:AVX"],
3316 x86_srcs = AVX_UKERNELS,
3317 deps = [
3318 ":tables",
3319 "@FP16",
3320 "@pthreadpool",
3321 ],
3322)
3323
3324xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003325 name = "xop_ukernels",
3326 hdrs = INTERNAL_HDRS,
3327 gcc_copts = xnnpack_gcc_std_copts(),
3328 gcc_x86_copts = ["-mxop"],
3329 msvc_copts = xnnpack_msvc_std_copts(),
3330 msvc_x86_32_copts = ["/arch:AVX"],
3331 msvc_x86_64_copts = ["/arch:AVX"],
3332 x86_srcs = XOP_UKERNELS,
3333 deps = [
3334 ":tables",
3335 "@FP16",
3336 "@pthreadpool",
3337 ],
3338)
3339
3340xnnpack_cc_library(
3341 name = "xop_ukernels_test_mode",
3342 hdrs = INTERNAL_HDRS,
3343 copts = [
3344 "-UNDEBUG",
3345 "-DXNN_TEST_MODE=1",
3346 ],
3347 gcc_copts = xnnpack_gcc_std_copts(),
3348 gcc_x86_copts = ["-mxop"],
3349 msvc_copts = xnnpack_msvc_std_copts(),
3350 msvc_x86_32_copts = ["/arch:AVX"],
3351 msvc_x86_64_copts = ["/arch:AVX"],
3352 x86_srcs = XOP_UKERNELS,
3353 deps = [
3354 ":tables",
3355 "@FP16",
3356 "@pthreadpool",
3357 ],
3358)
3359
3360xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003361 name = "fma3_ukernels",
3362 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003363 gcc_copts = xnnpack_gcc_std_copts(),
3364 gcc_x86_copts = ["-mfma"],
3365 msvc_copts = xnnpack_msvc_std_copts(),
3366 msvc_x86_32_copts = ["/arch:AVX"],
3367 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003368 x86_srcs = FMA3_UKERNELS,
3369 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003370 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003371 "@FP16",
3372 "@pthreadpool",
3373 ],
3374)
3375
3376xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003377 name = "fma3_ukernels_test_mode",
3378 hdrs = INTERNAL_HDRS,
3379 copts = [
3380 "-UNDEBUG",
3381 "-DXNN_TEST_MODE=1",
3382 ],
3383 gcc_copts = xnnpack_gcc_std_copts(),
3384 gcc_x86_copts = ["-mfma"],
3385 msvc_copts = xnnpack_msvc_std_copts(),
3386 msvc_x86_32_copts = ["/arch:AVX"],
3387 msvc_x86_64_copts = ["/arch:AVX"],
3388 x86_srcs = FMA3_UKERNELS,
3389 deps = [
3390 ":tables",
3391 "@FP16",
3392 "@pthreadpool",
3393 ],
3394)
3395
3396xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003397 name = "avx2_ukernels",
3398 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003399 gcc_copts = xnnpack_gcc_std_copts(),
3400 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003401 "-mfma",
3402 "-mavx2",
3403 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003404 msvc_copts = xnnpack_msvc_std_copts(),
3405 msvc_x86_32_copts = ["/arch:AVX2"],
3406 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003407 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003408 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003409 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003410 "@FP16",
3411 "@pthreadpool",
3412 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003413)
3414
3415xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003416 name = "avx2_ukernels_test_mode",
3417 hdrs = INTERNAL_HDRS,
3418 copts = [
3419 "-UNDEBUG",
3420 "-DXNN_TEST_MODE=1",
3421 ],
3422 gcc_copts = xnnpack_gcc_std_copts(),
3423 gcc_x86_copts = [
3424 "-mfma",
3425 "-mavx2",
3426 ],
3427 msvc_copts = xnnpack_msvc_std_copts(),
3428 msvc_x86_32_copts = ["/arch:AVX2"],
3429 msvc_x86_64_copts = ["/arch:AVX2"],
3430 x86_srcs = AVX2_UKERNELS,
3431 deps = [
3432 ":tables",
3433 "@FP16",
3434 "@pthreadpool",
3435 ],
3436)
3437
3438xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439 name = "avx512f_ukernels",
3440 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003441 gcc_copts = xnnpack_gcc_std_copts(),
3442 gcc_x86_copts = ["-mavx512f"],
3443 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3444 msvc_copts = xnnpack_msvc_std_copts(),
3445 msvc_x86_32_copts = ["/arch:AVX512"],
3446 msvc_x86_64_copts = ["/arch:AVX512"],
3447 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003449 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003450 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003451 "@FP16",
3452 "@pthreadpool",
3453 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003454)
3455
3456xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003457 name = "avx512f_ukernels_test_mode",
3458 hdrs = INTERNAL_HDRS,
3459 copts = [
3460 "-UNDEBUG",
3461 "-DXNN_TEST_MODE=1",
3462 ],
3463 gcc_copts = xnnpack_gcc_std_copts(),
3464 gcc_x86_copts = ["-mavx512f"],
3465 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3466 msvc_copts = xnnpack_msvc_std_copts(),
3467 msvc_x86_32_copts = ["/arch:AVX512"],
3468 msvc_x86_64_copts = ["/arch:AVX512"],
3469 msys_copts = ["-fno-asynchronous-unwind-tables"],
3470 x86_srcs = AVX512F_UKERNELS,
3471 deps = [
3472 ":tables",
3473 "@FP16",
3474 "@pthreadpool",
3475 ],
3476)
3477
3478xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003479 name = "avx512skx_ukernels",
3480 hdrs = INTERNAL_HDRS,
3481 gcc_copts = xnnpack_gcc_std_copts(),
3482 gcc_x86_copts = [
3483 "-mavx512f",
3484 "-mavx512cd",
3485 "-mavx512bw",
3486 "-mavx512dq",
3487 "-mavx512vl",
3488 ],
3489 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3490 msvc_copts = xnnpack_msvc_std_copts(),
3491 msvc_x86_32_copts = ["/arch:AVX512"],
3492 msvc_x86_64_copts = ["/arch:AVX512"],
3493 msys_copts = ["-fno-asynchronous-unwind-tables"],
3494 x86_srcs = AVX512SKX_UKERNELS,
3495 deps = [
3496 ":tables",
3497 "@FP16",
3498 "@pthreadpool",
3499 ],
3500)
3501
3502xnnpack_cc_library(
3503 name = "avx512skx_ukernels_test_mode",
3504 hdrs = INTERNAL_HDRS,
3505 copts = [
3506 "-UNDEBUG",
3507 "-DXNN_TEST_MODE=1",
3508 ],
3509 gcc_copts = xnnpack_gcc_std_copts(),
3510 gcc_x86_copts = [
3511 "-mavx512f",
3512 "-mavx512cd",
3513 "-mavx512bw",
3514 "-mavx512dq",
3515 "-mavx512vl",
3516 ],
3517 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3518 msvc_copts = xnnpack_msvc_std_copts(),
3519 msvc_x86_32_copts = ["/arch:AVX512"],
3520 msvc_x86_64_copts = ["/arch:AVX512"],
3521 msys_copts = ["-fno-asynchronous-unwind-tables"],
3522 x86_srcs = AVX512SKX_UKERNELS,
3523 deps = [
3524 ":tables",
3525 "@FP16",
3526 "@pthreadpool",
3527 ],
3528)
3529
3530xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003531 name = "asm_ukernels",
3532 hdrs = ["src/xnnpack/assembly.h"],
3533 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003534 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003535 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003536 wasm_srcs = WASM32_ASM_UKERNELS,
3537 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538)
3539
Marat Dukhan3b59de22020-06-03 20:15:19 -07003540xnnpack_cc_library(
3541 name = "logging_utils",
3542 srcs = LOGGING_SRCS,
3543 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3544 copts = LOGGING_COPTS + [
3545 "-Isrc",
3546 "-Iinclude",
3547 ] + select({
3548 ":debug_build": [],
3549 "//conditions:default": xnnpack_min_size_copts(),
3550 }),
3551 gcc_copts = xnnpack_gcc_std_copts(),
3552 msvc_copts = xnnpack_msvc_std_copts(),
3553 visibility = xnnpack_visibility(),
3554 deps = [
3555 "@FP16",
3556 "@clog",
3557 "@pthreadpool",
3558 ],
3559)
3560
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561xnnpack_aggregate_library(
3562 name = "ukernels",
3563 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 ":neon_ukernels",
3565 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003566 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003567 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 ":asm_ukernels",
3569 ],
3570 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571 ":neon_ukernels",
3572 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003573 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003575 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 ":asm_ukernels",
3577 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003578 generic_deps = [
3579 ":scalar_ukernels",
3580 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003581 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003582 ":psimd_fastmath_ukernels",
3583 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003585 wasm_deps = [
3586 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003587 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003588 ],
3589 wasmsimd_deps = [
3590 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003591 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003595 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003596 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003598 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003599 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003600 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003602 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 ],
3604)
3605
Marat Dukhan33fcf782020-05-24 14:27:15 -07003606xnnpack_aggregate_library(
3607 name = "ukernels_test_mode",
3608 aarch32_deps = [
3609 ":neon_ukernels_test_mode",
3610 ":neonfma_ukernels_test_mode",
3611 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003612 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003613 ":asm_ukernels",
3614 ],
3615 aarch64_deps = [
3616 ":neon_ukernels_test_mode",
3617 ":neonfma_ukernels_test_mode",
3618 ":neonv8_ukernels_test_mode",
3619 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003620 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003621 ":asm_ukernels",
3622 ],
3623 generic_deps = [
3624 ":scalar_ukernels_test_mode",
3625 ],
3626 psimd_deps = [
3627 ":psimd_fastmath_ukernels_test_mode",
3628 ":psimd_accmath_ukernels_test_mode",
3629 ],
3630 wasm_deps = [
3631 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003632 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003633 ],
3634 wasmsimd_deps = [
3635 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003636 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003637 ],
3638 x86_deps = [
3639 ":sse2_ukernels_test_mode",
3640 ":ssse3_ukernels_test_mode",
3641 ":sse41_ukernels_test_mode",
3642 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003643 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003644 ":fma3_ukernels_test_mode",
3645 ":avx2_ukernels_test_mode",
3646 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003647 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003648 ],
3649)
3650
Marat Dukhan86df3842020-06-17 14:03:31 -07003651# This target must be compatible with Android NDK r17 and compile with Clang 6.
3652# It is similar to :ukernels target, but excludes microkernels which require
3653# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003654xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003655 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003656 aarch32_deps = [
3657 ":neon_ukernels",
3658 ":neonfma_ukernels",
3659 ":neonv8_ukernels",
3660 ":asm_ukernels",
3661 ],
3662 aarch64_deps = [
3663 ":neon_ukernels",
3664 ":neonfma_ukernels",
3665 ":neonv8_ukernels",
3666 ":asm_ukernels",
3667 ],
3668 generic_deps = [
3669 ":scalar_ukernels",
3670 ],
3671 psimd_deps = [
3672 ":psimd_fastmath_ukernels",
3673 ":psimd_accmath_ukernels",
3674 ],
3675 wasm_deps = [
3676 ":wasm_ukernels",
3677 ],
3678 wasmsimd_deps = [
3679 ":wasm_ukernels",
3680 ],
3681 x86_deps = [
3682 ":sse2_ukernels",
3683 ":ssse3_ukernels",
3684 ":sse41_ukernels",
3685 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003686 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003687 ":fma3_ukernels",
3688 ":avx2_ukernels",
3689 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003690 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003691 ],
3692)
3693
Marat Dukhan08c4a432019-10-03 09:29:21 -07003694xnnpack_cc_library(
3695 name = "im2col",
3696 srcs = ["src/im2col.c"],
3697 hdrs = [
3698 "src/xnnpack/common.h",
3699 "src/xnnpack/im2col.h",
3700 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003701 gcc_copts = xnnpack_gcc_std_copts(),
3702 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703)
3704
3705xnnpack_cc_library(
3706 name = "indirection",
3707 srcs = ["src/indirection.c"],
3708 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003709 gcc_copts = xnnpack_gcc_std_copts(),
3710 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 deps = [
3712 "@FP16",
3713 "@FXdiv",
3714 "@pthreadpool",
3715 ],
3716)
3717
3718xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003719 name = "indirection_test_mode",
3720 srcs = ["src/indirection.c"],
3721 hdrs = INTERNAL_HDRS,
3722 copts = [
3723 "-UNDEBUG",
3724 "-DXNN_TEST_MODE=1",
3725 ],
3726 gcc_copts = xnnpack_gcc_std_copts(),
3727 msvc_copts = xnnpack_msvc_std_copts(),
3728 deps = [
3729 "@FP16",
3730 "@FXdiv",
3731 "@pthreadpool",
3732 ],
3733)
3734
3735xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003736 name = "packing",
3737 srcs = ["src/packing.c"],
3738 hdrs = INTERNAL_HDRS,
3739 gcc_copts = xnnpack_gcc_std_copts(),
3740 msvc_copts = xnnpack_msvc_std_copts(),
3741 deps = [
3742 "@FP16",
3743 "@FXdiv",
3744 "@pthreadpool",
3745 ],
3746)
3747
3748xnnpack_cc_library(
3749 name = "packing_test_mode",
3750 srcs = ["src/packing.c"],
3751 hdrs = INTERNAL_HDRS,
3752 copts = [
3753 "-UNDEBUG",
3754 "-DXNN_TEST_MODE=1",
3755 ],
3756 gcc_copts = xnnpack_gcc_std_copts(),
3757 msvc_copts = xnnpack_msvc_std_copts(),
3758 deps = [
3759 "@FP16",
3760 "@FXdiv",
3761 "@pthreadpool",
3762 ],
3763)
3764
3765xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766 name = "operator_run",
3767 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003768 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003769 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003770 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3771 "//conditions:default": [],
3772 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003773 gcc_copts = xnnpack_gcc_std_copts(),
3774 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003776 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003777 "@FP16",
3778 "@FXdiv",
3779 "@clog",
3780 "@pthreadpool",
3781 ],
3782)
3783
Chao Mei6ddfc602020-05-13 22:29:36 -07003784xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003785 name = "operator_run_test_mode",
3786 srcs = ["src/operator-run.c"],
3787 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3788 copts = LOGGING_COPTS + [
3789 "-UNDEBUG",
3790 "-DXNN_TEST_MODE=1",
3791 ] + select({
3792 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3793 "//conditions:default": [],
3794 }),
3795 gcc_copts = xnnpack_gcc_std_copts(),
3796 msvc_copts = xnnpack_msvc_std_copts(),
3797 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003798 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003799 "@FP16",
3800 "@FXdiv",
3801 "@clog",
3802 "@pthreadpool",
3803 ],
3804)
3805
3806xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003807 name = "memory_planner",
3808 srcs = ["src/memory-planner.c"],
3809 hdrs = INTERNAL_HDRS,
3810 defines = select({
3811 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3812 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3813 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3814 }),
3815 gcc_copts = xnnpack_gcc_std_copts(),
3816 msvc_copts = xnnpack_msvc_std_copts(),
3817 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003818 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003819 "@pthreadpool",
3820 ],
3821)
3822
Marat Dukhan33fcf782020-05-24 14:27:15 -07003823xnnpack_cc_library(
3824 name = "memory_planner_test_mode",
3825 srcs = ["src/memory-planner.c"],
3826 hdrs = INTERNAL_HDRS,
3827 copts = [
3828 "-UNDEBUG",
3829 "-DXNN_TEST_MODE=1",
3830 ],
3831 defines = select({
3832 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3833 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3834 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3835 }),
3836 gcc_copts = xnnpack_gcc_std_copts(),
3837 msvc_copts = xnnpack_msvc_std_copts(),
3838 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003839 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003840 "@pthreadpool",
3841 ],
3842)
3843
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844cc_library(
3845 name = "enable_assembly",
3846 defines = select({
3847 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3848 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003849 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003850 }),
3851)
3852
Marat Dukhan9de90e02020-06-18 16:04:12 -07003853cc_library(
3854 name = "enable_sparse",
3855 defines = select({
3856 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3857 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003858 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003859 }),
3860)
3861
Marat Dukhancf056b22019-10-07 10:26:29 -07003862xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863 name = "operators",
3864 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003865 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003867 ],
3868 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003869 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870 "-Isrc",
3871 "-Iinclude",
3872 ] + select({
3873 ":debug_build": [],
3874 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003875 }) + select({
3876 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3877 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003878 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003879 gcc_copts = xnnpack_gcc_std_copts(),
3880 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003881 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003882 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003883 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003884 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003885 "@FP16",
3886 "@FXdiv",
3887 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003888 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003889 ],
3890)
3891
Marat Dukhan10a38082020-04-17 03:58:35 -07003892xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003893 name = "operators_test_mode",
3894 srcs = OPERATOR_SRCS + [
3895 "src/memory.c",
3896 "src/operator-delete.c",
3897 ],
3898 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3899 copts = LOGGING_COPTS + [
3900 "-Isrc",
3901 "-Iinclude",
3902 "-UNDEBUG",
3903 "-DXNN_TEST_MODE=1",
3904 ] + select({
3905 ":debug_build": [],
3906 "//conditions:default": xnnpack_min_size_copts(),
3907 }) + select({
3908 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3909 "//conditions:default": [],
3910 }),
3911 gcc_copts = xnnpack_gcc_std_copts(),
3912 msvc_copts = xnnpack_msvc_std_copts(),
3913 deps = [
3914 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003915 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003916 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003917 "@FP16",
3918 "@FXdiv",
3919 "@clog",
3920 "@pthreadpool",
3921 ],
3922)
3923
3924xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003925 name = "XNNPACK",
3926 srcs = [
3927 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003928 "src/runtime.c",
3929 "src/subgraph.c",
3930 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003931 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003932 hdrs = ["include/xnnpack.h"],
3933 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003934 "-Isrc",
3935 "-Iinclude",
3936 ] + select({
3937 ":debug_build": [],
3938 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003939 }) + select({
3940 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3941 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003942 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003943 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003944 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003945 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003946 visibility = xnnpack_visibility(),
3947 deps = [
3948 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003949 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003950 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003951 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003952 ":operator_run",
3953 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003954 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003955 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003956 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003957 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003958 ] + select({
3959 ":emscripten": [],
3960 "//conditions:default": ["@cpuinfo"],
3961 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003962)
3963
Marat Dukhan10a38082020-04-17 03:58:35 -07003964xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003965 name = "XNNPACK_test_mode",
3966 srcs = [
3967 "src/init.c",
3968 "src/runtime.c",
3969 "src/subgraph.c",
3970 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003971 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003972 hdrs = ["include/xnnpack.h"],
3973 copts = LOGGING_COPTS + [
3974 "-Isrc",
3975 "-Iinclude",
3976 "-UNDEBUG",
3977 "-DXNN_TEST_MODE=1",
3978 ] + select({
3979 ":debug_build": [],
3980 "//conditions:default": xnnpack_min_size_copts(),
3981 }) + select({
3982 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3983 "//conditions:default": [],
3984 }),
3985 gcc_copts = xnnpack_gcc_std_copts(),
3986 includes = ["include"],
3987 msvc_copts = xnnpack_msvc_std_copts(),
3988 visibility = xnnpack_visibility(),
3989 deps = [
3990 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003991 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003992 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003993 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003994 ":operator_run_test_mode",
3995 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003996 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003997 "@clog",
3998 "@FP16",
3999 "@pthreadpool",
4000 ] + select({
4001 ":emscripten": [],
4002 "//conditions:default": ["@cpuinfo"],
4003 }),
4004)
4005
4006xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004007 name = "xnnpack_f32",
4008 srcs = [
4009 "src/init.c",
4010 "src/runtime.c",
4011 "src/subgraph.c",
4012 "src/tensor.c",
4013 ] + SUBGRAPH_SRCS,
4014 hdrs = ["include/xnnpack.h"],
4015 copts = LOGGING_COPTS + [
4016 "-Isrc",
4017 "-Iinclude",
4018 ] + select({
4019 ":debug_build": [],
4020 "//conditions:default": xnnpack_min_size_copts(),
4021 }) + select({
4022 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4023 "//conditions:default": [],
4024 }),
4025 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004026 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004027 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004028 "XNN_NO_U8_OPERATORS",
4029 "XNN_NO_X8_OPERATORS",
4030 "XNN_NO_F16_OPERATORS",
4031 "XNN_NO_X16_OPERATORS",
4032 ],
4033 gcc_copts = xnnpack_gcc_std_copts(),
4034 includes = ["include"],
4035 msvc_copts = xnnpack_msvc_std_copts(),
4036 visibility = xnnpack_visibility(),
4037 deps = [
4038 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004039 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004040 ":logging_utils",
4041 ":memory_planner",
4042 ":operator_run",
4043 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004044 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004045 "@clog",
4046 "@FP16",
4047 "@pthreadpool",
4048 ] + select({
4049 ":emscripten": [],
4050 "//conditions:default": ["@cpuinfo"],
4051 }),
4052)
4053
4054xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004055 name = "xnnpack_operators_nhwc_f32",
4056 srcs = [
4057 "src/init.c",
4058 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004059 hdrs = ["include/xnnpack.h"],
4060 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004061 "-Isrc",
4062 "-Iinclude",
4063 ] + select({
4064 ":debug_build": [],
4065 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004066 }) + select({
4067 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4068 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004069 }),
4070 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004071 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004072 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004073 "XNN_NO_U8_OPERATORS",
4074 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004075 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004076 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004077 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004079 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004080 visibility = xnnpack_visibility(),
4081 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004082 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004083 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084 ":operator_run",
4085 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004086 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004087 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004088 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004089 ] + select({
4090 ":emscripten": [],
4091 "//conditions:default": ["@cpuinfo"],
4092 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093)
4094
Marat Dukhancf056b22019-10-07 10:26:29 -07004095xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096 name = "bench_utils",
4097 srcs = ["bench/utils.cc"],
4098 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004099 deps = [
4100 "@com_google_benchmark//:benchmark",
4101 "@cpuinfo",
4102 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004103)
4104
Frank Barchard7e955972019-10-11 10:34:25 -07004105######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106
4107xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004108 name = "qs8_gemm_bench",
4109 srcs = [
4110 "bench/gemm.h",
4111 "bench/qs8-gemm.cc",
4112 "src/xnnpack/AlignedAllocator.h",
4113 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4114 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4115)
4116
4117xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004118 name = "qs8_requantization_bench",
4119 srcs = [
4120 "bench/qs8-requantization.cc",
4121 "src/xnnpack/requantization-stubs.h",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + MICROKERNEL_BENCHMARK_HDRS,
4124 deps = MICROKERNEL_BENCHMARK_DEPS,
4125)
4126
4127xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004128 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 srcs = [
4130 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004131 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004132 "src/xnnpack/AlignedAllocator.h",
4133 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004134 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004135 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004136)
4137
4138xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004139 name = "qu8_requantization_bench",
4140 srcs = [
4141 "bench/qu8-requantization.cc",
4142 "src/xnnpack/requantization-stubs.h",
4143 "src/xnnpack/AlignedAllocator.h",
4144 ] + MICROKERNEL_BENCHMARK_HDRS,
4145 deps = MICROKERNEL_BENCHMARK_DEPS,
4146)
4147
4148xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004149 name = "f16_igemm_bench",
4150 srcs = [
4151 "bench/f16-igemm.cc",
4152 "bench/conv.h",
4153 "bench/google/conv.h",
4154 "src/xnnpack/AlignedAllocator.h",
4155 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004156 deps = MICROKERNEL_BENCHMARK_DEPS + [
4157 ":indirection",
4158 ":packing",
4159 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004160)
4161
4162xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004163 name = "f16_gemm_bench",
4164 srcs = [
4165 "bench/f16-gemm.cc",
4166 "bench/gemm.h",
4167 "src/xnnpack/AlignedAllocator.h",
4168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004169 deps = MICROKERNEL_BENCHMARK_DEPS + [
4170 ":packing",
4171 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172)
4173
4174xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004175 name = "f16_spmm_bench",
4176 srcs = [
4177 "bench/f16-spmm.cc",
4178 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004179 "src/xnnpack/AlignedAllocator.h",
4180 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004181 deps = MICROKERNEL_BENCHMARK_DEPS,
4182)
4183
4184xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 name = "f32_igemm_bench",
4186 srcs = [
4187 "bench/f32-igemm.cc",
4188 "bench/conv.h",
4189 "src/xnnpack/AlignedAllocator.h",
4190 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004191 deps = MICROKERNEL_BENCHMARK_DEPS + [
4192 ":indirection",
4193 ":packing",
4194 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195)
4196
4197xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004198 name = "f16_relu_bench",
4199 srcs = [
4200 "bench/f16-relu.cc",
4201 "src/xnnpack/AlignedAllocator.h",
4202 ] + MICROKERNEL_BENCHMARK_HDRS,
4203 deps = MICROKERNEL_BENCHMARK_DEPS,
4204)
4205
4206xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207 name = "f32_conv_hwc_bench",
4208 srcs = [
4209 "bench/f32-conv-hwc.cc",
4210 "bench/dconv.h",
4211 "src/xnnpack/AlignedAllocator.h",
4212 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004213 deps = MICROKERNEL_BENCHMARK_DEPS + [
4214 ":packing",
4215 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216)
4217
4218xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004219 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004220 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004221 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004222 "bench/dconv.h",
4223 "src/xnnpack/AlignedAllocator.h",
4224 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004225 deps = MICROKERNEL_BENCHMARK_DEPS + [
4226 ":packing",
4227 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004228)
4229
4230xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004231 name = "f16_dwconv_bench",
4232 srcs = [
4233 "bench/f16-dwconv.cc",
4234 "bench/dwconv.h",
4235 "bench/google/dwconv.h",
4236 "src/xnnpack/AlignedAllocator.h",
4237 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004238 deps = MICROKERNEL_BENCHMARK_DEPS + [
4239 ":indirection",
4240 ":packing",
4241 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004242)
4243
4244xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 name = "f32_dwconv_bench",
4246 srcs = [
4247 "bench/f32-dwconv.cc",
4248 "bench/dwconv.h",
4249 "src/xnnpack/AlignedAllocator.h",
4250 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004251 deps = MICROKERNEL_BENCHMARK_DEPS + [
4252 ":indirection",
4253 ":packing",
4254 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255)
4256
4257xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004258 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004259 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004260 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004261 "bench/dwconv.h",
4262 "src/xnnpack/AlignedAllocator.h",
4263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_BENCHMARK_DEPS + [
4265 ":indirection",
4266 ":packing",
4267 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268)
4269
4270xnnpack_benchmark(
4271 name = "f32_gemm_bench",
4272 srcs = [
4273 "bench/f32-gemm.cc",
4274 "bench/gemm.h",
4275 "src/xnnpack/AlignedAllocator.h",
4276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004277 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004278 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279)
4280
4281xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004282 name = "f32_hswish_bench",
4283 srcs = [
4284 "bench/f32-hswish.cc",
4285 "src/xnnpack/AlignedAllocator.h",
4286 ] + MICROKERNEL_BENCHMARK_HDRS,
4287 deps = MICROKERNEL_BENCHMARK_DEPS,
4288)
4289
4290xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004291 name = "f32_raddexpminusmax_bench",
4292 srcs = [
4293 "bench/f32-raddexpminusmax.cc",
4294 "src/xnnpack/AlignedAllocator.h",
4295 ] + MICROKERNEL_BENCHMARK_HDRS,
4296 deps = MICROKERNEL_BENCHMARK_DEPS,
4297)
4298
4299xnnpack_benchmark(
4300 name = "f32_raddextexp_bench",
4301 srcs = [
4302 "bench/f32-raddextexp.cc",
4303 "src/xnnpack/AlignedAllocator.h",
4304 ] + MICROKERNEL_BENCHMARK_HDRS,
4305 deps = MICROKERNEL_BENCHMARK_DEPS,
4306)
4307
4308xnnpack_benchmark(
4309 name = "f32_raddstoreexpminusmax_bench",
4310 srcs = [
4311 "bench/f32-raddstoreexpminusmax.cc",
4312 "src/xnnpack/AlignedAllocator.h",
4313 ] + MICROKERNEL_BENCHMARK_HDRS,
4314 deps = MICROKERNEL_BENCHMARK_DEPS,
4315)
4316
4317xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004318 name = "f32_relu_bench",
4319 srcs = [
4320 "bench/f32-relu.cc",
4321 "src/xnnpack/AlignedAllocator.h",
4322 ] + MICROKERNEL_BENCHMARK_HDRS,
4323 deps = MICROKERNEL_BENCHMARK_DEPS,
4324)
4325
4326xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327 name = "f32_rmax_bench",
4328 srcs = [
4329 "bench/f32-rmax.cc",
4330 "src/xnnpack/AlignedAllocator.h",
4331 ] + MICROKERNEL_BENCHMARK_HDRS,
4332 deps = MICROKERNEL_BENCHMARK_DEPS,
4333)
4334
4335xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004336 name = "f32_sigmoid_bench",
4337 srcs = [
4338 "bench/f32-sigmoid.cc",
4339 "src/xnnpack/AlignedAllocator.h",
4340 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004341 deps = MICROKERNEL_BENCHMARK_DEPS,
4342)
4343
4344xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345 name = "f32_spmm_bench",
4346 srcs = [
4347 "bench/f32-spmm.cc",
4348 "bench/gemm.h",
4349 "src/xnnpack/AlignedAllocator.h",
4350 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351 deps = MICROKERNEL_BENCHMARK_DEPS,
4352)
4353
4354xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004355 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004356 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004357 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004358 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004359 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004360 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004361)
4362
4363xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004364 name = "f32_vscaleexpminusmax_bench",
4365 srcs = [
4366 "bench/f32-vscaleexpminusmax.cc",
4367 "src/xnnpack/AlignedAllocator.h",
4368 ] + MICROKERNEL_BENCHMARK_HDRS,
4369 deps = MICROKERNEL_BENCHMARK_DEPS,
4370)
4371
4372xnnpack_benchmark(
4373 name = "f32_vscaleextexp_bench",
4374 srcs = [
4375 "bench/f32-vscaleextexp.cc",
4376 "src/xnnpack/AlignedAllocator.h",
4377 ] + MICROKERNEL_BENCHMARK_HDRS,
4378 deps = MICROKERNEL_BENCHMARK_DEPS,
4379)
4380
4381xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004382 name = "f32_vsqrt_bench",
4383 srcs = [
4384 "bench/f32-vsqrt.cc",
4385 "src/xnnpack/AlignedAllocator.h",
4386 ] + MICROKERNEL_BENCHMARK_HDRS,
4387 deps = MICROKERNEL_BENCHMARK_DEPS,
4388)
4389
4390xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 name = "f32_im2col_gemm_bench",
4392 srcs = [
4393 "bench/f32-im2col-gemm.cc",
4394 "bench/conv.h",
4395 "src/xnnpack/AlignedAllocator.h",
4396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004397 deps = MICROKERNEL_BENCHMARK_DEPS + [
4398 ":im2col",
4399 ":packing",
4400 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401)
4402
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004403xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004404 name = "rounding_bench",
4405 srcs = [
4406 "bench/rounding.cc",
4407 "src/xnnpack/math-stubs.h",
4408 "src/xnnpack/AlignedAllocator.h",
4409 ] + MICROKERNEL_BENCHMARK_HDRS,
4410 deps = MICROKERNEL_BENCHMARK_DEPS,
4411)
4412
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413########################### Benchmarks for operators ###########################
4414
4415xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 name = "average_pooling_bench",
4417 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004418 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004419 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004420 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421)
4422
4423xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004424 name = "bankers_rounding_bench",
4425 srcs = ["bench/bankers-rounding.cc"],
4426 copts = xnnpack_optional_tflite_copts(),
4427 tags = ["nowin32"],
4428 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4429)
4430
4431xnnpack_benchmark(
4432 name = "ceiling_bench",
4433 srcs = ["bench/ceiling.cc"],
4434 copts = xnnpack_optional_tflite_copts(),
4435 tags = ["nowin32"],
4436 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4437)
4438
4439xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440 name = "channel_shuffle_bench",
4441 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004442 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443)
4444
4445xnnpack_benchmark(
4446 name = "convolution_bench",
4447 srcs = ["bench/convolution.cc"],
4448 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004449 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004450 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451)
4452
4453xnnpack_benchmark(
4454 name = "deconvolution_bench",
4455 srcs = ["bench/deconvolution.cc"],
4456 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004457 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004458 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459)
4460
4461xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004462 name = "floor_bench",
4463 srcs = ["bench/floor.cc"],
4464 copts = xnnpack_optional_tflite_copts(),
4465 tags = ["nowin32"],
4466 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4467)
4468
4469xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 name = "global_average_pooling_bench",
4471 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004472 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473)
4474
4475xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004476 name = "hardswish_bench",
4477 srcs = ["bench/hardswish.cc"],
4478 copts = xnnpack_optional_tflite_copts(),
4479 tags = ["nowin32"],
4480 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4481)
4482
4483xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004484 name = "max_pooling_bench",
4485 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004486 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487)
4488
4489xnnpack_benchmark(
4490 name = "sigmoid_bench",
4491 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004492 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004493 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004494 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495)
4496
4497xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004498 name = "prelu_bench",
4499 srcs = ["bench/prelu.cc"],
4500 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004501 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004502 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004503)
4504
4505xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004506 name = "softmax_bench",
4507 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004508 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004509 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004510 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004511)
4512
Marat Dukhan87727142020-06-24 15:24:10 -07004513xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004514 name = "square_root_bench",
4515 srcs = ["bench/square-root.cc"],
4516 copts = xnnpack_optional_tflite_copts(),
4517 tags = ["nowin32"],
4518 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4519)
4520
4521xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004522 name = "truncation_bench",
4523 srcs = ["bench/truncation.cc"],
4524 deps = OPERATOR_BENCHMARK_DEPS,
4525)
4526
Marat Dukhanc068bb62019-10-04 13:24:39 -07004527############################# End-to-end benchmarks ############################
4528
4529cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004530 name = "fp32_mobilenet_v1",
4531 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004532 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004533 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004534 linkstatic = True,
4535 deps = [
4536 ":XNNPACK",
4537 "@pthreadpool",
4538 ],
4539)
4540
4541cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004542 name = "fp16_mobilenet_v1",
4543 srcs = ["models/fp16-mobilenet-v1.cc"],
4544 hdrs = ["models/models.h"],
4545 copts = xnnpack_std_cxxopts(),
4546 linkstatic = True,
4547 deps = [
4548 ":XNNPACK",
4549 "@FP16",
4550 "@pthreadpool",
4551 ],
4552)
4553
4554cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004555 name = "qs8_mobilenet_v1",
4556 srcs = ["models/qs8-mobilenet-v1.cc"],
4557 hdrs = ["models/models.h"],
4558 copts = xnnpack_std_cxxopts(),
4559 linkstatic = True,
4560 deps = [
4561 ":XNNPACK",
4562 "@pthreadpool",
4563 ],
4564)
4565
4566cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004567 name = "fp32_mobilenet_v2",
4568 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004569 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004570 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004571 linkstatic = True,
4572 deps = [
4573 ":XNNPACK",
4574 "@pthreadpool",
4575 ],
4576)
4577
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004578cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004579 name = "fp16_mobilenet_v2",
4580 srcs = ["models/fp16-mobilenet-v2.cc"],
4581 hdrs = ["models/models.h"],
4582 copts = xnnpack_std_cxxopts(),
4583 linkstatic = True,
4584 deps = [
4585 ":XNNPACK",
4586 "@FP16",
4587 "@pthreadpool",
4588 ],
4589)
4590
4591cc_library(
4592 name = "fp32_mobilenet_v3_large",
4593 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004594 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004595 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004596 linkstatic = True,
4597 deps = [
4598 ":XNNPACK",
4599 "@pthreadpool",
4600 ],
4601)
4602
4603cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004604 name = "fp32_mobilenet_v3_small",
4605 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004606 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004607 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004608 linkstatic = True,
4609 deps = [
4610 ":XNNPACK",
4611 "@pthreadpool",
4612 ],
4613)
4614
Marat Dukhanc068bb62019-10-04 13:24:39 -07004615xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004616 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004617 srcs = [
4618 "bench/f32-dwconv-e2e.cc",
4619 "bench/end2end.h",
4620 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004621 deps = MICROKERNEL_BENCHMARK_DEPS + [
4622 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004623 ":fp32_mobilenet_v1",
4624 ":fp32_mobilenet_v2",
4625 ":fp32_mobilenet_v3_large",
4626 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004627 ],
4628)
4629
4630xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004631 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004632 srcs = [
4633 "bench/f32-gemm-e2e.cc",
4634 "bench/end2end.h",
4635 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004636 deps = MICROKERNEL_BENCHMARK_DEPS + [
4637 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004638 ":fp32_mobilenet_v1",
4639 ":fp32_mobilenet_v2",
4640 ":fp32_mobilenet_v3_large",
4641 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004642 ],
4643)
4644
4645xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004646 name = "end2end_bench",
4647 srcs = ["bench/end2end.cc"],
4648 deps = [
4649 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004650 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004651 ":fp16_mobilenet_v1",
4652 ":fp16_mobilenet_v2",
4653 ":fp32_mobilenet_v1",
4654 ":fp32_mobilenet_v2",
4655 ":fp32_mobilenet_v3_large",
4656 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004657 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004658 "@pthreadpool",
4659 ],
4660)
4661
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004662#################### Accuracy evaluation for math functions ####################
4663
4664xnnpack_benchmark(
4665 name = "f32_exp_eval",
4666 srcs = [
4667 "eval/f32-exp.cc",
4668 "src/xnnpack/AlignedAllocator.h",
4669 ] + ACCURACY_EVAL_HDRS,
4670 deps = ACCURACY_EVAL_DEPS,
4671)
4672
Marat Dukhan515c9772019-10-17 18:07:57 -07004673xnnpack_benchmark(
4674 name = "f32_expminus_eval",
4675 srcs = [
4676 "eval/f32-expminus.cc",
4677 "src/xnnpack/AlignedAllocator.h",
4678 ] + ACCURACY_EVAL_HDRS,
4679 deps = ACCURACY_EVAL_DEPS,
4680)
4681
Marat Dukhan98ba4412019-10-23 02:14:28 -07004682xnnpack_benchmark(
4683 name = "f32_extexp_eval",
4684 srcs = [
4685 "eval/f32-extexp.cc",
4686 "src/xnnpack/AlignedAllocator.h",
4687 ] + ACCURACY_EVAL_HDRS,
4688 deps = ACCURACY_EVAL_DEPS,
4689)
4690
Marat Dukhan8853b822020-05-07 12:19:01 -07004691xnnpack_unit_test(
4692 name = "f32_roundne_eval",
4693 srcs = [
4694 "eval/f32-roundne.cc",
4695 "src/xnnpack/AlignedAllocator.h",
4696 "src/xnnpack/math-stubs.h",
4697 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004698 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004699 deps = MICROKERNEL_TEST_DEPS,
4700)
4701
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004702xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004703 name = "f32_roundd_eval",
4704 srcs = [
4705 "eval/f32-roundd.cc",
4706 "src/xnnpack/AlignedAllocator.h",
4707 "src/xnnpack/math-stubs.h",
4708 ] + MICROKERNEL_TEST_HDRS,
4709 automatic = False,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
4714 name = "f32_roundu_eval",
4715 srcs = [
4716 "eval/f32-roundu.cc",
4717 "src/xnnpack/AlignedAllocator.h",
4718 "src/xnnpack/math-stubs.h",
4719 ] + MICROKERNEL_TEST_HDRS,
4720 automatic = False,
4721 deps = MICROKERNEL_TEST_DEPS,
4722)
4723
4724xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004725 name = "f32_roundz_eval",
4726 srcs = [
4727 "eval/f32-roundz.cc",
4728 "src/xnnpack/AlignedAllocator.h",
4729 "src/xnnpack/math-stubs.h",
4730 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004731 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
Marat Dukhan346a9e52019-11-15 09:06:30 -08004735xnnpack_benchmark(
4736 name = "f32_sigmoid_eval",
4737 srcs = [
4738 "eval/f32-sigmoid.cc",
4739 "src/xnnpack/AlignedAllocator.h",
4740 ] + ACCURACY_EVAL_HDRS,
4741 deps = ACCURACY_EVAL_DEPS,
4742)
4743
Marat Dukhan84000762020-06-29 18:38:43 -07004744xnnpack_benchmark(
4745 name = "f32_sqrt_eval",
4746 srcs = [
4747 "eval/f32-sqrt.cc",
4748 "src/xnnpack/AlignedAllocator.h",
4749 ] + ACCURACY_EVAL_HDRS,
4750 deps = ACCURACY_EVAL_DEPS + [
4751 ":bench_utils",
4752 ],
4753)
4754
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755######################### Unit tests for micro-kernels #########################
4756
4757xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004758 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004760 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004761 "test/gemm-microkernel-tester.h",
4762 "src/xnnpack/AlignedAllocator.h",
4763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765)
4766
4767xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004768 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004769 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004770 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004771 "test/spmm-microkernel-tester.h",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + MICROKERNEL_TEST_HDRS,
4774 deps = MICROKERNEL_TEST_DEPS,
4775)
4776
4777xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778 name = "f32_argmaxpool_test",
4779 srcs = [
4780 "test/f32-argmaxpool.cc",
4781 "test/argmaxpool-microkernel-tester.h",
4782 "src/xnnpack/AlignedAllocator.h",
4783 ] + MICROKERNEL_TEST_HDRS,
4784 deps = MICROKERNEL_TEST_DEPS,
4785)
4786
4787xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004788 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004790 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004791 "test/avgpool-microkernel-tester.h",
4792 "src/xnnpack/AlignedAllocator.h",
4793 ] + MICROKERNEL_TEST_HDRS,
4794 deps = MICROKERNEL_TEST_DEPS,
4795)
4796
4797xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004798 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004799 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004800 "test/f32-ibilinear.cc",
4801 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004802 "src/xnnpack/AlignedAllocator.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004808 name = "f16_clamp_test",
4809 srcs = [
4810 "test/f16-clamp.cc",
4811 "test/clamp-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004817 name = "f32_clamp_test",
4818 srcs = [
4819 "test/f32-clamp.cc",
4820 "test/clamp-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004826 name = "f32_igemm_test",
4827 srcs = [
4828 "test/f32-igemm.cc",
4829 "test/gemm-microkernel-tester.h",
4830 "src/xnnpack/AlignedAllocator.h",
4831 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004832 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004833)
4834
4835xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004836 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004837 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004838 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839 "test/gemm-microkernel-tester.h",
4840 "src/xnnpack/AlignedAllocator.h",
4841 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004842 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004843)
4844
4845xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004846 name = "f32_igemm_minmax_test",
4847 srcs = [
4848 "test/f32-igemm-minmax.cc",
4849 "test/gemm-microkernel-tester.h",
4850 "src/xnnpack/AlignedAllocator.h",
4851 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004852 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004853)
4854
4855xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004856 name = "f16_igemm_minmax_test",
4857 srcs = [
4858 "test/f16-igemm-minmax.cc",
4859 "test/gemm-microkernel-tester.h",
4860 "src/xnnpack/AlignedAllocator.h",
4861 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004862 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004863)
4864
4865xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866 name = "f32_conv_hwc_test",
4867 srcs = [
4868 "test/f32-conv-hwc.cc",
4869 "test/conv-hwc-microkernel-tester.h",
4870 "src/xnnpack/AlignedAllocator.h",
4871 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004872 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004873)
4874
4875xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004876 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004877 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004878 "test/f32-conv-hwc2chw.cc",
4879 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880 "src/xnnpack/AlignedAllocator.h",
4881 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004882 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883)
4884
4885xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004886 name = "f32_dwconv_test",
4887 srcs = [
4888 "test/f32-dwconv.cc",
4889 "test/dwconv-microkernel-tester.h",
4890 "src/xnnpack/AlignedAllocator.h",
4891 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004892 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004893)
4894
4895xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004896 name = "f16_dwconv_minmax_test",
4897 srcs = [
4898 "test/f16-dwconv-minmax.cc",
4899 "test/dwconv-microkernel-tester.h",
4900 "src/xnnpack/AlignedAllocator.h",
4901 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004902 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004903)
4904
4905xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004906 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004908 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004909 "test/dwconv-microkernel-tester.h",
4910 "src/xnnpack/AlignedAllocator.h",
4911 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004912 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004913)
4914
4915xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004916 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004917 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004918 "test/f32-dwconv-chw.cc",
4919 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004920 "src/xnnpack/AlignedAllocator.h",
4921 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004922 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004923)
4924
4925xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004926 name = "f16_gavgpool_minmax_test",
4927 srcs = [
4928 "test/f16-gavgpool-minmax.cc",
4929 "test/gavgpool-microkernel-tester.h",
4930 "src/xnnpack/AlignedAllocator.h",
4931 ] + MICROKERNEL_TEST_HDRS,
4932 deps = MICROKERNEL_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004936 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004938 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004939 "test/gavgpool-microkernel-tester.h",
4940 "src/xnnpack/AlignedAllocator.h",
4941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004946 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004947 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004948 "test/f32-gavgpool-cw.cc",
4949 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004950 "src/xnnpack/AlignedAllocator.h",
4951 ] + MICROKERNEL_TEST_HDRS,
4952 deps = MICROKERNEL_TEST_DEPS,
4953)
4954
4955xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004956 name = "f32_gemm_test",
4957 srcs = [
4958 "test/f32-gemm.cc",
4959 "test/gemm-microkernel-tester.h",
4960 "src/xnnpack/AlignedAllocator.h",
4961 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004962 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004963)
4964
4965xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004966 name = "f32_gemm_relu_test",
4967 srcs = [
4968 "test/f32-gemm-relu.cc",
4969 "test/gemm-microkernel-tester.h",
4970 "src/xnnpack/AlignedAllocator.h",
4971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004972 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004973)
4974
4975xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004976 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004977 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004978 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979 "test/gemm-microkernel-tester.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004982 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983)
4984
4985xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004986 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004988 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 "test/gemm-microkernel-tester.h",
4990 "src/xnnpack/AlignedAllocator.h",
4991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004992 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
4995xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004996 name = "f16_hswish_test",
4997 srcs = [
4998 "test/f16-hswish.cc",
4999 "test/hswish-microkernel-tester.h",
5000 ] + MICROKERNEL_TEST_HDRS,
5001 deps = MICROKERNEL_TEST_DEPS,
5002)
5003
5004xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 name = "f32_hswish_test",
5006 srcs = [
5007 "test/f32-hswish.cc",
5008 "test/hswish-microkernel-tester.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005014 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005016 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017 "test/maxpool-microkernel-tester.h",
5018 ] + MICROKERNEL_TEST_HDRS,
5019 deps = MICROKERNEL_TEST_DEPS,
5020)
5021
5022xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005023 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005024 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005025 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026 "test/avgpool-microkernel-tester.h",
5027 "src/xnnpack/AlignedAllocator.h",
5028 ] + MICROKERNEL_TEST_HDRS,
5029 deps = MICROKERNEL_TEST_DEPS,
5030)
5031
5032xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005033 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005034 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005035 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 "test/gemm-microkernel-tester.h",
5037 "src/xnnpack/AlignedAllocator.h",
5038 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005039 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040)
5041
5042xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005043 name = "f16_prelu_test",
5044 srcs = [
5045 "test/f16-prelu.cc",
5046 "test/prelu-microkernel-tester.h",
5047 "src/xnnpack/AlignedAllocator.h",
5048 ] + MICROKERNEL_TEST_HDRS,
5049 deps = MICROKERNEL_TEST_DEPS,
5050)
5051
5052xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053 name = "f32_prelu_test",
5054 srcs = [
5055 "test/f32-prelu.cc",
5056 "test/prelu-microkernel-tester.h",
5057 "src/xnnpack/AlignedAllocator.h",
5058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005063 name = "f32_raddexpminusmax_test",
5064 srcs = [
5065 "test/f32-raddexpminusmax.cc",
5066 "test/raddexpminusmax-microkernel-tester.h",
5067 ] + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS,
5069)
5070
5071xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005072 name = "f32_raddextexp_test",
5073 srcs = [
5074 "test/f32-raddextexp.cc",
5075 "test/raddextexp-microkernel-tester.h",
5076 ] + MICROKERNEL_TEST_HDRS,
5077 deps = MICROKERNEL_TEST_DEPS,
5078)
5079
5080xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005081 name = "f32_raddstoreexpminusmax_test",
5082 srcs = [
5083 "test/f32-raddstoreexpminusmax.cc",
5084 "test/raddstoreexpminusmax-microkernel-tester.h",
5085 ] + MICROKERNEL_TEST_HDRS,
5086 deps = MICROKERNEL_TEST_DEPS,
5087)
5088
5089xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005090 name = "f32_relu_test",
5091 srcs = [
5092 "test/f32-relu.cc",
5093 "test/vunary-microkernel-tester.h",
5094 ] + MICROKERNEL_TEST_HDRS,
5095 deps = MICROKERNEL_TEST_DEPS,
5096)
5097
5098xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005099 name = "f32_rmax_test",
5100 srcs = [
5101 "test/f32-rmax.cc",
5102 "test/rmax-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005108 name = "f32_sigmoid_test",
5109 srcs = [
5110 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005111 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005117 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005119 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120 "test/spmm-microkernel-tester.h",
5121 "src/xnnpack/AlignedAllocator.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 deps = MICROKERNEL_TEST_DEPS,
5124)
5125
5126xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005127 name = "f16_vadd_minmax_test",
5128 srcs = [
5129 "test/f16-vadd-minmax.cc",
5130 "test/vbinary-microkernel-tester.h",
5131 ] + MICROKERNEL_TEST_HDRS,
5132 deps = MICROKERNEL_TEST_DEPS,
5133)
5134
5135xnnpack_unit_test(
5136 name = "f16_vaddc_minmax_test",
5137 srcs = [
5138 "test/f16-vaddc-minmax.cc",
5139 "test/vbinaryc-microkernel-tester.h",
5140 ] + MICROKERNEL_TEST_HDRS,
5141 deps = MICROKERNEL_TEST_DEPS,
5142)
5143
5144xnnpack_unit_test(
5145 name = "f16_vdiv_minmax_test",
5146 srcs = [
5147 "test/f16-vdiv-minmax.cc",
5148 "test/vbinary-microkernel-tester.h",
5149 ] + MICROKERNEL_TEST_HDRS,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
5154 name = "f16_vdivc_minmax_test",
5155 srcs = [
5156 "test/f16-vdivc-minmax.cc",
5157 "test/vbinaryc-microkernel-tester.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
5163 name = "f16_vrdivc_minmax_test",
5164 srcs = [
5165 "test/f16-vrdivc-minmax.cc",
5166 "test/vbinaryc-microkernel-tester.h",
5167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
5172 name = "f16_vmax_test",
5173 srcs = [
5174 "test/f16-vmax.cc",
5175 "test/vbinary-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
5181 name = "f16_vmaxc_test",
5182 srcs = [
5183 "test/f16-vmaxc.cc",
5184 "test/vbinaryc-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
5190 name = "f16_vmin_test",
5191 srcs = [
5192 "test/f16-vmin.cc",
5193 "test/vbinary-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
5199 name = "f16_vminc_test",
5200 srcs = [
5201 "test/f16-vminc.cc",
5202 "test/vbinaryc-microkernel-tester.h",
5203 ] + MICROKERNEL_TEST_HDRS,
5204 deps = MICROKERNEL_TEST_DEPS,
5205)
5206
5207xnnpack_unit_test(
5208 name = "f16_vmul_minmax_test",
5209 srcs = [
5210 "test/f16-vmul-minmax.cc",
5211 "test/vbinary-microkernel-tester.h",
5212 ] + MICROKERNEL_TEST_HDRS,
5213 deps = MICROKERNEL_TEST_DEPS,
5214)
5215
5216xnnpack_unit_test(
5217 name = "f16_vmulc_minmax_test",
5218 srcs = [
5219 "test/f16-vmulc-minmax.cc",
5220 "test/vbinaryc-microkernel-tester.h",
5221 ] + MICROKERNEL_TEST_HDRS,
5222 deps = MICROKERNEL_TEST_DEPS,
5223)
5224
5225xnnpack_unit_test(
5226 name = "f16_vsub_minmax_test",
5227 srcs = [
5228 "test/f16-vsub-minmax.cc",
5229 "test/vbinary-microkernel-tester.h",
5230 ] + MICROKERNEL_TEST_HDRS,
5231 deps = MICROKERNEL_TEST_DEPS,
5232)
5233
5234xnnpack_unit_test(
5235 name = "f16_vsubc_minmax_test",
5236 srcs = [
5237 "test/f16-vsubc-minmax.cc",
5238 "test/vbinaryc-microkernel-tester.h",
5239 ] + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS,
5241)
5242
5243xnnpack_unit_test(
5244 name = "f16_vrsubc_minmax_test",
5245 srcs = [
5246 "test/f16-vrsubc-minmax.cc",
5247 "test/vbinaryc-microkernel-tester.h",
5248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005253 name = "f32_vabs_test",
5254 srcs = [
5255 "test/f32-vabs.cc",
5256 "test/vunary-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005262 name = "f32_vadd_test",
5263 srcs = [
5264 "test/f32-vadd.cc",
5265 "test/vbinary-microkernel-tester.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005271 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005272 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005273 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005274 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005275 ] + MICROKERNEL_TEST_HDRS,
5276 deps = MICROKERNEL_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005280 name = "f32_vadd_relu_test",
5281 srcs = [
5282 "test/f32-vadd-relu.cc",
5283 "test/vbinary-microkernel-tester.h",
5284 ] + MICROKERNEL_TEST_HDRS,
5285 deps = MICROKERNEL_TEST_DEPS,
5286)
5287
5288xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005289 name = "f32_vaddc_test",
5290 srcs = [
5291 "test/f32-vaddc.cc",
5292 "test/vbinaryc-microkernel-tester.h",
5293 ] + MICROKERNEL_TEST_HDRS,
5294 deps = MICROKERNEL_TEST_DEPS,
5295)
5296
5297xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005298 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005299 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005300 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005301 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 ] + MICROKERNEL_TEST_HDRS,
5303 deps = MICROKERNEL_TEST_DEPS,
5304)
5305
5306xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005307 name = "f32_vaddc_relu_test",
5308 srcs = [
5309 "test/f32-vaddc-relu.cc",
5310 "test/vbinaryc-microkernel-tester.h",
5311 ] + MICROKERNEL_TEST_HDRS,
5312 deps = MICROKERNEL_TEST_DEPS,
5313)
5314
5315xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005316 name = "f32_vdiv_test",
5317 srcs = [
5318 "test/f32-vdiv.cc",
5319 "test/vbinary-microkernel-tester.h",
5320 ] + MICROKERNEL_TEST_HDRS,
5321 deps = MICROKERNEL_TEST_DEPS,
5322)
5323
5324xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005325 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005326 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005327 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005328 "test/vbinary-microkernel-tester.h",
5329 ] + MICROKERNEL_TEST_HDRS,
5330 deps = MICROKERNEL_TEST_DEPS,
5331)
5332
5333xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005334 name = "f32_vdiv_relu_test",
5335 srcs = [
5336 "test/f32-vdiv-relu.cc",
5337 "test/vbinary-microkernel-tester.h",
5338 ] + MICROKERNEL_TEST_HDRS,
5339 deps = MICROKERNEL_TEST_DEPS,
5340)
5341
5342xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005343 name = "f32_vdivc_test",
5344 srcs = [
5345 "test/f32-vdivc.cc",
5346 "test/vbinaryc-microkernel-tester.h",
5347 ] + MICROKERNEL_TEST_HDRS,
5348 deps = MICROKERNEL_TEST_DEPS,
5349)
5350
5351xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005352 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005353 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005354 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005355 "test/vbinaryc-microkernel-tester.h",
5356 ] + MICROKERNEL_TEST_HDRS,
5357 deps = MICROKERNEL_TEST_DEPS,
5358)
5359
5360xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005361 name = "f32_vdivc_relu_test",
5362 srcs = [
5363 "test/f32-vdivc-relu.cc",
5364 "test/vbinaryc-microkernel-tester.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005370 name = "f32_vrdivc_test",
5371 srcs = [
5372 "test/f32-vrdivc.cc",
5373 "test/vbinaryc-microkernel-tester.h",
5374 ] + MICROKERNEL_TEST_HDRS,
5375 deps = MICROKERNEL_TEST_DEPS,
5376)
5377
5378xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005379 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005380 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005381 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005382 "test/vbinaryc-microkernel-tester.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005388 name = "f32_vrdivc_relu_test",
5389 srcs = [
5390 "test/f32-vrdivc-relu.cc",
5391 "test/vbinaryc-microkernel-tester.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 deps = MICROKERNEL_TEST_DEPS,
5394)
5395
5396xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005397 name = "f32_vmax_test",
5398 srcs = [
5399 "test/f32-vmax.cc",
5400 "test/vbinary-microkernel-tester.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
5406 name = "f32_vmaxc_test",
5407 srcs = [
5408 "test/f32-vmaxc.cc",
5409 "test/vbinaryc-microkernel-tester.h",
5410 ] + MICROKERNEL_TEST_HDRS,
5411 deps = MICROKERNEL_TEST_DEPS,
5412)
5413
5414xnnpack_unit_test(
5415 name = "f32_vmin_test",
5416 srcs = [
5417 "test/f32-vmin.cc",
5418 "test/vbinary-microkernel-tester.h",
5419 ] + MICROKERNEL_TEST_HDRS,
5420 deps = MICROKERNEL_TEST_DEPS,
5421)
5422
5423xnnpack_unit_test(
5424 name = "f32_vminc_test",
5425 srcs = [
5426 "test/f32-vminc.cc",
5427 "test/vbinaryc-microkernel-tester.h",
5428 ] + MICROKERNEL_TEST_HDRS,
5429 deps = MICROKERNEL_TEST_DEPS,
5430)
5431
5432xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005433 name = "f32_vmul_test",
5434 srcs = [
5435 "test/f32-vmul.cc",
5436 "test/vbinary-microkernel-tester.h",
5437 ] + MICROKERNEL_TEST_HDRS,
5438 deps = MICROKERNEL_TEST_DEPS,
5439)
5440
5441xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005442 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005444 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005445 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005446 ] + MICROKERNEL_TEST_HDRS,
5447 deps = MICROKERNEL_TEST_DEPS,
5448)
5449
5450xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005451 name = "f32_vmul_relu_test",
5452 srcs = [
5453 "test/f32-vmul-relu.cc",
5454 "test/vbinary-microkernel-tester.h",
5455 ] + MICROKERNEL_TEST_HDRS,
5456 deps = MICROKERNEL_TEST_DEPS,
5457)
5458
5459xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005460 name = "f32_vmulc_test",
5461 srcs = [
5462 "test/f32-vmulc.cc",
5463 "test/vbinaryc-microkernel-tester.h",
5464 ] + MICROKERNEL_TEST_HDRS,
5465 deps = MICROKERNEL_TEST_DEPS,
5466)
5467
5468xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005469 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005470 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005471 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005472 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473 ] + MICROKERNEL_TEST_HDRS,
5474 deps = MICROKERNEL_TEST_DEPS,
5475)
5476
5477xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005478 name = "f32_vmulc_relu_test",
5479 srcs = [
5480 "test/f32-vmulc-relu.cc",
5481 "test/vbinaryc-microkernel-tester.h",
5482 ] + MICROKERNEL_TEST_HDRS,
5483 deps = MICROKERNEL_TEST_DEPS,
5484)
5485
5486xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005487 name = "f16_vmulcaddc_minmax_test",
5488 srcs = [
5489 "test/f16-vmulcaddc-minmax.cc",
5490 "test/vmulcaddc-microkernel-tester.h",
5491 "src/xnnpack/AlignedAllocator.h",
5492 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005493 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005494)
5495
5496xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005497 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005499 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005500 "test/vmulcaddc-microkernel-tester.h",
5501 "src/xnnpack/AlignedAllocator.h",
5502 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005503 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504)
5505
5506xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005507 name = "f32_vlrelu_test",
5508 srcs = [
5509 "test/f32-vlrelu.cc",
5510 "test/vunary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005516 name = "f32_vneg_test",
5517 srcs = [
5518 "test/f32-vneg.cc",
5519 "test/vunary-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005525 name = "f32_vrndne_test",
5526 srcs = [
5527 "test/f32-vrndne.cc",
5528 "test/vunary-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
5534 name = "f32_vrndz_test",
5535 srcs = [
5536 "test/f32-vrndz.cc",
5537 "test/vunary-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
5543 name = "f32_vrndu_test",
5544 srcs = [
5545 "test/f32-vrndu.cc",
5546 "test/vunary-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
5552 name = "f32_vrndd_test",
5553 srcs = [
5554 "test/f32-vrndd.cc",
5555 "test/vunary-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005561 name = "f32_vscale_test",
5562 srcs = [
5563 "test/f32-vscale.cc",
5564 "test/vscale-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005570 name = "f32_vscaleexpminusmax_test",
5571 srcs = [
5572 "test/f32-vscaleexpminusmax.cc",
5573 "test/vscaleexpminusmax-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005579 name = "f32_vscaleextexp_test",
5580 srcs = [
5581 "test/f32-vscaleextexp.cc",
5582 "test/vscaleextexp-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005588 name = "f32_vsqr_test",
5589 srcs = [
5590 "test/f32-vsqr.cc",
5591 "test/vunary-microkernel-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005597 name = "f32_vsqrdiff_test",
5598 srcs = [
5599 "test/f32-vsqrdiff.cc",
5600 "test/vbinary-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
5606 name = "f32_vsqrdiffc_test",
5607 srcs = [
5608 "test/f32-vsqrdiffc.cc",
5609 "test/vbinaryc-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005615 name = "f32_vsqrt_test",
5616 srcs = [
5617 "test/f32-vsqrt.cc",
5618 "test/vunary-microkernel-tester.h",
5619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005624 name = "f32_vsub_test",
5625 srcs = [
5626 "test/f32-vsub.cc",
5627 "test/vbinary-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005633 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005634 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005635 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005636 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005642 name = "f32_vsub_relu_test",
5643 srcs = [
5644 "test/f32-vsub-relu.cc",
5645 "test/vbinary-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005651 name = "f32_vsubc_test",
5652 srcs = [
5653 "test/f32-vsubc.cc",
5654 "test/vbinaryc-microkernel-tester.h",
5655 ] + MICROKERNEL_TEST_HDRS,
5656 deps = MICROKERNEL_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005660 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005661 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005662 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005663 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005664 ] + MICROKERNEL_TEST_HDRS,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005669 name = "f32_vsubc_relu_test",
5670 srcs = [
5671 "test/f32-vsubc-relu.cc",
5672 "test/vbinaryc-microkernel-tester.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005678 name = "f32_vrsubc_test",
5679 srcs = [
5680 "test/f32-vrsubc.cc",
5681 "test/vbinaryc-microkernel-tester.h",
5682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005687 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005688 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005689 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005690 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005691 ] + MICROKERNEL_TEST_HDRS,
5692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
5695xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005696 name = "f32_vrsubc_relu_test",
5697 srcs = [
5698 "test/f32-vrsubc-relu.cc",
5699 "test/vbinaryc-microkernel-tester.h",
5700 ] + MICROKERNEL_TEST_HDRS,
5701 deps = MICROKERNEL_TEST_DEPS,
5702)
5703
5704xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005705 name = "qs8_dwconv_minmax_test",
5706 srcs = [
5707 "test/qs8-dwconv-minmax.cc",
5708 "test/dwconv-microkernel-tester.h",
5709 "src/xnnpack/AlignedAllocator.h",
5710 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5712)
5713
5714xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005715 name = "qs8_gavgpool_minmax_test",
5716 srcs = [
5717 "test/qs8-gavgpool-minmax.cc",
5718 "test/gavgpool-microkernel-tester.h",
5719 "src/xnnpack/AlignedAllocator.h",
5720 ] + MICROKERNEL_TEST_HDRS,
5721 deps = MICROKERNEL_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005725 name = "qs8_gemm_minmax_test",
5726 srcs = [
5727 "test/qs8-gemm-minmax.cc",
5728 "test/gemm-microkernel-tester.h",
5729 "src/xnnpack/AlignedAllocator.h",
5730 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5731 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5732)
5733
5734xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005735 name = "qs8_igemm_minmax_test",
5736 srcs = [
5737 "test/qs8-igemm-minmax.cc",
5738 "test/gemm-microkernel-tester.h",
5739 "src/xnnpack/AlignedAllocator.h",
5740 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5742)
5743
5744xnnpack_unit_test(
5745 name = "qs8_requantization_test",
5746 srcs = [
5747 "src/xnnpack/requantization-stubs.h",
5748 "test/qs8-requantization.cc",
5749 "test/requantization-tester.h",
5750 ] + MICROKERNEL_TEST_HDRS,
5751 deps = MICROKERNEL_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005755 name = "qs8_vadd_minmax_test",
5756 srcs = [
5757 "test/qs8-vadd-minmax.cc",
5758 "test/vadd-microkernel-tester.h",
5759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005764 name = "qs8_vaddc_minmax_test",
5765 srcs = [
5766 "test/qs8-vaddc-minmax.cc",
5767 "test/vaddc-microkernel-tester.h",
5768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005773 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005774 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005775 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005776 "test/avgpool-microkernel-tester.h",
5777 "src/xnnpack/AlignedAllocator.h",
5778 ] + MICROKERNEL_TEST_HDRS,
5779 deps = MICROKERNEL_TEST_DEPS,
5780)
5781
5782xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005783 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005784 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005785 "test/qu8-dwconv-minmax.cc",
5786 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005787 "src/xnnpack/AlignedAllocator.h",
5788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005789 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005790)
5791
5792xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005793 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005794 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005795 "test/qu8-igemm-minmax.cc",
5796 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005797 "src/xnnpack/AlignedAllocator.h",
5798 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005799 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005800)
5801
5802xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005803 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005804 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005805 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806 "test/gavgpool-microkernel-tester.h",
5807 "src/xnnpack/AlignedAllocator.h",
5808 ] + MICROKERNEL_TEST_HDRS,
5809 deps = MICROKERNEL_TEST_DEPS,
5810)
5811
5812xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005813 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005814 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005815 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005816 "test/gemm-microkernel-tester.h",
5817 "src/xnnpack/AlignedAllocator.h",
5818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005819 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005820)
5821
5822xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005823 name = "qu8_requantization_test",
5824 srcs = [
5825 "src/xnnpack/requantization-stubs.h",
5826 "test/qu8-requantization.cc",
5827 "test/requantization-tester.h",
5828 ] + MICROKERNEL_TEST_HDRS,
5829 deps = MICROKERNEL_TEST_DEPS,
5830)
5831
5832xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005833 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005834 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005835 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836 "test/vadd-microkernel-tester.h",
5837 ] + MICROKERNEL_TEST_HDRS,
5838 deps = MICROKERNEL_TEST_DEPS,
5839)
5840
5841xnnpack_unit_test(
5842 name = "u8_clamp_test",
5843 srcs = [
5844 "test/u8-clamp.cc",
5845 "test/clamp-microkernel-tester.h",
5846 ] + MICROKERNEL_TEST_HDRS,
5847 deps = MICROKERNEL_TEST_DEPS,
5848)
5849
5850xnnpack_unit_test(
5851 name = "u8_lut32norm_test",
5852 srcs = [
5853 "test/u8-lut32norm.cc",
5854 "test/lut-norm-microkernel-tester.h",
5855 ] + MICROKERNEL_TEST_HDRS,
5856 deps = MICROKERNEL_TEST_DEPS,
5857)
5858
5859xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005860 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005861 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005862 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005863 "test/maxpool-microkernel-tester.h",
5864 ] + MICROKERNEL_TEST_HDRS,
5865 deps = MICROKERNEL_TEST_DEPS,
5866)
5867
5868xnnpack_unit_test(
5869 name = "u8_rmax_test",
5870 srcs = [
5871 "test/u8-rmax.cc",
5872 "test/rmax-microkernel-tester.h",
5873 ] + MICROKERNEL_TEST_HDRS,
5874 deps = MICROKERNEL_TEST_DEPS,
5875)
5876
5877xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005878 name = "x32_fill_test",
5879 srcs = [
5880 "test/x32-fill.cc",
5881 "test/fill-microkernel-tester.h",
5882 ] + MICROKERNEL_TEST_HDRS,
5883 deps = MICROKERNEL_TEST_DEPS,
5884)
5885
5886xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005887 name = "x32_packx_test",
5888 srcs = [
5889 "test/x32-packx.cc",
5890 "test/pack-microkernel-tester.h",
5891 "src/xnnpack/AlignedAllocator.h",
5892 ] + MICROKERNEL_TEST_HDRS,
5893 deps = MICROKERNEL_TEST_DEPS,
5894)
5895
5896xnnpack_unit_test(
5897 name = "x32_pad_test",
5898 srcs = [
5899 "test/x32-pad.cc",
5900 "test/pad-microkernel-tester.h",
5901 ] + MICROKERNEL_TEST_HDRS,
5902 deps = MICROKERNEL_TEST_DEPS,
5903)
5904
5905xnnpack_unit_test(
5906 name = "x32_unpool_test",
5907 srcs = [
5908 "test/x32-unpool.cc",
5909 "test/unpool-microkernel-tester.h",
5910 ] + MICROKERNEL_TEST_HDRS,
5911 deps = MICROKERNEL_TEST_DEPS,
5912)
5913
5914xnnpack_unit_test(
5915 name = "x32_zip_test",
5916 srcs = [
5917 "test/x32-zip.cc",
5918 "test/zip-microkernel-tester.h",
5919 ] + MICROKERNEL_TEST_HDRS,
5920 deps = MICROKERNEL_TEST_DEPS,
5921)
5922
5923xnnpack_unit_test(
5924 name = "x8_lut_test",
5925 srcs = [
5926 "test/x8-lut.cc",
5927 "test/lut-microkernel-tester.h",
5928 ] + MICROKERNEL_TEST_HDRS,
5929 deps = MICROKERNEL_TEST_DEPS,
5930)
5931
5932xnnpack_unit_test(
5933 name = "x8_zip_test",
5934 srcs = [
5935 "test/x8-zip.cc",
5936 "test/zip-microkernel-tester.h",
5937 ] + MICROKERNEL_TEST_HDRS,
5938 deps = MICROKERNEL_TEST_DEPS,
5939)
5940
Marat Dukhan20c3b922020-03-10 03:45:06 -07005941########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005942
5943xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005944 name = "operator_size_test",
5945 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005946 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005947)
5948
Marat Dukhan20c3b922020-03-10 03:45:06 -07005949xnnpack_binary(
5950 name = "subgraph_size_test",
5951 srcs = ["test/subgraph-size.c"],
5952 deps = [":XNNPACK"],
5953)
5954
5955########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005956
5957xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005958 name = "abs_nc_test",
5959 srcs = [
5960 "test/abs-nc.cc",
5961 "test/abs-operator-tester.h",
5962 ],
5963 deps = OPERATOR_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005967 name = "add_nd_test",
5968 srcs = [
5969 "test/add-nd.cc",
5970 "test/binary-elementwise-operator-tester.h",
5971 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005972 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005973)
5974
5975xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005976 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005977 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005978 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005979 "test/argmax-pooling-operator-tester.h",
5980 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005981 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982)
5983
5984xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005985 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005986 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005987 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 "test/average-pooling-operator-tester.h",
5989 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005990 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005991)
5992
5993xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005994 name = "bankers_rounding_nc_test",
5995 srcs = [
5996 "test/bankers-rounding-nc.cc",
5997 "test/bankers-rounding-operator-tester.h",
5998 ],
5999 deps = OPERATOR_TEST_DEPS,
6000)
6001
6002xnnpack_unit_test(
6003 name = "ceiling_nc_test",
6004 srcs = [
6005 "test/ceiling-nc.cc",
6006 "test/ceiling-operator-tester.h",
6007 ],
6008 deps = OPERATOR_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006012 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006013 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006014 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006015 "test/channel-shuffle-operator-tester.h",
6016 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006017 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018)
6019
6020xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006021 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006022 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006023 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006024 "test/clamp-operator-tester.h",
6025 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006026 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006027)
6028
6029xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006030 name = "constant_pad_nd_test",
6031 srcs = [
6032 "test/constant-pad-nd.cc",
6033 "test/constant-pad-operator-tester.h",
6034 ],
6035 deps = OPERATOR_TEST_DEPS,
6036)
6037
6038xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006039 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006040 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006041 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006042 "test/convolution-operator-tester.h",
6043 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006044 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006045)
6046
6047xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006048 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006050 "test/convolution-nchw.cc",
6051 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006053 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054)
6055
6056xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006057 name = "copy_nc_test",
6058 srcs = [
6059 "test/copy-nc.cc",
6060 "test/copy-operator-tester.h",
6061 ],
6062 deps = OPERATOR_TEST_DEPS,
6063)
6064
6065xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006066 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006067 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006068 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006069 "test/deconvolution-operator-tester.h",
6070 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006071 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006072)
6073
6074xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006075 name = "divide_nd_test",
6076 srcs = [
6077 "test/binary-elementwise-operator-tester.h",
6078 "test/divide-nd.cc",
6079 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006080 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006081)
6082
6083xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006084 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006085 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006086 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006087 "test/fully-connected-operator-tester.h",
6088 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006089 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006090)
6091
6092xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006093 name = "floor_nc_test",
6094 srcs = [
6095 "test/floor-nc.cc",
6096 "test/floor-operator-tester.h",
6097 ],
6098 deps = OPERATOR_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006102 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006103 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006104 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006105 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006106 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006107 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006108)
6109
6110xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006111 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006112 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006113 "test/global-average-pooling-ncw.cc",
6114 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006115 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006116 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006117)
6118
6119xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006120 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006121 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006122 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006123 "test/hardswish-operator-tester.h",
6124 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006125 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006126)
6127
6128xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006129 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006130 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006131 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 "test/leaky-relu-operator-tester.h",
6133 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006134 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135)
6136
6137xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006138 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006139 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006140 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006141 "test/max-pooling-operator-tester.h",
6142 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006143 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144)
6145
6146xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006147 name = "maximum_nd_test",
6148 srcs = [
6149 "test/binary-elementwise-operator-tester.h",
6150 "test/maximum-nd.cc",
6151 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006152 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006153)
6154
6155xnnpack_unit_test(
6156 name = "minimum_nd_test",
6157 srcs = [
6158 "test/binary-elementwise-operator-tester.h",
6159 "test/minimum-nd.cc",
6160 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006161 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006162)
6163
6164xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006165 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006166 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006167 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006168 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006169 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006170 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006171)
6172
6173xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006174 name = "negate_nc_test",
6175 srcs = [
6176 "test/negate-nc.cc",
6177 "test/negate-operator-tester.h",
6178 ],
6179 deps = OPERATOR_TEST_DEPS,
6180)
6181
6182xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006183 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006184 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006185 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006186 "test/prelu-operator-tester.h",
6187 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006188 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006189)
6190
6191xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006192 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006193 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006194 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006195 "test/resize-bilinear-operator-tester.h",
6196 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006197 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006198)
6199
6200xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006201 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006202 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006203 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204 "test/sigmoid-operator-tester.h",
6205 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006206 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207)
6208
6209xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006210 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006211 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006212 "test/softmax-nc.cc",
6213 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006215 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216)
6217
6218xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006219 name = "square_nc_test",
6220 srcs = [
6221 "test/square-nc.cc",
6222 "test/square-operator-tester.h",
6223 ],
6224 deps = OPERATOR_TEST_DEPS,
6225)
6226
6227xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006228 name = "square_root_nc_test",
6229 srcs = [
6230 "test/square-root-nc.cc",
6231 "test/square-root-operator-tester.h",
6232 ],
6233 deps = OPERATOR_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006237 name = "squared_difference_nd_test",
6238 srcs = [
6239 "test/binary-elementwise-operator-tester.h",
6240 "test/squared-difference-nd.cc",
6241 ],
6242 deps = OPERATOR_TEST_DEPS,
6243)
6244
6245xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006246 name = "subtract_nd_test",
6247 srcs = [
6248 "test/binary-elementwise-operator-tester.h",
6249 "test/subtract-nd.cc",
6250 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006251 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006252)
6253
6254xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006255 name = "truncation_nc_test",
6256 srcs = [
6257 "test/truncation-nc.cc",
6258 "test/truncation-operator-tester.h",
6259 ],
6260 deps = OPERATOR_TEST_DEPS,
6261)
6262
6263xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006264 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006265 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006266 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006267 "test/unpooling-operator-tester.h",
6268 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006269 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270)
6271
Chao Mei6ddfc602020-05-13 22:29:36 -07006272############################### Misc unit tests ###############################
6273
6274xnnpack_unit_test(
6275 name = "memory_planner_test",
6276 srcs = [
6277 "test/memory-planner-test.cc",
6278 ],
6279 deps = [
6280 ":XNNPACK",
6281 ":memory_planner",
6282 ],
6283)
6284
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285############################# Build configurations #############################
6286
Marat Dukhanb8642352019-10-30 15:43:02 -07006287# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006289 name = "xnn_enable_assembly_explicit_true",
6290 define_values = {"xnn_enable_assembly": "true"},
6291)
6292
6293# Disables usage of assembly kernels.
6294config_setting(
6295 name = "xnn_enable_assembly_explicit_false",
6296 define_values = {"xnn_enable_assembly": "false"},
6297)
6298
Marat Dukhan9de90e02020-06-18 16:04:12 -07006299# Enables usage of sparse inference.
6300config_setting(
6301 name = "xnn_enable_sparse_explicit_true",
6302 define_values = {"xnn_enable_sparse": "true"},
6303)
6304
6305# Disables usage of sparse inference.
6306config_setting(
6307 name = "xnn_enable_sparse_explicit_false",
6308 define_values = {"xnn_enable_sparse": "false"},
6309)
6310
Marat Dukhan05702cf2020-03-26 15:41:33 -07006311# Disables usage of HMP-aware optimizations.
6312config_setting(
6313 name = "xnn_enable_hmp_explicit_false",
6314 define_values = {"xnn_enable_hmp": "false"},
6315)
6316
Chao Mei6ddfc602020-05-13 22:29:36 -07006317# Enable usage of optimized memory allocation
6318config_setting(
6319 name = "xnn_enable_memopt_explicit_true",
6320 define_values = {"xnnpack_enable_memopt": "true"},
6321)
6322
6323# Disable usage of optimized memory allocation
6324config_setting(
6325 name = "xnn_enable_memopt_explicit_false",
6326 define_values = {"xnnpack_enable_memopt": "false"},
6327)
6328
Marat Dukhanb8642352019-10-30 15:43:02 -07006329# Builds with -c dbg
6330config_setting(
6331 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006332 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006333 "compilation_mode": "dbg",
6334 },
6335)
6336
6337# Builds with -c opt
6338config_setting(
6339 name = "optimized_build",
6340 values = {
6341 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006342 },
6343)
6344
6345config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006346 name = "linux_k8",
6347 values = {"cpu": "k8"},
6348)
6349
6350config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006351 name = "linux_arm",
6352 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006353)
6354
6355config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006356 name = "linux_armeabi",
6357 values = {"cpu": "armeabi"},
6358)
6359
6360config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006361 name = "linux_armhf",
6362 values = {"cpu": "armhf"},
6363)
6364
6365config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006366 name = "linux_armv7a",
6367 values = {"cpu": "armv7a"},
6368)
6369
6370config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006371 name = "linux_aarch64",
6372 values = {"cpu": "aarch64"},
6373)
6374
6375config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006376 name = "android",
6377 values = {"crosstool_top": "//external:android/crosstool"},
6378)
6379
6380config_setting(
6381 name = "android_armv7",
6382 values = {
6383 "crosstool_top": "//external:android/crosstool",
6384 "cpu": "armeabi-v7a",
6385 },
6386)
6387
6388config_setting(
6389 name = "android_arm64",
6390 values = {
6391 "crosstool_top": "//external:android/crosstool",
6392 "cpu": "arm64-v8a",
6393 },
6394)
6395
6396config_setting(
6397 name = "android_x86",
6398 values = {
6399 "crosstool_top": "//external:android/crosstool",
6400 "cpu": "x86",
6401 },
6402)
6403
6404config_setting(
6405 name = "android_x86_64",
6406 values = {
6407 "crosstool_top": "//external:android/crosstool",
6408 "cpu": "x86_64",
6409 },
6410)
6411
6412config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006413 name = "windows_x86_64",
6414 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006415)
6416
6417config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006418 name = "windows_x86_64_clang",
6419 values = {
6420 "compiler": "clang-cl",
6421 "cpu": "x64_windows",
6422 },
6423)
6424
6425config_setting(
6426 name = "windows_x86_64_mingw",
6427 values = {
6428 "compiler": "mingw-gcc",
6429 "cpu": "x64_windows",
6430 },
6431)
6432
6433config_setting(
6434 name = "windows_x86_64_msys",
6435 values = {
6436 "compiler": "msys-gcc",
6437 "cpu": "x64_windows",
6438 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006439)
6440
6441config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006442 name = "macos_x86_64",
6443 values = {
6444 "apple_platform_type": "macos",
6445 "cpu": "darwin",
6446 },
6447)
6448
6449config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006451 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452)
6453
6454config_setting(
6455 name = "emscripten_wasm",
6456 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006457 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006458 "cpu": "wasm",
6459 },
6460)
6461
6462config_setting(
6463 name = "emscripten_wasmsimd",
6464 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006465 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006466 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006467 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006468 },
6469)
6470
6471config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006472 name = "ios_armv7",
6473 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006474 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006475 "cpu": "ios_armv7",
6476 },
6477)
6478
6479config_setting(
6480 name = "ios_arm64",
6481 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006482 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006483 "cpu": "ios_arm64",
6484 },
6485)
6486
6487config_setting(
6488 name = "ios_arm64e",
6489 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006490 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006491 "cpu": "ios_arm64e",
6492 },
6493)
6494
6495config_setting(
6496 name = "ios_x86",
6497 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006498 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006499 "cpu": "ios_i386",
6500 },
6501)
6502
6503config_setting(
6504 name = "ios_x86_64",
6505 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006506 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006507 "cpu": "ios_x86_64",
6508 },
6509)
6510
6511config_setting(
6512 name = "watchos_armv7k",
6513 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006514 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006515 "cpu": "watchos_armv7k",
6516 },
6517)
6518
6519config_setting(
6520 name = "watchos_arm64_32",
6521 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006522 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006523 "cpu": "watchos_arm64_32",
6524 },
6525)
6526
6527config_setting(
6528 name = "watchos_x86",
6529 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006530 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006531 "cpu": "watchos_i386",
6532 },
6533)
6534
6535config_setting(
6536 name = "watchos_x86_64",
6537 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006538 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006539 "cpu": "watchos_x86_64",
6540 },
6541)
6542
6543config_setting(
6544 name = "tvos_arm64",
6545 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006546 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006547 "cpu": "tvos_arm64",
6548 },
6549)
6550
6551config_setting(
6552 name = "tvos_x86_64",
6553 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006554 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006555 "cpu": "tvos_x86_64",
6556 },
6557)