blob: b85024efed518dff1c59c96ac4b0ad71560005b5 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700425 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001074 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1075 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1076 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1078 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1079 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001082 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001083 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001084 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001085 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001086 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001087 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001088 "src/x32-zip/x2-wasmsimd.c",
1089 "src/x32-zip/x3-wasmsimd.c",
1090 "src/x32-zip/x4-wasmsimd.c",
1091 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001092 "src/math/roundne-wasmsimd-addsub.c",
1093 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001094 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001095 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001099 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1100 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001101]
1102
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001103PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001104 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001105 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001106 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001107 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001108 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001109 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1110 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1111 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/precise-psimd.c",
1116 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001117 "src/qu8-requantization/precise-psimd.c",
1118 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121# ISA-specific micro-kernels
1122NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001123 "src/f32-argmaxpool/4x-neon-c4.c",
1124 "src/f32-argmaxpool/9p8x-neon-c4.c",
1125 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001126 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1127 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001128 "src/f32-clamp/gen/neon-x4.c",
1129 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001130 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1131 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001134 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1135 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1144 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1152 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1154 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1160 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1161 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1162 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1164 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1165 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1166 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1167 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1170 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1171 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1173 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1181 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1182 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001184 "src/f32-hswish/gen/hswish-neon-x4.c",
1185 "src/f32-hswish/gen/hswish-neon-x8.c",
1186 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001187 "src/f32-ibilinear/gen/neon-c4.c",
1188 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1190 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1191 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1193 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1195 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1196 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1197 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1198 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1200 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1201 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1202 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1203 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001205 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1206 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1207 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001210 "src/f32-prelu/gen/neon-1x4.c",
1211 "src/f32-prelu/gen/neon-1x8.c",
1212 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001213 "src/f32-prelu/gen/neon-2x4.c",
1214 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001215 "src/f32-prelu/gen/neon-2x16.c",
1216 "src/f32-prelu/gen/neon-4x4.c",
1217 "src/f32-prelu/gen/neon-4x8.c",
1218 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001243 "src/f32-relu/gen/neon-x4.c",
1244 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001246 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001247 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1248 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1267 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001269 "src/f32-vbinary/gen/vmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1272 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1273 "src/f32-vbinary/gen/vmin-neon-x4.c",
1274 "src/f32-vbinary/gen/vmin-neon-x8.c",
1275 "src/f32-vbinary/gen/vminc-neon-x4.c",
1276 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001283 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1285 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001291 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1292 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001293 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1294 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001295 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1296 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1297 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001303 "src/f32-vunary/gen/vabs-neon-x4.c",
1304 "src/f32-vunary/gen/vabs-neon-x8.c",
1305 "src/f32-vunary/gen/vneg-neon-x4.c",
1306 "src/f32-vunary/gen/vneg-neon-x8.c",
1307 "src/f32-vunary/gen/vsqr-neon-x4.c",
1308 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001309 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1310 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1311 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001313 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1314 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1315 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1317 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1318 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001321 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1322 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1323 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1325 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1326 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001329 "src/qs8-requantization/precise-neon.c",
1330 "src/qs8-requantization/fp32-neon.c",
1331 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001332 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1333 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1334 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1336 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1337 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001340 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1341 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1342 "src/qu8-dwconv/up8x9-minmax-neon.c",
1343 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1344 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1345 "src/qu8-gemm/4x8-minmax-neon.c",
1346 "src/qu8-gemm/8x8-minmax-neon.c",
1347 "src/qu8-igemm/4x8-minmax-neon.c",
1348 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001349 "src/qu8-requantization/precise-neon.c",
1350 "src/qu8-requantization/fp32-neon.c",
1351 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001352 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001355 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001356 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001358 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001359 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360 "src/x32-zip/x2-neon.c",
1361 "src/x32-zip/x3-neon.c",
1362 "src/x32-zip/x4-neon.c",
1363 "src/x32-zip/xm-neon.c",
1364 "src/x8-zip/x2-neon.c",
1365 "src/x8-zip/x3-neon.c",
1366 "src/x8-zip/x4-neon.c",
1367 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001368 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-neon-addsub.c",
1370 "src/math/roundd-neon-cvt.c",
1371 "src/math/roundu-neon-addsub.c",
1372 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neon-addsub.c",
1374 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001375 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001376 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1377 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1378 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1379 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1380 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001382 "src/math/sqrt-neon-nr1rsqrts.c",
1383 "src/math/sqrt-neon-nr2rsqrts.c",
1384 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001385]
1386
1387NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001388 "src/f32-ibilinear/gen/neonfma-c4.c",
1389 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001390 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1391 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1392 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1394 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1395 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1396 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1397 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1405 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1412 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1413 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1416 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1417 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1418 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1424 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1425 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1430 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001509 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1510 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001531 "src/math/exp-neonfma-lut64-p2.c",
1532 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001533 "src/math/expminus-neonfma-lut2048-p1.c",
1534 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001535 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001536 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1537 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1539 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1542 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1543 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1545 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1546 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1548 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1551 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1552 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001554 "src/math/sqrt-neonfma-nr1fma.c",
1555 "src/math/sqrt-neonfma-nr2fma.c",
1556 "src/math/sqrt-neonfma-nr3fma.c",
1557 "src/math/sqrt-neonfma-nr2fma1adj.c",
1558 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001559]
1560
1561AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1569 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1573 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1575 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1579 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1581 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1583 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1585 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1587 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1589 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001592 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1593 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001596 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1598 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1599 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001601 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001619 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1620 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1621 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1622 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1623 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1625 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1626 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1627 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1628 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1630 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1631 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1632 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1633 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1635 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1636 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001637 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1638 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1640 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1641 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1643 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1644 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645]
1646
Marat Dukhan8853b822020-05-07 12:19:01 -07001647NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001648 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1649 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1650 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001656 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001657 "src/math/roundd-neonv8.c",
1658 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001659 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001660]
1661
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001663 "src/f16-clamp/gen/neonfp16arith-x16.c",
1664 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001665 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001677 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1678 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001679 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1680 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001681 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001682 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001683 "src/f16-relu/gen/neonfp16arith-x16.c",
1684 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001685 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001713 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1714 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001715 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001723 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001727 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001739 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1740 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1741 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Benoit Jacoba9644732020-08-13 12:48:55 -07001749NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001750 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001751 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001753 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001754 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001755 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001756 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001757 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1758 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001759 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001760 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001761 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001762]
1763
Marat Dukhan08c4a432019-10-03 09:29:21 -07001764SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001765 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1766 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001767 "src/f32-clamp/gen/sse-x4.c",
1768 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001769 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1770 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001771 "src/f32-dwconv-chw/3x3p1-sse.c",
1772 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001773 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1775 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1777 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1779 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1780 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1781 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1783 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1784 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001785 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001786 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1787 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001788 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1789 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1790 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1791 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1792 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1793 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1794 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1795 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1796 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1797 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1798 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1799 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1800 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001801 "src/f32-hswish/gen/hswish-sse-x4.c",
1802 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001803 "src/f32-ibilinear/gen/sse-c4.c",
1804 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001805 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1806 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1807 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1808 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1809 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1810 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1811 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001812 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1813 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1814 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001815 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001816 "src/f32-prelu/gen/sse-2x4.c",
1817 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001818 "src/f32-relu/gen/sse-x4.c",
1819 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001821 "src/f32-spmm/gen/4x1-minmax-sse.c",
1822 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001823 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001824 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1826 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1829 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1830 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001832 "src/f32-vbinary/gen/vmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vmax-sse-x8.c",
1834 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1835 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1836 "src/f32-vbinary/gen/vmin-sse-x4.c",
1837 "src/f32-vbinary/gen/vmin-sse-x8.c",
1838 "src/f32-vbinary/gen/vminc-sse-x4.c",
1839 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001840 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1843 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1844 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1845 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1846 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001848 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1849 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1850 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1851 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001852 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1854 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1855 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001856 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1857 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001858 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1859 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001860 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1861 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001862 "src/f32-vunary/gen/vabs-sse-x4.c",
1863 "src/f32-vunary/gen/vabs-sse-x8.c",
1864 "src/f32-vunary/gen/vneg-sse-x4.c",
1865 "src/f32-vunary/gen/vneg-sse-x8.c",
1866 "src/f32-vunary/gen/vsqr-sse-x4.c",
1867 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001868 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001869 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001870 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001871 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001872 "src/math/roundd-sse-addsub.c",
1873 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001874 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001875 "src/math/sqrt-sse-hh1mac.c",
1876 "src/math/sqrt-sse-nr1mac.c",
1877 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878]
1879
1880SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001881 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1882 "src/f32-argmaxpool/4x-sse2-c4.c",
1883 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001884 "src/f32-prelu/gen/sse2-2x4.c",
1885 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1896 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1897 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001898 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1899 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1900 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1901 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1902 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1903 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001904 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001905 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001906 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001907 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001908 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1909 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001910 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1911 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001912 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1913 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1914 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1915 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1916 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1917 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1918 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1919 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001920 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1921 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1922 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001923 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1924 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1925 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001926 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1927 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1928 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001929 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001931 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1932 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001933 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1934 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001935 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1936 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001937 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001938 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1939 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001940 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001941 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1942 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1943 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001944 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1945 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001946 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1947 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001948 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1949 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001950 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001951 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1952 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001953 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001954 "src/qs8-requantization/precise-sse2.c",
1955 "src/qs8-requantization/fp32-sse2.c",
1956 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001957 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1958 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1959 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1960 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001961 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1962 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1963 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1964 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001965 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1966 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1967 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1968 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1969 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1970 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1971 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1972 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001973 "src/qu8-requantization/precise-sse2.c",
1974 "src/qu8-requantization/fp32-sse2.c",
1975 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001976 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001977 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001978 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001980 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 "src/x32-zip/x2-sse2.c",
1982 "src/x32-zip/x3-sse2.c",
1983 "src/x32-zip/x4-sse2.c",
1984 "src/x32-zip/xm-sse2.c",
1985 "src/x8-zip/x2-sse2.c",
1986 "src/x8-zip/x3-sse2.c",
1987 "src/x8-zip/x4-sse2.c",
1988 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001989 "src/math/exp-sse2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07001990 "src/math/exp-sse2-lut64-p2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001991 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001992 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001993 "src/math/roundd-sse2-cvt.c",
1994 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001995 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07001996 "src/math/sigmoid-sse2-rr2-p5-div.c",
1997 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001998]
1999
2000SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002001 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2002 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2003 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002004 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2005 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2006 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002007 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2008 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2009 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002010 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2011 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002012 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2013 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002014 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2015 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002016 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2017 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002018 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002019 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2020 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002021 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002022 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2023 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2024 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002025 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2026 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002027 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2028 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002029 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2030 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002031 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002032 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2033 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002034 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002035 "src/qs8-requantization/precise-ssse3.c",
2036 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002037 "src/qu8-requantization/precise-ssse3.c",
2038 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039]
2040
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002041SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002042 "src/f32-prelu/gen/sse41-2x4.c",
2043 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002044 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2045 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2046 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2047 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2048 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2049 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002050 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002051 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002052 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002053 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002054 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2055 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002056 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2057 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002058 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2059 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2060 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2061 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2062 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2063 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2064 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2065 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002066 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2067 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2068 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002069 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2070 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2071 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002072 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2073 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2074 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002075 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2076 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002077 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2078 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002079 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2080 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002081 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2082 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002083 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002084 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2085 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002086 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002087 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2088 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2089 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002090 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2091 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002092 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2093 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002094 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2095 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002096 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002097 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2098 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002099 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002100 "src/qs8-requantization/fp32-sse4.c",
2101 "src/qs8-requantization/precise-sse4.c",
2102 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002103 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2104 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2105 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2106 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002107 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2108 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2109 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2110 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002111 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2112 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2113 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2114 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002115 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2116 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2117 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2118 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002119 "src/qu8-requantization/precise-sse4.c",
2120 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002121 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002122 "src/math/roundd-sse41.c",
2123 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002124 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002125]
2126
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002128 "src/f32-clamp/gen/avx-x8.c",
2129 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002130 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2131 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2132 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2133 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2134 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2135 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2136 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2137 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2138 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2139 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2140 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2141 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2142 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2143 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2144 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2145 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2146 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2147 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2148 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2149 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2150 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2151 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2152 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2153 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2154 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2157 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2158 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2159 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002160 "src/f32-hswish/gen/hswish-avx-x8.c",
2161 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002162 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2163 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2164 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2165 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2166 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2167 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2168 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2169 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2170 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002171 "src/f32-prelu/gen/avx-2x8.c",
2172 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002173 "src/f32-relu/gen/avx-x16.c",
2174 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002176 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2177 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2178 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2179 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2180 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2181 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2182 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2183 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002184 "src/f32-vbinary/gen/vmax-avx-x8.c",
2185 "src/f32-vbinary/gen/vmax-avx-x16.c",
2186 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2187 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2188 "src/f32-vbinary/gen/vmin-avx-x8.c",
2189 "src/f32-vbinary/gen/vmin-avx-x16.c",
2190 "src/f32-vbinary/gen/vminc-avx-x8.c",
2191 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002192 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2193 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2194 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2195 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2196 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2197 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2198 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2199 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002200 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2201 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2202 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2203 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002204 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2205 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2206 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2207 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002208 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2209 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002210 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002211 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2212 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2213 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2214 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2215 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2216 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2217 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2218 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002219 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2220 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002221 "src/f32-vunary/gen/vabs-avx-x8.c",
2222 "src/f32-vunary/gen/vabs-avx-x16.c",
2223 "src/f32-vunary/gen/vneg-avx-x8.c",
2224 "src/f32-vunary/gen/vneg-avx-x16.c",
2225 "src/f32-vunary/gen/vsqr-avx-x8.c",
2226 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002227 "src/math/exp-avx-p5.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002228 "src/math/sigmoid-avx-rr2-p5-div.c",
2229 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230]
2231
Marat Dukhan1566fee2020-08-02 21:55:41 -07002232XOP_UKERNELS = [
2233 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2234 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2235 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2236 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002237 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2238 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002239 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2240 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002241 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002242 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2243 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002244 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002245 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2246 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2247 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002248 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2249 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2250 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2251 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2252 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2253 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002254 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002255 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2256 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002257 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002258 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2259 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2260 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2261 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2262 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2263 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2264 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2265 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002266]
2267
Marat Dukhanfda12b82019-11-21 12:27:59 -08002268FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002269 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2270 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2271 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2272 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2273 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2274 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2275 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2276 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2277 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2278 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2279 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2280 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2281 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2282 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2283 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2284 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2285 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2286 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2287 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2288 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2289 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2291 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2292 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2293 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2294 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2295 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2296 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2297 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2298 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2299 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2300 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2301 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2302 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2303 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2304 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2305 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2306 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2307 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2308 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002309 "src/f32-hswish/gen/hswish-fma3-x8.c",
2310 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002311 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2312 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2313 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2314 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2315 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2316 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2317 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2318 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2319 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2320 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2321 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2322 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2323 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2324 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002325 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2326 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2327 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2328 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2329 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2330 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2331 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2332 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002333 "src/math/sqrt-fma3-nr1fma.c",
2334 "src/math/sqrt-fma3-nr2fma.c",
2335 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002336]
2337
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002338AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002339 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2340 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2341 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2342 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2343 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2344 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2345 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2346 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2347 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2348 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2349 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2350 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2351 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2352 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2353 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2354 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2355 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2356 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2357 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2358 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2359 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2360 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2361 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2362 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2363 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2364 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2365 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2366 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2367 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2368 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2369 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002375 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2376 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2377 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2378 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2379 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2380 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2381 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002405 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2406 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2407 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2408 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2409 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2410 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2411 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2417 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2418 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2419 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2420 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2421 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2422 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2423 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2424 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2425 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2426 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2427 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2428 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002429 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2430 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002431 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2432 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2433 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2434 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002435 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2436 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2437 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002438 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2439 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2440 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002441 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2442 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2443 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002444 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2445 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2446 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2447 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2448 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2449 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2450 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2451 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002452 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002453 "src/math/exp-avx2-lut8-p3-perm.c",
2454 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002455 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002456 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002457 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002458 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002459 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002460 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002461 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002462 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002463]
2464
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002466 "src/f32-clamp/gen/avx512f-x16.c",
2467 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002468 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2469 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2470 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2471 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2472 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2473 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2474 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2475 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2476 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2477 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2478 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2479 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2480 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2481 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2482 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2483 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2484 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2485 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2486 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2487 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2488 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2489 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2490 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2491 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002492 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2493 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002494 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2495 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2496 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2497 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2498 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2499 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002500 "src/f32-prelu/gen/avx512f-2x16.c",
2501 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002502 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2503 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2504 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2505 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2506 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2507 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2508 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2509 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2510 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2511 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2512 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2513 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2514 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2515 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2516 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2517 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2518 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2519 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2520 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2521 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2522 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2523 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2524 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2525 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2526 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2527 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2528 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2529 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2530 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2531 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2532 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2533 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2534 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2535 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2536 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2537 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002538 "src/f32-relu/gen/avx512f-x16.c",
2539 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002540 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002541 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2542 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2543 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2544 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2545 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2546 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2547 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2548 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2549 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2550 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2551 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2552 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2553 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2554 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2555 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2556 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2557 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2558 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2559 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2560 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2561 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2562 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2563 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2564 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2565 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2566 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2567 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2568 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2569 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2570 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2571 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2572 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2573 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2574 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2575 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2576 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2577 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2578 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2579 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2580 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2581 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2582 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2583 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2584 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2585 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2586 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2587 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2588 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002589 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2590 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2591 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2592 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2593 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2594 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2595 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2596 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002597 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2598 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2599 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2600 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2601 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2602 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2603 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2604 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002605 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2606 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2607 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2608 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2609 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2610 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2611 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2612 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002613 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2614 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2615 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2616 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002617 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2618 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2619 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2620 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002621 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2622 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002623 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002624 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2625 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2626 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2627 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2628 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2629 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2630 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2631 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2632 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2633 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2634 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2635 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2636 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2637 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2638 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2639 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2640 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2641 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2642 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2643 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2644 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2645 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2646 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2647 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002648 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2649 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2650 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2651 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2652 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2653 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2654 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2655 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002656 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2657 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2658 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2659 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2660 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2661 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2662 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2663 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002664 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2665 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2666 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2667 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2668 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2669 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002670 "src/math/exp-avx512f-p5-scalef.c",
2671 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002672 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002673 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002674 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002675 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002676 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002677 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002678 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002679 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002680 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002681 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002682 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002683 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2684 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2685 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2686 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2687 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2688 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2689 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2690 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2691 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2692 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2693 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2694 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002695 "src/math/sqrt-avx512f-nr1fma.c",
2696 "src/math/sqrt-avx512f-nr2fma.c",
2697 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698]
2699
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002700AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002701 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2702 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002703 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2704 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2705 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2706 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2707 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2708 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2709 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2710 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2711]
2712
Frank Barchardbcedc082020-08-17 18:00:51 -07002713WASM32_ASM_UKERNELS = [
2714 "src/f32-relu/wasm_shr_x1.S",
2715 "src/f32-relu/wasm_shr_x2.S",
2716 "src/f32-relu/wasm_shr_x4.S",
2717]
2718
Marat Dukhan08c4a432019-10-03 09:29:21 -07002719AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002720 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002721 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002722 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2723 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002724 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002725 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002726 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002727 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002728 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2729 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002730 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2731 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2732 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2733 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002734]
2735
2736AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002737 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002738 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2739 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002740 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002741 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2742 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002743 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2744 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2745 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002746 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002747 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2748 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2749 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002750 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2752 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002753 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002754 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2755 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2756 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2757 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2758 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2759 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2760 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2761 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2762 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2763 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2764 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2765 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2766 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2767 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2768 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2769 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2770 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2771 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002772 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2773 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002774 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002775 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2776 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2777 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2778 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2779 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2780 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2781 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2782 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2783 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2784 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2785 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2786 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2787 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2788 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2789 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2790 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2791 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2792 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002793 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2794 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2795 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2796 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2797 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2798 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2799 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2800 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2801 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2802 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2803 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2804 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2805 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2806 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2807 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2808 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2809 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2810 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002811]
2812
Marat Dukhan1b354632020-03-23 12:50:22 -07002813INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 "src/xnnpack/argmaxpool.h",
2815 "src/xnnpack/avgpool.h",
2816 "src/xnnpack/clamp.h",
2817 "src/xnnpack/common.h",
2818 "src/xnnpack/conv.h",
2819 "src/xnnpack/dwconv.h",
2820 "src/xnnpack/gavgpool.h",
2821 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002822 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002824 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002825 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002826 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002827 "src/xnnpack/lut.h",
2828 "src/xnnpack/math.h",
2829 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002830 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831 "src/xnnpack/packx.h",
2832 "src/xnnpack/pad.h",
2833 "src/xnnpack/params.h",
2834 "src/xnnpack/pavgpool.h",
2835 "src/xnnpack/ppmm.h",
2836 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002837 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002838 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002839 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 "src/xnnpack/rmax.h",
2841 "src/xnnpack/scalar-utils.h",
2842 "src/xnnpack/spmm.h",
2843 "src/xnnpack/unpool.h",
2844 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002845 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002846 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002847 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002848 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002849 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002850 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002852]
2853
2854INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002855 "include/xnnpack.h",
2856 "src/xnnpack/allocator.h",
2857 "src/xnnpack/compute.h",
2858 "src/xnnpack/im2col.h",
2859 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002860 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002861 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862 "src/xnnpack/operator.h",
2863 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002864 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002866 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002867 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002868]
2869
Marat Dukhan1b354632020-03-23 12:50:22 -07002870ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002871 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872]
2873
Marat Dukhan1b354632020-03-23 12:50:22 -07002874MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002875 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002876 "include/xnnpack.h",
2877]
2878
Marat Dukhan1b354632020-03-23 12:50:22 -07002879MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002881 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 "src/xnnpack/requantization.h",
2883 "include/xnnpack.h",
2884]
2885
2886OPERATOR_TEST_PARAMS_HDRS = [
2887 "src/xnnpack/params.h",
2888 "src/xnnpack/common.h",
2889]
2890
2891WEIGHTS_PACK_HDRS = [
2892 "src/xnnpack/pack.h",
2893 "src/xnnpack/operator.h",
2894 "src/xnnpack/compute.h",
2895]
2896
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002897LOGGING_COPTS = select({
2898 # No logging in optimized mode
2899 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2900 # Full logging in debug mode
2901 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2902 # Error-only logging in default (fastbuild) mode
2903 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2904})
2905
Marat Dukhan3b59de22020-06-03 20:15:19 -07002906LOGGING_SRCS = select({
2907 # No logging in optimized mode
2908 ":optimized_build": [],
2909 "//conditions:default": [
2910 "src/operator-strings.c",
2911 "src/subgraph-strings.c",
2912 ],
2913})
2914
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002915LOGGING_HDRS = [
2916 "src/xnnpack/log.h",
2917]
2918
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002920 name = "tables",
2921 srcs = TABLE_SRCS,
2922 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002923 gcc_copts = xnnpack_gcc_std_copts(),
2924 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002925)
2926
2927xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 name = "scalar_ukernels",
2929 srcs = SCALAR_UKERNELS,
2930 hdrs = INTERNAL_HDRS,
2931 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002932 gcc_copts = xnnpack_gcc_std_copts(),
2933 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002935 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "@FP16",
2937 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002938 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 ],
2940)
2941
2942xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002943 name = "scalar_ukernels_test_mode",
2944 srcs = SCALAR_UKERNELS,
2945 hdrs = INTERNAL_HDRS,
2946 aarch32_copts = ["-marm"],
2947 copts = [
2948 "-UNDEBUG",
2949 "-DXNN_TEST_MODE=1",
2950 ],
2951 gcc_copts = xnnpack_gcc_std_copts(),
2952 msvc_copts = xnnpack_msvc_std_copts(),
2953 deps = [
2954 ":tables",
2955 "@FP16",
2956 "@FXdiv",
2957 "@pthreadpool",
2958 ],
2959)
2960
2961xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002962 name = "wasm_ukernels",
2963 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002964 gcc_copts = xnnpack_gcc_std_copts(),
2965 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002966 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002967 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002968 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002969 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002970 "@FP16",
2971 "@FXdiv",
2972 "@pthreadpool",
2973 ],
2974)
2975
2976xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002977 name = "wasm_ukernels_test_mode",
2978 hdrs = INTERNAL_HDRS,
2979 copts = [
2980 "-UNDEBUG",
2981 "-DXNN_TEST_MODE=1",
2982 ],
2983 gcc_copts = xnnpack_gcc_std_copts(),
2984 msvc_copts = xnnpack_msvc_std_copts(),
2985 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002986 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002987 deps = [
2988 ":tables",
2989 "@FP16",
2990 "@FXdiv",
2991 "@pthreadpool",
2992 ],
2993)
2994
2995xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002996 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002997 hdrs = INTERNAL_HDRS,
2998 aarch32_copts = [
2999 "-marm",
3000 "-mfpu=neon",
3001 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003002 gcc_copts = xnnpack_gcc_std_copts(),
3003 gcc_x86_copts = ["-msse2"],
3004 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 optimized_copts = [
3006 "-O3",
3007 "-ffast-math",
3008 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003009 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003011 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003012 "@FP16",
3013 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003014 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 ],
3016)
3017
3018xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003019 name = "psimd_fastmath_ukernels_test_mode",
3020 hdrs = INTERNAL_HDRS,
3021 aarch32_copts = [
3022 "-marm",
3023 "-mfpu=neon",
3024 ],
3025 copts = [
3026 "-UNDEBUG",
3027 "-DXNN_TEST_MODE=1",
3028 ],
3029 gcc_copts = xnnpack_gcc_std_copts(),
3030 gcc_x86_copts = ["-msse2"],
3031 msvc_copts = xnnpack_msvc_std_copts(),
3032 optimized_copts = [
3033 "-O3",
3034 "-ffast-math",
3035 ],
3036 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3037 deps = [
3038 ":tables",
3039 "@FP16",
3040 "@psimd",
3041 "@pthreadpool",
3042 ],
3043)
3044
3045xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003046 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003047 hdrs = INTERNAL_HDRS,
3048 aarch32_copts = [
3049 "-marm",
3050 "-mfpu=neon",
3051 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003052 gcc_copts = xnnpack_gcc_std_copts(),
3053 gcc_x86_copts = ["-msse2"],
3054 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003055 optimized_copts = [
3056 "-O3",
3057 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003058 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003059 deps = [
3060 ":tables",
3061 "@FP16",
3062 "@psimd",
3063 "@pthreadpool",
3064 ],
3065)
3066
3067xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003068 name = "psimd_accmath_ukernels_test_mode",
3069 hdrs = INTERNAL_HDRS,
3070 aarch32_copts = [
3071 "-marm",
3072 "-mfpu=neon",
3073 ],
3074 copts = [
3075 "-UNDEBUG",
3076 "-DXNN_TEST_MODE=1",
3077 ],
3078 gcc_copts = xnnpack_gcc_std_copts(),
3079 gcc_x86_copts = ["-msse2"],
3080 msvc_copts = xnnpack_msvc_std_copts(),
3081 optimized_copts = [
3082 "-O3",
3083 ],
3084 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3085 deps = [
3086 ":tables",
3087 "@FP16",
3088 "@psimd",
3089 "@pthreadpool",
3090 ],
3091)
3092
3093xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 name = "neon_ukernels",
3095 hdrs = INTERNAL_HDRS,
3096 aarch32_copts = [
3097 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003098 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003099 "-mfpu=neon",
3100 ],
3101 aarch32_srcs = NEON_UKERNELS,
3102 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003103 gcc_copts = xnnpack_gcc_std_copts(),
3104 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003105 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003106 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003107 "@FP16",
3108 "@pthreadpool",
3109 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110)
3111
3112xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003113 name = "neon_ukernels_test_mode",
3114 hdrs = INTERNAL_HDRS,
3115 aarch32_copts = [
3116 "-marm",
3117 "-march=armv7-a",
3118 "-mfpu=neon",
3119 ],
3120 aarch32_srcs = NEON_UKERNELS,
3121 aarch64_srcs = NEON_UKERNELS,
3122 copts = [
3123 "-UNDEBUG",
3124 "-DXNN_TEST_MODE=1",
3125 ],
3126 gcc_copts = xnnpack_gcc_std_copts(),
3127 msvc_copts = xnnpack_msvc_std_copts(),
3128 deps = [
3129 ":tables",
3130 "@FP16",
3131 "@pthreadpool",
3132 ],
3133)
3134
3135xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 name = "neonfma_ukernels",
3137 hdrs = INTERNAL_HDRS,
3138 aarch32_copts = [
3139 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003140 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003141 "-mfpu=neon-vfpv4",
3142 ],
3143 aarch32_srcs = NEONFMA_UKERNELS,
3144 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003145 apple_aarch32_copts = [
3146 "-mcpu=swift",
3147 "-mtune=generic",
3148 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003149 gcc_copts = xnnpack_gcc_std_copts(),
3150 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003151 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003152 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003153 "@FP16",
3154 "@pthreadpool",
3155 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156)
3157
3158xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003159 name = "neonfma_ukernels_test_mode",
3160 hdrs = INTERNAL_HDRS,
3161 aarch32_copts = [
3162 "-marm",
3163 "-march=armv7-a",
3164 "-mfpu=neon-vfpv4",
3165 ],
3166 aarch32_srcs = NEONFMA_UKERNELS,
3167 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003168 apple_aarch32_copts = [
3169 "-mcpu=swift",
3170 "-mtune=generic",
3171 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003172 copts = [
3173 "-UNDEBUG",
3174 "-DXNN_TEST_MODE=1",
3175 ],
3176 gcc_copts = xnnpack_gcc_std_copts(),
3177 msvc_copts = xnnpack_msvc_std_copts(),
3178 deps = [
3179 ":tables",
3180 "@FP16",
3181 "@pthreadpool",
3182 ],
3183)
3184
3185xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003186 name = "neonv8_ukernels",
3187 hdrs = INTERNAL_HDRS,
3188 aarch32_copts = [
3189 "-marm",
3190 "-march=armv8-a",
3191 "-mfpu=neon-fp-armv8",
3192 ],
3193 aarch32_srcs = NEONV8_UKERNELS,
3194 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003195 apple_aarch32_copts = [
3196 "-mcpu=cyclone",
3197 "-mtune=generic",
3198 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003199 gcc_copts = xnnpack_gcc_std_copts(),
3200 msvc_copts = xnnpack_msvc_std_copts(),
3201 deps = [
3202 ":tables",
3203 "@FP16",
3204 "@pthreadpool",
3205 ],
3206)
3207
3208xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003209 name = "neonv8_ukernels_test_mode",
3210 hdrs = INTERNAL_HDRS,
3211 aarch32_copts = [
3212 "-marm",
3213 "-march=armv8-a",
3214 "-mfpu=neon-fp-armv8",
3215 ],
3216 aarch32_srcs = NEONV8_UKERNELS,
3217 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003218 apple_aarch32_copts = [
3219 "-mcpu=cyclone",
3220 "-mtune=generic",
3221 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003222 copts = [
3223 "-UNDEBUG",
3224 "-DXNN_TEST_MODE=1",
3225 ],
3226 gcc_copts = xnnpack_gcc_std_copts(),
3227 msvc_copts = xnnpack_msvc_std_copts(),
3228 deps = [
3229 ":tables",
3230 "@FP16",
3231 "@pthreadpool",
3232 ],
3233)
3234
3235xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236 name = "neonfp16arith_ukernels",
3237 hdrs = INTERNAL_HDRS,
3238 aarch64_copts = ["-march=armv8.2-a+fp16"],
3239 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003240 gcc_copts = xnnpack_gcc_std_copts(),
3241 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003242 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003243 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003244 "@FP16",
3245 "@pthreadpool",
3246 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247)
3248
3249xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003250 name = "neonfp16arith_ukernels_test_mode",
3251 hdrs = INTERNAL_HDRS,
3252 aarch64_copts = ["-march=armv8.2-a+fp16"],
3253 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3254 copts = [
3255 "-UNDEBUG",
3256 "-DXNN_TEST_MODE=1",
3257 ],
3258 gcc_copts = xnnpack_gcc_std_copts(),
3259 msvc_copts = xnnpack_msvc_std_copts(),
3260 deps = [
3261 ":tables",
3262 "@FP16",
3263 "@pthreadpool",
3264 ],
3265)
3266
3267xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003268 name = "neondot_ukernels",
3269 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003270 aarch32_copts = [
3271 "-marm",
3272 "-march=armv8.2-a+dotprod",
3273 "-mfpu=neon-fp-armv8",
3274 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003275 aarch32_srcs = NEONDOT_UKERNELS,
3276 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3277 aarch64_srcs = NEONDOT_UKERNELS,
3278 gcc_copts = xnnpack_gcc_std_copts(),
3279 msvc_copts = xnnpack_msvc_std_copts(),
3280 deps = [
3281 ":tables",
3282 "@FP16",
3283 "@pthreadpool",
3284 ],
3285)
3286
3287xnnpack_cc_library(
3288 name = "neondot_ukernels_test_mode",
3289 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003290 aarch32_copts = [
3291 "-marm",
3292 "-march=armv8.2-a+dotprod",
3293 "-mfpu=neon-fp-armv8",
3294 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003295 aarch32_srcs = NEONDOT_UKERNELS,
3296 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3297 aarch64_srcs = NEONDOT_UKERNELS,
3298 copts = [
3299 "-UNDEBUG",
3300 "-DXNN_TEST_MODE=1",
3301 ],
3302 gcc_copts = xnnpack_gcc_std_copts(),
3303 msvc_copts = xnnpack_msvc_std_copts(),
3304 deps = [
3305 ":tables",
3306 "@FP16",
3307 "@pthreadpool",
3308 ],
3309)
3310
3311xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003312 name = "sse2_ukernels",
3313 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003314 gcc_copts = xnnpack_gcc_std_copts(),
3315 gcc_x86_copts = ["-msse2"],
3316 msvc_copts = xnnpack_msvc_std_copts(),
3317 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003318 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003319 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003320 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003321 "@FP16",
3322 "@pthreadpool",
3323 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003324)
3325
3326xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003327 name = "sse2_ukernels_test_mode",
3328 hdrs = INTERNAL_HDRS,
3329 copts = [
3330 "-UNDEBUG",
3331 "-DXNN_TEST_MODE=1",
3332 ],
3333 gcc_copts = xnnpack_gcc_std_copts(),
3334 gcc_x86_copts = ["-msse2"],
3335 msvc_copts = xnnpack_msvc_std_copts(),
3336 msvc_x86_32_copts = ["/arch:SSE2"],
3337 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3338 deps = [
3339 ":tables",
3340 "@FP16",
3341 "@pthreadpool",
3342 ],
3343)
3344
3345xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003346 name = "ssse3_ukernels",
3347 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003348 gcc_copts = xnnpack_gcc_std_copts(),
3349 gcc_x86_copts = ["-mssse3"],
3350 msvc_copts = xnnpack_msvc_std_copts(),
3351 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003352 x86_srcs = SSSE3_UKERNELS,
3353 deps = [
3354 ":tables",
3355 "@FP16",
3356 "@pthreadpool",
3357 ],
3358)
3359
3360xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003361 name = "ssse3_ukernels_test_mode",
3362 hdrs = INTERNAL_HDRS,
3363 copts = [
3364 "-UNDEBUG",
3365 "-DXNN_TEST_MODE=1",
3366 ],
3367 gcc_copts = xnnpack_gcc_std_copts(),
3368 gcc_x86_copts = ["-mssse3"],
3369 msvc_copts = xnnpack_msvc_std_copts(),
3370 msvc_x86_32_copts = ["/arch:SSE2"],
3371 x86_srcs = SSSE3_UKERNELS,
3372 deps = [
3373 ":tables",
3374 "@FP16",
3375 "@pthreadpool",
3376 ],
3377)
3378
3379xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003380 name = "sse41_ukernels",
3381 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003382 gcc_copts = xnnpack_gcc_std_copts(),
3383 gcc_x86_copts = ["-msse4.1"],
3384 msvc_copts = xnnpack_msvc_std_copts(),
3385 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003386 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003387 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003388 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003389 "@FP16",
3390 "@pthreadpool",
3391 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003392)
3393
3394xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003395 name = "sse41_ukernels_test_mode",
3396 hdrs = INTERNAL_HDRS,
3397 copts = [
3398 "-UNDEBUG",
3399 "-DXNN_TEST_MODE=1",
3400 ],
3401 gcc_copts = xnnpack_gcc_std_copts(),
3402 gcc_x86_copts = ["-msse4.1"],
3403 msvc_copts = xnnpack_msvc_std_copts(),
3404 msvc_x86_32_copts = ["/arch:SSE2"],
3405 x86_srcs = SSE41_UKERNELS,
3406 deps = [
3407 ":tables",
3408 "@FP16",
3409 "@pthreadpool",
3410 ],
3411)
3412
3413xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414 name = "avx_ukernels",
3415 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003416 gcc_copts = xnnpack_gcc_std_copts(),
3417 gcc_x86_copts = ["-mavx"],
3418 msvc_copts = xnnpack_msvc_std_copts(),
3419 msvc_x86_32_copts = ["/arch:AVX"],
3420 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003422 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003423 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003424 "@FP16",
3425 "@pthreadpool",
3426 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427)
3428
3429xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003430 name = "avx_ukernels_test_mode",
3431 hdrs = INTERNAL_HDRS,
3432 copts = [
3433 "-UNDEBUG",
3434 "-DXNN_TEST_MODE=1",
3435 ],
3436 gcc_copts = xnnpack_gcc_std_copts(),
3437 gcc_x86_copts = ["-mavx"],
3438 msvc_copts = xnnpack_msvc_std_copts(),
3439 msvc_x86_32_copts = ["/arch:AVX"],
3440 msvc_x86_64_copts = ["/arch:AVX"],
3441 x86_srcs = AVX_UKERNELS,
3442 deps = [
3443 ":tables",
3444 "@FP16",
3445 "@pthreadpool",
3446 ],
3447)
3448
3449xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003450 name = "xop_ukernels",
3451 hdrs = INTERNAL_HDRS,
3452 gcc_copts = xnnpack_gcc_std_copts(),
3453 gcc_x86_copts = ["-mxop"],
3454 msvc_copts = xnnpack_msvc_std_copts(),
3455 msvc_x86_32_copts = ["/arch:AVX"],
3456 msvc_x86_64_copts = ["/arch:AVX"],
3457 x86_srcs = XOP_UKERNELS,
3458 deps = [
3459 ":tables",
3460 "@FP16",
3461 "@pthreadpool",
3462 ],
3463)
3464
3465xnnpack_cc_library(
3466 name = "xop_ukernels_test_mode",
3467 hdrs = INTERNAL_HDRS,
3468 copts = [
3469 "-UNDEBUG",
3470 "-DXNN_TEST_MODE=1",
3471 ],
3472 gcc_copts = xnnpack_gcc_std_copts(),
3473 gcc_x86_copts = ["-mxop"],
3474 msvc_copts = xnnpack_msvc_std_copts(),
3475 msvc_x86_32_copts = ["/arch:AVX"],
3476 msvc_x86_64_copts = ["/arch:AVX"],
3477 x86_srcs = XOP_UKERNELS,
3478 deps = [
3479 ":tables",
3480 "@FP16",
3481 "@pthreadpool",
3482 ],
3483)
3484
3485xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003486 name = "fma3_ukernels",
3487 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003488 gcc_copts = xnnpack_gcc_std_copts(),
3489 gcc_x86_copts = ["-mfma"],
3490 msvc_copts = xnnpack_msvc_std_copts(),
3491 msvc_x86_32_copts = ["/arch:AVX"],
3492 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003493 x86_srcs = FMA3_UKERNELS,
3494 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003495 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003496 "@FP16",
3497 "@pthreadpool",
3498 ],
3499)
3500
3501xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003502 name = "fma3_ukernels_test_mode",
3503 hdrs = INTERNAL_HDRS,
3504 copts = [
3505 "-UNDEBUG",
3506 "-DXNN_TEST_MODE=1",
3507 ],
3508 gcc_copts = xnnpack_gcc_std_copts(),
3509 gcc_x86_copts = ["-mfma"],
3510 msvc_copts = xnnpack_msvc_std_copts(),
3511 msvc_x86_32_copts = ["/arch:AVX"],
3512 msvc_x86_64_copts = ["/arch:AVX"],
3513 x86_srcs = FMA3_UKERNELS,
3514 deps = [
3515 ":tables",
3516 "@FP16",
3517 "@pthreadpool",
3518 ],
3519)
3520
3521xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003522 name = "avx2_ukernels",
3523 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003524 gcc_copts = xnnpack_gcc_std_copts(),
3525 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003526 "-mfma",
3527 "-mavx2",
3528 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003529 msvc_copts = xnnpack_msvc_std_copts(),
3530 msvc_x86_32_copts = ["/arch:AVX2"],
3531 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003532 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003533 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003534 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003535 "@FP16",
3536 "@pthreadpool",
3537 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003538)
3539
3540xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003541 name = "avx2_ukernels_test_mode",
3542 hdrs = INTERNAL_HDRS,
3543 copts = [
3544 "-UNDEBUG",
3545 "-DXNN_TEST_MODE=1",
3546 ],
3547 gcc_copts = xnnpack_gcc_std_copts(),
3548 gcc_x86_copts = [
3549 "-mfma",
3550 "-mavx2",
3551 ],
3552 msvc_copts = xnnpack_msvc_std_copts(),
3553 msvc_x86_32_copts = ["/arch:AVX2"],
3554 msvc_x86_64_copts = ["/arch:AVX2"],
3555 x86_srcs = AVX2_UKERNELS,
3556 deps = [
3557 ":tables",
3558 "@FP16",
3559 "@pthreadpool",
3560 ],
3561)
3562
3563xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 name = "avx512f_ukernels",
3565 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003566 gcc_copts = xnnpack_gcc_std_copts(),
3567 gcc_x86_copts = ["-mavx512f"],
3568 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3569 msvc_copts = xnnpack_msvc_std_copts(),
3570 msvc_x86_32_copts = ["/arch:AVX512"],
3571 msvc_x86_64_copts = ["/arch:AVX512"],
3572 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003574 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003575 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003576 "@FP16",
3577 "@pthreadpool",
3578 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579)
3580
3581xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003582 name = "avx512f_ukernels_test_mode",
3583 hdrs = INTERNAL_HDRS,
3584 copts = [
3585 "-UNDEBUG",
3586 "-DXNN_TEST_MODE=1",
3587 ],
3588 gcc_copts = xnnpack_gcc_std_copts(),
3589 gcc_x86_copts = ["-mavx512f"],
3590 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3591 msvc_copts = xnnpack_msvc_std_copts(),
3592 msvc_x86_32_copts = ["/arch:AVX512"],
3593 msvc_x86_64_copts = ["/arch:AVX512"],
3594 msys_copts = ["-fno-asynchronous-unwind-tables"],
3595 x86_srcs = AVX512F_UKERNELS,
3596 deps = [
3597 ":tables",
3598 "@FP16",
3599 "@pthreadpool",
3600 ],
3601)
3602
3603xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003604 name = "avx512skx_ukernels",
3605 hdrs = INTERNAL_HDRS,
3606 gcc_copts = xnnpack_gcc_std_copts(),
3607 gcc_x86_copts = [
3608 "-mavx512f",
3609 "-mavx512cd",
3610 "-mavx512bw",
3611 "-mavx512dq",
3612 "-mavx512vl",
3613 ],
3614 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3615 msvc_copts = xnnpack_msvc_std_copts(),
3616 msvc_x86_32_copts = ["/arch:AVX512"],
3617 msvc_x86_64_copts = ["/arch:AVX512"],
3618 msys_copts = ["-fno-asynchronous-unwind-tables"],
3619 x86_srcs = AVX512SKX_UKERNELS,
3620 deps = [
3621 ":tables",
3622 "@FP16",
3623 "@pthreadpool",
3624 ],
3625)
3626
3627xnnpack_cc_library(
3628 name = "avx512skx_ukernels_test_mode",
3629 hdrs = INTERNAL_HDRS,
3630 copts = [
3631 "-UNDEBUG",
3632 "-DXNN_TEST_MODE=1",
3633 ],
3634 gcc_copts = xnnpack_gcc_std_copts(),
3635 gcc_x86_copts = [
3636 "-mavx512f",
3637 "-mavx512cd",
3638 "-mavx512bw",
3639 "-mavx512dq",
3640 "-mavx512vl",
3641 ],
3642 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3643 msvc_copts = xnnpack_msvc_std_copts(),
3644 msvc_x86_32_copts = ["/arch:AVX512"],
3645 msvc_x86_64_copts = ["/arch:AVX512"],
3646 msys_copts = ["-fno-asynchronous-unwind-tables"],
3647 x86_srcs = AVX512SKX_UKERNELS,
3648 deps = [
3649 ":tables",
3650 "@FP16",
3651 "@pthreadpool",
3652 ],
3653)
3654
3655xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656 name = "asm_ukernels",
3657 hdrs = ["src/xnnpack/assembly.h"],
3658 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003659 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003661 wasm_srcs = WASM32_ASM_UKERNELS,
3662 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663)
3664
Marat Dukhan3b59de22020-06-03 20:15:19 -07003665xnnpack_cc_library(
3666 name = "logging_utils",
3667 srcs = LOGGING_SRCS,
3668 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3669 copts = LOGGING_COPTS + [
3670 "-Isrc",
3671 "-Iinclude",
3672 ] + select({
3673 ":debug_build": [],
3674 "//conditions:default": xnnpack_min_size_copts(),
3675 }),
3676 gcc_copts = xnnpack_gcc_std_copts(),
3677 msvc_copts = xnnpack_msvc_std_copts(),
3678 visibility = xnnpack_visibility(),
3679 deps = [
3680 "@FP16",
3681 "@clog",
3682 "@pthreadpool",
3683 ],
3684)
3685
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686xnnpack_aggregate_library(
3687 name = "ukernels",
3688 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003689 ":neon_ukernels",
3690 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003691 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003692 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693 ":asm_ukernels",
3694 ],
3695 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 ":neon_ukernels",
3697 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003698 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003700 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701 ":asm_ukernels",
3702 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003703 generic_deps = [
3704 ":scalar_ukernels",
3705 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003706 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003707 ":psimd_fastmath_ukernels",
3708 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003710 wasm_deps = [
3711 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003712 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003713 ],
3714 wasmsimd_deps = [
3715 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003716 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003717 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003720 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003721 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003722 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003723 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003724 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003725 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003727 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 ],
3729)
3730
Marat Dukhan33fcf782020-05-24 14:27:15 -07003731xnnpack_aggregate_library(
3732 name = "ukernels_test_mode",
3733 aarch32_deps = [
3734 ":neon_ukernels_test_mode",
3735 ":neonfma_ukernels_test_mode",
3736 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003737 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003738 ":asm_ukernels",
3739 ],
3740 aarch64_deps = [
3741 ":neon_ukernels_test_mode",
3742 ":neonfma_ukernels_test_mode",
3743 ":neonv8_ukernels_test_mode",
3744 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003745 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003746 ":asm_ukernels",
3747 ],
3748 generic_deps = [
3749 ":scalar_ukernels_test_mode",
3750 ],
3751 psimd_deps = [
3752 ":psimd_fastmath_ukernels_test_mode",
3753 ":psimd_accmath_ukernels_test_mode",
3754 ],
3755 wasm_deps = [
3756 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003757 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003758 ],
3759 wasmsimd_deps = [
3760 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003761 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003762 ],
3763 x86_deps = [
3764 ":sse2_ukernels_test_mode",
3765 ":ssse3_ukernels_test_mode",
3766 ":sse41_ukernels_test_mode",
3767 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003768 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003769 ":fma3_ukernels_test_mode",
3770 ":avx2_ukernels_test_mode",
3771 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003772 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003773 ],
3774)
3775
Marat Dukhan86df3842020-06-17 14:03:31 -07003776# This target must be compatible with Android NDK r17 and compile with Clang 6.
3777# It is similar to :ukernels target, but excludes microkernels which require
3778# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003779xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003780 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003781 aarch32_deps = [
3782 ":neon_ukernels",
3783 ":neonfma_ukernels",
3784 ":neonv8_ukernels",
3785 ":asm_ukernels",
3786 ],
3787 aarch64_deps = [
3788 ":neon_ukernels",
3789 ":neonfma_ukernels",
3790 ":neonv8_ukernels",
3791 ":asm_ukernels",
3792 ],
3793 generic_deps = [
3794 ":scalar_ukernels",
3795 ],
3796 psimd_deps = [
3797 ":psimd_fastmath_ukernels",
3798 ":psimd_accmath_ukernels",
3799 ],
3800 wasm_deps = [
3801 ":wasm_ukernels",
3802 ],
3803 wasmsimd_deps = [
3804 ":wasm_ukernels",
3805 ],
3806 x86_deps = [
3807 ":sse2_ukernels",
3808 ":ssse3_ukernels",
3809 ":sse41_ukernels",
3810 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003811 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003812 ":fma3_ukernels",
3813 ":avx2_ukernels",
3814 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003815 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003816 ],
3817)
3818
Marat Dukhan08c4a432019-10-03 09:29:21 -07003819xnnpack_cc_library(
3820 name = "im2col",
3821 srcs = ["src/im2col.c"],
3822 hdrs = [
3823 "src/xnnpack/common.h",
3824 "src/xnnpack/im2col.h",
3825 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003826 gcc_copts = xnnpack_gcc_std_copts(),
3827 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828)
3829
3830xnnpack_cc_library(
3831 name = "indirection",
3832 srcs = ["src/indirection.c"],
3833 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003834 gcc_copts = xnnpack_gcc_std_copts(),
3835 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 deps = [
3837 "@FP16",
3838 "@FXdiv",
3839 "@pthreadpool",
3840 ],
3841)
3842
3843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003844 name = "indirection_test_mode",
3845 srcs = ["src/indirection.c"],
3846 hdrs = INTERNAL_HDRS,
3847 copts = [
3848 "-UNDEBUG",
3849 "-DXNN_TEST_MODE=1",
3850 ],
3851 gcc_copts = xnnpack_gcc_std_copts(),
3852 msvc_copts = xnnpack_msvc_std_copts(),
3853 deps = [
3854 "@FP16",
3855 "@FXdiv",
3856 "@pthreadpool",
3857 ],
3858)
3859
3860xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003861 name = "packing",
3862 srcs = ["src/packing.c"],
3863 hdrs = INTERNAL_HDRS,
3864 gcc_copts = xnnpack_gcc_std_copts(),
3865 msvc_copts = xnnpack_msvc_std_copts(),
3866 deps = [
3867 "@FP16",
3868 "@FXdiv",
3869 "@pthreadpool",
3870 ],
3871)
3872
3873xnnpack_cc_library(
3874 name = "packing_test_mode",
3875 srcs = ["src/packing.c"],
3876 hdrs = INTERNAL_HDRS,
3877 copts = [
3878 "-UNDEBUG",
3879 "-DXNN_TEST_MODE=1",
3880 ],
3881 gcc_copts = xnnpack_gcc_std_copts(),
3882 msvc_copts = xnnpack_msvc_std_copts(),
3883 deps = [
3884 "@FP16",
3885 "@FXdiv",
3886 "@pthreadpool",
3887 ],
3888)
3889
3890xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003891 name = "operator_run",
3892 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003893 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003894 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003895 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3896 "//conditions:default": [],
3897 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003898 gcc_copts = xnnpack_gcc_std_copts(),
3899 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003900 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003901 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003902 "@FP16",
3903 "@FXdiv",
3904 "@clog",
3905 "@pthreadpool",
3906 ],
3907)
3908
Chao Mei6ddfc602020-05-13 22:29:36 -07003909xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003910 name = "operator_run_test_mode",
3911 srcs = ["src/operator-run.c"],
3912 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3913 copts = LOGGING_COPTS + [
3914 "-UNDEBUG",
3915 "-DXNN_TEST_MODE=1",
3916 ] + select({
3917 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3918 "//conditions:default": [],
3919 }),
3920 gcc_copts = xnnpack_gcc_std_copts(),
3921 msvc_copts = xnnpack_msvc_std_copts(),
3922 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003923 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003924 "@FP16",
3925 "@FXdiv",
3926 "@clog",
3927 "@pthreadpool",
3928 ],
3929)
3930
3931xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003932 name = "memory_planner",
3933 srcs = ["src/memory-planner.c"],
3934 hdrs = INTERNAL_HDRS,
3935 defines = select({
3936 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3937 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3938 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3939 }),
3940 gcc_copts = xnnpack_gcc_std_copts(),
3941 msvc_copts = xnnpack_msvc_std_copts(),
3942 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003943 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003944 "@pthreadpool",
3945 ],
3946)
3947
Marat Dukhan33fcf782020-05-24 14:27:15 -07003948xnnpack_cc_library(
3949 name = "memory_planner_test_mode",
3950 srcs = ["src/memory-planner.c"],
3951 hdrs = INTERNAL_HDRS,
3952 copts = [
3953 "-UNDEBUG",
3954 "-DXNN_TEST_MODE=1",
3955 ],
3956 defines = select({
3957 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3958 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3959 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3960 }),
3961 gcc_copts = xnnpack_gcc_std_copts(),
3962 msvc_copts = xnnpack_msvc_std_copts(),
3963 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003964 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003965 "@pthreadpool",
3966 ],
3967)
3968
Marat Dukhan08c4a432019-10-03 09:29:21 -07003969cc_library(
3970 name = "enable_assembly",
3971 defines = select({
3972 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3973 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003974 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003975 }),
3976)
3977
Marat Dukhan9de90e02020-06-18 16:04:12 -07003978cc_library(
3979 name = "enable_sparse",
3980 defines = select({
3981 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3982 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003983 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003984 }),
3985)
3986
Marat Dukhancf056b22019-10-07 10:26:29 -07003987xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988 name = "operators",
3989 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003990 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003991 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003992 ],
3993 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003994 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003995 "-Isrc",
3996 "-Iinclude",
3997 ] + select({
3998 ":debug_build": [],
3999 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004000 }) + select({
4001 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4002 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004003 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004004 gcc_copts = xnnpack_gcc_std_copts(),
4005 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004006 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004007 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004008 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004009 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004010 "@FP16",
4011 "@FXdiv",
4012 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004013 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004014 ],
4015)
4016
Marat Dukhan10a38082020-04-17 03:58:35 -07004017xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004018 name = "operators_test_mode",
4019 srcs = OPERATOR_SRCS + [
4020 "src/memory.c",
4021 "src/operator-delete.c",
4022 ],
4023 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4024 copts = LOGGING_COPTS + [
4025 "-Isrc",
4026 "-Iinclude",
4027 "-UNDEBUG",
4028 "-DXNN_TEST_MODE=1",
4029 ] + select({
4030 ":debug_build": [],
4031 "//conditions:default": xnnpack_min_size_copts(),
4032 }) + select({
4033 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4034 "//conditions:default": [],
4035 }),
4036 gcc_copts = xnnpack_gcc_std_copts(),
4037 msvc_copts = xnnpack_msvc_std_copts(),
4038 deps = [
4039 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004040 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004041 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004042 "@FP16",
4043 "@FXdiv",
4044 "@clog",
4045 "@pthreadpool",
4046 ],
4047)
4048
4049xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004050 name = "XNNPACK",
4051 srcs = [
4052 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004053 "src/runtime.c",
4054 "src/subgraph.c",
4055 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004056 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004057 hdrs = ["include/xnnpack.h"],
4058 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004059 "-Isrc",
4060 "-Iinclude",
4061 ] + select({
4062 ":debug_build": [],
4063 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004064 }) + select({
4065 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4066 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004067 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004068 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004069 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004070 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004071 visibility = xnnpack_visibility(),
4072 deps = [
4073 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004074 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004075 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004076 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004077 ":operator_run",
4078 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004079 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004080 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004081 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004082 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004083 ] + select({
4084 ":emscripten": [],
4085 "//conditions:default": ["@cpuinfo"],
4086 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004087)
4088
Marat Dukhan10a38082020-04-17 03:58:35 -07004089xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004090 name = "XNNPACK_test_mode",
4091 srcs = [
4092 "src/init.c",
4093 "src/runtime.c",
4094 "src/subgraph.c",
4095 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004096 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004097 hdrs = ["include/xnnpack.h"],
4098 copts = LOGGING_COPTS + [
4099 "-Isrc",
4100 "-Iinclude",
4101 "-UNDEBUG",
4102 "-DXNN_TEST_MODE=1",
4103 ] + select({
4104 ":debug_build": [],
4105 "//conditions:default": xnnpack_min_size_copts(),
4106 }) + select({
4107 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4108 "//conditions:default": [],
4109 }),
4110 gcc_copts = xnnpack_gcc_std_copts(),
4111 includes = ["include"],
4112 msvc_copts = xnnpack_msvc_std_copts(),
4113 visibility = xnnpack_visibility(),
4114 deps = [
4115 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004116 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004117 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004118 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004119 ":operator_run_test_mode",
4120 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004121 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004122 "@clog",
4123 "@FP16",
4124 "@pthreadpool",
4125 ] + select({
4126 ":emscripten": [],
4127 "//conditions:default": ["@cpuinfo"],
4128 }),
4129)
4130
4131xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004132 name = "xnnpack_f32",
4133 srcs = [
4134 "src/init.c",
4135 "src/runtime.c",
4136 "src/subgraph.c",
4137 "src/tensor.c",
4138 ] + SUBGRAPH_SRCS,
4139 hdrs = ["include/xnnpack.h"],
4140 copts = LOGGING_COPTS + [
4141 "-Isrc",
4142 "-Iinclude",
4143 ] + select({
4144 ":debug_build": [],
4145 "//conditions:default": xnnpack_min_size_copts(),
4146 }) + select({
4147 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4148 "//conditions:default": [],
4149 }),
4150 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004151 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004152 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004153 "XNN_NO_U8_OPERATORS",
4154 "XNN_NO_X8_OPERATORS",
4155 "XNN_NO_F16_OPERATORS",
4156 "XNN_NO_X16_OPERATORS",
4157 ],
4158 gcc_copts = xnnpack_gcc_std_copts(),
4159 includes = ["include"],
4160 msvc_copts = xnnpack_msvc_std_copts(),
4161 visibility = xnnpack_visibility(),
4162 deps = [
4163 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004164 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004165 ":logging_utils",
4166 ":memory_planner",
4167 ":operator_run",
4168 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004169 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004170 "@clog",
4171 "@FP16",
4172 "@pthreadpool",
4173 ] + select({
4174 ":emscripten": [],
4175 "//conditions:default": ["@cpuinfo"],
4176 }),
4177)
4178
4179xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004180 name = "xnnpack_operators_nhwc_f32",
4181 srcs = [
4182 "src/init.c",
4183 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004184 hdrs = ["include/xnnpack.h"],
4185 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004186 "-Isrc",
4187 "-Iinclude",
4188 ] + select({
4189 ":debug_build": [],
4190 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004191 }) + select({
4192 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4193 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004194 }),
4195 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004196 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004197 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004198 "XNN_NO_U8_OPERATORS",
4199 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004200 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004201 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004202 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004204 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205 visibility = xnnpack_visibility(),
4206 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004207 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004208 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209 ":operator_run",
4210 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004211 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004212 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004214 ] + select({
4215 ":emscripten": [],
4216 "//conditions:default": ["@cpuinfo"],
4217 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218)
4219
Marat Dukhancf056b22019-10-07 10:26:29 -07004220xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221 name = "bench_utils",
4222 srcs = ["bench/utils.cc"],
4223 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004224 deps = [
4225 "@com_google_benchmark//:benchmark",
4226 "@cpuinfo",
4227 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228)
4229
Frank Barchard7e955972019-10-11 10:34:25 -07004230######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004231
4232xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004233 name = "qs8_gemm_bench",
4234 srcs = [
4235 "bench/gemm.h",
4236 "bench/qs8-gemm.cc",
4237 "src/xnnpack/AlignedAllocator.h",
4238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4239 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4240)
4241
4242xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004243 name = "qs8_requantization_bench",
4244 srcs = [
4245 "bench/qs8-requantization.cc",
4246 "src/xnnpack/requantization-stubs.h",
4247 "src/xnnpack/AlignedAllocator.h",
4248 ] + MICROKERNEL_BENCHMARK_HDRS,
4249 deps = MICROKERNEL_BENCHMARK_DEPS,
4250)
4251
4252xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004253 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004254 srcs = [
4255 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004256 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004257 "src/xnnpack/AlignedAllocator.h",
4258 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004259 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004260 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004261)
4262
4263xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004264 name = "qu8_requantization_bench",
4265 srcs = [
4266 "bench/qu8-requantization.cc",
4267 "src/xnnpack/requantization-stubs.h",
4268 "src/xnnpack/AlignedAllocator.h",
4269 ] + MICROKERNEL_BENCHMARK_HDRS,
4270 deps = MICROKERNEL_BENCHMARK_DEPS,
4271)
4272
4273xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004274 name = "f16_igemm_bench",
4275 srcs = [
4276 "bench/f16-igemm.cc",
4277 "bench/conv.h",
4278 "bench/google/conv.h",
4279 "src/xnnpack/AlignedAllocator.h",
4280 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004281 deps = MICROKERNEL_BENCHMARK_DEPS + [
4282 ":indirection",
4283 ":packing",
4284 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004285)
4286
4287xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004288 name = "f16_gemm_bench",
4289 srcs = [
4290 "bench/f16-gemm.cc",
4291 "bench/gemm.h",
4292 "src/xnnpack/AlignedAllocator.h",
4293 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004294 deps = MICROKERNEL_BENCHMARK_DEPS + [
4295 ":packing",
4296 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297)
4298
4299xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004300 name = "f16_spmm_bench",
4301 srcs = [
4302 "bench/f16-spmm.cc",
4303 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004304 "src/xnnpack/AlignedAllocator.h",
4305 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004306 deps = MICROKERNEL_BENCHMARK_DEPS,
4307)
4308
4309xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 name = "f32_igemm_bench",
4311 srcs = [
4312 "bench/f32-igemm.cc",
4313 "bench/conv.h",
4314 "src/xnnpack/AlignedAllocator.h",
4315 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004316 deps = MICROKERNEL_BENCHMARK_DEPS + [
4317 ":indirection",
4318 ":packing",
4319 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004320)
4321
4322xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004323 name = "f16_relu_bench",
4324 srcs = [
4325 "bench/f16-relu.cc",
4326 "src/xnnpack/AlignedAllocator.h",
4327 ] + MICROKERNEL_BENCHMARK_HDRS,
4328 deps = MICROKERNEL_BENCHMARK_DEPS,
4329)
4330
4331xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 name = "f32_conv_hwc_bench",
4333 srcs = [
4334 "bench/f32-conv-hwc.cc",
4335 "bench/dconv.h",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004338 deps = MICROKERNEL_BENCHMARK_DEPS + [
4339 ":packing",
4340 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341)
4342
4343xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004344 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004345 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004346 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004347 "bench/dconv.h",
4348 "src/xnnpack/AlignedAllocator.h",
4349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004350 deps = MICROKERNEL_BENCHMARK_DEPS + [
4351 ":packing",
4352 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004353)
4354
4355xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004356 name = "f16_dwconv_bench",
4357 srcs = [
4358 "bench/f16-dwconv.cc",
4359 "bench/dwconv.h",
4360 "bench/google/dwconv.h",
4361 "src/xnnpack/AlignedAllocator.h",
4362 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004363 deps = MICROKERNEL_BENCHMARK_DEPS + [
4364 ":indirection",
4365 ":packing",
4366 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004367)
4368
4369xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004370 name = "f32_dwconv_bench",
4371 srcs = [
4372 "bench/f32-dwconv.cc",
4373 "bench/dwconv.h",
4374 "src/xnnpack/AlignedAllocator.h",
4375 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004376 deps = MICROKERNEL_BENCHMARK_DEPS + [
4377 ":indirection",
4378 ":packing",
4379 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380)
4381
4382xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004383 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004385 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 "bench/dwconv.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004389 deps = MICROKERNEL_BENCHMARK_DEPS + [
4390 ":indirection",
4391 ":packing",
4392 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393)
4394
4395xnnpack_benchmark(
4396 name = "f32_gemm_bench",
4397 srcs = [
4398 "bench/f32-gemm.cc",
4399 "bench/gemm.h",
4400 "src/xnnpack/AlignedAllocator.h",
4401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004402 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004403 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004404)
4405
4406xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004407 name = "f32_hswish_bench",
4408 srcs = [
4409 "bench/f32-hswish.cc",
4410 "src/xnnpack/AlignedAllocator.h",
4411 ] + MICROKERNEL_BENCHMARK_HDRS,
4412 deps = MICROKERNEL_BENCHMARK_DEPS,
4413)
4414
4415xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004416 name = "f32_raddexpminusmax_bench",
4417 srcs = [
4418 "bench/f32-raddexpminusmax.cc",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + MICROKERNEL_BENCHMARK_HDRS,
4421 deps = MICROKERNEL_BENCHMARK_DEPS,
4422)
4423
4424xnnpack_benchmark(
4425 name = "f32_raddextexp_bench",
4426 srcs = [
4427 "bench/f32-raddextexp.cc",
4428 "src/xnnpack/AlignedAllocator.h",
4429 ] + MICROKERNEL_BENCHMARK_HDRS,
4430 deps = MICROKERNEL_BENCHMARK_DEPS,
4431)
4432
4433xnnpack_benchmark(
4434 name = "f32_raddstoreexpminusmax_bench",
4435 srcs = [
4436 "bench/f32-raddstoreexpminusmax.cc",
4437 "src/xnnpack/AlignedAllocator.h",
4438 ] + MICROKERNEL_BENCHMARK_HDRS,
4439 deps = MICROKERNEL_BENCHMARK_DEPS,
4440)
4441
4442xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004443 name = "f32_relu_bench",
4444 srcs = [
4445 "bench/f32-relu.cc",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + MICROKERNEL_BENCHMARK_HDRS,
4448 deps = MICROKERNEL_BENCHMARK_DEPS,
4449)
4450
4451xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 name = "f32_rmax_bench",
4453 srcs = [
4454 "bench/f32-rmax.cc",
4455 "src/xnnpack/AlignedAllocator.h",
4456 ] + MICROKERNEL_BENCHMARK_HDRS,
4457 deps = MICROKERNEL_BENCHMARK_DEPS,
4458)
4459
4460xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004461 name = "f32_sigmoid_bench",
4462 srcs = [
4463 "bench/f32-sigmoid.cc",
4464 "src/xnnpack/AlignedAllocator.h",
4465 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004466 deps = MICROKERNEL_BENCHMARK_DEPS,
4467)
4468
4469xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 name = "f32_spmm_bench",
4471 srcs = [
4472 "bench/f32-spmm.cc",
4473 "bench/gemm.h",
4474 "src/xnnpack/AlignedAllocator.h",
4475 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476 deps = MICROKERNEL_BENCHMARK_DEPS,
4477)
4478
4479xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004480 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004481 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004482 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004483 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004484 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004485 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004486)
4487
4488xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004489 name = "f32_vscaleexpminusmax_bench",
4490 srcs = [
4491 "bench/f32-vscaleexpminusmax.cc",
4492 "src/xnnpack/AlignedAllocator.h",
4493 ] + MICROKERNEL_BENCHMARK_HDRS,
4494 deps = MICROKERNEL_BENCHMARK_DEPS,
4495)
4496
4497xnnpack_benchmark(
4498 name = "f32_vscaleextexp_bench",
4499 srcs = [
4500 "bench/f32-vscaleextexp.cc",
4501 "src/xnnpack/AlignedAllocator.h",
4502 ] + MICROKERNEL_BENCHMARK_HDRS,
4503 deps = MICROKERNEL_BENCHMARK_DEPS,
4504)
4505
4506xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004507 name = "f32_vsqrt_bench",
4508 srcs = [
4509 "bench/f32-vsqrt.cc",
4510 "src/xnnpack/AlignedAllocator.h",
4511 ] + MICROKERNEL_BENCHMARK_HDRS,
4512 deps = MICROKERNEL_BENCHMARK_DEPS,
4513)
4514
4515xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516 name = "f32_im2col_gemm_bench",
4517 srcs = [
4518 "bench/f32-im2col-gemm.cc",
4519 "bench/conv.h",
4520 "src/xnnpack/AlignedAllocator.h",
4521 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004522 deps = MICROKERNEL_BENCHMARK_DEPS + [
4523 ":im2col",
4524 ":packing",
4525 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526)
4527
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004528xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004529 name = "rounding_bench",
4530 srcs = [
4531 "bench/rounding.cc",
4532 "src/xnnpack/math-stubs.h",
4533 "src/xnnpack/AlignedAllocator.h",
4534 ] + MICROKERNEL_BENCHMARK_HDRS,
4535 deps = MICROKERNEL_BENCHMARK_DEPS,
4536)
4537
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538########################### Benchmarks for operators ###########################
4539
4540xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004541 name = "average_pooling_bench",
4542 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004543 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004544 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004545 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546)
4547
4548xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004549 name = "bankers_rounding_bench",
4550 srcs = ["bench/bankers-rounding.cc"],
4551 copts = xnnpack_optional_tflite_copts(),
4552 tags = ["nowin32"],
4553 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4554)
4555
4556xnnpack_benchmark(
4557 name = "ceiling_bench",
4558 srcs = ["bench/ceiling.cc"],
4559 copts = xnnpack_optional_tflite_copts(),
4560 tags = ["nowin32"],
4561 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4562)
4563
4564xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004565 name = "channel_shuffle_bench",
4566 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004567 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004568)
4569
4570xnnpack_benchmark(
4571 name = "convolution_bench",
4572 srcs = ["bench/convolution.cc"],
4573 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004574 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004575 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576)
4577
4578xnnpack_benchmark(
4579 name = "deconvolution_bench",
4580 srcs = ["bench/deconvolution.cc"],
4581 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004582 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004583 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004584)
4585
4586xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004587 name = "floor_bench",
4588 srcs = ["bench/floor.cc"],
4589 copts = xnnpack_optional_tflite_copts(),
4590 tags = ["nowin32"],
4591 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4592)
4593
4594xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004595 name = "global_average_pooling_bench",
4596 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004597 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598)
4599
4600xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004601 name = "hardswish_bench",
4602 srcs = ["bench/hardswish.cc"],
4603 copts = xnnpack_optional_tflite_copts(),
4604 tags = ["nowin32"],
4605 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4606)
4607
4608xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004609 name = "max_pooling_bench",
4610 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004611 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004612)
4613
4614xnnpack_benchmark(
4615 name = "sigmoid_bench",
4616 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004617 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004618 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004619 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620)
4621
4622xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004623 name = "prelu_bench",
4624 srcs = ["bench/prelu.cc"],
4625 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004626 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004627 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004628)
4629
4630xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004631 name = "softmax_bench",
4632 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004633 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004634 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004635 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636)
4637
Marat Dukhan87727142020-06-24 15:24:10 -07004638xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004639 name = "square_root_bench",
4640 srcs = ["bench/square-root.cc"],
4641 copts = xnnpack_optional_tflite_copts(),
4642 tags = ["nowin32"],
4643 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4644)
4645
4646xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004647 name = "truncation_bench",
4648 srcs = ["bench/truncation.cc"],
4649 deps = OPERATOR_BENCHMARK_DEPS,
4650)
4651
Marat Dukhanc068bb62019-10-04 13:24:39 -07004652############################# End-to-end benchmarks ############################
4653
4654cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004655 name = "fp32_mobilenet_v1",
4656 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004657 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004658 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004659 linkstatic = True,
4660 deps = [
4661 ":XNNPACK",
4662 "@pthreadpool",
4663 ],
4664)
4665
4666cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004667 name = "fp16_mobilenet_v1",
4668 srcs = ["models/fp16-mobilenet-v1.cc"],
4669 hdrs = ["models/models.h"],
4670 copts = xnnpack_std_cxxopts(),
4671 linkstatic = True,
4672 deps = [
4673 ":XNNPACK",
4674 "@FP16",
4675 "@pthreadpool",
4676 ],
4677)
4678
4679cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004680 name = "qs8_mobilenet_v1",
4681 srcs = ["models/qs8-mobilenet-v1.cc"],
4682 hdrs = ["models/models.h"],
4683 copts = xnnpack_std_cxxopts(),
4684 linkstatic = True,
4685 deps = [
4686 ":XNNPACK",
4687 "@pthreadpool",
4688 ],
4689)
4690
4691cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004692 name = "qs8_mobilenet_v2",
4693 srcs = ["models/qs8-mobilenet-v2.cc"],
4694 hdrs = ["models/models.h"],
4695 copts = xnnpack_std_cxxopts(),
4696 linkstatic = True,
4697 deps = [
4698 ":XNNPACK",
4699 "@pthreadpool",
4700 ],
4701)
4702
4703cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004704 name = "fp32_mobilenet_v2",
4705 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004706 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004707 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004708 linkstatic = True,
4709 deps = [
4710 ":XNNPACK",
4711 "@pthreadpool",
4712 ],
4713)
4714
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004715cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004716 name = "fp16_mobilenet_v2",
4717 srcs = ["models/fp16-mobilenet-v2.cc"],
4718 hdrs = ["models/models.h"],
4719 copts = xnnpack_std_cxxopts(),
4720 linkstatic = True,
4721 deps = [
4722 ":XNNPACK",
4723 "@FP16",
4724 "@pthreadpool",
4725 ],
4726)
4727
4728cc_library(
4729 name = "fp32_mobilenet_v3_large",
4730 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004731 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004732 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004733 linkstatic = True,
4734 deps = [
4735 ":XNNPACK",
4736 "@pthreadpool",
4737 ],
4738)
4739
4740cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004741 name = "fp16_mobilenet_v3_large",
4742 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4743 hdrs = ["models/models.h"],
4744 copts = xnnpack_std_cxxopts(),
4745 linkstatic = True,
4746 deps = [
4747 ":XNNPACK",
4748 "@FP16",
4749 "@pthreadpool",
4750 ],
4751)
4752
4753cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004754 name = "fp32_mobilenet_v3_small",
4755 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004756 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004757 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004758 linkstatic = True,
4759 deps = [
4760 ":XNNPACK",
4761 "@pthreadpool",
4762 ],
4763)
4764
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004765cc_library(
4766 name = "fp16_mobilenet_v3_small",
4767 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4768 hdrs = ["models/models.h"],
4769 copts = xnnpack_std_cxxopts(),
4770 linkstatic = True,
4771 deps = [
4772 ":XNNPACK",
4773 "@FP16",
4774 "@pthreadpool",
4775 ],
4776)
4777
Marat Dukhanc068bb62019-10-04 13:24:39 -07004778xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004779 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004780 srcs = [
4781 "bench/f32-dwconv-e2e.cc",
4782 "bench/end2end.h",
4783 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004784 deps = MICROKERNEL_BENCHMARK_DEPS + [
4785 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004786 ":fp32_mobilenet_v1",
4787 ":fp32_mobilenet_v2",
4788 ":fp32_mobilenet_v3_large",
4789 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004790 ],
4791)
4792
4793xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004794 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004795 srcs = [
4796 "bench/f32-gemm-e2e.cc",
4797 "bench/end2end.h",
4798 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004799 deps = MICROKERNEL_BENCHMARK_DEPS + [
4800 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004801 ":fp32_mobilenet_v1",
4802 ":fp32_mobilenet_v2",
4803 ":fp32_mobilenet_v3_large",
4804 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004805 ],
4806)
4807
4808xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004809 name = "end2end_bench",
4810 srcs = ["bench/end2end.cc"],
4811 deps = [
4812 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004813 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004814 ":fp16_mobilenet_v1",
4815 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004816 ":fp16_mobilenet_v3_large",
4817 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004818 ":fp32_mobilenet_v1",
4819 ":fp32_mobilenet_v2",
4820 ":fp32_mobilenet_v3_large",
4821 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004822 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004823 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004824 "@pthreadpool",
4825 ],
4826)
4827
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004828#################### Accuracy evaluation for math functions ####################
4829
4830xnnpack_benchmark(
4831 name = "f32_exp_eval",
4832 srcs = [
4833 "eval/f32-exp.cc",
4834 "src/xnnpack/AlignedAllocator.h",
4835 ] + ACCURACY_EVAL_HDRS,
4836 deps = ACCURACY_EVAL_DEPS,
4837)
4838
Marat Dukhan515c9772019-10-17 18:07:57 -07004839xnnpack_benchmark(
4840 name = "f32_expminus_eval",
4841 srcs = [
4842 "eval/f32-expminus.cc",
4843 "src/xnnpack/AlignedAllocator.h",
4844 ] + ACCURACY_EVAL_HDRS,
4845 deps = ACCURACY_EVAL_DEPS,
4846)
4847
Marat Dukhan98ba4412019-10-23 02:14:28 -07004848xnnpack_benchmark(
4849 name = "f32_extexp_eval",
4850 srcs = [
4851 "eval/f32-extexp.cc",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + ACCURACY_EVAL_HDRS,
4854 deps = ACCURACY_EVAL_DEPS,
4855)
4856
Marat Dukhan8853b822020-05-07 12:19:01 -07004857xnnpack_unit_test(
4858 name = "f32_roundne_eval",
4859 srcs = [
4860 "eval/f32-roundne.cc",
4861 "src/xnnpack/AlignedAllocator.h",
4862 "src/xnnpack/math-stubs.h",
4863 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004864 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004865 deps = MICROKERNEL_TEST_DEPS,
4866)
4867
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004868xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004869 name = "f32_roundd_eval",
4870 srcs = [
4871 "eval/f32-roundd.cc",
4872 "src/xnnpack/AlignedAllocator.h",
4873 "src/xnnpack/math-stubs.h",
4874 ] + MICROKERNEL_TEST_HDRS,
4875 automatic = False,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
4880 name = "f32_roundu_eval",
4881 srcs = [
4882 "eval/f32-roundu.cc",
4883 "src/xnnpack/AlignedAllocator.h",
4884 "src/xnnpack/math-stubs.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 automatic = False,
4887 deps = MICROKERNEL_TEST_DEPS,
4888)
4889
4890xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004891 name = "f32_roundz_eval",
4892 srcs = [
4893 "eval/f32-roundz.cc",
4894 "src/xnnpack/AlignedAllocator.h",
4895 "src/xnnpack/math-stubs.h",
4896 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004897 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004898 deps = MICROKERNEL_TEST_DEPS,
4899)
4900
Marat Dukhan346a9e52019-11-15 09:06:30 -08004901xnnpack_benchmark(
4902 name = "f32_sigmoid_eval",
4903 srcs = [
4904 "eval/f32-sigmoid.cc",
4905 "src/xnnpack/AlignedAllocator.h",
4906 ] + ACCURACY_EVAL_HDRS,
4907 deps = ACCURACY_EVAL_DEPS,
4908)
4909
Marat Dukhan84000762020-06-29 18:38:43 -07004910xnnpack_benchmark(
4911 name = "f32_sqrt_eval",
4912 srcs = [
4913 "eval/f32-sqrt.cc",
4914 "src/xnnpack/AlignedAllocator.h",
4915 ] + ACCURACY_EVAL_HDRS,
4916 deps = ACCURACY_EVAL_DEPS + [
4917 ":bench_utils",
4918 ],
4919)
4920
Marat Dukhan08c4a432019-10-03 09:29:21 -07004921######################### Unit tests for micro-kernels #########################
4922
4923xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004924 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004925 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004926 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004927 "test/gemm-microkernel-tester.h",
4928 "src/xnnpack/AlignedAllocator.h",
4929 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004930 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931)
4932
4933xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004934 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004935 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004936 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004937 "test/spmm-microkernel-tester.h",
4938 "src/xnnpack/AlignedAllocator.h",
4939 ] + MICROKERNEL_TEST_HDRS,
4940 deps = MICROKERNEL_TEST_DEPS,
4941)
4942
4943xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004944 name = "f32_argmaxpool_test",
4945 srcs = [
4946 "test/f32-argmaxpool.cc",
4947 "test/argmaxpool-microkernel-tester.h",
4948 "src/xnnpack/AlignedAllocator.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004954 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004955 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004956 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004957 "test/avgpool-microkernel-tester.h",
4958 "src/xnnpack/AlignedAllocator.h",
4959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004964 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004965 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004966 "test/f32-ibilinear.cc",
4967 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004968 "src/xnnpack/AlignedAllocator.h",
4969 ] + MICROKERNEL_TEST_HDRS,
4970 deps = MICROKERNEL_TEST_DEPS,
4971)
4972
4973xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004974 name = "f16_clamp_test",
4975 srcs = [
4976 "test/f16-clamp.cc",
4977 "test/clamp-microkernel-tester.h",
4978 ] + MICROKERNEL_TEST_HDRS,
4979 deps = MICROKERNEL_TEST_DEPS,
4980)
4981
4982xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 name = "f32_clamp_test",
4984 srcs = [
4985 "test/f32-clamp.cc",
4986 "test/clamp-microkernel-tester.h",
4987 ] + MICROKERNEL_TEST_HDRS,
4988 deps = MICROKERNEL_TEST_DEPS,
4989)
4990
4991xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004992 name = "f32_igemm_test",
4993 srcs = [
4994 "test/f32-igemm.cc",
4995 "test/gemm-microkernel-tester.h",
4996 "src/xnnpack/AlignedAllocator.h",
4997 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004998 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004999)
5000
5001xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005002 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005003 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005004 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 "test/gemm-microkernel-tester.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005008 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009)
5010
5011xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005012 name = "f32_igemm_minmax_test",
5013 srcs = [
5014 "test/f32-igemm-minmax.cc",
5015 "test/gemm-microkernel-tester.h",
5016 "src/xnnpack/AlignedAllocator.h",
5017 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005018 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005019)
5020
5021xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005022 name = "f16_igemm_minmax_test",
5023 srcs = [
5024 "test/f16-igemm-minmax.cc",
5025 "test/gemm-microkernel-tester.h",
5026 "src/xnnpack/AlignedAllocator.h",
5027 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005028 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005029)
5030
5031xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032 name = "f32_conv_hwc_test",
5033 srcs = [
5034 "test/f32-conv-hwc.cc",
5035 "test/conv-hwc-microkernel-tester.h",
5036 "src/xnnpack/AlignedAllocator.h",
5037 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005038 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039)
5040
5041xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005042 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005043 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005044 "test/f32-conv-hwc2chw.cc",
5045 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 "src/xnnpack/AlignedAllocator.h",
5047 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005048 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049)
5050
5051xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005052 name = "f32_dwconv_test",
5053 srcs = [
5054 "test/f32-dwconv.cc",
5055 "test/dwconv-microkernel-tester.h",
5056 "src/xnnpack/AlignedAllocator.h",
5057 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005058 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005059)
5060
5061xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005062 name = "f16_dwconv_minmax_test",
5063 srcs = [
5064 "test/f16-dwconv-minmax.cc",
5065 "test/dwconv-microkernel-tester.h",
5066 "src/xnnpack/AlignedAllocator.h",
5067 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005068 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005069)
5070
5071xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005072 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005074 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075 "test/dwconv-microkernel-tester.h",
5076 "src/xnnpack/AlignedAllocator.h",
5077 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005078 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005079)
5080
5081xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005082 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005084 "test/f32-dwconv-chw.cc",
5085 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005086 "src/xnnpack/AlignedAllocator.h",
5087 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005088 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089)
5090
5091xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005092 name = "f16_gavgpool_minmax_test",
5093 srcs = [
5094 "test/f16-gavgpool-minmax.cc",
5095 "test/gavgpool-microkernel-tester.h",
5096 "src/xnnpack/AlignedAllocator.h",
5097 ] + MICROKERNEL_TEST_HDRS,
5098 deps = MICROKERNEL_TEST_DEPS,
5099)
5100
5101xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005102 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005103 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005104 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005105 "test/gavgpool-microkernel-tester.h",
5106 "src/xnnpack/AlignedAllocator.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005112 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005113 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005114 "test/f32-gavgpool-cw.cc",
5115 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005116 "src/xnnpack/AlignedAllocator.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005122 name = "f32_gemm_test",
5123 srcs = [
5124 "test/f32-gemm.cc",
5125 "test/gemm-microkernel-tester.h",
5126 "src/xnnpack/AlignedAllocator.h",
5127 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005128 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005129)
5130
5131xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005132 name = "f32_gemm_relu_test",
5133 srcs = [
5134 "test/f32-gemm-relu.cc",
5135 "test/gemm-microkernel-tester.h",
5136 "src/xnnpack/AlignedAllocator.h",
5137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005138 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005139)
5140
5141xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005142 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005144 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005145 "test/gemm-microkernel-tester.h",
5146 "src/xnnpack/AlignedAllocator.h",
5147 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005148 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149)
5150
5151xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005152 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005153 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005154 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155 "test/gemm-microkernel-tester.h",
5156 "src/xnnpack/AlignedAllocator.h",
5157 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005158 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159)
5160
5161xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005162 name = "f16_hswish_test",
5163 srcs = [
5164 "test/f16-hswish.cc",
5165 "test/hswish-microkernel-tester.h",
5166 ] + MICROKERNEL_TEST_HDRS,
5167 deps = MICROKERNEL_TEST_DEPS,
5168)
5169
5170xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171 name = "f32_hswish_test",
5172 srcs = [
5173 "test/f32-hswish.cc",
5174 "test/hswish-microkernel-tester.h",
5175 ] + MICROKERNEL_TEST_HDRS,
5176 deps = MICROKERNEL_TEST_DEPS,
5177)
5178
5179xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005180 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005182 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 "test/maxpool-microkernel-tester.h",
5184 ] + MICROKERNEL_TEST_HDRS,
5185 deps = MICROKERNEL_TEST_DEPS,
5186)
5187
5188xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005189 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005191 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192 "test/avgpool-microkernel-tester.h",
5193 "src/xnnpack/AlignedAllocator.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005199 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005201 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005202 "test/gemm-microkernel-tester.h",
5203 "src/xnnpack/AlignedAllocator.h",
5204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005205 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206)
5207
5208xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005209 name = "f16_prelu_test",
5210 srcs = [
5211 "test/f16-prelu.cc",
5212 "test/prelu-microkernel-tester.h",
5213 "src/xnnpack/AlignedAllocator.h",
5214 ] + MICROKERNEL_TEST_HDRS,
5215 deps = MICROKERNEL_TEST_DEPS,
5216)
5217
5218xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 name = "f32_prelu_test",
5220 srcs = [
5221 "test/f32-prelu.cc",
5222 "test/prelu-microkernel-tester.h",
5223 "src/xnnpack/AlignedAllocator.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005229 name = "f32_raddexpminusmax_test",
5230 srcs = [
5231 "test/f32-raddexpminusmax.cc",
5232 "test/raddexpminusmax-microkernel-tester.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005238 name = "f32_raddextexp_test",
5239 srcs = [
5240 "test/f32-raddextexp.cc",
5241 "test/raddextexp-microkernel-tester.h",
5242 ] + MICROKERNEL_TEST_HDRS,
5243 deps = MICROKERNEL_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005247 name = "f32_raddstoreexpminusmax_test",
5248 srcs = [
5249 "test/f32-raddstoreexpminusmax.cc",
5250 "test/raddstoreexpminusmax-microkernel-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005256 name = "f32_relu_test",
5257 srcs = [
5258 "test/f32-relu.cc",
5259 "test/vunary-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 name = "f32_rmax_test",
5266 srcs = [
5267 "test/f32-rmax.cc",
5268 "test/rmax-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005274 name = "f32_sigmoid_test",
5275 srcs = [
5276 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005277 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005283 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005284 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005285 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005286 "test/spmm-microkernel-tester.h",
5287 "src/xnnpack/AlignedAllocator.h",
5288 ] + MICROKERNEL_TEST_HDRS,
5289 deps = MICROKERNEL_TEST_DEPS,
5290)
5291
5292xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005293 name = "f16_vadd_minmax_test",
5294 srcs = [
5295 "test/f16-vadd-minmax.cc",
5296 "test/vbinary-microkernel-tester.h",
5297 ] + MICROKERNEL_TEST_HDRS,
5298 deps = MICROKERNEL_TEST_DEPS,
5299)
5300
5301xnnpack_unit_test(
5302 name = "f16_vaddc_minmax_test",
5303 srcs = [
5304 "test/f16-vaddc-minmax.cc",
5305 "test/vbinaryc-microkernel-tester.h",
5306 ] + MICROKERNEL_TEST_HDRS,
5307 deps = MICROKERNEL_TEST_DEPS,
5308)
5309
5310xnnpack_unit_test(
5311 name = "f16_vdiv_minmax_test",
5312 srcs = [
5313 "test/f16-vdiv-minmax.cc",
5314 "test/vbinary-microkernel-tester.h",
5315 ] + MICROKERNEL_TEST_HDRS,
5316 deps = MICROKERNEL_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
5320 name = "f16_vdivc_minmax_test",
5321 srcs = [
5322 "test/f16-vdivc-minmax.cc",
5323 "test/vbinaryc-microkernel-tester.h",
5324 ] + MICROKERNEL_TEST_HDRS,
5325 deps = MICROKERNEL_TEST_DEPS,
5326)
5327
5328xnnpack_unit_test(
5329 name = "f16_vrdivc_minmax_test",
5330 srcs = [
5331 "test/f16-vrdivc-minmax.cc",
5332 "test/vbinaryc-microkernel-tester.h",
5333 ] + MICROKERNEL_TEST_HDRS,
5334 deps = MICROKERNEL_TEST_DEPS,
5335)
5336
5337xnnpack_unit_test(
5338 name = "f16_vmax_test",
5339 srcs = [
5340 "test/f16-vmax.cc",
5341 "test/vbinary-microkernel-tester.h",
5342 ] + MICROKERNEL_TEST_HDRS,
5343 deps = MICROKERNEL_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
5347 name = "f16_vmaxc_test",
5348 srcs = [
5349 "test/f16-vmaxc.cc",
5350 "test/vbinaryc-microkernel-tester.h",
5351 ] + MICROKERNEL_TEST_HDRS,
5352 deps = MICROKERNEL_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
5356 name = "f16_vmin_test",
5357 srcs = [
5358 "test/f16-vmin.cc",
5359 "test/vbinary-microkernel-tester.h",
5360 ] + MICROKERNEL_TEST_HDRS,
5361 deps = MICROKERNEL_TEST_DEPS,
5362)
5363
5364xnnpack_unit_test(
5365 name = "f16_vminc_test",
5366 srcs = [
5367 "test/f16-vminc.cc",
5368 "test/vbinaryc-microkernel-tester.h",
5369 ] + MICROKERNEL_TEST_HDRS,
5370 deps = MICROKERNEL_TEST_DEPS,
5371)
5372
5373xnnpack_unit_test(
5374 name = "f16_vmul_minmax_test",
5375 srcs = [
5376 "test/f16-vmul-minmax.cc",
5377 "test/vbinary-microkernel-tester.h",
5378 ] + MICROKERNEL_TEST_HDRS,
5379 deps = MICROKERNEL_TEST_DEPS,
5380)
5381
5382xnnpack_unit_test(
5383 name = "f16_vmulc_minmax_test",
5384 srcs = [
5385 "test/f16-vmulc-minmax.cc",
5386 "test/vbinaryc-microkernel-tester.h",
5387 ] + MICROKERNEL_TEST_HDRS,
5388 deps = MICROKERNEL_TEST_DEPS,
5389)
5390
5391xnnpack_unit_test(
5392 name = "f16_vsub_minmax_test",
5393 srcs = [
5394 "test/f16-vsub-minmax.cc",
5395 "test/vbinary-microkernel-tester.h",
5396 ] + MICROKERNEL_TEST_HDRS,
5397 deps = MICROKERNEL_TEST_DEPS,
5398)
5399
5400xnnpack_unit_test(
5401 name = "f16_vsubc_minmax_test",
5402 srcs = [
5403 "test/f16-vsubc-minmax.cc",
5404 "test/vbinaryc-microkernel-tester.h",
5405 ] + MICROKERNEL_TEST_HDRS,
5406 deps = MICROKERNEL_TEST_DEPS,
5407)
5408
5409xnnpack_unit_test(
5410 name = "f16_vrsubc_minmax_test",
5411 srcs = [
5412 "test/f16-vrsubc-minmax.cc",
5413 "test/vbinaryc-microkernel-tester.h",
5414 ] + MICROKERNEL_TEST_HDRS,
5415 deps = MICROKERNEL_TEST_DEPS,
5416)
5417
5418xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005419 name = "f32_vabs_test",
5420 srcs = [
5421 "test/f32-vabs.cc",
5422 "test/vunary-microkernel-tester.h",
5423 ] + MICROKERNEL_TEST_HDRS,
5424 deps = MICROKERNEL_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005428 name = "f32_vadd_test",
5429 srcs = [
5430 "test/f32-vadd.cc",
5431 "test/vbinary-microkernel-tester.h",
5432 ] + MICROKERNEL_TEST_HDRS,
5433 deps = MICROKERNEL_TEST_DEPS,
5434)
5435
5436xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005437 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005439 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005440 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005441 ] + MICROKERNEL_TEST_HDRS,
5442 deps = MICROKERNEL_TEST_DEPS,
5443)
5444
5445xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005446 name = "f32_vadd_relu_test",
5447 srcs = [
5448 "test/f32-vadd-relu.cc",
5449 "test/vbinary-microkernel-tester.h",
5450 ] + MICROKERNEL_TEST_HDRS,
5451 deps = MICROKERNEL_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005455 name = "f32_vaddc_test",
5456 srcs = [
5457 "test/f32-vaddc.cc",
5458 "test/vbinaryc-microkernel-tester.h",
5459 ] + MICROKERNEL_TEST_HDRS,
5460 deps = MICROKERNEL_TEST_DEPS,
5461)
5462
5463xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005464 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005465 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005466 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005467 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468 ] + MICROKERNEL_TEST_HDRS,
5469 deps = MICROKERNEL_TEST_DEPS,
5470)
5471
5472xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005473 name = "f32_vaddc_relu_test",
5474 srcs = [
5475 "test/f32-vaddc-relu.cc",
5476 "test/vbinaryc-microkernel-tester.h",
5477 ] + MICROKERNEL_TEST_HDRS,
5478 deps = MICROKERNEL_TEST_DEPS,
5479)
5480
5481xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005482 name = "f32_vdiv_test",
5483 srcs = [
5484 "test/f32-vdiv.cc",
5485 "test/vbinary-microkernel-tester.h",
5486 ] + MICROKERNEL_TEST_HDRS,
5487 deps = MICROKERNEL_TEST_DEPS,
5488)
5489
5490xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005491 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005492 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005493 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005494 "test/vbinary-microkernel-tester.h",
5495 ] + MICROKERNEL_TEST_HDRS,
5496 deps = MICROKERNEL_TEST_DEPS,
5497)
5498
5499xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005500 name = "f32_vdiv_relu_test",
5501 srcs = [
5502 "test/f32-vdiv-relu.cc",
5503 "test/vbinary-microkernel-tester.h",
5504 ] + MICROKERNEL_TEST_HDRS,
5505 deps = MICROKERNEL_TEST_DEPS,
5506)
5507
5508xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005509 name = "f32_vdivc_test",
5510 srcs = [
5511 "test/f32-vdivc.cc",
5512 "test/vbinaryc-microkernel-tester.h",
5513 ] + MICROKERNEL_TEST_HDRS,
5514 deps = MICROKERNEL_TEST_DEPS,
5515)
5516
5517xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005518 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005519 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005520 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005521 "test/vbinaryc-microkernel-tester.h",
5522 ] + MICROKERNEL_TEST_HDRS,
5523 deps = MICROKERNEL_TEST_DEPS,
5524)
5525
5526xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005527 name = "f32_vdivc_relu_test",
5528 srcs = [
5529 "test/f32-vdivc-relu.cc",
5530 "test/vbinaryc-microkernel-tester.h",
5531 ] + MICROKERNEL_TEST_HDRS,
5532 deps = MICROKERNEL_TEST_DEPS,
5533)
5534
5535xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005536 name = "f32_vrdivc_test",
5537 srcs = [
5538 "test/f32-vrdivc.cc",
5539 "test/vbinaryc-microkernel-tester.h",
5540 ] + MICROKERNEL_TEST_HDRS,
5541 deps = MICROKERNEL_TEST_DEPS,
5542)
5543
5544xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005545 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005546 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005547 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005548 "test/vbinaryc-microkernel-tester.h",
5549 ] + MICROKERNEL_TEST_HDRS,
5550 deps = MICROKERNEL_TEST_DEPS,
5551)
5552
5553xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005554 name = "f32_vrdivc_relu_test",
5555 srcs = [
5556 "test/f32-vrdivc-relu.cc",
5557 "test/vbinaryc-microkernel-tester.h",
5558 ] + MICROKERNEL_TEST_HDRS,
5559 deps = MICROKERNEL_TEST_DEPS,
5560)
5561
5562xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005563 name = "f32_vmax_test",
5564 srcs = [
5565 "test/f32-vmax.cc",
5566 "test/vbinary-microkernel-tester.h",
5567 ] + MICROKERNEL_TEST_HDRS,
5568 deps = MICROKERNEL_TEST_DEPS,
5569)
5570
5571xnnpack_unit_test(
5572 name = "f32_vmaxc_test",
5573 srcs = [
5574 "test/f32-vmaxc.cc",
5575 "test/vbinaryc-microkernel-tester.h",
5576 ] + MICROKERNEL_TEST_HDRS,
5577 deps = MICROKERNEL_TEST_DEPS,
5578)
5579
5580xnnpack_unit_test(
5581 name = "f32_vmin_test",
5582 srcs = [
5583 "test/f32-vmin.cc",
5584 "test/vbinary-microkernel-tester.h",
5585 ] + MICROKERNEL_TEST_HDRS,
5586 deps = MICROKERNEL_TEST_DEPS,
5587)
5588
5589xnnpack_unit_test(
5590 name = "f32_vminc_test",
5591 srcs = [
5592 "test/f32-vminc.cc",
5593 "test/vbinaryc-microkernel-tester.h",
5594 ] + MICROKERNEL_TEST_HDRS,
5595 deps = MICROKERNEL_TEST_DEPS,
5596)
5597
5598xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005599 name = "f32_vmul_test",
5600 srcs = [
5601 "test/f32-vmul.cc",
5602 "test/vbinary-microkernel-tester.h",
5603 ] + MICROKERNEL_TEST_HDRS,
5604 deps = MICROKERNEL_TEST_DEPS,
5605)
5606
5607xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005608 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005609 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005610 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005611 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005612 ] + MICROKERNEL_TEST_HDRS,
5613 deps = MICROKERNEL_TEST_DEPS,
5614)
5615
5616xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005617 name = "f32_vmul_relu_test",
5618 srcs = [
5619 "test/f32-vmul-relu.cc",
5620 "test/vbinary-microkernel-tester.h",
5621 ] + MICROKERNEL_TEST_HDRS,
5622 deps = MICROKERNEL_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005626 name = "f32_vmulc_test",
5627 srcs = [
5628 "test/f32-vmulc.cc",
5629 "test/vbinaryc-microkernel-tester.h",
5630 ] + MICROKERNEL_TEST_HDRS,
5631 deps = MICROKERNEL_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005635 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005636 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005637 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005638 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639 ] + MICROKERNEL_TEST_HDRS,
5640 deps = MICROKERNEL_TEST_DEPS,
5641)
5642
5643xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005644 name = "f32_vmulc_relu_test",
5645 srcs = [
5646 "test/f32-vmulc-relu.cc",
5647 "test/vbinaryc-microkernel-tester.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005653 name = "f16_vmulcaddc_minmax_test",
5654 srcs = [
5655 "test/f16-vmulcaddc-minmax.cc",
5656 "test/vmulcaddc-microkernel-tester.h",
5657 "src/xnnpack/AlignedAllocator.h",
5658 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005659 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005660)
5661
5662xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005663 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005664 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005665 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666 "test/vmulcaddc-microkernel-tester.h",
5667 "src/xnnpack/AlignedAllocator.h",
5668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005669 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670)
5671
5672xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005673 name = "f32_vlrelu_test",
5674 srcs = [
5675 "test/f32-vlrelu.cc",
5676 "test/vunary-microkernel-tester.h",
5677 ] + MICROKERNEL_TEST_HDRS,
5678 deps = MICROKERNEL_TEST_DEPS,
5679)
5680
5681xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005682 name = "f32_vneg_test",
5683 srcs = [
5684 "test/f32-vneg.cc",
5685 "test/vunary-microkernel-tester.h",
5686 ] + MICROKERNEL_TEST_HDRS,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005691 name = "f32_vrndne_test",
5692 srcs = [
5693 "test/f32-vrndne.cc",
5694 "test/vunary-microkernel-tester.h",
5695 ] + MICROKERNEL_TEST_HDRS,
5696 deps = MICROKERNEL_TEST_DEPS,
5697)
5698
5699xnnpack_unit_test(
5700 name = "f32_vrndz_test",
5701 srcs = [
5702 "test/f32-vrndz.cc",
5703 "test/vunary-microkernel-tester.h",
5704 ] + MICROKERNEL_TEST_HDRS,
5705 deps = MICROKERNEL_TEST_DEPS,
5706)
5707
5708xnnpack_unit_test(
5709 name = "f32_vrndu_test",
5710 srcs = [
5711 "test/f32-vrndu.cc",
5712 "test/vunary-microkernel-tester.h",
5713 ] + MICROKERNEL_TEST_HDRS,
5714 deps = MICROKERNEL_TEST_DEPS,
5715)
5716
5717xnnpack_unit_test(
5718 name = "f32_vrndd_test",
5719 srcs = [
5720 "test/f32-vrndd.cc",
5721 "test/vunary-microkernel-tester.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005727 name = "f32_vscale_test",
5728 srcs = [
5729 "test/f32-vscale.cc",
5730 "test/vscale-microkernel-tester.h",
5731 ] + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS,
5733)
5734
5735xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005736 name = "f32_vscaleexpminusmax_test",
5737 srcs = [
5738 "test/f32-vscaleexpminusmax.cc",
5739 "test/vscaleexpminusmax-microkernel-tester.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS,
5742)
5743
5744xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005745 name = "f32_vscaleextexp_test",
5746 srcs = [
5747 "test/f32-vscaleextexp.cc",
5748 "test/vscaleextexp-microkernel-tester.h",
5749 ] + MICROKERNEL_TEST_HDRS,
5750 deps = MICROKERNEL_TEST_DEPS,
5751)
5752
5753xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005754 name = "f32_vsqr_test",
5755 srcs = [
5756 "test/f32-vsqr.cc",
5757 "test/vunary-microkernel-tester.h",
5758 ] + MICROKERNEL_TEST_HDRS,
5759 deps = MICROKERNEL_TEST_DEPS,
5760)
5761
5762xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005763 name = "f32_vsqrdiff_test",
5764 srcs = [
5765 "test/f32-vsqrdiff.cc",
5766 "test/vbinary-microkernel-tester.h",
5767 ] + MICROKERNEL_TEST_HDRS,
5768 deps = MICROKERNEL_TEST_DEPS,
5769)
5770
5771xnnpack_unit_test(
5772 name = "f32_vsqrdiffc_test",
5773 srcs = [
5774 "test/f32-vsqrdiffc.cc",
5775 "test/vbinaryc-microkernel-tester.h",
5776 ] + MICROKERNEL_TEST_HDRS,
5777 deps = MICROKERNEL_TEST_DEPS,
5778)
5779
5780xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005781 name = "f32_vsqrt_test",
5782 srcs = [
5783 "test/f32-vsqrt.cc",
5784 "test/vunary-microkernel-tester.h",
5785 ] + MICROKERNEL_TEST_HDRS,
5786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
5789xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005790 name = "f32_vsub_test",
5791 srcs = [
5792 "test/f32-vsub.cc",
5793 "test/vbinary-microkernel-tester.h",
5794 ] + MICROKERNEL_TEST_HDRS,
5795 deps = MICROKERNEL_TEST_DEPS,
5796)
5797
5798xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005799 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005800 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005801 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005802 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005803 ] + MICROKERNEL_TEST_HDRS,
5804 deps = MICROKERNEL_TEST_DEPS,
5805)
5806
5807xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005808 name = "f32_vsub_relu_test",
5809 srcs = [
5810 "test/f32-vsub-relu.cc",
5811 "test/vbinary-microkernel-tester.h",
5812 ] + MICROKERNEL_TEST_HDRS,
5813 deps = MICROKERNEL_TEST_DEPS,
5814)
5815
5816xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005817 name = "f32_vsubc_test",
5818 srcs = [
5819 "test/f32-vsubc.cc",
5820 "test/vbinaryc-microkernel-tester.h",
5821 ] + MICROKERNEL_TEST_HDRS,
5822 deps = MICROKERNEL_TEST_DEPS,
5823)
5824
5825xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005826 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005827 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005828 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005829 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005835 name = "f32_vsubc_relu_test",
5836 srcs = [
5837 "test/f32-vsubc-relu.cc",
5838 "test/vbinaryc-microkernel-tester.h",
5839 ] + MICROKERNEL_TEST_HDRS,
5840 deps = MICROKERNEL_TEST_DEPS,
5841)
5842
5843xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005844 name = "f32_vrsubc_test",
5845 srcs = [
5846 "test/f32-vrsubc.cc",
5847 "test/vbinaryc-microkernel-tester.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 deps = MICROKERNEL_TEST_DEPS,
5850)
5851
5852xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005853 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005854 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005855 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005856 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005857 ] + MICROKERNEL_TEST_HDRS,
5858 deps = MICROKERNEL_TEST_DEPS,
5859)
5860
5861xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005862 name = "f32_vrsubc_relu_test",
5863 srcs = [
5864 "test/f32-vrsubc-relu.cc",
5865 "test/vbinaryc-microkernel-tester.h",
5866 ] + MICROKERNEL_TEST_HDRS,
5867 deps = MICROKERNEL_TEST_DEPS,
5868)
5869
5870xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005871 name = "qs8_dwconv_minmax_test",
5872 srcs = [
5873 "test/qs8-dwconv-minmax.cc",
5874 "test/dwconv-microkernel-tester.h",
5875 "src/xnnpack/AlignedAllocator.h",
5876 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5878)
5879
5880xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005881 name = "qs8_gavgpool_minmax_test",
5882 srcs = [
5883 "test/qs8-gavgpool-minmax.cc",
5884 "test/gavgpool-microkernel-tester.h",
5885 "src/xnnpack/AlignedAllocator.h",
5886 ] + MICROKERNEL_TEST_HDRS,
5887 deps = MICROKERNEL_TEST_DEPS,
5888)
5889
5890xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005891 name = "qs8_gemm_minmax_test",
5892 srcs = [
5893 "test/qs8-gemm-minmax.cc",
5894 "test/gemm-microkernel-tester.h",
5895 "src/xnnpack/AlignedAllocator.h",
5896 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5898)
5899
5900xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005901 name = "qs8_igemm_minmax_test",
5902 srcs = [
5903 "test/qs8-igemm-minmax.cc",
5904 "test/gemm-microkernel-tester.h",
5905 "src/xnnpack/AlignedAllocator.h",
5906 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5907 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5908)
5909
5910xnnpack_unit_test(
5911 name = "qs8_requantization_test",
5912 srcs = [
5913 "src/xnnpack/requantization-stubs.h",
5914 "test/qs8-requantization.cc",
5915 "test/requantization-tester.h",
5916 ] + MICROKERNEL_TEST_HDRS,
5917 deps = MICROKERNEL_TEST_DEPS,
5918)
5919
5920xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005921 name = "qs8_vadd_minmax_test",
5922 srcs = [
5923 "test/qs8-vadd-minmax.cc",
5924 "test/vadd-microkernel-tester.h",
5925 ] + MICROKERNEL_TEST_HDRS,
5926 deps = MICROKERNEL_TEST_DEPS,
5927)
5928
5929xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005930 name = "qs8_vaddc_minmax_test",
5931 srcs = [
5932 "test/qs8-vaddc-minmax.cc",
5933 "test/vaddc-microkernel-tester.h",
5934 ] + MICROKERNEL_TEST_HDRS,
5935 deps = MICROKERNEL_TEST_DEPS,
5936)
5937
5938xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005939 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005940 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005941 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005942 "test/avgpool-microkernel-tester.h",
5943 "src/xnnpack/AlignedAllocator.h",
5944 ] + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS,
5946)
5947
5948xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005949 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005951 "test/qu8-dwconv-minmax.cc",
5952 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 "src/xnnpack/AlignedAllocator.h",
5954 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005955 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005956)
5957
5958xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005959 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005960 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005961 "test/qu8-igemm-minmax.cc",
5962 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005963 "src/xnnpack/AlignedAllocator.h",
5964 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005965 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966)
5967
5968xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005969 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005971 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972 "test/gavgpool-microkernel-tester.h",
5973 "src/xnnpack/AlignedAllocator.h",
5974 ] + MICROKERNEL_TEST_HDRS,
5975 deps = MICROKERNEL_TEST_DEPS,
5976)
5977
5978xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005979 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005981 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982 "test/gemm-microkernel-tester.h",
5983 "src/xnnpack/AlignedAllocator.h",
5984 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005985 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005986)
5987
5988xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005989 name = "qu8_requantization_test",
5990 srcs = [
5991 "src/xnnpack/requantization-stubs.h",
5992 "test/qu8-requantization.cc",
5993 "test/requantization-tester.h",
5994 ] + MICROKERNEL_TEST_HDRS,
5995 deps = MICROKERNEL_TEST_DEPS,
5996)
5997
5998xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005999 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006001 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006002 "test/vadd-microkernel-tester.h",
6003 ] + MICROKERNEL_TEST_HDRS,
6004 deps = MICROKERNEL_TEST_DEPS,
6005)
6006
6007xnnpack_unit_test(
6008 name = "u8_clamp_test",
6009 srcs = [
6010 "test/u8-clamp.cc",
6011 "test/clamp-microkernel-tester.h",
6012 ] + MICROKERNEL_TEST_HDRS,
6013 deps = MICROKERNEL_TEST_DEPS,
6014)
6015
6016xnnpack_unit_test(
6017 name = "u8_lut32norm_test",
6018 srcs = [
6019 "test/u8-lut32norm.cc",
6020 "test/lut-norm-microkernel-tester.h",
6021 ] + MICROKERNEL_TEST_HDRS,
6022 deps = MICROKERNEL_TEST_DEPS,
6023)
6024
6025xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006026 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006027 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006028 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006029 "test/maxpool-microkernel-tester.h",
6030 ] + MICROKERNEL_TEST_HDRS,
6031 deps = MICROKERNEL_TEST_DEPS,
6032)
6033
6034xnnpack_unit_test(
6035 name = "u8_rmax_test",
6036 srcs = [
6037 "test/u8-rmax.cc",
6038 "test/rmax-microkernel-tester.h",
6039 ] + MICROKERNEL_TEST_HDRS,
6040 deps = MICROKERNEL_TEST_DEPS,
6041)
6042
6043xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006044 name = "x32_fill_test",
6045 srcs = [
6046 "test/x32-fill.cc",
6047 "test/fill-microkernel-tester.h",
6048 ] + MICROKERNEL_TEST_HDRS,
6049 deps = MICROKERNEL_TEST_DEPS,
6050)
6051
6052xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006053 name = "x32_packx_test",
6054 srcs = [
6055 "test/x32-packx.cc",
6056 "test/pack-microkernel-tester.h",
6057 "src/xnnpack/AlignedAllocator.h",
6058 ] + MICROKERNEL_TEST_HDRS,
6059 deps = MICROKERNEL_TEST_DEPS,
6060)
6061
6062xnnpack_unit_test(
6063 name = "x32_pad_test",
6064 srcs = [
6065 "test/x32-pad.cc",
6066 "test/pad-microkernel-tester.h",
6067 ] + MICROKERNEL_TEST_HDRS,
6068 deps = MICROKERNEL_TEST_DEPS,
6069)
6070
6071xnnpack_unit_test(
6072 name = "x32_unpool_test",
6073 srcs = [
6074 "test/x32-unpool.cc",
6075 "test/unpool-microkernel-tester.h",
6076 ] + MICROKERNEL_TEST_HDRS,
6077 deps = MICROKERNEL_TEST_DEPS,
6078)
6079
6080xnnpack_unit_test(
6081 name = "x32_zip_test",
6082 srcs = [
6083 "test/x32-zip.cc",
6084 "test/zip-microkernel-tester.h",
6085 ] + MICROKERNEL_TEST_HDRS,
6086 deps = MICROKERNEL_TEST_DEPS,
6087)
6088
6089xnnpack_unit_test(
6090 name = "x8_lut_test",
6091 srcs = [
6092 "test/x8-lut.cc",
6093 "test/lut-microkernel-tester.h",
6094 ] + MICROKERNEL_TEST_HDRS,
6095 deps = MICROKERNEL_TEST_DEPS,
6096)
6097
6098xnnpack_unit_test(
6099 name = "x8_zip_test",
6100 srcs = [
6101 "test/x8-zip.cc",
6102 "test/zip-microkernel-tester.h",
6103 ] + MICROKERNEL_TEST_HDRS,
6104 deps = MICROKERNEL_TEST_DEPS,
6105)
6106
Marat Dukhan20c3b922020-03-10 03:45:06 -07006107########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006108
6109xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006110 name = "operator_size_test",
6111 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006112 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006113)
6114
Marat Dukhan20c3b922020-03-10 03:45:06 -07006115xnnpack_binary(
6116 name = "subgraph_size_test",
6117 srcs = ["test/subgraph-size.c"],
6118 deps = [":XNNPACK"],
6119)
6120
6121########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006122
6123xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006124 name = "abs_nc_test",
6125 srcs = [
6126 "test/abs-nc.cc",
6127 "test/abs-operator-tester.h",
6128 ],
6129 deps = OPERATOR_TEST_DEPS,
6130)
6131
6132xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006133 name = "add_nd_test",
6134 srcs = [
6135 "test/add-nd.cc",
6136 "test/binary-elementwise-operator-tester.h",
6137 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006138 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006139)
6140
6141xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006142 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006143 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006144 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006145 "test/argmax-pooling-operator-tester.h",
6146 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006147 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006148)
6149
6150xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006151 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006153 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154 "test/average-pooling-operator-tester.h",
6155 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006156 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006157)
6158
6159xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006160 name = "bankers_rounding_nc_test",
6161 srcs = [
6162 "test/bankers-rounding-nc.cc",
6163 "test/bankers-rounding-operator-tester.h",
6164 ],
6165 deps = OPERATOR_TEST_DEPS,
6166)
6167
6168xnnpack_unit_test(
6169 name = "ceiling_nc_test",
6170 srcs = [
6171 "test/ceiling-nc.cc",
6172 "test/ceiling-operator-tester.h",
6173 ],
6174 deps = OPERATOR_TEST_DEPS,
6175)
6176
6177xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006178 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006179 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006180 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006181 "test/channel-shuffle-operator-tester.h",
6182 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006183 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006184)
6185
6186xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006187 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006188 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006189 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006190 "test/clamp-operator-tester.h",
6191 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006192 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006193)
6194
6195xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006196 name = "constant_pad_nd_test",
6197 srcs = [
6198 "test/constant-pad-nd.cc",
6199 "test/constant-pad-operator-tester.h",
6200 ],
6201 deps = OPERATOR_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006205 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006206 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006207 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006208 "test/convolution-operator-tester.h",
6209 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006210 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006211)
6212
6213xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006214 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006216 "test/convolution-nchw.cc",
6217 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006219 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006220)
6221
6222xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006223 name = "copy_nc_test",
6224 srcs = [
6225 "test/copy-nc.cc",
6226 "test/copy-operator-tester.h",
6227 ],
6228 deps = OPERATOR_TEST_DEPS,
6229)
6230
6231xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006232 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006234 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006235 "test/deconvolution-operator-tester.h",
6236 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006237 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006238)
6239
6240xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006241 name = "divide_nd_test",
6242 srcs = [
6243 "test/binary-elementwise-operator-tester.h",
6244 "test/divide-nd.cc",
6245 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006246 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006247)
6248
6249xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006250 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006251 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006252 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 "test/fully-connected-operator-tester.h",
6254 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006255 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256)
6257
6258xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006259 name = "floor_nc_test",
6260 srcs = [
6261 "test/floor-nc.cc",
6262 "test/floor-operator-tester.h",
6263 ],
6264 deps = OPERATOR_TEST_DEPS,
6265)
6266
6267xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006268 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006269 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006270 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006271 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006272 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006273 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274)
6275
6276xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006277 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006278 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006279 "test/global-average-pooling-ncw.cc",
6280 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006282 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283)
6284
6285xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006286 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006288 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006289 "test/hardswish-operator-tester.h",
6290 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006291 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006292)
6293
6294xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006295 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006297 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006298 "test/leaky-relu-operator-tester.h",
6299 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006300 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006301)
6302
6303xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006304 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006306 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006307 "test/max-pooling-operator-tester.h",
6308 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006309 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006310)
6311
6312xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006313 name = "maximum_nd_test",
6314 srcs = [
6315 "test/binary-elementwise-operator-tester.h",
6316 "test/maximum-nd.cc",
6317 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006319)
6320
6321xnnpack_unit_test(
6322 name = "minimum_nd_test",
6323 srcs = [
6324 "test/binary-elementwise-operator-tester.h",
6325 "test/minimum-nd.cc",
6326 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006327 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006328)
6329
6330xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006331 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006332 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006333 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006334 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006335 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006336 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006337)
6338
6339xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006340 name = "negate_nc_test",
6341 srcs = [
6342 "test/negate-nc.cc",
6343 "test/negate-operator-tester.h",
6344 ],
6345 deps = OPERATOR_TEST_DEPS,
6346)
6347
6348xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006349 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006350 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006351 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006352 "test/prelu-operator-tester.h",
6353 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006355)
6356
6357xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006359 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006360 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006361 "test/resize-bilinear-operator-tester.h",
6362 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006363 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006364)
6365
6366xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006367 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006368 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006369 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006370 "test/sigmoid-operator-tester.h",
6371 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006372 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006373)
6374
6375xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006376 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006377 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006378 "test/softmax-nc.cc",
6379 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006381 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006382)
6383
6384xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006385 name = "square_nc_test",
6386 srcs = [
6387 "test/square-nc.cc",
6388 "test/square-operator-tester.h",
6389 ],
6390 deps = OPERATOR_TEST_DEPS,
6391)
6392
6393xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006394 name = "square_root_nc_test",
6395 srcs = [
6396 "test/square-root-nc.cc",
6397 "test/square-root-operator-tester.h",
6398 ],
6399 deps = OPERATOR_TEST_DEPS,
6400)
6401
6402xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006403 name = "squared_difference_nd_test",
6404 srcs = [
6405 "test/binary-elementwise-operator-tester.h",
6406 "test/squared-difference-nd.cc",
6407 ],
6408 deps = OPERATOR_TEST_DEPS,
6409)
6410
6411xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006412 name = "subtract_nd_test",
6413 srcs = [
6414 "test/binary-elementwise-operator-tester.h",
6415 "test/subtract-nd.cc",
6416 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006417 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006418)
6419
6420xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006421 name = "truncation_nc_test",
6422 srcs = [
6423 "test/truncation-nc.cc",
6424 "test/truncation-operator-tester.h",
6425 ],
6426 deps = OPERATOR_TEST_DEPS,
6427)
6428
6429xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006430 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006431 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006432 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006433 "test/unpooling-operator-tester.h",
6434 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006435 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006436)
6437
Chao Mei6ddfc602020-05-13 22:29:36 -07006438############################### Misc unit tests ###############################
6439
6440xnnpack_unit_test(
6441 name = "memory_planner_test",
6442 srcs = [
6443 "test/memory-planner-test.cc",
6444 ],
6445 deps = [
6446 ":XNNPACK",
6447 ":memory_planner",
6448 ],
6449)
6450
Marat Dukhan08c4a432019-10-03 09:29:21 -07006451############################# Build configurations #############################
6452
Marat Dukhanb8642352019-10-30 15:43:02 -07006453# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006454config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006455 name = "xnn_enable_assembly_explicit_true",
6456 define_values = {"xnn_enable_assembly": "true"},
6457)
6458
6459# Disables usage of assembly kernels.
6460config_setting(
6461 name = "xnn_enable_assembly_explicit_false",
6462 define_values = {"xnn_enable_assembly": "false"},
6463)
6464
Marat Dukhan9de90e02020-06-18 16:04:12 -07006465# Enables usage of sparse inference.
6466config_setting(
6467 name = "xnn_enable_sparse_explicit_true",
6468 define_values = {"xnn_enable_sparse": "true"},
6469)
6470
6471# Disables usage of sparse inference.
6472config_setting(
6473 name = "xnn_enable_sparse_explicit_false",
6474 define_values = {"xnn_enable_sparse": "false"},
6475)
6476
Marat Dukhan05702cf2020-03-26 15:41:33 -07006477# Disables usage of HMP-aware optimizations.
6478config_setting(
6479 name = "xnn_enable_hmp_explicit_false",
6480 define_values = {"xnn_enable_hmp": "false"},
6481)
6482
Chao Mei6ddfc602020-05-13 22:29:36 -07006483# Enable usage of optimized memory allocation
6484config_setting(
6485 name = "xnn_enable_memopt_explicit_true",
6486 define_values = {"xnnpack_enable_memopt": "true"},
6487)
6488
6489# Disable usage of optimized memory allocation
6490config_setting(
6491 name = "xnn_enable_memopt_explicit_false",
6492 define_values = {"xnnpack_enable_memopt": "false"},
6493)
6494
Marat Dukhanb8642352019-10-30 15:43:02 -07006495# Builds with -c dbg
6496config_setting(
6497 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006498 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006499 "compilation_mode": "dbg",
6500 },
6501)
6502
6503# Builds with -c opt
6504config_setting(
6505 name = "optimized_build",
6506 values = {
6507 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006508 },
6509)
6510
6511config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006512 name = "linux_k8",
6513 values = {"cpu": "k8"},
6514)
6515
6516config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006517 name = "linux_arm",
6518 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006519)
6520
6521config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006522 name = "linux_armeabi",
6523 values = {"cpu": "armeabi"},
6524)
6525
6526config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006527 name = "linux_armhf",
6528 values = {"cpu": "armhf"},
6529)
6530
6531config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006532 name = "linux_armv7a",
6533 values = {"cpu": "armv7a"},
6534)
6535
6536config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006537 name = "linux_aarch64",
6538 values = {"cpu": "aarch64"},
6539)
6540
6541config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006542 name = "android",
6543 values = {"crosstool_top": "//external:android/crosstool"},
6544)
6545
6546config_setting(
6547 name = "android_armv7",
6548 values = {
6549 "crosstool_top": "//external:android/crosstool",
6550 "cpu": "armeabi-v7a",
6551 },
6552)
6553
6554config_setting(
6555 name = "android_arm64",
6556 values = {
6557 "crosstool_top": "//external:android/crosstool",
6558 "cpu": "arm64-v8a",
6559 },
6560)
6561
6562config_setting(
6563 name = "android_x86",
6564 values = {
6565 "crosstool_top": "//external:android/crosstool",
6566 "cpu": "x86",
6567 },
6568)
6569
6570config_setting(
6571 name = "android_x86_64",
6572 values = {
6573 "crosstool_top": "//external:android/crosstool",
6574 "cpu": "x86_64",
6575 },
6576)
6577
6578config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006579 name = "windows_x86_64",
6580 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006581)
6582
6583config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006584 name = "windows_x86_64_clang",
6585 values = {
6586 "compiler": "clang-cl",
6587 "cpu": "x64_windows",
6588 },
6589)
6590
6591config_setting(
6592 name = "windows_x86_64_mingw",
6593 values = {
6594 "compiler": "mingw-gcc",
6595 "cpu": "x64_windows",
6596 },
6597)
6598
6599config_setting(
6600 name = "windows_x86_64_msys",
6601 values = {
6602 "compiler": "msys-gcc",
6603 "cpu": "x64_windows",
6604 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006605)
6606
6607config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006608 name = "macos_x86_64",
6609 values = {
6610 "apple_platform_type": "macos",
6611 "cpu": "darwin",
6612 },
6613)
6614
6615config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006616 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006617 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006618)
6619
6620config_setting(
6621 name = "emscripten_wasm",
6622 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006623 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006624 "cpu": "wasm",
6625 },
6626)
6627
6628config_setting(
6629 name = "emscripten_wasmsimd",
6630 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006631 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006632 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006633 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006634 },
6635)
6636
6637config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006638 name = "ios_armv7",
6639 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006640 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006641 "cpu": "ios_armv7",
6642 },
6643)
6644
6645config_setting(
6646 name = "ios_arm64",
6647 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006648 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006649 "cpu": "ios_arm64",
6650 },
6651)
6652
6653config_setting(
6654 name = "ios_arm64e",
6655 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006656 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006657 "cpu": "ios_arm64e",
6658 },
6659)
6660
6661config_setting(
6662 name = "ios_x86",
6663 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006664 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006665 "cpu": "ios_i386",
6666 },
6667)
6668
6669config_setting(
6670 name = "ios_x86_64",
6671 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006672 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006673 "cpu": "ios_x86_64",
6674 },
6675)
6676
6677config_setting(
6678 name = "watchos_armv7k",
6679 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006680 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006681 "cpu": "watchos_armv7k",
6682 },
6683)
6684
6685config_setting(
6686 name = "watchos_arm64_32",
6687 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006688 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006689 "cpu": "watchos_arm64_32",
6690 },
6691)
6692
6693config_setting(
6694 name = "watchos_x86",
6695 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006696 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006697 "cpu": "watchos_i386",
6698 },
6699)
6700
6701config_setting(
6702 name = "watchos_x86_64",
6703 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006704 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006705 "cpu": "watchos_x86_64",
6706 },
6707)
6708
6709config_setting(
6710 name = "tvos_arm64",
6711 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006712 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006713 "cpu": "tvos_arm64",
6714 },
6715)
6716
6717config_setting(
6718 name = "tvos_x86_64",
6719 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006720 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006721 "cpu": "tvos_x86_64",
6722 },
6723)