blob: fb0d2de0c5b9f20d510cae43f2ef6983d02a2980 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800425 "src/math/sigmoid-scalar-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat 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 = [
1750 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1751 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001753 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1754 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1755 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001756]
1757
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001759 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1760 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001761 "src/f32-clamp/gen/sse-x4.c",
1762 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001763 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1764 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001765 "src/f32-dwconv-chw/3x3p1-sse.c",
1766 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001767 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1768 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1769 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1770 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1771 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1772 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1773 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1775 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1777 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001779 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001780 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1781 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001782 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1783 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1784 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1785 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1786 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1787 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1788 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1789 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1790 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1791 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1792 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1793 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1794 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001795 "src/f32-hswish/gen/hswish-sse-x4.c",
1796 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001797 "src/f32-ibilinear/gen/sse-c4.c",
1798 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001799 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1800 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1801 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1802 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1803 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1804 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1805 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001806 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1807 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1808 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001809 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001810 "src/f32-prelu/gen/sse-2x4.c",
1811 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001812 "src/f32-relu/gen/sse-x4.c",
1813 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001815 "src/f32-spmm/gen/4x1-minmax-sse.c",
1816 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001817 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001818 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1819 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1820 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1821 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1822 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001826 "src/f32-vbinary/gen/vmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1829 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1830 "src/f32-vbinary/gen/vmin-sse-x4.c",
1831 "src/f32-vbinary/gen/vmin-sse-x8.c",
1832 "src/f32-vbinary/gen/vminc-sse-x4.c",
1833 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001834 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1835 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1836 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1837 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1838 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001842 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1843 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1844 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1845 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001846 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1848 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1849 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001850 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1851 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001852 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1853 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001854 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1855 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001856 "src/f32-vunary/gen/vabs-sse-x4.c",
1857 "src/f32-vunary/gen/vabs-sse-x8.c",
1858 "src/f32-vunary/gen/vneg-sse-x4.c",
1859 "src/f32-vunary/gen/vneg-sse-x8.c",
1860 "src/f32-vunary/gen/vsqr-sse-x4.c",
1861 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001862 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001864 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001865 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001866 "src/math/roundd-sse-addsub.c",
1867 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001868 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001869 "src/math/sqrt-sse-hh1mac.c",
1870 "src/math/sqrt-sse-nr1mac.c",
1871 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872]
1873
1874SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001875 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1876 "src/f32-argmaxpool/4x-sse2-c4.c",
1877 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001878 "src/f32-prelu/gen/sse2-2x4.c",
1879 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001880 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1881 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1882 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1883 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1884 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1885 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001892 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001893 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001894 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001895 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001896 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1897 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001898 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1899 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001900 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1901 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1902 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1903 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1904 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1905 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1906 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1907 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001908 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1909 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1910 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001911 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1912 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1913 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001914 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1915 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1916 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001917 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1918 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001919 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1920 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001921 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1922 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001923 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1924 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001925 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001926 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1927 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001928 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001929 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1930 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1931 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001932 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1933 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001934 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1935 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001936 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1937 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001938 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001939 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1940 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001941 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001942 "src/qs8-requantization/precise-sse2.c",
1943 "src/qs8-requantization/fp32-sse2.c",
1944 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001945 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1946 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1947 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1948 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001949 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1950 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1951 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1952 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001953 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1954 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1955 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1956 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1957 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1958 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1959 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1960 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001961 "src/qu8-requantization/precise-sse2.c",
1962 "src/qu8-requantization/fp32-sse2.c",
1963 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001964 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001965 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001966 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001968 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "src/x32-zip/x2-sse2.c",
1970 "src/x32-zip/x3-sse2.c",
1971 "src/x32-zip/x4-sse2.c",
1972 "src/x32-zip/xm-sse2.c",
1973 "src/x8-zip/x2-sse2.c",
1974 "src/x8-zip/x3-sse2.c",
1975 "src/x8-zip/x4-sse2.c",
1976 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001977 "src/math/exp-sse2-p5.c",
1978 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001979 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001980 "src/math/roundd-sse2-cvt.c",
1981 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001982 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001983 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001984]
1985
1986SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001987 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1988 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1989 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001990 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1991 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1992 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001993 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1994 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1995 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001996 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1997 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001998 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1999 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002000 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2001 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002002 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2003 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002004 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002005 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2006 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002007 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002008 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2009 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2010 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002011 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2012 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002013 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2014 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002015 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2016 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002017 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002018 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2019 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002020 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002021 "src/qs8-requantization/precise-ssse3.c",
2022 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002023 "src/qu8-requantization/precise-ssse3.c",
2024 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025]
2026
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002027SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002028 "src/f32-prelu/gen/sse41-2x4.c",
2029 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002030 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002031 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002032 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002033 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002034 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2035 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002036 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2037 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002038 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2039 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2040 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2041 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2042 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2043 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2044 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2045 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002046 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2047 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2048 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002049 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2050 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2051 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002052 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2053 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2054 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002055 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2056 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002057 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2058 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002059 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2060 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002061 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2062 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002063 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002064 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2065 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002066 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002067 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2068 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2069 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002070 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2071 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002072 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2073 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002074 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2075 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002076 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002077 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2078 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002079 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002080 "src/qs8-requantization/fp32-sse4.c",
2081 "src/qs8-requantization/precise-sse4.c",
2082 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002083 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2084 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2085 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2086 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002087 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2088 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2089 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2090 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002091 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2092 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2093 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2094 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002095 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2096 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2097 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2098 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002099 "src/qu8-requantization/precise-sse4.c",
2100 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002101 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002102 "src/math/roundd-sse41.c",
2103 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002104 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002105]
2106
Marat Dukhan08c4a432019-10-03 09:29:21 -07002107AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002108 "src/f32-clamp/gen/avx-x8.c",
2109 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002110 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2111 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2112 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2113 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2114 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2115 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2116 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2117 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2118 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2119 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2120 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2121 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2122 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2123 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2124 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2125 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2126 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2127 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2128 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2129 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2130 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2131 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2132 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2133 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2134 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2135 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2136 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2137 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2138 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2139 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002140 "src/f32-hswish/gen/hswish-avx-x8.c",
2141 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002142 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2143 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2144 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2145 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2146 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2147 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2148 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2149 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2150 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002151 "src/f32-prelu/gen/avx-2x8.c",
2152 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002153 "src/f32-relu/gen/avx-x16.c",
2154 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002155 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002156 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2157 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2158 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2159 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2160 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2161 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2162 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2163 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002164 "src/f32-vbinary/gen/vmax-avx-x8.c",
2165 "src/f32-vbinary/gen/vmax-avx-x16.c",
2166 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2167 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2168 "src/f32-vbinary/gen/vmin-avx-x8.c",
2169 "src/f32-vbinary/gen/vmin-avx-x16.c",
2170 "src/f32-vbinary/gen/vminc-avx-x8.c",
2171 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002172 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2173 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2174 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2175 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2176 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2177 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2178 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2179 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002180 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2181 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2182 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2183 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002184 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2185 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2186 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2187 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002188 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2189 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002190 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002191 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2192 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2193 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2194 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2195 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2196 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2197 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2198 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002199 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2200 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002201 "src/f32-vunary/gen/vabs-avx-x8.c",
2202 "src/f32-vunary/gen/vabs-avx-x16.c",
2203 "src/f32-vunary/gen/vneg-avx-x8.c",
2204 "src/f32-vunary/gen/vneg-avx-x16.c",
2205 "src/f32-vunary/gen/vsqr-avx-x8.c",
2206 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002207]
2208
Marat Dukhan1566fee2020-08-02 21:55:41 -07002209XOP_UKERNELS = [
2210 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2211 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2212 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2213 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002214 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2215 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002216 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2217 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002218 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002219 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2220 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002221 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002222 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2223 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2224 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002225 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2226 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2227 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2228 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2229 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2230 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002231 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002232 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2233 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002234 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002235 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2236 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2237 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2238 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2239 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2240 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2241 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2242 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002243]
2244
Marat Dukhanfda12b82019-11-21 12:27:59 -08002245FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002246 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2247 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2248 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2249 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2250 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2251 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2252 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2253 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2254 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2255 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2256 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2257 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2258 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2259 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2260 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2261 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2262 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2263 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2264 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2265 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2266 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2267 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2268 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2269 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2270 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2271 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2272 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2273 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2274 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2275 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2276 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2277 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2278 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2279 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2280 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2281 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2282 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2283 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2284 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2285 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002286 "src/f32-hswish/gen/hswish-fma3-x8.c",
2287 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002288 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2289 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2290 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2291 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2292 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2293 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2294 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2295 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2296 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2297 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2298 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2299 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2300 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2301 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002302 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2303 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2304 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2305 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2306 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2307 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2308 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2309 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002310 "src/math/sqrt-fma3-nr1fma.c",
2311 "src/math/sqrt-fma3-nr2fma.c",
2312 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002313]
2314
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002315AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002316 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2317 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2318 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2319 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2320 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2321 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2322 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2323 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2324 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2325 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2326 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2327 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2328 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2329 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2330 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2331 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2332 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2333 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2334 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2335 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2336 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2337 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2338 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2339 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2340 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2341 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2342 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2343 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2344 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2345 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2346 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2347 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2348 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2349 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2350 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2351 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002352 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2353 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2354 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2355 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2356 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2357 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2358 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2359 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2360 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2361 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2362 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2363 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2364 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2365 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2366 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2367 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2368 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2369 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2370 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2371 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2372 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2373 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2374 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2375 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2376 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2377 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2378 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2379 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2380 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2381 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002382 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2383 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2384 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2385 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2386 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2387 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2388 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2389 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2390 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2391 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2392 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2393 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2394 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2395 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2396 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2397 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2398 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2399 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2400 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2401 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2402 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2403 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2404 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2405 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002406 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2407 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002408 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2409 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2410 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2411 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002412 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2413 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2414 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002415 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2416 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2417 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002418 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2419 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2420 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002421 "src/math/exp-avx2-p5.c",
2422 "src/math/exp-avx2-perm-p3.c",
2423 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002424 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002425 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002426 "src/math/sigmoid-avx2-rr2-p5-div.c",
2427 "src/math/sigmoid-avx2-rr1-p5-div.c",
2428 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2429 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2430 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2431 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002432]
2433
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002435 "src/f32-clamp/gen/avx512f-x16.c",
2436 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002437 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2438 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2439 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2440 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2441 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2442 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2443 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2444 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2445 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2446 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2447 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2448 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2449 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2450 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2451 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2452 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2453 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2454 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2455 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2456 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2457 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2458 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2459 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2460 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002461 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2462 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002463 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2464 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2465 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2466 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2467 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2468 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002469 "src/f32-prelu/gen/avx512f-2x16.c",
2470 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002471 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2472 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2473 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2474 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2475 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2476 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2477 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2478 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2479 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2480 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2481 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2482 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2483 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2484 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2485 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2486 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2487 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2488 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2489 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2490 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2491 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2492 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2493 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2494 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2495 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2496 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2497 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2498 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2499 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2500 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2501 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2502 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2503 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2504 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2505 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2506 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002507 "src/f32-relu/gen/avx512f-x16.c",
2508 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002509 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002510 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2511 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2512 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2513 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2514 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2515 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2516 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2517 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002518 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2519 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2520 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2521 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2522 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2523 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2524 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2525 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002526 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2527 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2528 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2529 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2530 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2531 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2532 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2533 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002534 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2535 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2536 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2537 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002538 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2539 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2540 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2541 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002542 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2543 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002544 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002545 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2546 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2547 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2548 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2549 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2550 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2551 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2552 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2553 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2554 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2555 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2556 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2557 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2558 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2559 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2560 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2561 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2562 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2563 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2564 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2565 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2566 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2567 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2568 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002569 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2570 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2571 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2572 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2573 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2574 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2575 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2576 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002577 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2578 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2579 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2580 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2581 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2582 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2583 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2584 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002585 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2586 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2587 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2588 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2589 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2590 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002591 "src/math/exp-avx512f-p5-scalef.c",
2592 "src/math/exp-avx512f-p5.c",
2593 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002594 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002595 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002596 "src/math/sqrt-avx512f-nr1fma.c",
2597 "src/math/sqrt-avx512f-nr2fma.c",
2598 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002599]
2600
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002601AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002602 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2603 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002604 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2605 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2606 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2607 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2608 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2609 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2610 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2611 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2612]
2613
Frank Barchardbcedc082020-08-17 18:00:51 -07002614WASM32_ASM_UKERNELS = [
2615 "src/f32-relu/wasm_shr_x1.S",
2616 "src/f32-relu/wasm_shr_x2.S",
2617 "src/f32-relu/wasm_shr_x4.S",
2618]
2619
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002621 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002622 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002623 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2624 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002625 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002626 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002627 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002628 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002629 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2630 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002631 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2632 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2633 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2634 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002635]
2636
2637AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002638 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002639 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2640 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002641 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002642 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2643 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002644 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2645 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2646 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002647 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002648 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2649 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2650 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002651 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002652 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2653 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002654 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002655 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2656 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2657 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2658 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2659 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2660 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2661 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2662 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2663 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2664 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2665 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2666 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2667 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2668 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2669 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2670 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2671 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2672 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002673 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2674 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002675 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002676 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2677 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2678 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2679 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2680 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2681 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2682 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2683 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2684 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2685 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2686 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2687 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2688 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2689 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2690 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2691 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2692 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2693 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002694 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2695 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2696 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2697 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2698 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2699 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2700 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2701 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2702 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2703 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2704 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2705 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2706 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2707 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2708 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2709 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2710 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2711 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712]
2713
Marat Dukhan1b354632020-03-23 12:50:22 -07002714INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002715 "src/xnnpack/argmaxpool.h",
2716 "src/xnnpack/avgpool.h",
2717 "src/xnnpack/clamp.h",
2718 "src/xnnpack/common.h",
2719 "src/xnnpack/conv.h",
2720 "src/xnnpack/dwconv.h",
2721 "src/xnnpack/gavgpool.h",
2722 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002723 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002725 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002726 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002727 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002728 "src/xnnpack/lut.h",
2729 "src/xnnpack/math.h",
2730 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002731 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 "src/xnnpack/packx.h",
2733 "src/xnnpack/pad.h",
2734 "src/xnnpack/params.h",
2735 "src/xnnpack/pavgpool.h",
2736 "src/xnnpack/ppmm.h",
2737 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002738 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002739 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002740 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 "src/xnnpack/rmax.h",
2742 "src/xnnpack/scalar-utils.h",
2743 "src/xnnpack/spmm.h",
2744 "src/xnnpack/unpool.h",
2745 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002746 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002748 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002749 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002750 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002751 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002753]
2754
2755INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 "include/xnnpack.h",
2757 "src/xnnpack/allocator.h",
2758 "src/xnnpack/compute.h",
2759 "src/xnnpack/im2col.h",
2760 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002761 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002762 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002763 "src/xnnpack/operator.h",
2764 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002765 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002766 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002767 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002768 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002769]
2770
Marat Dukhan1b354632020-03-23 12:50:22 -07002771ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002772 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002773]
2774
Marat Dukhan1b354632020-03-23 12:50:22 -07002775MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002776 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002777 "include/xnnpack.h",
2778]
2779
Marat Dukhan1b354632020-03-23 12:50:22 -07002780MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002781 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002782 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 "src/xnnpack/requantization.h",
2784 "include/xnnpack.h",
2785]
2786
2787OPERATOR_TEST_PARAMS_HDRS = [
2788 "src/xnnpack/params.h",
2789 "src/xnnpack/common.h",
2790]
2791
2792WEIGHTS_PACK_HDRS = [
2793 "src/xnnpack/pack.h",
2794 "src/xnnpack/operator.h",
2795 "src/xnnpack/compute.h",
2796]
2797
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002798LOGGING_COPTS = select({
2799 # No logging in optimized mode
2800 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2801 # Full logging in debug mode
2802 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2803 # Error-only logging in default (fastbuild) mode
2804 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2805})
2806
Marat Dukhan3b59de22020-06-03 20:15:19 -07002807LOGGING_SRCS = select({
2808 # No logging in optimized mode
2809 ":optimized_build": [],
2810 "//conditions:default": [
2811 "src/operator-strings.c",
2812 "src/subgraph-strings.c",
2813 ],
2814})
2815
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002816LOGGING_HDRS = [
2817 "src/xnnpack/log.h",
2818]
2819
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002821 name = "tables",
2822 srcs = TABLE_SRCS,
2823 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002824 gcc_copts = xnnpack_gcc_std_copts(),
2825 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002826)
2827
2828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002829 name = "scalar_ukernels",
2830 srcs = SCALAR_UKERNELS,
2831 hdrs = INTERNAL_HDRS,
2832 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002833 gcc_copts = xnnpack_gcc_std_copts(),
2834 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002836 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 "@FP16",
2838 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002839 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 ],
2841)
2842
2843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002844 name = "scalar_ukernels_test_mode",
2845 srcs = SCALAR_UKERNELS,
2846 hdrs = INTERNAL_HDRS,
2847 aarch32_copts = ["-marm"],
2848 copts = [
2849 "-UNDEBUG",
2850 "-DXNN_TEST_MODE=1",
2851 ],
2852 gcc_copts = xnnpack_gcc_std_copts(),
2853 msvc_copts = xnnpack_msvc_std_copts(),
2854 deps = [
2855 ":tables",
2856 "@FP16",
2857 "@FXdiv",
2858 "@pthreadpool",
2859 ],
2860)
2861
2862xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002863 name = "wasm_ukernels",
2864 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002865 gcc_copts = xnnpack_gcc_std_copts(),
2866 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002867 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002868 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002869 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002870 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002871 "@FP16",
2872 "@FXdiv",
2873 "@pthreadpool",
2874 ],
2875)
2876
2877xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002878 name = "wasm_ukernels_test_mode",
2879 hdrs = INTERNAL_HDRS,
2880 copts = [
2881 "-UNDEBUG",
2882 "-DXNN_TEST_MODE=1",
2883 ],
2884 gcc_copts = xnnpack_gcc_std_copts(),
2885 msvc_copts = xnnpack_msvc_std_copts(),
2886 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002887 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002888 deps = [
2889 ":tables",
2890 "@FP16",
2891 "@FXdiv",
2892 "@pthreadpool",
2893 ],
2894)
2895
2896xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002897 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 hdrs = INTERNAL_HDRS,
2899 aarch32_copts = [
2900 "-marm",
2901 "-mfpu=neon",
2902 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002903 gcc_copts = xnnpack_gcc_std_copts(),
2904 gcc_x86_copts = ["-msse2"],
2905 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906 optimized_copts = [
2907 "-O3",
2908 "-ffast-math",
2909 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002910 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002912 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 "@FP16",
2914 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002915 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002916 ],
2917)
2918
2919xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002920 name = "psimd_fastmath_ukernels_test_mode",
2921 hdrs = INTERNAL_HDRS,
2922 aarch32_copts = [
2923 "-marm",
2924 "-mfpu=neon",
2925 ],
2926 copts = [
2927 "-UNDEBUG",
2928 "-DXNN_TEST_MODE=1",
2929 ],
2930 gcc_copts = xnnpack_gcc_std_copts(),
2931 gcc_x86_copts = ["-msse2"],
2932 msvc_copts = xnnpack_msvc_std_copts(),
2933 optimized_copts = [
2934 "-O3",
2935 "-ffast-math",
2936 ],
2937 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2938 deps = [
2939 ":tables",
2940 "@FP16",
2941 "@psimd",
2942 "@pthreadpool",
2943 ],
2944)
2945
2946xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002947 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002948 hdrs = INTERNAL_HDRS,
2949 aarch32_copts = [
2950 "-marm",
2951 "-mfpu=neon",
2952 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002953 gcc_copts = xnnpack_gcc_std_copts(),
2954 gcc_x86_copts = ["-msse2"],
2955 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002956 optimized_copts = [
2957 "-O3",
2958 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002959 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002960 deps = [
2961 ":tables",
2962 "@FP16",
2963 "@psimd",
2964 "@pthreadpool",
2965 ],
2966)
2967
2968xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002969 name = "psimd_accmath_ukernels_test_mode",
2970 hdrs = INTERNAL_HDRS,
2971 aarch32_copts = [
2972 "-marm",
2973 "-mfpu=neon",
2974 ],
2975 copts = [
2976 "-UNDEBUG",
2977 "-DXNN_TEST_MODE=1",
2978 ],
2979 gcc_copts = xnnpack_gcc_std_copts(),
2980 gcc_x86_copts = ["-msse2"],
2981 msvc_copts = xnnpack_msvc_std_copts(),
2982 optimized_copts = [
2983 "-O3",
2984 ],
2985 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2986 deps = [
2987 ":tables",
2988 "@FP16",
2989 "@psimd",
2990 "@pthreadpool",
2991 ],
2992)
2993
2994xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 name = "neon_ukernels",
2996 hdrs = INTERNAL_HDRS,
2997 aarch32_copts = [
2998 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002999 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 "-mfpu=neon",
3001 ],
3002 aarch32_srcs = NEON_UKERNELS,
3003 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003004 gcc_copts = xnnpack_gcc_std_copts(),
3005 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003006 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003007 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003008 "@FP16",
3009 "@pthreadpool",
3010 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011)
3012
3013xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003014 name = "neon_ukernels_test_mode",
3015 hdrs = INTERNAL_HDRS,
3016 aarch32_copts = [
3017 "-marm",
3018 "-march=armv7-a",
3019 "-mfpu=neon",
3020 ],
3021 aarch32_srcs = NEON_UKERNELS,
3022 aarch64_srcs = NEON_UKERNELS,
3023 copts = [
3024 "-UNDEBUG",
3025 "-DXNN_TEST_MODE=1",
3026 ],
3027 gcc_copts = xnnpack_gcc_std_copts(),
3028 msvc_copts = xnnpack_msvc_std_copts(),
3029 deps = [
3030 ":tables",
3031 "@FP16",
3032 "@pthreadpool",
3033 ],
3034)
3035
3036xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003037 name = "neonfma_ukernels",
3038 hdrs = INTERNAL_HDRS,
3039 aarch32_copts = [
3040 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003041 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003042 "-mfpu=neon-vfpv4",
3043 ],
3044 aarch32_srcs = NEONFMA_UKERNELS,
3045 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003046 apple_aarch32_copts = [
3047 "-mcpu=swift",
3048 "-mtune=generic",
3049 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003050 gcc_copts = xnnpack_gcc_std_copts(),
3051 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003052 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003053 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003054 "@FP16",
3055 "@pthreadpool",
3056 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003057)
3058
3059xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003060 name = "neonfma_ukernels_test_mode",
3061 hdrs = INTERNAL_HDRS,
3062 aarch32_copts = [
3063 "-marm",
3064 "-march=armv7-a",
3065 "-mfpu=neon-vfpv4",
3066 ],
3067 aarch32_srcs = NEONFMA_UKERNELS,
3068 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003069 apple_aarch32_copts = [
3070 "-mcpu=swift",
3071 "-mtune=generic",
3072 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003073 copts = [
3074 "-UNDEBUG",
3075 "-DXNN_TEST_MODE=1",
3076 ],
3077 gcc_copts = xnnpack_gcc_std_copts(),
3078 msvc_copts = xnnpack_msvc_std_copts(),
3079 deps = [
3080 ":tables",
3081 "@FP16",
3082 "@pthreadpool",
3083 ],
3084)
3085
3086xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003087 name = "neonv8_ukernels",
3088 hdrs = INTERNAL_HDRS,
3089 aarch32_copts = [
3090 "-marm",
3091 "-march=armv8-a",
3092 "-mfpu=neon-fp-armv8",
3093 ],
3094 aarch32_srcs = NEONV8_UKERNELS,
3095 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003096 apple_aarch32_copts = [
3097 "-mcpu=cyclone",
3098 "-mtune=generic",
3099 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003100 gcc_copts = xnnpack_gcc_std_copts(),
3101 msvc_copts = xnnpack_msvc_std_copts(),
3102 deps = [
3103 ":tables",
3104 "@FP16",
3105 "@pthreadpool",
3106 ],
3107)
3108
3109xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003110 name = "neonv8_ukernels_test_mode",
3111 hdrs = INTERNAL_HDRS,
3112 aarch32_copts = [
3113 "-marm",
3114 "-march=armv8-a",
3115 "-mfpu=neon-fp-armv8",
3116 ],
3117 aarch32_srcs = NEONV8_UKERNELS,
3118 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003119 apple_aarch32_copts = [
3120 "-mcpu=cyclone",
3121 "-mtune=generic",
3122 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003123 copts = [
3124 "-UNDEBUG",
3125 "-DXNN_TEST_MODE=1",
3126 ],
3127 gcc_copts = xnnpack_gcc_std_copts(),
3128 msvc_copts = xnnpack_msvc_std_copts(),
3129 deps = [
3130 ":tables",
3131 "@FP16",
3132 "@pthreadpool",
3133 ],
3134)
3135
3136xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 name = "neonfp16arith_ukernels",
3138 hdrs = INTERNAL_HDRS,
3139 aarch64_copts = ["-march=armv8.2-a+fp16"],
3140 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003141 gcc_copts = xnnpack_gcc_std_copts(),
3142 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003143 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003144 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003145 "@FP16",
3146 "@pthreadpool",
3147 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148)
3149
3150xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003151 name = "neonfp16arith_ukernels_test_mode",
3152 hdrs = INTERNAL_HDRS,
3153 aarch64_copts = ["-march=armv8.2-a+fp16"],
3154 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3155 copts = [
3156 "-UNDEBUG",
3157 "-DXNN_TEST_MODE=1",
3158 ],
3159 gcc_copts = xnnpack_gcc_std_copts(),
3160 msvc_copts = xnnpack_msvc_std_copts(),
3161 deps = [
3162 ":tables",
3163 "@FP16",
3164 "@pthreadpool",
3165 ],
3166)
3167
3168xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003169 name = "neondot_ukernels",
3170 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003171 aarch32_copts = [
3172 "-marm",
3173 "-march=armv8.2-a+dotprod",
3174 "-mfpu=neon-fp-armv8",
3175 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003176 aarch32_srcs = NEONDOT_UKERNELS,
3177 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3178 aarch64_srcs = NEONDOT_UKERNELS,
3179 gcc_copts = xnnpack_gcc_std_copts(),
3180 msvc_copts = xnnpack_msvc_std_copts(),
3181 deps = [
3182 ":tables",
3183 "@FP16",
3184 "@pthreadpool",
3185 ],
3186)
3187
3188xnnpack_cc_library(
3189 name = "neondot_ukernels_test_mode",
3190 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003191 aarch32_copts = [
3192 "-marm",
3193 "-march=armv8.2-a+dotprod",
3194 "-mfpu=neon-fp-armv8",
3195 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003196 aarch32_srcs = NEONDOT_UKERNELS,
3197 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3198 aarch64_srcs = NEONDOT_UKERNELS,
3199 copts = [
3200 "-UNDEBUG",
3201 "-DXNN_TEST_MODE=1",
3202 ],
3203 gcc_copts = xnnpack_gcc_std_copts(),
3204 msvc_copts = xnnpack_msvc_std_copts(),
3205 deps = [
3206 ":tables",
3207 "@FP16",
3208 "@pthreadpool",
3209 ],
3210)
3211
3212xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213 name = "sse2_ukernels",
3214 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003215 gcc_copts = xnnpack_gcc_std_copts(),
3216 gcc_x86_copts = ["-msse2"],
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003220 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003221 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003222 "@FP16",
3223 "@pthreadpool",
3224 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003225)
3226
3227xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003228 name = "sse2_ukernels_test_mode",
3229 hdrs = INTERNAL_HDRS,
3230 copts = [
3231 "-UNDEBUG",
3232 "-DXNN_TEST_MODE=1",
3233 ],
3234 gcc_copts = xnnpack_gcc_std_copts(),
3235 gcc_x86_copts = ["-msse2"],
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 msvc_x86_32_copts = ["/arch:SSE2"],
3238 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3239 deps = [
3240 ":tables",
3241 "@FP16",
3242 "@pthreadpool",
3243 ],
3244)
3245
3246xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003247 name = "ssse3_ukernels",
3248 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003249 gcc_copts = xnnpack_gcc_std_copts(),
3250 gcc_x86_copts = ["-mssse3"],
3251 msvc_copts = xnnpack_msvc_std_copts(),
3252 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003253 x86_srcs = SSSE3_UKERNELS,
3254 deps = [
3255 ":tables",
3256 "@FP16",
3257 "@pthreadpool",
3258 ],
3259)
3260
3261xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003262 name = "ssse3_ukernels_test_mode",
3263 hdrs = INTERNAL_HDRS,
3264 copts = [
3265 "-UNDEBUG",
3266 "-DXNN_TEST_MODE=1",
3267 ],
3268 gcc_copts = xnnpack_gcc_std_copts(),
3269 gcc_x86_copts = ["-mssse3"],
3270 msvc_copts = xnnpack_msvc_std_copts(),
3271 msvc_x86_32_copts = ["/arch:SSE2"],
3272 x86_srcs = SSSE3_UKERNELS,
3273 deps = [
3274 ":tables",
3275 "@FP16",
3276 "@pthreadpool",
3277 ],
3278)
3279
3280xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003281 name = "sse41_ukernels",
3282 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003283 gcc_copts = xnnpack_gcc_std_copts(),
3284 gcc_x86_copts = ["-msse4.1"],
3285 msvc_copts = xnnpack_msvc_std_copts(),
3286 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003287 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003288 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003289 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003290 "@FP16",
3291 "@pthreadpool",
3292 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003293)
3294
3295xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003296 name = "sse41_ukernels_test_mode",
3297 hdrs = INTERNAL_HDRS,
3298 copts = [
3299 "-UNDEBUG",
3300 "-DXNN_TEST_MODE=1",
3301 ],
3302 gcc_copts = xnnpack_gcc_std_copts(),
3303 gcc_x86_copts = ["-msse4.1"],
3304 msvc_copts = xnnpack_msvc_std_copts(),
3305 msvc_x86_32_copts = ["/arch:SSE2"],
3306 x86_srcs = SSE41_UKERNELS,
3307 deps = [
3308 ":tables",
3309 "@FP16",
3310 "@pthreadpool",
3311 ],
3312)
3313
3314xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003315 name = "avx_ukernels",
3316 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003317 gcc_copts = xnnpack_gcc_std_copts(),
3318 gcc_x86_copts = ["-mavx"],
3319 msvc_copts = xnnpack_msvc_std_copts(),
3320 msvc_x86_32_copts = ["/arch:AVX"],
3321 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003323 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003324 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003325 "@FP16",
3326 "@pthreadpool",
3327 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328)
3329
3330xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003331 name = "avx_ukernels_test_mode",
3332 hdrs = INTERNAL_HDRS,
3333 copts = [
3334 "-UNDEBUG",
3335 "-DXNN_TEST_MODE=1",
3336 ],
3337 gcc_copts = xnnpack_gcc_std_copts(),
3338 gcc_x86_copts = ["-mavx"],
3339 msvc_copts = xnnpack_msvc_std_copts(),
3340 msvc_x86_32_copts = ["/arch:AVX"],
3341 msvc_x86_64_copts = ["/arch:AVX"],
3342 x86_srcs = AVX_UKERNELS,
3343 deps = [
3344 ":tables",
3345 "@FP16",
3346 "@pthreadpool",
3347 ],
3348)
3349
3350xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003351 name = "xop_ukernels",
3352 hdrs = INTERNAL_HDRS,
3353 gcc_copts = xnnpack_gcc_std_copts(),
3354 gcc_x86_copts = ["-mxop"],
3355 msvc_copts = xnnpack_msvc_std_copts(),
3356 msvc_x86_32_copts = ["/arch:AVX"],
3357 msvc_x86_64_copts = ["/arch:AVX"],
3358 x86_srcs = XOP_UKERNELS,
3359 deps = [
3360 ":tables",
3361 "@FP16",
3362 "@pthreadpool",
3363 ],
3364)
3365
3366xnnpack_cc_library(
3367 name = "xop_ukernels_test_mode",
3368 hdrs = INTERNAL_HDRS,
3369 copts = [
3370 "-UNDEBUG",
3371 "-DXNN_TEST_MODE=1",
3372 ],
3373 gcc_copts = xnnpack_gcc_std_copts(),
3374 gcc_x86_copts = ["-mxop"],
3375 msvc_copts = xnnpack_msvc_std_copts(),
3376 msvc_x86_32_copts = ["/arch:AVX"],
3377 msvc_x86_64_copts = ["/arch:AVX"],
3378 x86_srcs = XOP_UKERNELS,
3379 deps = [
3380 ":tables",
3381 "@FP16",
3382 "@pthreadpool",
3383 ],
3384)
3385
3386xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003387 name = "fma3_ukernels",
3388 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003389 gcc_copts = xnnpack_gcc_std_copts(),
3390 gcc_x86_copts = ["-mfma"],
3391 msvc_copts = xnnpack_msvc_std_copts(),
3392 msvc_x86_32_copts = ["/arch:AVX"],
3393 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003394 x86_srcs = FMA3_UKERNELS,
3395 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003396 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003397 "@FP16",
3398 "@pthreadpool",
3399 ],
3400)
3401
3402xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003403 name = "fma3_ukernels_test_mode",
3404 hdrs = INTERNAL_HDRS,
3405 copts = [
3406 "-UNDEBUG",
3407 "-DXNN_TEST_MODE=1",
3408 ],
3409 gcc_copts = xnnpack_gcc_std_copts(),
3410 gcc_x86_copts = ["-mfma"],
3411 msvc_copts = xnnpack_msvc_std_copts(),
3412 msvc_x86_32_copts = ["/arch:AVX"],
3413 msvc_x86_64_copts = ["/arch:AVX"],
3414 x86_srcs = FMA3_UKERNELS,
3415 deps = [
3416 ":tables",
3417 "@FP16",
3418 "@pthreadpool",
3419 ],
3420)
3421
3422xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003423 name = "avx2_ukernels",
3424 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003425 gcc_copts = xnnpack_gcc_std_copts(),
3426 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003427 "-mfma",
3428 "-mavx2",
3429 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003430 msvc_copts = xnnpack_msvc_std_copts(),
3431 msvc_x86_32_copts = ["/arch:AVX2"],
3432 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003433 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003434 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003435 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003436 "@FP16",
3437 "@pthreadpool",
3438 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003439)
3440
3441xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003442 name = "avx2_ukernels_test_mode",
3443 hdrs = INTERNAL_HDRS,
3444 copts = [
3445 "-UNDEBUG",
3446 "-DXNN_TEST_MODE=1",
3447 ],
3448 gcc_copts = xnnpack_gcc_std_copts(),
3449 gcc_x86_copts = [
3450 "-mfma",
3451 "-mavx2",
3452 ],
3453 msvc_copts = xnnpack_msvc_std_copts(),
3454 msvc_x86_32_copts = ["/arch:AVX2"],
3455 msvc_x86_64_copts = ["/arch:AVX2"],
3456 x86_srcs = AVX2_UKERNELS,
3457 deps = [
3458 ":tables",
3459 "@FP16",
3460 "@pthreadpool",
3461 ],
3462)
3463
3464xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003465 name = "avx512f_ukernels",
3466 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003467 gcc_copts = xnnpack_gcc_std_copts(),
3468 gcc_x86_copts = ["-mavx512f"],
3469 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3470 msvc_copts = xnnpack_msvc_std_copts(),
3471 msvc_x86_32_copts = ["/arch:AVX512"],
3472 msvc_x86_64_copts = ["/arch:AVX512"],
3473 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003474 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003475 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003476 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003477 "@FP16",
3478 "@pthreadpool",
3479 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003480)
3481
3482xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003483 name = "avx512f_ukernels_test_mode",
3484 hdrs = INTERNAL_HDRS,
3485 copts = [
3486 "-UNDEBUG",
3487 "-DXNN_TEST_MODE=1",
3488 ],
3489 gcc_copts = xnnpack_gcc_std_copts(),
3490 gcc_x86_copts = ["-mavx512f"],
3491 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3492 msvc_copts = xnnpack_msvc_std_copts(),
3493 msvc_x86_32_copts = ["/arch:AVX512"],
3494 msvc_x86_64_copts = ["/arch:AVX512"],
3495 msys_copts = ["-fno-asynchronous-unwind-tables"],
3496 x86_srcs = AVX512F_UKERNELS,
3497 deps = [
3498 ":tables",
3499 "@FP16",
3500 "@pthreadpool",
3501 ],
3502)
3503
3504xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003505 name = "avx512skx_ukernels",
3506 hdrs = INTERNAL_HDRS,
3507 gcc_copts = xnnpack_gcc_std_copts(),
3508 gcc_x86_copts = [
3509 "-mavx512f",
3510 "-mavx512cd",
3511 "-mavx512bw",
3512 "-mavx512dq",
3513 "-mavx512vl",
3514 ],
3515 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3516 msvc_copts = xnnpack_msvc_std_copts(),
3517 msvc_x86_32_copts = ["/arch:AVX512"],
3518 msvc_x86_64_copts = ["/arch:AVX512"],
3519 msys_copts = ["-fno-asynchronous-unwind-tables"],
3520 x86_srcs = AVX512SKX_UKERNELS,
3521 deps = [
3522 ":tables",
3523 "@FP16",
3524 "@pthreadpool",
3525 ],
3526)
3527
3528xnnpack_cc_library(
3529 name = "avx512skx_ukernels_test_mode",
3530 hdrs = INTERNAL_HDRS,
3531 copts = [
3532 "-UNDEBUG",
3533 "-DXNN_TEST_MODE=1",
3534 ],
3535 gcc_copts = xnnpack_gcc_std_copts(),
3536 gcc_x86_copts = [
3537 "-mavx512f",
3538 "-mavx512cd",
3539 "-mavx512bw",
3540 "-mavx512dq",
3541 "-mavx512vl",
3542 ],
3543 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 msvc_x86_32_copts = ["/arch:AVX512"],
3546 msvc_x86_64_copts = ["/arch:AVX512"],
3547 msys_copts = ["-fno-asynchronous-unwind-tables"],
3548 x86_srcs = AVX512SKX_UKERNELS,
3549 deps = [
3550 ":tables",
3551 "@FP16",
3552 "@pthreadpool",
3553 ],
3554)
3555
3556xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 name = "asm_ukernels",
3558 hdrs = ["src/xnnpack/assembly.h"],
3559 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003560 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003562 wasm_srcs = WASM32_ASM_UKERNELS,
3563 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564)
3565
Marat Dukhan3b59de22020-06-03 20:15:19 -07003566xnnpack_cc_library(
3567 name = "logging_utils",
3568 srcs = LOGGING_SRCS,
3569 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3570 copts = LOGGING_COPTS + [
3571 "-Isrc",
3572 "-Iinclude",
3573 ] + select({
3574 ":debug_build": [],
3575 "//conditions:default": xnnpack_min_size_copts(),
3576 }),
3577 gcc_copts = xnnpack_gcc_std_copts(),
3578 msvc_copts = xnnpack_msvc_std_copts(),
3579 visibility = xnnpack_visibility(),
3580 deps = [
3581 "@FP16",
3582 "@clog",
3583 "@pthreadpool",
3584 ],
3585)
3586
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587xnnpack_aggregate_library(
3588 name = "ukernels",
3589 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590 ":neon_ukernels",
3591 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003592 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003593 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 ":asm_ukernels",
3595 ],
3596 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597 ":neon_ukernels",
3598 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003599 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003601 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 ":asm_ukernels",
3603 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003604 generic_deps = [
3605 ":scalar_ukernels",
3606 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003607 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003608 ":psimd_fastmath_ukernels",
3609 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003611 wasm_deps = [
3612 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003613 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003614 ],
3615 wasmsimd_deps = [
3616 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003617 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003618 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003621 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003622 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003624 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003625 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003626 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003628 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629 ],
3630)
3631
Marat Dukhan33fcf782020-05-24 14:27:15 -07003632xnnpack_aggregate_library(
3633 name = "ukernels_test_mode",
3634 aarch32_deps = [
3635 ":neon_ukernels_test_mode",
3636 ":neonfma_ukernels_test_mode",
3637 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003638 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003639 ":asm_ukernels",
3640 ],
3641 aarch64_deps = [
3642 ":neon_ukernels_test_mode",
3643 ":neonfma_ukernels_test_mode",
3644 ":neonv8_ukernels_test_mode",
3645 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003646 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003647 ":asm_ukernels",
3648 ],
3649 generic_deps = [
3650 ":scalar_ukernels_test_mode",
3651 ],
3652 psimd_deps = [
3653 ":psimd_fastmath_ukernels_test_mode",
3654 ":psimd_accmath_ukernels_test_mode",
3655 ],
3656 wasm_deps = [
3657 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003658 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003659 ],
3660 wasmsimd_deps = [
3661 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003662 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003663 ],
3664 x86_deps = [
3665 ":sse2_ukernels_test_mode",
3666 ":ssse3_ukernels_test_mode",
3667 ":sse41_ukernels_test_mode",
3668 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003669 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003670 ":fma3_ukernels_test_mode",
3671 ":avx2_ukernels_test_mode",
3672 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003673 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003674 ],
3675)
3676
Marat Dukhan86df3842020-06-17 14:03:31 -07003677# This target must be compatible with Android NDK r17 and compile with Clang 6.
3678# It is similar to :ukernels target, but excludes microkernels which require
3679# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003680xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003681 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003682 aarch32_deps = [
3683 ":neon_ukernels",
3684 ":neonfma_ukernels",
3685 ":neonv8_ukernels",
3686 ":asm_ukernels",
3687 ],
3688 aarch64_deps = [
3689 ":neon_ukernels",
3690 ":neonfma_ukernels",
3691 ":neonv8_ukernels",
3692 ":asm_ukernels",
3693 ],
3694 generic_deps = [
3695 ":scalar_ukernels",
3696 ],
3697 psimd_deps = [
3698 ":psimd_fastmath_ukernels",
3699 ":psimd_accmath_ukernels",
3700 ],
3701 wasm_deps = [
3702 ":wasm_ukernels",
3703 ],
3704 wasmsimd_deps = [
3705 ":wasm_ukernels",
3706 ],
3707 x86_deps = [
3708 ":sse2_ukernels",
3709 ":ssse3_ukernels",
3710 ":sse41_ukernels",
3711 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003712 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003713 ":fma3_ukernels",
3714 ":avx2_ukernels",
3715 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003716 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003717 ],
3718)
3719
Marat Dukhan08c4a432019-10-03 09:29:21 -07003720xnnpack_cc_library(
3721 name = "im2col",
3722 srcs = ["src/im2col.c"],
3723 hdrs = [
3724 "src/xnnpack/common.h",
3725 "src/xnnpack/im2col.h",
3726 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003727 gcc_copts = xnnpack_gcc_std_copts(),
3728 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003729)
3730
3731xnnpack_cc_library(
3732 name = "indirection",
3733 srcs = ["src/indirection.c"],
3734 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003735 gcc_copts = xnnpack_gcc_std_copts(),
3736 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 deps = [
3738 "@FP16",
3739 "@FXdiv",
3740 "@pthreadpool",
3741 ],
3742)
3743
3744xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003745 name = "indirection_test_mode",
3746 srcs = ["src/indirection.c"],
3747 hdrs = INTERNAL_HDRS,
3748 copts = [
3749 "-UNDEBUG",
3750 "-DXNN_TEST_MODE=1",
3751 ],
3752 gcc_copts = xnnpack_gcc_std_copts(),
3753 msvc_copts = xnnpack_msvc_std_copts(),
3754 deps = [
3755 "@FP16",
3756 "@FXdiv",
3757 "@pthreadpool",
3758 ],
3759)
3760
3761xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003762 name = "packing",
3763 srcs = ["src/packing.c"],
3764 hdrs = INTERNAL_HDRS,
3765 gcc_copts = xnnpack_gcc_std_copts(),
3766 msvc_copts = xnnpack_msvc_std_copts(),
3767 deps = [
3768 "@FP16",
3769 "@FXdiv",
3770 "@pthreadpool",
3771 ],
3772)
3773
3774xnnpack_cc_library(
3775 name = "packing_test_mode",
3776 srcs = ["src/packing.c"],
3777 hdrs = INTERNAL_HDRS,
3778 copts = [
3779 "-UNDEBUG",
3780 "-DXNN_TEST_MODE=1",
3781 ],
3782 gcc_copts = xnnpack_gcc_std_copts(),
3783 msvc_copts = xnnpack_msvc_std_copts(),
3784 deps = [
3785 "@FP16",
3786 "@FXdiv",
3787 "@pthreadpool",
3788 ],
3789)
3790
3791xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003792 name = "operator_run",
3793 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003794 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003795 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003796 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3797 "//conditions:default": [],
3798 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003799 gcc_copts = xnnpack_gcc_std_copts(),
3800 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003801 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003802 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003803 "@FP16",
3804 "@FXdiv",
3805 "@clog",
3806 "@pthreadpool",
3807 ],
3808)
3809
Chao Mei6ddfc602020-05-13 22:29:36 -07003810xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003811 name = "operator_run_test_mode",
3812 srcs = ["src/operator-run.c"],
3813 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3814 copts = LOGGING_COPTS + [
3815 "-UNDEBUG",
3816 "-DXNN_TEST_MODE=1",
3817 ] + select({
3818 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3819 "//conditions:default": [],
3820 }),
3821 gcc_copts = xnnpack_gcc_std_copts(),
3822 msvc_copts = xnnpack_msvc_std_copts(),
3823 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003824 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003825 "@FP16",
3826 "@FXdiv",
3827 "@clog",
3828 "@pthreadpool",
3829 ],
3830)
3831
3832xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003833 name = "memory_planner",
3834 srcs = ["src/memory-planner.c"],
3835 hdrs = INTERNAL_HDRS,
3836 defines = select({
3837 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3838 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3839 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3840 }),
3841 gcc_copts = xnnpack_gcc_std_copts(),
3842 msvc_copts = xnnpack_msvc_std_copts(),
3843 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003844 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003845 "@pthreadpool",
3846 ],
3847)
3848
Marat Dukhan33fcf782020-05-24 14:27:15 -07003849xnnpack_cc_library(
3850 name = "memory_planner_test_mode",
3851 srcs = ["src/memory-planner.c"],
3852 hdrs = INTERNAL_HDRS,
3853 copts = [
3854 "-UNDEBUG",
3855 "-DXNN_TEST_MODE=1",
3856 ],
3857 defines = select({
3858 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3859 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3860 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3861 }),
3862 gcc_copts = xnnpack_gcc_std_copts(),
3863 msvc_copts = xnnpack_msvc_std_copts(),
3864 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003865 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003866 "@pthreadpool",
3867 ],
3868)
3869
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870cc_library(
3871 name = "enable_assembly",
3872 defines = select({
3873 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3874 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003875 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876 }),
3877)
3878
Marat Dukhan9de90e02020-06-18 16:04:12 -07003879cc_library(
3880 name = "enable_sparse",
3881 defines = select({
3882 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3883 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003884 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003885 }),
3886)
3887
Marat Dukhancf056b22019-10-07 10:26:29 -07003888xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889 name = "operators",
3890 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003891 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003893 ],
3894 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003895 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003896 "-Isrc",
3897 "-Iinclude",
3898 ] + select({
3899 ":debug_build": [],
3900 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003901 }) + select({
3902 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3903 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003905 gcc_copts = xnnpack_gcc_std_copts(),
3906 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003907 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003909 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003910 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911 "@FP16",
3912 "@FXdiv",
3913 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003915 ],
3916)
3917
Marat Dukhan10a38082020-04-17 03:58:35 -07003918xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003919 name = "operators_test_mode",
3920 srcs = OPERATOR_SRCS + [
3921 "src/memory.c",
3922 "src/operator-delete.c",
3923 ],
3924 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3925 copts = LOGGING_COPTS + [
3926 "-Isrc",
3927 "-Iinclude",
3928 "-UNDEBUG",
3929 "-DXNN_TEST_MODE=1",
3930 ] + select({
3931 ":debug_build": [],
3932 "//conditions:default": xnnpack_min_size_copts(),
3933 }) + select({
3934 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3935 "//conditions:default": [],
3936 }),
3937 gcc_copts = xnnpack_gcc_std_copts(),
3938 msvc_copts = xnnpack_msvc_std_copts(),
3939 deps = [
3940 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003941 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003942 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003943 "@FP16",
3944 "@FXdiv",
3945 "@clog",
3946 "@pthreadpool",
3947 ],
3948)
3949
3950xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003951 name = "XNNPACK",
3952 srcs = [
3953 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003954 "src/runtime.c",
3955 "src/subgraph.c",
3956 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003957 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003958 hdrs = ["include/xnnpack.h"],
3959 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003960 "-Isrc",
3961 "-Iinclude",
3962 ] + select({
3963 ":debug_build": [],
3964 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003965 }) + select({
3966 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3967 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003968 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003969 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003970 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003971 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003972 visibility = xnnpack_visibility(),
3973 deps = [
3974 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003975 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003976 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003977 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003978 ":operator_run",
3979 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003980 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003981 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003982 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003983 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003984 ] + select({
3985 ":emscripten": [],
3986 "//conditions:default": ["@cpuinfo"],
3987 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988)
3989
Marat Dukhan10a38082020-04-17 03:58:35 -07003990xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003991 name = "XNNPACK_test_mode",
3992 srcs = [
3993 "src/init.c",
3994 "src/runtime.c",
3995 "src/subgraph.c",
3996 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003997 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003998 hdrs = ["include/xnnpack.h"],
3999 copts = LOGGING_COPTS + [
4000 "-Isrc",
4001 "-Iinclude",
4002 "-UNDEBUG",
4003 "-DXNN_TEST_MODE=1",
4004 ] + select({
4005 ":debug_build": [],
4006 "//conditions:default": xnnpack_min_size_copts(),
4007 }) + select({
4008 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4009 "//conditions:default": [],
4010 }),
4011 gcc_copts = xnnpack_gcc_std_copts(),
4012 includes = ["include"],
4013 msvc_copts = xnnpack_msvc_std_copts(),
4014 visibility = xnnpack_visibility(),
4015 deps = [
4016 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004017 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004018 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004019 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004020 ":operator_run_test_mode",
4021 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004022 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004023 "@clog",
4024 "@FP16",
4025 "@pthreadpool",
4026 ] + select({
4027 ":emscripten": [],
4028 "//conditions:default": ["@cpuinfo"],
4029 }),
4030)
4031
4032xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004033 name = "xnnpack_f32",
4034 srcs = [
4035 "src/init.c",
4036 "src/runtime.c",
4037 "src/subgraph.c",
4038 "src/tensor.c",
4039 ] + SUBGRAPH_SRCS,
4040 hdrs = ["include/xnnpack.h"],
4041 copts = LOGGING_COPTS + [
4042 "-Isrc",
4043 "-Iinclude",
4044 ] + select({
4045 ":debug_build": [],
4046 "//conditions:default": xnnpack_min_size_copts(),
4047 }) + select({
4048 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4049 "//conditions:default": [],
4050 }),
4051 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004052 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004053 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004054 "XNN_NO_U8_OPERATORS",
4055 "XNN_NO_X8_OPERATORS",
4056 "XNN_NO_F16_OPERATORS",
4057 "XNN_NO_X16_OPERATORS",
4058 ],
4059 gcc_copts = xnnpack_gcc_std_copts(),
4060 includes = ["include"],
4061 msvc_copts = xnnpack_msvc_std_copts(),
4062 visibility = xnnpack_visibility(),
4063 deps = [
4064 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004065 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004066 ":logging_utils",
4067 ":memory_planner",
4068 ":operator_run",
4069 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004070 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004071 "@clog",
4072 "@FP16",
4073 "@pthreadpool",
4074 ] + select({
4075 ":emscripten": [],
4076 "//conditions:default": ["@cpuinfo"],
4077 }),
4078)
4079
4080xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004081 name = "xnnpack_operators_nhwc_f32",
4082 srcs = [
4083 "src/init.c",
4084 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004085 hdrs = ["include/xnnpack.h"],
4086 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004087 "-Isrc",
4088 "-Iinclude",
4089 ] + select({
4090 ":debug_build": [],
4091 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004092 }) + select({
4093 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4094 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004095 }),
4096 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004097 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004098 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004099 "XNN_NO_U8_OPERATORS",
4100 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004101 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004102 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004103 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004104 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004105 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106 visibility = xnnpack_visibility(),
4107 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004108 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004109 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004110 ":operator_run",
4111 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004112 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004113 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004114 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004115 ] + select({
4116 ":emscripten": [],
4117 "//conditions:default": ["@cpuinfo"],
4118 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119)
4120
Marat Dukhancf056b22019-10-07 10:26:29 -07004121xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122 name = "bench_utils",
4123 srcs = ["bench/utils.cc"],
4124 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004125 deps = [
4126 "@com_google_benchmark//:benchmark",
4127 "@cpuinfo",
4128 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129)
4130
Frank Barchard7e955972019-10-11 10:34:25 -07004131######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004132
4133xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004134 name = "qs8_gemm_bench",
4135 srcs = [
4136 "bench/gemm.h",
4137 "bench/qs8-gemm.cc",
4138 "src/xnnpack/AlignedAllocator.h",
4139 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4140 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4141)
4142
4143xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004144 name = "qs8_requantization_bench",
4145 srcs = [
4146 "bench/qs8-requantization.cc",
4147 "src/xnnpack/requantization-stubs.h",
4148 "src/xnnpack/AlignedAllocator.h",
4149 ] + MICROKERNEL_BENCHMARK_HDRS,
4150 deps = MICROKERNEL_BENCHMARK_DEPS,
4151)
4152
4153xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004154 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004155 srcs = [
4156 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004157 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004158 "src/xnnpack/AlignedAllocator.h",
4159 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004160 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004161 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004162)
4163
4164xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004165 name = "qu8_requantization_bench",
4166 srcs = [
4167 "bench/qu8-requantization.cc",
4168 "src/xnnpack/requantization-stubs.h",
4169 "src/xnnpack/AlignedAllocator.h",
4170 ] + MICROKERNEL_BENCHMARK_HDRS,
4171 deps = MICROKERNEL_BENCHMARK_DEPS,
4172)
4173
4174xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004175 name = "f16_igemm_bench",
4176 srcs = [
4177 "bench/f16-igemm.cc",
4178 "bench/conv.h",
4179 "bench/google/conv.h",
4180 "src/xnnpack/AlignedAllocator.h",
4181 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004182 deps = MICROKERNEL_BENCHMARK_DEPS + [
4183 ":indirection",
4184 ":packing",
4185 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004186)
4187
4188xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004189 name = "f16_gemm_bench",
4190 srcs = [
4191 "bench/f16-gemm.cc",
4192 "bench/gemm.h",
4193 "src/xnnpack/AlignedAllocator.h",
4194 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004195 deps = MICROKERNEL_BENCHMARK_DEPS + [
4196 ":packing",
4197 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198)
4199
4200xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004201 name = "f16_spmm_bench",
4202 srcs = [
4203 "bench/f16-spmm.cc",
4204 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004205 "src/xnnpack/AlignedAllocator.h",
4206 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004207 deps = MICROKERNEL_BENCHMARK_DEPS,
4208)
4209
4210xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211 name = "f32_igemm_bench",
4212 srcs = [
4213 "bench/f32-igemm.cc",
4214 "bench/conv.h",
4215 "src/xnnpack/AlignedAllocator.h",
4216 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004217 deps = MICROKERNEL_BENCHMARK_DEPS + [
4218 ":indirection",
4219 ":packing",
4220 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221)
4222
4223xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004224 name = "f16_relu_bench",
4225 srcs = [
4226 "bench/f16-relu.cc",
4227 "src/xnnpack/AlignedAllocator.h",
4228 ] + MICROKERNEL_BENCHMARK_HDRS,
4229 deps = MICROKERNEL_BENCHMARK_DEPS,
4230)
4231
4232xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 name = "f32_conv_hwc_bench",
4234 srcs = [
4235 "bench/f32-conv-hwc.cc",
4236 "bench/dconv.h",
4237 "src/xnnpack/AlignedAllocator.h",
4238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004239 deps = MICROKERNEL_BENCHMARK_DEPS + [
4240 ":packing",
4241 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242)
4243
4244xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004245 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004246 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004247 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004248 "bench/dconv.h",
4249 "src/xnnpack/AlignedAllocator.h",
4250 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004251 deps = MICROKERNEL_BENCHMARK_DEPS + [
4252 ":packing",
4253 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004254)
4255
4256xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004257 name = "f16_dwconv_bench",
4258 srcs = [
4259 "bench/f16-dwconv.cc",
4260 "bench/dwconv.h",
4261 "bench/google/dwconv.h",
4262 "src/xnnpack/AlignedAllocator.h",
4263 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_BENCHMARK_DEPS + [
4265 ":indirection",
4266 ":packing",
4267 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004268)
4269
4270xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271 name = "f32_dwconv_bench",
4272 srcs = [
4273 "bench/f32-dwconv.cc",
4274 "bench/dwconv.h",
4275 "src/xnnpack/AlignedAllocator.h",
4276 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004277 deps = MICROKERNEL_BENCHMARK_DEPS + [
4278 ":indirection",
4279 ":packing",
4280 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281)
4282
4283xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004284 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004286 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287 "bench/dwconv.h",
4288 "src/xnnpack/AlignedAllocator.h",
4289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004290 deps = MICROKERNEL_BENCHMARK_DEPS + [
4291 ":indirection",
4292 ":packing",
4293 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004294)
4295
4296xnnpack_benchmark(
4297 name = "f32_gemm_bench",
4298 srcs = [
4299 "bench/f32-gemm.cc",
4300 "bench/gemm.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004303 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004304 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305)
4306
4307xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004308 name = "f32_hswish_bench",
4309 srcs = [
4310 "bench/f32-hswish.cc",
4311 "src/xnnpack/AlignedAllocator.h",
4312 ] + MICROKERNEL_BENCHMARK_HDRS,
4313 deps = MICROKERNEL_BENCHMARK_DEPS,
4314)
4315
4316xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004317 name = "f32_raddexpminusmax_bench",
4318 srcs = [
4319 "bench/f32-raddexpminusmax.cc",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + MICROKERNEL_BENCHMARK_HDRS,
4322 deps = MICROKERNEL_BENCHMARK_DEPS,
4323)
4324
4325xnnpack_benchmark(
4326 name = "f32_raddextexp_bench",
4327 srcs = [
4328 "bench/f32-raddextexp.cc",
4329 "src/xnnpack/AlignedAllocator.h",
4330 ] + MICROKERNEL_BENCHMARK_HDRS,
4331 deps = MICROKERNEL_BENCHMARK_DEPS,
4332)
4333
4334xnnpack_benchmark(
4335 name = "f32_raddstoreexpminusmax_bench",
4336 srcs = [
4337 "bench/f32-raddstoreexpminusmax.cc",
4338 "src/xnnpack/AlignedAllocator.h",
4339 ] + MICROKERNEL_BENCHMARK_HDRS,
4340 deps = MICROKERNEL_BENCHMARK_DEPS,
4341)
4342
4343xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004344 name = "f32_relu_bench",
4345 srcs = [
4346 "bench/f32-relu.cc",
4347 "src/xnnpack/AlignedAllocator.h",
4348 ] + MICROKERNEL_BENCHMARK_HDRS,
4349 deps = MICROKERNEL_BENCHMARK_DEPS,
4350)
4351
4352xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 name = "f32_rmax_bench",
4354 srcs = [
4355 "bench/f32-rmax.cc",
4356 "src/xnnpack/AlignedAllocator.h",
4357 ] + MICROKERNEL_BENCHMARK_HDRS,
4358 deps = MICROKERNEL_BENCHMARK_DEPS,
4359)
4360
4361xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004362 name = "f32_sigmoid_bench",
4363 srcs = [
4364 "bench/f32-sigmoid.cc",
4365 "src/xnnpack/AlignedAllocator.h",
4366 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004367 deps = MICROKERNEL_BENCHMARK_DEPS,
4368)
4369
4370xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371 name = "f32_spmm_bench",
4372 srcs = [
4373 "bench/f32-spmm.cc",
4374 "bench/gemm.h",
4375 "src/xnnpack/AlignedAllocator.h",
4376 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004377 deps = MICROKERNEL_BENCHMARK_DEPS,
4378)
4379
4380xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004381 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004382 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004383 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004384 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004385 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004386 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004387)
4388
4389xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004390 name = "f32_vscaleexpminusmax_bench",
4391 srcs = [
4392 "bench/f32-vscaleexpminusmax.cc",
4393 "src/xnnpack/AlignedAllocator.h",
4394 ] + MICROKERNEL_BENCHMARK_HDRS,
4395 deps = MICROKERNEL_BENCHMARK_DEPS,
4396)
4397
4398xnnpack_benchmark(
4399 name = "f32_vscaleextexp_bench",
4400 srcs = [
4401 "bench/f32-vscaleextexp.cc",
4402 "src/xnnpack/AlignedAllocator.h",
4403 ] + MICROKERNEL_BENCHMARK_HDRS,
4404 deps = MICROKERNEL_BENCHMARK_DEPS,
4405)
4406
4407xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004408 name = "f32_vsqrt_bench",
4409 srcs = [
4410 "bench/f32-vsqrt.cc",
4411 "src/xnnpack/AlignedAllocator.h",
4412 ] + MICROKERNEL_BENCHMARK_HDRS,
4413 deps = MICROKERNEL_BENCHMARK_DEPS,
4414)
4415
4416xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417 name = "f32_im2col_gemm_bench",
4418 srcs = [
4419 "bench/f32-im2col-gemm.cc",
4420 "bench/conv.h",
4421 "src/xnnpack/AlignedAllocator.h",
4422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004423 deps = MICROKERNEL_BENCHMARK_DEPS + [
4424 ":im2col",
4425 ":packing",
4426 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004427)
4428
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004429xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004430 name = "rounding_bench",
4431 srcs = [
4432 "bench/rounding.cc",
4433 "src/xnnpack/math-stubs.h",
4434 "src/xnnpack/AlignedAllocator.h",
4435 ] + MICROKERNEL_BENCHMARK_HDRS,
4436 deps = MICROKERNEL_BENCHMARK_DEPS,
4437)
4438
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439########################### Benchmarks for operators ###########################
4440
4441xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 name = "average_pooling_bench",
4443 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004444 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004445 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004446 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004447)
4448
4449xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004450 name = "bankers_rounding_bench",
4451 srcs = ["bench/bankers-rounding.cc"],
4452 copts = xnnpack_optional_tflite_copts(),
4453 tags = ["nowin32"],
4454 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4455)
4456
4457xnnpack_benchmark(
4458 name = "ceiling_bench",
4459 srcs = ["bench/ceiling.cc"],
4460 copts = xnnpack_optional_tflite_copts(),
4461 tags = ["nowin32"],
4462 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4463)
4464
4465xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 name = "channel_shuffle_bench",
4467 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004468 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004469)
4470
4471xnnpack_benchmark(
4472 name = "convolution_bench",
4473 srcs = ["bench/convolution.cc"],
4474 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004475 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004476 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477)
4478
4479xnnpack_benchmark(
4480 name = "deconvolution_bench",
4481 srcs = ["bench/deconvolution.cc"],
4482 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004483 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004484 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485)
4486
4487xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004488 name = "floor_bench",
4489 srcs = ["bench/floor.cc"],
4490 copts = xnnpack_optional_tflite_copts(),
4491 tags = ["nowin32"],
4492 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4493)
4494
4495xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 name = "global_average_pooling_bench",
4497 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004498 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499)
4500
4501xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004502 name = "hardswish_bench",
4503 srcs = ["bench/hardswish.cc"],
4504 copts = xnnpack_optional_tflite_copts(),
4505 tags = ["nowin32"],
4506 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4507)
4508
4509xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510 name = "max_pooling_bench",
4511 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004512 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004513)
4514
4515xnnpack_benchmark(
4516 name = "sigmoid_bench",
4517 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004518 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004519 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004520 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521)
4522
4523xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004524 name = "prelu_bench",
4525 srcs = ["bench/prelu.cc"],
4526 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004527 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004528 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004529)
4530
4531xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004532 name = "softmax_bench",
4533 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004534 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004535 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004536 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537)
4538
Marat Dukhan87727142020-06-24 15:24:10 -07004539xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004540 name = "square_root_bench",
4541 srcs = ["bench/square-root.cc"],
4542 copts = xnnpack_optional_tflite_copts(),
4543 tags = ["nowin32"],
4544 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4545)
4546
4547xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004548 name = "truncation_bench",
4549 srcs = ["bench/truncation.cc"],
4550 deps = OPERATOR_BENCHMARK_DEPS,
4551)
4552
Marat Dukhanc068bb62019-10-04 13:24:39 -07004553############################# End-to-end benchmarks ############################
4554
4555cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004556 name = "fp32_mobilenet_v1",
4557 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004558 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004559 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004560 linkstatic = True,
4561 deps = [
4562 ":XNNPACK",
4563 "@pthreadpool",
4564 ],
4565)
4566
4567cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004568 name = "fp16_mobilenet_v1",
4569 srcs = ["models/fp16-mobilenet-v1.cc"],
4570 hdrs = ["models/models.h"],
4571 copts = xnnpack_std_cxxopts(),
4572 linkstatic = True,
4573 deps = [
4574 ":XNNPACK",
4575 "@FP16",
4576 "@pthreadpool",
4577 ],
4578)
4579
4580cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004581 name = "qs8_mobilenet_v1",
4582 srcs = ["models/qs8-mobilenet-v1.cc"],
4583 hdrs = ["models/models.h"],
4584 copts = xnnpack_std_cxxopts(),
4585 linkstatic = True,
4586 deps = [
4587 ":XNNPACK",
4588 "@pthreadpool",
4589 ],
4590)
4591
4592cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004593 name = "qs8_mobilenet_v2",
4594 srcs = ["models/qs8-mobilenet-v2.cc"],
4595 hdrs = ["models/models.h"],
4596 copts = xnnpack_std_cxxopts(),
4597 linkstatic = True,
4598 deps = [
4599 ":XNNPACK",
4600 "@pthreadpool",
4601 ],
4602)
4603
4604cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004605 name = "fp32_mobilenet_v2",
4606 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004607 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004608 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004609 linkstatic = True,
4610 deps = [
4611 ":XNNPACK",
4612 "@pthreadpool",
4613 ],
4614)
4615
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004616cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004617 name = "fp16_mobilenet_v2",
4618 srcs = ["models/fp16-mobilenet-v2.cc"],
4619 hdrs = ["models/models.h"],
4620 copts = xnnpack_std_cxxopts(),
4621 linkstatic = True,
4622 deps = [
4623 ":XNNPACK",
4624 "@FP16",
4625 "@pthreadpool",
4626 ],
4627)
4628
4629cc_library(
4630 name = "fp32_mobilenet_v3_large",
4631 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004632 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004633 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004634 linkstatic = True,
4635 deps = [
4636 ":XNNPACK",
4637 "@pthreadpool",
4638 ],
4639)
4640
4641cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004642 name = "fp32_mobilenet_v3_small",
4643 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004644 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004645 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004646 linkstatic = True,
4647 deps = [
4648 ":XNNPACK",
4649 "@pthreadpool",
4650 ],
4651)
4652
Marat Dukhanc068bb62019-10-04 13:24:39 -07004653xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004654 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004655 srcs = [
4656 "bench/f32-dwconv-e2e.cc",
4657 "bench/end2end.h",
4658 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004659 deps = MICROKERNEL_BENCHMARK_DEPS + [
4660 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004661 ":fp32_mobilenet_v1",
4662 ":fp32_mobilenet_v2",
4663 ":fp32_mobilenet_v3_large",
4664 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004665 ],
4666)
4667
4668xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004669 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004670 srcs = [
4671 "bench/f32-gemm-e2e.cc",
4672 "bench/end2end.h",
4673 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004674 deps = MICROKERNEL_BENCHMARK_DEPS + [
4675 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004676 ":fp32_mobilenet_v1",
4677 ":fp32_mobilenet_v2",
4678 ":fp32_mobilenet_v3_large",
4679 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004680 ],
4681)
4682
4683xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004684 name = "end2end_bench",
4685 srcs = ["bench/end2end.cc"],
4686 deps = [
4687 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004688 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004689 ":fp16_mobilenet_v1",
4690 ":fp16_mobilenet_v2",
4691 ":fp32_mobilenet_v1",
4692 ":fp32_mobilenet_v2",
4693 ":fp32_mobilenet_v3_large",
4694 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004695 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004696 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004697 "@pthreadpool",
4698 ],
4699)
4700
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004701#################### Accuracy evaluation for math functions ####################
4702
4703xnnpack_benchmark(
4704 name = "f32_exp_eval",
4705 srcs = [
4706 "eval/f32-exp.cc",
4707 "src/xnnpack/AlignedAllocator.h",
4708 ] + ACCURACY_EVAL_HDRS,
4709 deps = ACCURACY_EVAL_DEPS,
4710)
4711
Marat Dukhan515c9772019-10-17 18:07:57 -07004712xnnpack_benchmark(
4713 name = "f32_expminus_eval",
4714 srcs = [
4715 "eval/f32-expminus.cc",
4716 "src/xnnpack/AlignedAllocator.h",
4717 ] + ACCURACY_EVAL_HDRS,
4718 deps = ACCURACY_EVAL_DEPS,
4719)
4720
Marat Dukhan98ba4412019-10-23 02:14:28 -07004721xnnpack_benchmark(
4722 name = "f32_extexp_eval",
4723 srcs = [
4724 "eval/f32-extexp.cc",
4725 "src/xnnpack/AlignedAllocator.h",
4726 ] + ACCURACY_EVAL_HDRS,
4727 deps = ACCURACY_EVAL_DEPS,
4728)
4729
Marat Dukhan8853b822020-05-07 12:19:01 -07004730xnnpack_unit_test(
4731 name = "f32_roundne_eval",
4732 srcs = [
4733 "eval/f32-roundne.cc",
4734 "src/xnnpack/AlignedAllocator.h",
4735 "src/xnnpack/math-stubs.h",
4736 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004737 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004738 deps = MICROKERNEL_TEST_DEPS,
4739)
4740
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004741xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004742 name = "f32_roundd_eval",
4743 srcs = [
4744 "eval/f32-roundd.cc",
4745 "src/xnnpack/AlignedAllocator.h",
4746 "src/xnnpack/math-stubs.h",
4747 ] + MICROKERNEL_TEST_HDRS,
4748 automatic = False,
4749 deps = MICROKERNEL_TEST_DEPS,
4750)
4751
4752xnnpack_unit_test(
4753 name = "f32_roundu_eval",
4754 srcs = [
4755 "eval/f32-roundu.cc",
4756 "src/xnnpack/AlignedAllocator.h",
4757 "src/xnnpack/math-stubs.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 automatic = False,
4760 deps = MICROKERNEL_TEST_DEPS,
4761)
4762
4763xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004764 name = "f32_roundz_eval",
4765 srcs = [
4766 "eval/f32-roundz.cc",
4767 "src/xnnpack/AlignedAllocator.h",
4768 "src/xnnpack/math-stubs.h",
4769 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004770 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004771 deps = MICROKERNEL_TEST_DEPS,
4772)
4773
Marat Dukhan346a9e52019-11-15 09:06:30 -08004774xnnpack_benchmark(
4775 name = "f32_sigmoid_eval",
4776 srcs = [
4777 "eval/f32-sigmoid.cc",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + ACCURACY_EVAL_HDRS,
4780 deps = ACCURACY_EVAL_DEPS,
4781)
4782
Marat Dukhan84000762020-06-29 18:38:43 -07004783xnnpack_benchmark(
4784 name = "f32_sqrt_eval",
4785 srcs = [
4786 "eval/f32-sqrt.cc",
4787 "src/xnnpack/AlignedAllocator.h",
4788 ] + ACCURACY_EVAL_HDRS,
4789 deps = ACCURACY_EVAL_DEPS + [
4790 ":bench_utils",
4791 ],
4792)
4793
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794######################### Unit tests for micro-kernels #########################
4795
4796xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004797 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004798 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004799 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004800 "test/gemm-microkernel-tester.h",
4801 "src/xnnpack/AlignedAllocator.h",
4802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004803 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804)
4805
4806xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004807 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004808 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004809 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004810 "test/spmm-microkernel-tester.h",
4811 "src/xnnpack/AlignedAllocator.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004817 name = "f32_argmaxpool_test",
4818 srcs = [
4819 "test/f32-argmaxpool.cc",
4820 "test/argmaxpool-microkernel-tester.h",
4821 "src/xnnpack/AlignedAllocator.h",
4822 ] + MICROKERNEL_TEST_HDRS,
4823 deps = MICROKERNEL_TEST_DEPS,
4824)
4825
4826xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004827 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004828 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004829 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004830 "test/avgpool-microkernel-tester.h",
4831 "src/xnnpack/AlignedAllocator.h",
4832 ] + MICROKERNEL_TEST_HDRS,
4833 deps = MICROKERNEL_TEST_DEPS,
4834)
4835
4836xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004837 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004838 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004839 "test/f32-ibilinear.cc",
4840 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004841 "src/xnnpack/AlignedAllocator.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004847 name = "f16_clamp_test",
4848 srcs = [
4849 "test/f16-clamp.cc",
4850 "test/clamp-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
4855xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004856 name = "f32_clamp_test",
4857 srcs = [
4858 "test/f32-clamp.cc",
4859 "test/clamp-microkernel-tester.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004865 name = "f32_igemm_test",
4866 srcs = [
4867 "test/f32-igemm.cc",
4868 "test/gemm-microkernel-tester.h",
4869 "src/xnnpack/AlignedAllocator.h",
4870 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004871 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004872)
4873
4874xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004875 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004876 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004877 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878 "test/gemm-microkernel-tester.h",
4879 "src/xnnpack/AlignedAllocator.h",
4880 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004881 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882)
4883
4884xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004885 name = "f32_igemm_minmax_test",
4886 srcs = [
4887 "test/f32-igemm-minmax.cc",
4888 "test/gemm-microkernel-tester.h",
4889 "src/xnnpack/AlignedAllocator.h",
4890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004891 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004892)
4893
4894xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004895 name = "f16_igemm_minmax_test",
4896 srcs = [
4897 "test/f16-igemm-minmax.cc",
4898 "test/gemm-microkernel-tester.h",
4899 "src/xnnpack/AlignedAllocator.h",
4900 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004901 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004902)
4903
4904xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004905 name = "f32_conv_hwc_test",
4906 srcs = [
4907 "test/f32-conv-hwc.cc",
4908 "test/conv-hwc-microkernel-tester.h",
4909 "src/xnnpack/AlignedAllocator.h",
4910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004911 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004912)
4913
4914xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004915 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004916 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004917 "test/f32-conv-hwc2chw.cc",
4918 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004919 "src/xnnpack/AlignedAllocator.h",
4920 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004921 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004922)
4923
4924xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004925 name = "f32_dwconv_test",
4926 srcs = [
4927 "test/f32-dwconv.cc",
4928 "test/dwconv-microkernel-tester.h",
4929 "src/xnnpack/AlignedAllocator.h",
4930 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004931 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004932)
4933
4934xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004935 name = "f16_dwconv_minmax_test",
4936 srcs = [
4937 "test/f16-dwconv-minmax.cc",
4938 "test/dwconv-microkernel-tester.h",
4939 "src/xnnpack/AlignedAllocator.h",
4940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004941 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004942)
4943
4944xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004945 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004947 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004948 "test/dwconv-microkernel-tester.h",
4949 "src/xnnpack/AlignedAllocator.h",
4950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004951 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004952)
4953
4954xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004955 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004956 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004957 "test/f32-dwconv-chw.cc",
4958 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004959 "src/xnnpack/AlignedAllocator.h",
4960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004961 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004962)
4963
4964xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004965 name = "f16_gavgpool_minmax_test",
4966 srcs = [
4967 "test/f16-gavgpool-minmax.cc",
4968 "test/gavgpool-microkernel-tester.h",
4969 "src/xnnpack/AlignedAllocator.h",
4970 ] + MICROKERNEL_TEST_HDRS,
4971 deps = MICROKERNEL_TEST_DEPS,
4972)
4973
4974xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004975 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004977 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978 "test/gavgpool-microkernel-tester.h",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004985 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004987 "test/f32-gavgpool-cw.cc",
4988 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 "src/xnnpack/AlignedAllocator.h",
4990 ] + MICROKERNEL_TEST_HDRS,
4991 deps = MICROKERNEL_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004995 name = "f32_gemm_test",
4996 srcs = [
4997 "test/f32-gemm.cc",
4998 "test/gemm-microkernel-tester.h",
4999 "src/xnnpack/AlignedAllocator.h",
5000 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005001 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005002)
5003
5004xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005005 name = "f32_gemm_relu_test",
5006 srcs = [
5007 "test/f32-gemm-relu.cc",
5008 "test/gemm-microkernel-tester.h",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005011 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005012)
5013
5014xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005015 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005017 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018 "test/gemm-microkernel-tester.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005021 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005022)
5023
5024xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005025 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005027 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028 "test/gemm-microkernel-tester.h",
5029 "src/xnnpack/AlignedAllocator.h",
5030 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005031 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032)
5033
5034xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005035 name = "f16_hswish_test",
5036 srcs = [
5037 "test/f16-hswish.cc",
5038 "test/hswish-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005044 name = "f32_hswish_test",
5045 srcs = [
5046 "test/f32-hswish.cc",
5047 "test/hswish-microkernel-tester.h",
5048 ] + MICROKERNEL_TEST_HDRS,
5049 deps = MICROKERNEL_TEST_DEPS,
5050)
5051
5052xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005053 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005055 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005056 "test/maxpool-microkernel-tester.h",
5057 ] + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS,
5059)
5060
5061xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005062 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005064 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 "test/avgpool-microkernel-tester.h",
5066 "src/xnnpack/AlignedAllocator.h",
5067 ] + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS,
5069)
5070
5071xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005072 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005074 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075 "test/gemm-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(
Frank Barchardb1966592020-05-12 13:47:06 -07005082 name = "f16_prelu_test",
5083 srcs = [
5084 "test/f16-prelu.cc",
5085 "test/prelu-microkernel-tester.h",
5086 "src/xnnpack/AlignedAllocator.h",
5087 ] + MICROKERNEL_TEST_HDRS,
5088 deps = MICROKERNEL_TEST_DEPS,
5089)
5090
5091xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 name = "f32_prelu_test",
5093 srcs = [
5094 "test/f32-prelu.cc",
5095 "test/prelu-microkernel-tester.h",
5096 "src/xnnpack/AlignedAllocator.h",
5097 ] + MICROKERNEL_TEST_HDRS,
5098 deps = MICROKERNEL_TEST_DEPS,
5099)
5100
5101xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005102 name = "f32_raddexpminusmax_test",
5103 srcs = [
5104 "test/f32-raddexpminusmax.cc",
5105 "test/raddexpminusmax-microkernel-tester.h",
5106 ] + MICROKERNEL_TEST_HDRS,
5107 deps = MICROKERNEL_TEST_DEPS,
5108)
5109
5110xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005111 name = "f32_raddextexp_test",
5112 srcs = [
5113 "test/f32-raddextexp.cc",
5114 "test/raddextexp-microkernel-tester.h",
5115 ] + MICROKERNEL_TEST_HDRS,
5116 deps = MICROKERNEL_TEST_DEPS,
5117)
5118
5119xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005120 name = "f32_raddstoreexpminusmax_test",
5121 srcs = [
5122 "test/f32-raddstoreexpminusmax.cc",
5123 "test/raddstoreexpminusmax-microkernel-tester.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
5128xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005129 name = "f32_relu_test",
5130 srcs = [
5131 "test/f32-relu.cc",
5132 "test/vunary-microkernel-tester.h",
5133 ] + MICROKERNEL_TEST_HDRS,
5134 deps = MICROKERNEL_TEST_DEPS,
5135)
5136
5137xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005138 name = "f32_rmax_test",
5139 srcs = [
5140 "test/f32-rmax.cc",
5141 "test/rmax-microkernel-tester.h",
5142 ] + MICROKERNEL_TEST_HDRS,
5143 deps = MICROKERNEL_TEST_DEPS,
5144)
5145
5146xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005147 name = "f32_sigmoid_test",
5148 srcs = [
5149 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005150 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005151 ] + MICROKERNEL_TEST_HDRS,
5152 deps = MICROKERNEL_TEST_DEPS,
5153)
5154
5155xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005156 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005158 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 "test/spmm-microkernel-tester.h",
5160 "src/xnnpack/AlignedAllocator.h",
5161 ] + MICROKERNEL_TEST_HDRS,
5162 deps = MICROKERNEL_TEST_DEPS,
5163)
5164
5165xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005166 name = "f16_vadd_minmax_test",
5167 srcs = [
5168 "test/f16-vadd-minmax.cc",
5169 "test/vbinary-microkernel-tester.h",
5170 ] + MICROKERNEL_TEST_HDRS,
5171 deps = MICROKERNEL_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
5175 name = "f16_vaddc_minmax_test",
5176 srcs = [
5177 "test/f16-vaddc-minmax.cc",
5178 "test/vbinaryc-microkernel-tester.h",
5179 ] + MICROKERNEL_TEST_HDRS,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
5184 name = "f16_vdiv_minmax_test",
5185 srcs = [
5186 "test/f16-vdiv-minmax.cc",
5187 "test/vbinary-microkernel-tester.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
5193 name = "f16_vdivc_minmax_test",
5194 srcs = [
5195 "test/f16-vdivc-minmax.cc",
5196 "test/vbinaryc-microkernel-tester.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
5202 name = "f16_vrdivc_minmax_test",
5203 srcs = [
5204 "test/f16-vrdivc-minmax.cc",
5205 "test/vbinaryc-microkernel-tester.h",
5206 ] + MICROKERNEL_TEST_HDRS,
5207 deps = MICROKERNEL_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
5211 name = "f16_vmax_test",
5212 srcs = [
5213 "test/f16-vmax.cc",
5214 "test/vbinary-microkernel-tester.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
5220 name = "f16_vmaxc_test",
5221 srcs = [
5222 "test/f16-vmaxc.cc",
5223 "test/vbinaryc-microkernel-tester.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
5229 name = "f16_vmin_test",
5230 srcs = [
5231 "test/f16-vmin.cc",
5232 "test/vbinary-microkernel-tester.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
5238 name = "f16_vminc_test",
5239 srcs = [
5240 "test/f16-vminc.cc",
5241 "test/vbinaryc-microkernel-tester.h",
5242 ] + MICROKERNEL_TEST_HDRS,
5243 deps = MICROKERNEL_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
5247 name = "f16_vmul_minmax_test",
5248 srcs = [
5249 "test/f16-vmul-minmax.cc",
5250 "test/vbinary-microkernel-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
5256 name = "f16_vmulc_minmax_test",
5257 srcs = [
5258 "test/f16-vmulc-minmax.cc",
5259 "test/vbinaryc-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
5265 name = "f16_vsub_minmax_test",
5266 srcs = [
5267 "test/f16-vsub-minmax.cc",
5268 "test/vbinary-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
5274 name = "f16_vsubc_minmax_test",
5275 srcs = [
5276 "test/f16-vsubc-minmax.cc",
5277 "test/vbinaryc-microkernel-tester.h",
5278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
5283 name = "f16_vrsubc_minmax_test",
5284 srcs = [
5285 "test/f16-vrsubc-minmax.cc",
5286 "test/vbinaryc-microkernel-tester.h",
5287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005292 name = "f32_vabs_test",
5293 srcs = [
5294 "test/f32-vabs.cc",
5295 "test/vunary-microkernel-tester.h",
5296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005301 name = "f32_vadd_test",
5302 srcs = [
5303 "test/f32-vadd.cc",
5304 "test/vbinary-microkernel-tester.h",
5305 ] + MICROKERNEL_TEST_HDRS,
5306 deps = MICROKERNEL_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005310 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005312 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005313 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005314 ] + MICROKERNEL_TEST_HDRS,
5315 deps = MICROKERNEL_TEST_DEPS,
5316)
5317
5318xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005319 name = "f32_vadd_relu_test",
5320 srcs = [
5321 "test/f32-vadd-relu.cc",
5322 "test/vbinary-microkernel-tester.h",
5323 ] + MICROKERNEL_TEST_HDRS,
5324 deps = MICROKERNEL_TEST_DEPS,
5325)
5326
5327xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005328 name = "f32_vaddc_test",
5329 srcs = [
5330 "test/f32-vaddc.cc",
5331 "test/vbinaryc-microkernel-tester.h",
5332 ] + MICROKERNEL_TEST_HDRS,
5333 deps = MICROKERNEL_TEST_DEPS,
5334)
5335
5336xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005337 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005338 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005339 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005340 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 ] + MICROKERNEL_TEST_HDRS,
5342 deps = MICROKERNEL_TEST_DEPS,
5343)
5344
5345xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005346 name = "f32_vaddc_relu_test",
5347 srcs = [
5348 "test/f32-vaddc-relu.cc",
5349 "test/vbinaryc-microkernel-tester.h",
5350 ] + MICROKERNEL_TEST_HDRS,
5351 deps = MICROKERNEL_TEST_DEPS,
5352)
5353
5354xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005355 name = "f32_vdiv_test",
5356 srcs = [
5357 "test/f32-vdiv.cc",
5358 "test/vbinary-microkernel-tester.h",
5359 ] + MICROKERNEL_TEST_HDRS,
5360 deps = MICROKERNEL_TEST_DEPS,
5361)
5362
5363xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005364 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005365 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005366 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005367 "test/vbinary-microkernel-tester.h",
5368 ] + MICROKERNEL_TEST_HDRS,
5369 deps = MICROKERNEL_TEST_DEPS,
5370)
5371
5372xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005373 name = "f32_vdiv_relu_test",
5374 srcs = [
5375 "test/f32-vdiv-relu.cc",
5376 "test/vbinary-microkernel-tester.h",
5377 ] + MICROKERNEL_TEST_HDRS,
5378 deps = MICROKERNEL_TEST_DEPS,
5379)
5380
5381xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005382 name = "f32_vdivc_test",
5383 srcs = [
5384 "test/f32-vdivc.cc",
5385 "test/vbinaryc-microkernel-tester.h",
5386 ] + MICROKERNEL_TEST_HDRS,
5387 deps = MICROKERNEL_TEST_DEPS,
5388)
5389
5390xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005391 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005392 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005393 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005394 "test/vbinaryc-microkernel-tester.h",
5395 ] + MICROKERNEL_TEST_HDRS,
5396 deps = MICROKERNEL_TEST_DEPS,
5397)
5398
5399xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005400 name = "f32_vdivc_relu_test",
5401 srcs = [
5402 "test/f32-vdivc-relu.cc",
5403 "test/vbinaryc-microkernel-tester.h",
5404 ] + MICROKERNEL_TEST_HDRS,
5405 deps = MICROKERNEL_TEST_DEPS,
5406)
5407
5408xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005409 name = "f32_vrdivc_test",
5410 srcs = [
5411 "test/f32-vrdivc.cc",
5412 "test/vbinaryc-microkernel-tester.h",
5413 ] + MICROKERNEL_TEST_HDRS,
5414 deps = MICROKERNEL_TEST_DEPS,
5415)
5416
5417xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005418 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005419 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005420 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005421 "test/vbinaryc-microkernel-tester.h",
5422 ] + MICROKERNEL_TEST_HDRS,
5423 deps = MICROKERNEL_TEST_DEPS,
5424)
5425
5426xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005427 name = "f32_vrdivc_relu_test",
5428 srcs = [
5429 "test/f32-vrdivc-relu.cc",
5430 "test/vbinaryc-microkernel-tester.h",
5431 ] + MICROKERNEL_TEST_HDRS,
5432 deps = MICROKERNEL_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005436 name = "f32_vmax_test",
5437 srcs = [
5438 "test/f32-vmax.cc",
5439 "test/vbinary-microkernel-tester.h",
5440 ] + MICROKERNEL_TEST_HDRS,
5441 deps = MICROKERNEL_TEST_DEPS,
5442)
5443
5444xnnpack_unit_test(
5445 name = "f32_vmaxc_test",
5446 srcs = [
5447 "test/f32-vmaxc.cc",
5448 "test/vbinaryc-microkernel-tester.h",
5449 ] + MICROKERNEL_TEST_HDRS,
5450 deps = MICROKERNEL_TEST_DEPS,
5451)
5452
5453xnnpack_unit_test(
5454 name = "f32_vmin_test",
5455 srcs = [
5456 "test/f32-vmin.cc",
5457 "test/vbinary-microkernel-tester.h",
5458 ] + MICROKERNEL_TEST_HDRS,
5459 deps = MICROKERNEL_TEST_DEPS,
5460)
5461
5462xnnpack_unit_test(
5463 name = "f32_vminc_test",
5464 srcs = [
5465 "test/f32-vminc.cc",
5466 "test/vbinaryc-microkernel-tester.h",
5467 ] + MICROKERNEL_TEST_HDRS,
5468 deps = MICROKERNEL_TEST_DEPS,
5469)
5470
5471xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005472 name = "f32_vmul_test",
5473 srcs = [
5474 "test/f32-vmul.cc",
5475 "test/vbinary-microkernel-tester.h",
5476 ] + MICROKERNEL_TEST_HDRS,
5477 deps = MICROKERNEL_TEST_DEPS,
5478)
5479
5480xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005481 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005483 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005484 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005485 ] + MICROKERNEL_TEST_HDRS,
5486 deps = MICROKERNEL_TEST_DEPS,
5487)
5488
5489xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005490 name = "f32_vmul_relu_test",
5491 srcs = [
5492 "test/f32-vmul-relu.cc",
5493 "test/vbinary-microkernel-tester.h",
5494 ] + MICROKERNEL_TEST_HDRS,
5495 deps = MICROKERNEL_TEST_DEPS,
5496)
5497
5498xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005499 name = "f32_vmulc_test",
5500 srcs = [
5501 "test/f32-vmulc.cc",
5502 "test/vbinaryc-microkernel-tester.h",
5503 ] + MICROKERNEL_TEST_HDRS,
5504 deps = MICROKERNEL_TEST_DEPS,
5505)
5506
5507xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005508 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005509 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005510 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005511 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005512 ] + MICROKERNEL_TEST_HDRS,
5513 deps = MICROKERNEL_TEST_DEPS,
5514)
5515
5516xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005517 name = "f32_vmulc_relu_test",
5518 srcs = [
5519 "test/f32-vmulc-relu.cc",
5520 "test/vbinaryc-microkernel-tester.h",
5521 ] + MICROKERNEL_TEST_HDRS,
5522 deps = MICROKERNEL_TEST_DEPS,
5523)
5524
5525xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005526 name = "f16_vmulcaddc_minmax_test",
5527 srcs = [
5528 "test/f16-vmulcaddc-minmax.cc",
5529 "test/vmulcaddc-microkernel-tester.h",
5530 "src/xnnpack/AlignedAllocator.h",
5531 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005532 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005533)
5534
5535xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005536 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005538 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005539 "test/vmulcaddc-microkernel-tester.h",
5540 "src/xnnpack/AlignedAllocator.h",
5541 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005542 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543)
5544
5545xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005546 name = "f32_vlrelu_test",
5547 srcs = [
5548 "test/f32-vlrelu.cc",
5549 "test/vunary-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005555 name = "f32_vneg_test",
5556 srcs = [
5557 "test/f32-vneg.cc",
5558 "test/vunary-microkernel-tester.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005564 name = "f32_vrndne_test",
5565 srcs = [
5566 "test/f32-vrndne.cc",
5567 "test/vunary-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
5573 name = "f32_vrndz_test",
5574 srcs = [
5575 "test/f32-vrndz.cc",
5576 "test/vunary-microkernel-tester.h",
5577 ] + MICROKERNEL_TEST_HDRS,
5578 deps = MICROKERNEL_TEST_DEPS,
5579)
5580
5581xnnpack_unit_test(
5582 name = "f32_vrndu_test",
5583 srcs = [
5584 "test/f32-vrndu.cc",
5585 "test/vunary-microkernel-tester.h",
5586 ] + MICROKERNEL_TEST_HDRS,
5587 deps = MICROKERNEL_TEST_DEPS,
5588)
5589
5590xnnpack_unit_test(
5591 name = "f32_vrndd_test",
5592 srcs = [
5593 "test/f32-vrndd.cc",
5594 "test/vunary-microkernel-tester.h",
5595 ] + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS,
5597)
5598
5599xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005600 name = "f32_vscale_test",
5601 srcs = [
5602 "test/f32-vscale.cc",
5603 "test/vscale-microkernel-tester.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005609 name = "f32_vscaleexpminusmax_test",
5610 srcs = [
5611 "test/f32-vscaleexpminusmax.cc",
5612 "test/vscaleexpminusmax-microkernel-tester.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005618 name = "f32_vscaleextexp_test",
5619 srcs = [
5620 "test/f32-vscaleextexp.cc",
5621 "test/vscaleextexp-microkernel-tester.h",
5622 ] + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005627 name = "f32_vsqr_test",
5628 srcs = [
5629 "test/f32-vsqr.cc",
5630 "test/vunary-microkernel-tester.h",
5631 ] + MICROKERNEL_TEST_HDRS,
5632 deps = MICROKERNEL_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005636 name = "f32_vsqrdiff_test",
5637 srcs = [
5638 "test/f32-vsqrdiff.cc",
5639 "test/vbinary-microkernel-tester.h",
5640 ] + MICROKERNEL_TEST_HDRS,
5641 deps = MICROKERNEL_TEST_DEPS,
5642)
5643
5644xnnpack_unit_test(
5645 name = "f32_vsqrdiffc_test",
5646 srcs = [
5647 "test/f32-vsqrdiffc.cc",
5648 "test/vbinaryc-microkernel-tester.h",
5649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005654 name = "f32_vsqrt_test",
5655 srcs = [
5656 "test/f32-vsqrt.cc",
5657 "test/vunary-microkernel-tester.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005663 name = "f32_vsub_test",
5664 srcs = [
5665 "test/f32-vsub.cc",
5666 "test/vbinary-microkernel-tester.h",
5667 ] + MICROKERNEL_TEST_HDRS,
5668 deps = MICROKERNEL_TEST_DEPS,
5669)
5670
5671xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005672 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005673 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005674 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005675 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005676 ] + MICROKERNEL_TEST_HDRS,
5677 deps = MICROKERNEL_TEST_DEPS,
5678)
5679
5680xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005681 name = "f32_vsub_relu_test",
5682 srcs = [
5683 "test/f32-vsub-relu.cc",
5684 "test/vbinary-microkernel-tester.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 deps = MICROKERNEL_TEST_DEPS,
5687)
5688
5689xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005690 name = "f32_vsubc_test",
5691 srcs = [
5692 "test/f32-vsubc.cc",
5693 "test/vbinaryc-microkernel-tester.h",
5694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005699 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005700 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005701 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005702 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005708 name = "f32_vsubc_relu_test",
5709 srcs = [
5710 "test/f32-vsubc-relu.cc",
5711 "test/vbinaryc-microkernel-tester.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005717 name = "f32_vrsubc_test",
5718 srcs = [
5719 "test/f32-vrsubc.cc",
5720 "test/vbinaryc-microkernel-tester.h",
5721 ] + MICROKERNEL_TEST_HDRS,
5722 deps = MICROKERNEL_TEST_DEPS,
5723)
5724
5725xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005726 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005727 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005728 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005729 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005730 ] + MICROKERNEL_TEST_HDRS,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
5734xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005735 name = "f32_vrsubc_relu_test",
5736 srcs = [
5737 "test/f32-vrsubc-relu.cc",
5738 "test/vbinaryc-microkernel-tester.h",
5739 ] + MICROKERNEL_TEST_HDRS,
5740 deps = MICROKERNEL_TEST_DEPS,
5741)
5742
5743xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005744 name = "qs8_dwconv_minmax_test",
5745 srcs = [
5746 "test/qs8-dwconv-minmax.cc",
5747 "test/dwconv-microkernel-tester.h",
5748 "src/xnnpack/AlignedAllocator.h",
5749 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5750 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5751)
5752
5753xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005754 name = "qs8_gavgpool_minmax_test",
5755 srcs = [
5756 "test/qs8-gavgpool-minmax.cc",
5757 "test/gavgpool-microkernel-tester.h",
5758 "src/xnnpack/AlignedAllocator.h",
5759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005764 name = "qs8_gemm_minmax_test",
5765 srcs = [
5766 "test/qs8-gemm-minmax.cc",
5767 "test/gemm-microkernel-tester.h",
5768 "src/xnnpack/AlignedAllocator.h",
5769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5770 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5771)
5772
5773xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005774 name = "qs8_igemm_minmax_test",
5775 srcs = [
5776 "test/qs8-igemm-minmax.cc",
5777 "test/gemm-microkernel-tester.h",
5778 "src/xnnpack/AlignedAllocator.h",
5779 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5780 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5781)
5782
5783xnnpack_unit_test(
5784 name = "qs8_requantization_test",
5785 srcs = [
5786 "src/xnnpack/requantization-stubs.h",
5787 "test/qs8-requantization.cc",
5788 "test/requantization-tester.h",
5789 ] + MICROKERNEL_TEST_HDRS,
5790 deps = MICROKERNEL_TEST_DEPS,
5791)
5792
5793xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005794 name = "qs8_vadd_minmax_test",
5795 srcs = [
5796 "test/qs8-vadd-minmax.cc",
5797 "test/vadd-microkernel-tester.h",
5798 ] + MICROKERNEL_TEST_HDRS,
5799 deps = MICROKERNEL_TEST_DEPS,
5800)
5801
5802xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005803 name = "qs8_vaddc_minmax_test",
5804 srcs = [
5805 "test/qs8-vaddc-minmax.cc",
5806 "test/vaddc-microkernel-tester.h",
5807 ] + MICROKERNEL_TEST_HDRS,
5808 deps = MICROKERNEL_TEST_DEPS,
5809)
5810
5811xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005812 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005813 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005814 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815 "test/avgpool-microkernel-tester.h",
5816 "src/xnnpack/AlignedAllocator.h",
5817 ] + MICROKERNEL_TEST_HDRS,
5818 deps = MICROKERNEL_TEST_DEPS,
5819)
5820
5821xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005822 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005824 "test/qu8-dwconv-minmax.cc",
5825 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005826 "src/xnnpack/AlignedAllocator.h",
5827 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005828 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005829)
5830
5831xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005832 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005833 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005834 "test/qu8-igemm-minmax.cc",
5835 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005836 "src/xnnpack/AlignedAllocator.h",
5837 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005838 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005839)
5840
5841xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005842 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005843 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005844 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005845 "test/gavgpool-microkernel-tester.h",
5846 "src/xnnpack/AlignedAllocator.h",
5847 ] + MICROKERNEL_TEST_HDRS,
5848 deps = MICROKERNEL_TEST_DEPS,
5849)
5850
5851xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005852 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005853 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005854 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855 "test/gemm-microkernel-tester.h",
5856 "src/xnnpack/AlignedAllocator.h",
5857 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005858 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005859)
5860
5861xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005862 name = "qu8_requantization_test",
5863 srcs = [
5864 "src/xnnpack/requantization-stubs.h",
5865 "test/qu8-requantization.cc",
5866 "test/requantization-tester.h",
5867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005872 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005873 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005874 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005875 "test/vadd-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
5881 name = "u8_clamp_test",
5882 srcs = [
5883 "test/u8-clamp.cc",
5884 "test/clamp-microkernel-tester.h",
5885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
5890 name = "u8_lut32norm_test",
5891 srcs = [
5892 "test/u8-lut32norm.cc",
5893 "test/lut-norm-microkernel-tester.h",
5894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005899 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005901 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005902 "test/maxpool-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
5908 name = "u8_rmax_test",
5909 srcs = [
5910 "test/u8-rmax.cc",
5911 "test/rmax-microkernel-tester.h",
5912 ] + MICROKERNEL_TEST_HDRS,
5913 deps = MICROKERNEL_TEST_DEPS,
5914)
5915
5916xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005917 name = "x32_fill_test",
5918 srcs = [
5919 "test/x32-fill.cc",
5920 "test/fill-microkernel-tester.h",
5921 ] + MICROKERNEL_TEST_HDRS,
5922 deps = MICROKERNEL_TEST_DEPS,
5923)
5924
5925xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926 name = "x32_packx_test",
5927 srcs = [
5928 "test/x32-packx.cc",
5929 "test/pack-microkernel-tester.h",
5930 "src/xnnpack/AlignedAllocator.h",
5931 ] + MICROKERNEL_TEST_HDRS,
5932 deps = MICROKERNEL_TEST_DEPS,
5933)
5934
5935xnnpack_unit_test(
5936 name = "x32_pad_test",
5937 srcs = [
5938 "test/x32-pad.cc",
5939 "test/pad-microkernel-tester.h",
5940 ] + MICROKERNEL_TEST_HDRS,
5941 deps = MICROKERNEL_TEST_DEPS,
5942)
5943
5944xnnpack_unit_test(
5945 name = "x32_unpool_test",
5946 srcs = [
5947 "test/x32-unpool.cc",
5948 "test/unpool-microkernel-tester.h",
5949 ] + MICROKERNEL_TEST_HDRS,
5950 deps = MICROKERNEL_TEST_DEPS,
5951)
5952
5953xnnpack_unit_test(
5954 name = "x32_zip_test",
5955 srcs = [
5956 "test/x32-zip.cc",
5957 "test/zip-microkernel-tester.h",
5958 ] + MICROKERNEL_TEST_HDRS,
5959 deps = MICROKERNEL_TEST_DEPS,
5960)
5961
5962xnnpack_unit_test(
5963 name = "x8_lut_test",
5964 srcs = [
5965 "test/x8-lut.cc",
5966 "test/lut-microkernel-tester.h",
5967 ] + MICROKERNEL_TEST_HDRS,
5968 deps = MICROKERNEL_TEST_DEPS,
5969)
5970
5971xnnpack_unit_test(
5972 name = "x8_zip_test",
5973 srcs = [
5974 "test/x8-zip.cc",
5975 "test/zip-microkernel-tester.h",
5976 ] + MICROKERNEL_TEST_HDRS,
5977 deps = MICROKERNEL_TEST_DEPS,
5978)
5979
Marat Dukhan20c3b922020-03-10 03:45:06 -07005980########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005981
5982xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005983 name = "operator_size_test",
5984 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005985 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005986)
5987
Marat Dukhan20c3b922020-03-10 03:45:06 -07005988xnnpack_binary(
5989 name = "subgraph_size_test",
5990 srcs = ["test/subgraph-size.c"],
5991 deps = [":XNNPACK"],
5992)
5993
5994########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005995
5996xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005997 name = "abs_nc_test",
5998 srcs = [
5999 "test/abs-nc.cc",
6000 "test/abs-operator-tester.h",
6001 ],
6002 deps = OPERATOR_TEST_DEPS,
6003)
6004
6005xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006006 name = "add_nd_test",
6007 srcs = [
6008 "test/add-nd.cc",
6009 "test/binary-elementwise-operator-tester.h",
6010 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006011 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006012)
6013
6014xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006015 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006016 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006017 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018 "test/argmax-pooling-operator-tester.h",
6019 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006020 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006021)
6022
6023xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006024 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006025 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006026 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006027 "test/average-pooling-operator-tester.h",
6028 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006029 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006030)
6031
6032xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006033 name = "bankers_rounding_nc_test",
6034 srcs = [
6035 "test/bankers-rounding-nc.cc",
6036 "test/bankers-rounding-operator-tester.h",
6037 ],
6038 deps = OPERATOR_TEST_DEPS,
6039)
6040
6041xnnpack_unit_test(
6042 name = "ceiling_nc_test",
6043 srcs = [
6044 "test/ceiling-nc.cc",
6045 "test/ceiling-operator-tester.h",
6046 ],
6047 deps = OPERATOR_TEST_DEPS,
6048)
6049
6050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006051 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006053 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054 "test/channel-shuffle-operator-tester.h",
6055 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006056 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006057)
6058
6059xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006060 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006062 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063 "test/clamp-operator-tester.h",
6064 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006065 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066)
6067
6068xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006069 name = "constant_pad_nd_test",
6070 srcs = [
6071 "test/constant-pad-nd.cc",
6072 "test/constant-pad-operator-tester.h",
6073 ],
6074 deps = OPERATOR_TEST_DEPS,
6075)
6076
6077xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006078 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006080 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006081 "test/convolution-operator-tester.h",
6082 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006083 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006084)
6085
6086xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006087 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006088 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006089 "test/convolution-nchw.cc",
6090 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006091 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006092 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093)
6094
6095xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006096 name = "copy_nc_test",
6097 srcs = [
6098 "test/copy-nc.cc",
6099 "test/copy-operator-tester.h",
6100 ],
6101 deps = OPERATOR_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006105 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006106 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006107 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006108 "test/deconvolution-operator-tester.h",
6109 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006110 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006111)
6112
6113xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006114 name = "divide_nd_test",
6115 srcs = [
6116 "test/binary-elementwise-operator-tester.h",
6117 "test/divide-nd.cc",
6118 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006119 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006120)
6121
6122xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006123 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006124 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006125 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006126 "test/fully-connected-operator-tester.h",
6127 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006128 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006129)
6130
6131xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006132 name = "floor_nc_test",
6133 srcs = [
6134 "test/floor-nc.cc",
6135 "test/floor-operator-tester.h",
6136 ],
6137 deps = OPERATOR_TEST_DEPS,
6138)
6139
6140xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006141 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006142 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006143 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006145 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006146 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006147)
6148
6149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006150 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006152 "test/global-average-pooling-ncw.cc",
6153 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006155 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006156)
6157
6158xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006159 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006160 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006161 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 "test/hardswish-operator-tester.h",
6163 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006164 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006165)
6166
6167xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006168 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006169 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006170 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006171 "test/leaky-relu-operator-tester.h",
6172 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006173 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006174)
6175
6176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006177 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006179 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006180 "test/max-pooling-operator-tester.h",
6181 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006183)
6184
6185xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006186 name = "maximum_nd_test",
6187 srcs = [
6188 "test/binary-elementwise-operator-tester.h",
6189 "test/maximum-nd.cc",
6190 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006191 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006192)
6193
6194xnnpack_unit_test(
6195 name = "minimum_nd_test",
6196 srcs = [
6197 "test/binary-elementwise-operator-tester.h",
6198 "test/minimum-nd.cc",
6199 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006200 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006201)
6202
6203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006204 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006205 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006206 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006207 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006208 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006209 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006210)
6211
6212xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006213 name = "negate_nc_test",
6214 srcs = [
6215 "test/negate-nc.cc",
6216 "test/negate-operator-tester.h",
6217 ],
6218 deps = OPERATOR_TEST_DEPS,
6219)
6220
6221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006222 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006224 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 "test/prelu-operator-tester.h",
6226 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228)
6229
6230xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006231 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006232 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006233 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006234 "test/resize-bilinear-operator-tester.h",
6235 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006236 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006237)
6238
6239xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006240 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006241 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006242 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 "test/sigmoid-operator-tester.h",
6244 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006245 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246)
6247
6248xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006249 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006250 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006251 "test/softmax-nc.cc",
6252 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006254 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006255)
6256
6257xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006258 name = "square_nc_test",
6259 srcs = [
6260 "test/square-nc.cc",
6261 "test/square-operator-tester.h",
6262 ],
6263 deps = OPERATOR_TEST_DEPS,
6264)
6265
6266xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006267 name = "square_root_nc_test",
6268 srcs = [
6269 "test/square-root-nc.cc",
6270 "test/square-root-operator-tester.h",
6271 ],
6272 deps = OPERATOR_TEST_DEPS,
6273)
6274
6275xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006276 name = "squared_difference_nd_test",
6277 srcs = [
6278 "test/binary-elementwise-operator-tester.h",
6279 "test/squared-difference-nd.cc",
6280 ],
6281 deps = OPERATOR_TEST_DEPS,
6282)
6283
6284xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006285 name = "subtract_nd_test",
6286 srcs = [
6287 "test/binary-elementwise-operator-tester.h",
6288 "test/subtract-nd.cc",
6289 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006291)
6292
6293xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006294 name = "truncation_nc_test",
6295 srcs = [
6296 "test/truncation-nc.cc",
6297 "test/truncation-operator-tester.h",
6298 ],
6299 deps = OPERATOR_TEST_DEPS,
6300)
6301
6302xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006303 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006304 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006305 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306 "test/unpooling-operator-tester.h",
6307 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006308 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309)
6310
Chao Mei6ddfc602020-05-13 22:29:36 -07006311############################### Misc unit tests ###############################
6312
6313xnnpack_unit_test(
6314 name = "memory_planner_test",
6315 srcs = [
6316 "test/memory-planner-test.cc",
6317 ],
6318 deps = [
6319 ":XNNPACK",
6320 ":memory_planner",
6321 ],
6322)
6323
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324############################# Build configurations #############################
6325
Marat Dukhanb8642352019-10-30 15:43:02 -07006326# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006327config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006328 name = "xnn_enable_assembly_explicit_true",
6329 define_values = {"xnn_enable_assembly": "true"},
6330)
6331
6332# Disables usage of assembly kernels.
6333config_setting(
6334 name = "xnn_enable_assembly_explicit_false",
6335 define_values = {"xnn_enable_assembly": "false"},
6336)
6337
Marat Dukhan9de90e02020-06-18 16:04:12 -07006338# Enables usage of sparse inference.
6339config_setting(
6340 name = "xnn_enable_sparse_explicit_true",
6341 define_values = {"xnn_enable_sparse": "true"},
6342)
6343
6344# Disables usage of sparse inference.
6345config_setting(
6346 name = "xnn_enable_sparse_explicit_false",
6347 define_values = {"xnn_enable_sparse": "false"},
6348)
6349
Marat Dukhan05702cf2020-03-26 15:41:33 -07006350# Disables usage of HMP-aware optimizations.
6351config_setting(
6352 name = "xnn_enable_hmp_explicit_false",
6353 define_values = {"xnn_enable_hmp": "false"},
6354)
6355
Chao Mei6ddfc602020-05-13 22:29:36 -07006356# Enable usage of optimized memory allocation
6357config_setting(
6358 name = "xnn_enable_memopt_explicit_true",
6359 define_values = {"xnnpack_enable_memopt": "true"},
6360)
6361
6362# Disable usage of optimized memory allocation
6363config_setting(
6364 name = "xnn_enable_memopt_explicit_false",
6365 define_values = {"xnnpack_enable_memopt": "false"},
6366)
6367
Marat Dukhanb8642352019-10-30 15:43:02 -07006368# Builds with -c dbg
6369config_setting(
6370 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006372 "compilation_mode": "dbg",
6373 },
6374)
6375
6376# Builds with -c opt
6377config_setting(
6378 name = "optimized_build",
6379 values = {
6380 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006381 },
6382)
6383
6384config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006385 name = "linux_k8",
6386 values = {"cpu": "k8"},
6387)
6388
6389config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006390 name = "linux_arm",
6391 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006392)
6393
6394config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006395 name = "linux_armeabi",
6396 values = {"cpu": "armeabi"},
6397)
6398
6399config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006400 name = "linux_armhf",
6401 values = {"cpu": "armhf"},
6402)
6403
6404config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006405 name = "linux_armv7a",
6406 values = {"cpu": "armv7a"},
6407)
6408
6409config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006410 name = "linux_aarch64",
6411 values = {"cpu": "aarch64"},
6412)
6413
6414config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006415 name = "android",
6416 values = {"crosstool_top": "//external:android/crosstool"},
6417)
6418
6419config_setting(
6420 name = "android_armv7",
6421 values = {
6422 "crosstool_top": "//external:android/crosstool",
6423 "cpu": "armeabi-v7a",
6424 },
6425)
6426
6427config_setting(
6428 name = "android_arm64",
6429 values = {
6430 "crosstool_top": "//external:android/crosstool",
6431 "cpu": "arm64-v8a",
6432 },
6433)
6434
6435config_setting(
6436 name = "android_x86",
6437 values = {
6438 "crosstool_top": "//external:android/crosstool",
6439 "cpu": "x86",
6440 },
6441)
6442
6443config_setting(
6444 name = "android_x86_64",
6445 values = {
6446 "crosstool_top": "//external:android/crosstool",
6447 "cpu": "x86_64",
6448 },
6449)
6450
6451config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006452 name = "windows_x86_64",
6453 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006454)
6455
6456config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006457 name = "windows_x86_64_clang",
6458 values = {
6459 "compiler": "clang-cl",
6460 "cpu": "x64_windows",
6461 },
6462)
6463
6464config_setting(
6465 name = "windows_x86_64_mingw",
6466 values = {
6467 "compiler": "mingw-gcc",
6468 "cpu": "x64_windows",
6469 },
6470)
6471
6472config_setting(
6473 name = "windows_x86_64_msys",
6474 values = {
6475 "compiler": "msys-gcc",
6476 "cpu": "x64_windows",
6477 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006478)
6479
6480config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006481 name = "macos_x86_64",
6482 values = {
6483 "apple_platform_type": "macos",
6484 "cpu": "darwin",
6485 },
6486)
6487
6488config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006490 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006491)
6492
6493config_setting(
6494 name = "emscripten_wasm",
6495 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006496 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006497 "cpu": "wasm",
6498 },
6499)
6500
6501config_setting(
6502 name = "emscripten_wasmsimd",
6503 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006504 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006505 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006506 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 },
6508)
6509
6510config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006511 name = "ios_armv7",
6512 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006513 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006514 "cpu": "ios_armv7",
6515 },
6516)
6517
6518config_setting(
6519 name = "ios_arm64",
6520 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006521 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006522 "cpu": "ios_arm64",
6523 },
6524)
6525
6526config_setting(
6527 name = "ios_arm64e",
6528 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006529 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006530 "cpu": "ios_arm64e",
6531 },
6532)
6533
6534config_setting(
6535 name = "ios_x86",
6536 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006537 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006538 "cpu": "ios_i386",
6539 },
6540)
6541
6542config_setting(
6543 name = "ios_x86_64",
6544 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006545 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006546 "cpu": "ios_x86_64",
6547 },
6548)
6549
6550config_setting(
6551 name = "watchos_armv7k",
6552 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006553 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006554 "cpu": "watchos_armv7k",
6555 },
6556)
6557
6558config_setting(
6559 name = "watchos_arm64_32",
6560 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006561 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006562 "cpu": "watchos_arm64_32",
6563 },
6564)
6565
6566config_setting(
6567 name = "watchos_x86",
6568 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006569 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006570 "cpu": "watchos_i386",
6571 },
6572)
6573
6574config_setting(
6575 name = "watchos_x86_64",
6576 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006577 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006578 "cpu": "watchos_x86_64",
6579 },
6580)
6581
6582config_setting(
6583 name = "tvos_arm64",
6584 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006585 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006586 "cpu": "tvos_arm64",
6587 },
6588)
6589
6590config_setting(
6591 name = "tvos_x86_64",
6592 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006593 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006594 "cpu": "tvos_x86_64",
6595 },
6596)