blob: e3d2485b42d759aa8ce52bd34fbbd1500b56ca39 [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",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700113 "src/tables/exp2minus-k-over-64.c",
114 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115]
116
Marat Dukhan08c4a432019-10-03 09:29:21 -0700117SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800119 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700121 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
122 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700123 "src/f32-clamp/gen/scalar-x1.c",
124 "src/f32-clamp/gen/scalar-x2.c",
125 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800126 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700127 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700128 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
129 "src/f32-dwconv-chw/3x3p1-scalar.c",
130 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
131 "src/f32-dwconv-chw/5x5p2-scalar.c",
132 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700143 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700147 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700149 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700155 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700156 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700157 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700158 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
159 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700172 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
173 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
174 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700175 "src/f32-hswish/gen/hswish-scalar-x1.c",
176 "src/f32-hswish/gen/hswish-scalar-x2.c",
177 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700178 "src/f32-ibilinear/gen/scalar-c1.c",
179 "src/f32-ibilinear/gen/scalar-c2.c",
180 "src/f32-ibilinear/gen/scalar-c4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700181 "src/f32-ibilinear-chw/gen/scalar-p1.c",
182 "src/f32-ibilinear-chw/gen/scalar-p2.c",
183 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700184 "src/f32-igemm/gen/1x4-scalar.c",
185 "src/f32-igemm/gen/2x4-scalar.c",
186 "src/f32-igemm/gen/4x2-scalar.c",
187 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700188 "src/f32-igemm/gen/1x4-relu-scalar.c",
189 "src/f32-igemm/gen/2x4-relu-scalar.c",
190 "src/f32-igemm/gen/4x2-relu-scalar.c",
191 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700192 "src/f32-igemm/gen/1x4-minmax-scalar.c",
193 "src/f32-igemm/gen/2x4-minmax-scalar.c",
194 "src/f32-igemm/gen/4x2-minmax-scalar.c",
195 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700196 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
197 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
198 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700199 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
200 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
201 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700202 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
203 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
204 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
205 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800206 "src/f32-prelu/gen/scalar-2x1.c",
207 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
215 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
216 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
217 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
218 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
219 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700220 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700221 "src/f32-relu/gen/scalar-x1.c",
222 "src/f32-relu/gen/scalar-x2.c",
223 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700224 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800225 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
228 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
229 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
230 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
231 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
232 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
233 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700234 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
235 "src/f32-spmm/gen/1x1-minmax-scalar.c",
236 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
237 "src/f32-spmm/gen/2x1-minmax-scalar.c",
238 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
239 "src/f32-spmm/gen/4x1-minmax-scalar.c",
240 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
241 "src/f32-spmm/gen/8x1-minmax-scalar.c",
242 "src/f32-spmm/gen/8x2-minmax-scalar.c",
243 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700244 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
245 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
246 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700247 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700248 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
249 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
250 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700251 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700252 "src/f32-vbinary/gen/vadd-scalar-x1.c",
253 "src/f32-vbinary/gen/vadd-scalar-x2.c",
254 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700255 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700259 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700260 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
261 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
262 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700263 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700264 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
265 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
266 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700267 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700268 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
269 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
270 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700271 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700272 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
273 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
274 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700275 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700276 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
277 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
278 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700279 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700280 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
281 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
282 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700283 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700284 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
285 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
286 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700287 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700288 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
289 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
290 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700291 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800292 "src/f32-vbinary/gen/vmax-scalar-x1.c",
293 "src/f32-vbinary/gen/vmax-scalar-x2.c",
294 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700295 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800296 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
297 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
298 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700299 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800300 "src/f32-vbinary/gen/vmin-scalar-x1.c",
301 "src/f32-vbinary/gen/vmin-scalar-x2.c",
302 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700303 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800304 "src/f32-vbinary/gen/vminc-scalar-x1.c",
305 "src/f32-vbinary/gen/vminc-scalar-x2.c",
306 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700307 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700308 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
309 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
310 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700311 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700312 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
313 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
314 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700315 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700316 "src/f32-vbinary/gen/vmul-scalar-x1.c",
317 "src/f32-vbinary/gen/vmul-scalar-x2.c",
318 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700319 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700320 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
321 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
322 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700323 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700324 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
325 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
326 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700327 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700328 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
329 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
330 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700331 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700332 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
333 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
334 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700335 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700336 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
337 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
338 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700339 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700340 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
341 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
342 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700343 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700344 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
345 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
346 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700347 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700348 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
349 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
350 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700351 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700352 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
353 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
354 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700355 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700356 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
357 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
358 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700359 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700360 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
361 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
362 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700363 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700364 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
365 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
366 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700367 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700368 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
369 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
370 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700371 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700372 "src/f32-vbinary/gen/vsub-scalar-x1.c",
373 "src/f32-vbinary/gen/vsub-scalar-x2.c",
374 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700375 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700376 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
377 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
378 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700379 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700380 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
381 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
382 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700383 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700384 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
385 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
386 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700387 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700388 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
389 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
390 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700391 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
392 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
393 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700394 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
395 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
396 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
397 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
398 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
399 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
400 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
401 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
402 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700403 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
404 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
405 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700406 "src/f32-vunary/gen/vabs-scalar-x1.c",
407 "src/f32-vunary/gen/vabs-scalar-x2.c",
408 "src/f32-vunary/gen/vabs-scalar-x4.c",
409 "src/f32-vunary/gen/vneg-scalar-x1.c",
410 "src/f32-vunary/gen/vneg-scalar-x2.c",
411 "src/f32-vunary/gen/vneg-scalar-x4.c",
412 "src/f32-vunary/gen/vsqr-scalar-x1.c",
413 "src/f32-vunary/gen/vsqr-scalar-x2.c",
414 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800415 "src/math/expminus-scalar-lut2048-p1.c",
416 "src/math/expminus-scalar-lut64-p2.c",
417 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700418 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700419 "src/math/roundne-scalar-nearbyint.c",
420 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundd-scalar-addsub.c",
422 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700423 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700424 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700425 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700426 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700427 "src/math/roundz-scalar-addsub.c",
428 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700429 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700430 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
431 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
432 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700433 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
434 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700435 "src/qs8-requantization/fp32-scalar-lrintf.c",
436 "src/qs8-requantization/fp32-scalar-magic.c",
437 "src/qs8-requantization/precise-scalar-signed64.c",
438 "src/qs8-requantization/precise-scalar-unsigned32.c",
439 "src/qs8-requantization/precise-scalar-unsigned64.c",
440 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700441 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
442 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
443 "src/qu8-dwconv/up1x9-minmax-scalar.c",
444 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
445 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700446 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
447 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700448 "src/qu8-gemm/2x2-minmax-scalar.c",
449 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700450 "src/qu8-requantization/fp32-scalar-lrintf.c",
451 "src/qu8-requantization/fp32-scalar-magic.c",
452 "src/qu8-requantization/precise-scalar-signed64.c",
453 "src/qu8-requantization/precise-scalar-unsigned32.c",
454 "src/qu8-requantization/precise-scalar-unsigned64.c",
455 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700456 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700457 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700459 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700461 "src/x32-fill/scalar-float.c",
462 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-packx/x2-scalar.c",
464 "src/x32-packx/x3-scalar.c",
465 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700466 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700467 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700468 "src/x32-unpool/scalar.c",
469 "src/x32-zip/x2-scalar.c",
470 "src/x32-zip/x3-scalar.c",
471 "src/x32-zip/x4-scalar.c",
472 "src/x32-zip/xm-scalar.c",
473 "src/x8-lut/scalar.c",
474 "src/x8-zip/x2-scalar.c",
475 "src/x8-zip/x3-scalar.c",
476 "src/x8-zip/x4-scalar.c",
477 "src/x8-zip/xm-scalar.c",
478]
479
Marat Dukhan436ebe62019-12-04 15:10:12 -0800480WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700481 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
482 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700483 "src/f32-clamp/gen/wasm-x1.c",
484 "src/f32-clamp/gen/wasm-x2.c",
485 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700486 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
487 "src/f32-dwconv/gen/up1x4-wasm.c",
488 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
489 "src/f32-dwconv/gen/up1x9-wasm.c",
490 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
491 "src/f32-dwconv/gen/up1x25-wasm.c",
492 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
493 "src/f32-dwconv/gen/up2x4-wasm.c",
494 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
495 "src/f32-dwconv/gen/up2x9-wasm.c",
496 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
497 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700498 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
499 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
500 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
501 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700502 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
503 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700504 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
505 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
506 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
507 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700508 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
509 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
511 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700512 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
513 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
514 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700515 "src/f32-gemm/gen/1x4-wasm.c",
516 "src/f32-gemm/gen/2x4-wasm.c",
517 "src/f32-gemm/gen/4x2-wasm.c",
518 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700519 "src/f32-gemm/gen/1x4-relu-wasm.c",
520 "src/f32-gemm/gen/2x4-relu-wasm.c",
521 "src/f32-gemm/gen/4x2-relu-wasm.c",
522 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700523 "src/f32-gemm/gen/1x4-minmax-wasm.c",
524 "src/f32-gemm/gen/2x4-minmax-wasm.c",
525 "src/f32-gemm/gen/4x2-minmax-wasm.c",
526 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700527 "src/f32-hswish/gen/hswish-wasm-x1.c",
528 "src/f32-hswish/gen/hswish-wasm-x2.c",
529 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700530 "src/f32-igemm/gen/1x4-wasm.c",
531 "src/f32-igemm/gen/2x4-wasm.c",
532 "src/f32-igemm/gen/4x2-wasm.c",
533 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700534 "src/f32-igemm/gen/1x4-relu-wasm.c",
535 "src/f32-igemm/gen/2x4-relu-wasm.c",
536 "src/f32-igemm/gen/4x2-relu-wasm.c",
537 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700538 "src/f32-igemm/gen/1x4-minmax-wasm.c",
539 "src/f32-igemm/gen/2x4-minmax-wasm.c",
540 "src/f32-igemm/gen/4x2-minmax-wasm.c",
541 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700542 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
543 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
544 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700545 "src/f32-prelu/gen/wasm-2x1.c",
546 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700547 "src/f32-relu/gen/wasm-x1.c",
548 "src/f32-relu/gen/wasm-x2.c",
549 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700550 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700551 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
552 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
553 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700554 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700555 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
556 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
557 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700558 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700559 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
560 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
561 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700562 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700563 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
564 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
565 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700566 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700567 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
568 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
569 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700570 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700571 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
572 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
573 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700574 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700575 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
576 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
577 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700578 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700579 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
580 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
581 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700582 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800583 "src/f32-vbinary/gen/vmax-wasm-x1.c",
584 "src/f32-vbinary/gen/vmax-wasm-x2.c",
585 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700586 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800587 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
588 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
589 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700590 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800591 "src/f32-vbinary/gen/vmin-wasm-x1.c",
592 "src/f32-vbinary/gen/vmin-wasm-x2.c",
593 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700594 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800595 "src/f32-vbinary/gen/vminc-wasm-x1.c",
596 "src/f32-vbinary/gen/vminc-wasm-x2.c",
597 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700598 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700599 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
600 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
601 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700602 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700603 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
604 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
605 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700606 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700607 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
608 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
609 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700610 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700611 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
612 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
613 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700614 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700615 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
616 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
617 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700618 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700619 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
620 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
621 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700622 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700623 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
624 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
625 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700626 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700627 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
628 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
629 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700630 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700631 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
632 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
633 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700634 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700635 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
636 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
637 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700638 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700639 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
640 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
641 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700642 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700643 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
644 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
645 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700646 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700647 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
648 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
649 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
651 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
652 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800653]
654
Marat Dukhan290055c2020-06-09 12:24:29 -0700655WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700656 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
657 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
658 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700659 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
660 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
661 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
662 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700663 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
664 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
665 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
666 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700667 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
668 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
669 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
670 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
671 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
672 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700673 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
674 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
675 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
676 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
677 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
678 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
679 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
680 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
681 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
682 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
683 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
684 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
685 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
686 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
687 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
688 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
689 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
690 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
691 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
692 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
693 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
694 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
695 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
696 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700697 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
698 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
699 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700700 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700701 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
702 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
703 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
704 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700705 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
706 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
707 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
708 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
709 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
710 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
711 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
712 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
713 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
714 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
715 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
716 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
717 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
718 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
719 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
720 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
721 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
722 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
723 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
724 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
725 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
726 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
727 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
728 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
729 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
730 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
731 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
732 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
733 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
734 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700735 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
736 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700737 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
738 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
739 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
740 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
741 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
742 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
743 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
744 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
745 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
746 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
747 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
748 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
749 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
750 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
751 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
752 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
753 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
754 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
755 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
756 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
757 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
758 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
759 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
760 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
761 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
762 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
763 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
764 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
765 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
766 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700767 "src/f32-ibilinear/gen/wasmsimd-c4.c",
768 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700769 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
770 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
771 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700772 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700773 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
774 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
775 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
776 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700777 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
778 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
779 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
780 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
782 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
783 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
784 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
785 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
786 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
787 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
788 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
789 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
790 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
791 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
792 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
793 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
794 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
795 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
796 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
797 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
798 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
799 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
800 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
801 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
802 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
803 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
804 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
805 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
806 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700807 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
808 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700809 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
810 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
811 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
812 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700813 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
814 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700815 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
816 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
817 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700818 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
819 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700820 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
821 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
822 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
823 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard9e053402020-10-19 15:29:08 -0700824 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
825 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
826 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
827 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
828 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
829 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700830 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
831 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700832 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
833 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
834 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700835 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
836 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700837 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
838 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
839 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
840 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
841 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
842 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
843 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700844 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
845 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700846 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
847 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
848 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
849 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
851 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
852 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
853 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
854 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
855 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
856 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
857 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
858 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
859 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
860 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
861 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700862 "src/f32-relu/gen/wasmsimd-x4.c",
863 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700864 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700865 "src/f32-rmax/wasmsimd-arm.c",
866 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700867 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
868 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
869 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
870 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
871 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
872 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700873 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
874 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
875 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
876 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
877 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
878 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700879 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
880 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700881 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700882 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
883 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700884 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700885 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
886 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700887 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700888 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
889 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700890 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700891 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
892 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700893 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700894 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
895 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700896 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700897 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
898 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700899 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700900 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
901 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700902 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700903 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
904 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700905 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700906 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
907 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700908 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700909 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
910 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700911 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700912 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
913 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700914 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700915 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
916 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700917 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700918 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
919 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700920 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700921 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
922 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700923 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700924 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
925 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700926 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700927 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
928 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700929 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700930 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
931 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700932 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700933 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
934 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700935 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700936 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
937 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700938 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700939 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
940 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700941 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700942 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
943 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700944 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700945 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
946 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700947 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700948 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
949 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700950 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700951 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
952 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700953 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700954 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
955 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700956 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700957 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
958 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700959 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700960 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
961 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700962 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700963 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
964 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700965 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700966 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
967 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700968 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700969 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
970 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700971 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700972 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
973 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700974 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700975 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
976 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700977 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700978 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
979 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700980 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700981 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
982 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700983 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700984 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
985 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700986 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700987 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
988 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700989 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700990 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
991 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700992 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700993 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
994 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700995 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700996 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
997 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700998 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700999 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1000 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001001 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001002 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1003 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001004 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001005 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1006 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001007 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001008 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1009 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001010 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001011 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1012 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001013 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001014 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1015 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001016 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001017 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1018 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001019 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001020 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1021 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001022 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001023 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1024 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001025 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001026 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1027 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001028 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001029 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1030 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1031 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1032 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001033 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1034 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1035 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1036 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001037 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1038 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1039 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1040 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1041 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1042 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1043 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1044 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1045 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1046 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1047 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1048 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1049 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1050 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001051 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1052 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001053 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1054 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1055 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1056 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1057 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1058 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001059 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1060 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1061 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001062 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1063 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1064 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001065 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1066 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1067 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001068 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1070 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1071 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1072 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1073 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1074 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1075 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1076 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1077 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1078 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1079 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1080 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1081 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1082 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001083 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001084 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001085 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1086 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1087 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1088 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1089 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1090 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1091 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1092 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001093 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001094 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001095 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001096 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001097 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001098 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001099 "src/x32-zip/x2-wasmsimd.c",
1100 "src/x32-zip/x3-wasmsimd.c",
1101 "src/x32-zip/x4-wasmsimd.c",
1102 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001103 "src/math/roundne-wasmsimd-addsub.c",
1104 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001105 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001106 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001107 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001108 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001109 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001110 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1111 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001115 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001116 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001117 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001118 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001119 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001120]
1121
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001122PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001123 "src/qs8-requantization/precise-psimd.c",
1124 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001125 "src/qu8-requantization/precise-psimd.c",
1126 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001127]
1128
Marat Dukhan08c4a432019-10-03 09:29:21 -07001129# ISA-specific micro-kernels
1130NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001131 "src/f32-argmaxpool/4x-neon-c4.c",
1132 "src/f32-argmaxpool/9p8x-neon-c4.c",
1133 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001134 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1135 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001136 "src/f32-clamp/gen/neon-x4.c",
1137 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001138 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1139 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001140 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1141 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001142 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1143 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001144 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1145 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001150 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1151 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1152 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1153 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001154 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1155 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1156 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1157 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001158 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001159 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1160 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001161 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1162 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1163 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1164 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1165 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1166 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1167 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1168 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1169 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1170 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1171 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1172 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1173 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1174 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1175 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1176 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1177 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1178 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1179 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1180 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1181 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1182 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1183 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1184 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1185 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1186 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1187 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1188 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1189 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1190 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1191 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001192 "src/f32-hswish/gen/hswish-neon-x4.c",
1193 "src/f32-hswish/gen/hswish-neon-x8.c",
1194 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001195 "src/f32-ibilinear/gen/neon-c4.c",
1196 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001197 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1198 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1199 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1200 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1201 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1202 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1203 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1204 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1205 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1206 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1207 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1208 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1209 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1210 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1211 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1212 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001213 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1214 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1215 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001216 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1217 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001218 "src/f32-prelu/gen/neon-1x4.c",
1219 "src/f32-prelu/gen/neon-1x8.c",
1220 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001221 "src/f32-prelu/gen/neon-2x4.c",
1222 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001223 "src/f32-prelu/gen/neon-2x16.c",
1224 "src/f32-prelu/gen/neon-4x4.c",
1225 "src/f32-prelu/gen/neon-4x8.c",
1226 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1243 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1245 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1246 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1247 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1248 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1249 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1250 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001251 "src/f32-relu/gen/neon-x4.c",
1252 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001253 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001254 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001255 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1256 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1257 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1258 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1259 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1260 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1265 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1266 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1267 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1268 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1269 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1270 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1271 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1272 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001273 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1274 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1275 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1276 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001277 "src/f32-vbinary/gen/vmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1280 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1281 "src/f32-vbinary/gen/vmin-neon-x4.c",
1282 "src/f32-vbinary/gen/vmin-neon-x8.c",
1283 "src/f32-vbinary/gen/vminc-neon-x4.c",
1284 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001285 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1286 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1287 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001291 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1292 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1293 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1294 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001295 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1296 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1297 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1298 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001299 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1300 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001301 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1302 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001303 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1304 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1305 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1306 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1307 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1308 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1309 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1310 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001311 "src/f32-vunary/gen/vabs-neon-x4.c",
1312 "src/f32-vunary/gen/vabs-neon-x8.c",
1313 "src/f32-vunary/gen/vneg-neon-x4.c",
1314 "src/f32-vunary/gen/vneg-neon-x8.c",
1315 "src/f32-vunary/gen/vsqr-neon-x4.c",
1316 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001317 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1318 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1319 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1320 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001321 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1322 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1323 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1324 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1325 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1326 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1327 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1328 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001329 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard146e9992020-10-13 13:38:54 -07001330 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001331 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1332 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1333 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard146e9992020-10-13 13:38:54 -07001334 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001335 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1336 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001337 "src/qs8-requantization/precise-neon.c",
1338 "src/qs8-requantization/fp32-neon.c",
1339 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001340 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1341 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1342 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1343 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1344 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1345 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1346 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1347 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001348 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1349 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1350 "src/qu8-dwconv/up8x9-minmax-neon.c",
1351 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1352 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1353 "src/qu8-gemm/4x8-minmax-neon.c",
1354 "src/qu8-gemm/8x8-minmax-neon.c",
1355 "src/qu8-igemm/4x8-minmax-neon.c",
1356 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001357 "src/qu8-requantization/precise-neon.c",
1358 "src/qu8-requantization/fp32-neon.c",
1359 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001360 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001361 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001362 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001363 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001364 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001365 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001366 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001367 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368 "src/x32-zip/x2-neon.c",
1369 "src/x32-zip/x3-neon.c",
1370 "src/x32-zip/x4-neon.c",
1371 "src/x32-zip/xm-neon.c",
1372 "src/x8-zip/x2-neon.c",
1373 "src/x8-zip/x3-neon.c",
1374 "src/x8-zip/x4-neon.c",
1375 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001376 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001377 "src/math/roundd-neon-addsub.c",
1378 "src/math/roundd-neon-cvt.c",
1379 "src/math/roundu-neon-addsub.c",
1380 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001381 "src/math/roundz-neon-addsub.c",
1382 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001383 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001384 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1385 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1386 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1387 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1388 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1389 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001390 "src/math/sqrt-neon-nr1rsqrts.c",
1391 "src/math/sqrt-neon-nr2rsqrts.c",
1392 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001393]
1394
1395NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001396 "src/f32-ibilinear/gen/neonfma-c4.c",
1397 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001398 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1399 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1400 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1401 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1402 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1403 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1404 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1405 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1406 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1412 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1413 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1414 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001415 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1416 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1417 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1418 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001419 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1420 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1421 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1422 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1424 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1425 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1426 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1427 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1429 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1430 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1431 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1432 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1433 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1434 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1435 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1436 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001437 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1438 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1455 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1456 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1457 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1458 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1459 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1460 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1461 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1462 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1510 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1511 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1512 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1513 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1514 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1515 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1516 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001517 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1518 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1521 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1522 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1523 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1524 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1525 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1526 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1527 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1528 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1531 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1532 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1533 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1534 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1535 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1536 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1537 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1538 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001539 "src/math/exp-neonfma-lut64-p2.c",
1540 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001541 "src/math/expminus-neonfma-lut2048-p1.c",
1542 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001543 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001544 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1545 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1546 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1547 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1548 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1549 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1550 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1551 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1552 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1553 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1554 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1555 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1556 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1557 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1558 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1559 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1560 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1561 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001562 "src/math/sqrt-neonfma-nr1fma.c",
1563 "src/math/sqrt-neonfma-nr2fma.c",
1564 "src/math/sqrt-neonfma-nr3fma.c",
1565 "src/math/sqrt-neonfma-nr2fma1adj.c",
1566 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001567]
1568
1569AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001570 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1571 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1572 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1573 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1574 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1575 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001576 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1577 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1579 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1581 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1583 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1584 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1585 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1586 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1587 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1588 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1589 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1590 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1591 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1592 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1593 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1594 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1595 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001596 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1597 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001598 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1599 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001600 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1601 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001602 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1603 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001604 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1605 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1606 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1607 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1608 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001609 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001627 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1628 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1629 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1630 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1631 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1632 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1633 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1634 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1635 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1636 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1637 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1638 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1639 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1640 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1641 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1642 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1643 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1644 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001645 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1646 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001647 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1648 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1649 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1650 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1651 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1652 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653]
1654
Marat Dukhan8853b822020-05-07 12:19:01 -07001655NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1657 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1658 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1659 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1662 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1663 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001664 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001665 "src/math/roundd-neonv8.c",
1666 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001667 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001668]
1669
Marat Dukhan08c4a432019-10-03 09:29:21 -07001670AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001671 "src/f16-clamp/gen/neonfp16arith-x16.c",
1672 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001673 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1677 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1678 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1679 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1680 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1681 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1682 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1683 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1684 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001685 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1686 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001687 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1688 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001689 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001690 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001691 "src/f16-relu/gen/neonfp16arith-x16.c",
1692 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001693 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1713 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1714 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1715 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1716 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1717 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1718 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1719 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1720 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001721 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1722 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001723 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1727 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001731 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1732 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1733 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1734 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001735 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1739 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1740 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1741 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1742 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1743 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1744 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1745 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1746 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001747 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1748 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1749 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1750 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1751 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1752 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1753 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1754 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001755]
1756
Benoit Jacoba9644732020-08-13 12:48:55 -07001757NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001758 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001759 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
1760 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001761 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1762 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001763 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001764 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001765 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001766 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001767 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001768 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
1769 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001770 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1771 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001772 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001773 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001774 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001775 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001776]
1777
Marat Dukhan08c4a432019-10-03 09:29:21 -07001778SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001779 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1780 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001781 "src/f32-clamp/gen/sse-x4.c",
1782 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001783 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1784 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001785 "src/f32-dwconv-chw/3x3p1-sse.c",
1786 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001787 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1788 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1789 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1790 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1791 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1792 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1793 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1794 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1795 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1796 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1797 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1798 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001799 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001800 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1801 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001802 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1803 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1804 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1805 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1806 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1807 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1808 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1809 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1810 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1811 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1812 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1813 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1814 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001815 "src/f32-hswish/gen/hswish-sse-x4.c",
1816 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001817 "src/f32-ibilinear/gen/sse-c4.c",
1818 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001819 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1820 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1821 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1822 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1823 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1824 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1825 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001826 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1827 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1828 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001829 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001830 "src/f32-prelu/gen/sse-2x4.c",
1831 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001832 "src/f32-relu/gen/sse-x4.c",
1833 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001834 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001835 "src/f32-spmm/gen/4x1-minmax-sse.c",
1836 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001837 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1843 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1844 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1845 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001846 "src/f32-vbinary/gen/vmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vmax-sse-x8.c",
1848 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1849 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1850 "src/f32-vbinary/gen/vmin-sse-x4.c",
1851 "src/f32-vbinary/gen/vmin-sse-x8.c",
1852 "src/f32-vbinary/gen/vminc-sse-x4.c",
1853 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001854 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1855 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1856 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1857 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1858 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1859 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1860 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1861 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001862 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1863 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1864 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1865 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001866 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1867 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1868 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1869 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001870 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1871 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001872 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1873 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001874 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1875 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001876 "src/f32-vunary/gen/vabs-sse-x4.c",
1877 "src/f32-vunary/gen/vabs-sse-x8.c",
1878 "src/f32-vunary/gen/vneg-sse-x4.c",
1879 "src/f32-vunary/gen/vneg-sse-x8.c",
1880 "src/f32-vunary/gen/vsqr-sse-x4.c",
1881 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001882 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001883 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001884 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001885 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001886 "src/math/roundd-sse-addsub.c",
1887 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001888 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001889 "src/math/sqrt-sse-hh1mac.c",
1890 "src/math/sqrt-sse-nr1mac.c",
1891 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001892]
1893
1894SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001895 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1896 "src/f32-argmaxpool/4x-sse2-c4.c",
1897 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001898 "src/f32-prelu/gen/sse2-2x4.c",
1899 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001900 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1901 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1902 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1904 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1905 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1906 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1907 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1908 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1909 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1910 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1911 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001912 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1913 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1914 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1915 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1916 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1917 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001918 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001919 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001920 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001921 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001922 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1923 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001924 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1925 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001926 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1927 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1928 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1929 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1930 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1931 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1932 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1933 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001934 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1935 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1936 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001937 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1938 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1939 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001940 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1941 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1942 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001943 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1944 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001945 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1946 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001947 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1948 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001949 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1950 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001951 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001952 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1953 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001954 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001955 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1956 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1957 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001958 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1959 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001960 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1961 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001962 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1963 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001964 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001965 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1966 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001967 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001968 "src/qs8-requantization/precise-sse2.c",
1969 "src/qs8-requantization/fp32-sse2.c",
1970 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001971 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1972 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1973 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1974 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001975 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1976 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1977 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1978 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001979 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1980 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1981 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1982 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1983 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1984 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1985 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1986 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001987 "src/qu8-requantization/precise-sse2.c",
1988 "src/qu8-requantization/fp32-sse2.c",
1989 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001990 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001991 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001992 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001993 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001994 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995 "src/x32-zip/x2-sse2.c",
1996 "src/x32-zip/x3-sse2.c",
1997 "src/x32-zip/x4-sse2.c",
1998 "src/x32-zip/xm-sse2.c",
1999 "src/x8-zip/x2-sse2.c",
2000 "src/x8-zip/x3-sse2.c",
2001 "src/x8-zip/x4-sse2.c",
2002 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08002003 "src/math/exp-sse2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002004 "src/math/exp-sse2-lut64-p2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08002005 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07002006 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002007 "src/math/roundd-sse2-cvt.c",
2008 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002009 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07002010 "src/math/sigmoid-sse2-rr2-p5-div.c",
Marat Dukhanac5f1602020-10-15 21:00:45 -07002011 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2012 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07002013 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanac5f1602020-10-15 21:00:45 -07002014 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2015 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002016]
2017
2018SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002019 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2020 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2021 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002022 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2023 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2024 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002025 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2026 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2027 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002028 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2029 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002030 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2031 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002032 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2033 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002034 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2035 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002036 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002037 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2038 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002039 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002040 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2041 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2042 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002043 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2044 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002045 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2046 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002047 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2048 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002049 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002050 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2051 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002052 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002053 "src/qs8-requantization/precise-ssse3.c",
2054 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002055 "src/qu8-requantization/precise-ssse3.c",
2056 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057]
2058
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002059SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002060 "src/f32-prelu/gen/sse41-2x4.c",
2061 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002062 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2063 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2064 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2065 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2066 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2067 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002068 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002069 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002070 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002071 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002072 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2073 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002074 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2075 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002076 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2077 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2078 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2079 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2080 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2081 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2082 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2083 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002084 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2085 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2086 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002087 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2088 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2089 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002090 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2091 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2092 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002093 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2094 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002095 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2096 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002097 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2098 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002099 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2100 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002101 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002102 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2103 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002104 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002105 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2106 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2107 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002108 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2109 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002110 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2111 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002112 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2113 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002114 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002115 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2116 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002117 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002118 "src/qs8-requantization/fp32-sse4.c",
2119 "src/qs8-requantization/precise-sse4.c",
2120 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002121 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2122 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2123 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2124 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002125 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2126 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2127 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2128 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002129 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2130 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2131 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2132 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002133 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2134 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2135 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2136 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002137 "src/qu8-requantization/precise-sse4.c",
2138 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002139 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002140 "src/math/roundd-sse41.c",
2141 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002142 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002143]
2144
Marat Dukhan08c4a432019-10-03 09:29:21 -07002145AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002146 "src/f32-clamp/gen/avx-x8.c",
2147 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002148 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2149 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2150 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2151 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2152 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2153 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2154 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2155 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2156 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2157 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2158 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2159 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2160 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2161 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2162 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2163 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2164 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2165 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2166 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2167 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2168 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2169 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2170 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2171 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2172 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2173 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2174 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2175 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2176 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2177 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002178 "src/f32-hswish/gen/hswish-avx-x8.c",
2179 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002180 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2181 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2182 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2183 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2184 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2185 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2186 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2187 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2188 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002189 "src/f32-prelu/gen/avx-2x8.c",
2190 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002191 "src/f32-relu/gen/avx-x16.c",
2192 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002193 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002194 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2195 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2196 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2197 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2198 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2199 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2200 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2201 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2202 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2203 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2204 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2205 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2206 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2207 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2208 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2209 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2210 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2211 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2212 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2213 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002214 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2215 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2216 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2217 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2218 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2219 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2220 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2221 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002222 "src/f32-vbinary/gen/vmax-avx-x8.c",
2223 "src/f32-vbinary/gen/vmax-avx-x16.c",
2224 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2225 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2226 "src/f32-vbinary/gen/vmin-avx-x8.c",
2227 "src/f32-vbinary/gen/vmin-avx-x16.c",
2228 "src/f32-vbinary/gen/vminc-avx-x8.c",
2229 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002230 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2231 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2232 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2233 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2234 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2235 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2236 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2237 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002238 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2239 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2240 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2241 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002242 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2243 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2244 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2245 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002246 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2247 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002248 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002249 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2250 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2251 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2252 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2253 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2254 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2255 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2256 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002257 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2258 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002259 "src/f32-vunary/gen/vabs-avx-x8.c",
2260 "src/f32-vunary/gen/vabs-avx-x16.c",
2261 "src/f32-vunary/gen/vneg-avx-x8.c",
2262 "src/f32-vunary/gen/vneg-avx-x16.c",
2263 "src/f32-vunary/gen/vsqr-avx-x8.c",
2264 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002265 "src/math/exp-avx-p5.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002266 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002267 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2268 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002269 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002270]
2271
Marat Dukhan1566fee2020-08-02 21:55:41 -07002272XOP_UKERNELS = [
2273 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2274 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2275 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2276 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002277 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2278 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002279 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2280 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002281 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002282 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2283 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002284 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002285 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2286 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2287 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002288 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2289 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2290 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2291 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2292 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2293 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002294 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002295 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2296 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002297 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002298 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2299 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2300 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2301 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2302 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2303 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2304 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2305 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002306]
2307
Marat Dukhanfda12b82019-11-21 12:27:59 -08002308FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002309 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2310 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2311 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2312 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2313 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2314 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2315 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2316 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2317 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2318 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2319 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2320 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2321 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2322 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2323 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2324 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2325 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2326 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2327 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2328 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2329 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2330 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2331 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2332 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2333 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2334 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2335 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2336 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2337 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2338 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2339 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2340 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2341 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2342 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2343 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2344 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2345 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2346 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2347 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2348 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002349 "src/f32-hswish/gen/hswish-fma3-x8.c",
2350 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002351 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2352 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2353 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2354 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2355 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2356 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2357 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2358 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2359 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2360 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2361 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2362 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2363 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2364 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002365 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2366 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2367 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2368 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2369 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2370 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2371 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2372 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002373 "src/math/sqrt-fma3-nr1fma.c",
2374 "src/math/sqrt-fma3-nr2fma.c",
2375 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002376]
2377
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002378AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002379 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2380 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2381 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2382 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2383 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2384 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2385 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2386 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2387 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2388 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2389 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2390 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2391 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2392 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2393 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2394 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2395 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2396 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2397 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2398 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2399 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2400 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2401 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2402 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2403 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2404 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2405 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2406 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2407 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2408 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2409 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2410 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2411 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2412 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2413 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2414 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002415 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2416 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2417 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2418 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2419 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2420 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2421 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2422 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2423 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2424 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2425 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2426 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2427 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2428 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2429 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2430 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2431 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2432 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2433 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2434 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2435 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2436 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2437 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2438 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2439 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2440 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2441 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2442 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2443 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2444 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002445 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2446 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2447 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2448 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2449 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2450 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2451 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2452 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2453 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2454 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2455 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2456 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2457 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2458 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2459 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2460 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2461 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2462 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2463 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2464 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2465 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2466 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2467 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2468 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002469 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2470 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002471 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2472 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2473 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2474 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002475 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2476 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2477 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002478 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2479 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2480 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002481 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2482 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2483 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002484 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2485 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2486 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2487 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2488 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2489 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2490 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2491 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002492 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002493 "src/math/exp-avx2-lut8-p3-perm.c",
2494 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002495 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002496 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002497 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002498 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002499 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002500 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002501 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002502 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002503 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2504 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2505 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2506 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2507 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2508 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2509 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2510 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002511]
2512
Marat Dukhan08c4a432019-10-03 09:29:21 -07002513AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002514 "src/f32-clamp/gen/avx512f-x16.c",
2515 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002516 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2517 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2518 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2519 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2520 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2521 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2522 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2523 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2524 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2525 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2526 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2527 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2528 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2529 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2530 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2531 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2532 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2533 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2534 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2535 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2536 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2537 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2538 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2539 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002540 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2541 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002542 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2543 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2544 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2545 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2546 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2547 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002548 "src/f32-prelu/gen/avx512f-2x16.c",
2549 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002550 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2551 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2552 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2553 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2554 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2555 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2556 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2557 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2558 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2559 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2560 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2561 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2562 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2563 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2564 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2565 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2566 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2567 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2568 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2569 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2570 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2571 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2572 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2573 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2574 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2575 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2576 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2577 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2578 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2579 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2580 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2581 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2582 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2583 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2584 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2585 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002586 "src/f32-relu/gen/avx512f-x16.c",
2587 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002588 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002589 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2590 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2591 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2592 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2593 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2594 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2595 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2596 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2597 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2598 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2599 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2600 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2601 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2602 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2603 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2604 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2605 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2606 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2607 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2608 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2609 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2610 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2611 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2612 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2613 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2614 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2615 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2616 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2617 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2618 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2619 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2620 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2621 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2622 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2623 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2624 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2625 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2626 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2627 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2628 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2629 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2630 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2631 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2632 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2633 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2634 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2635 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2636 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002637 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2638 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2639 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2640 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2641 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2642 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2643 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2644 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002645 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2646 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2647 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2648 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2649 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2650 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2651 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2652 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002653 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2654 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2655 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2656 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2657 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2658 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2659 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2660 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002661 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2662 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2663 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2664 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002665 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2666 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2667 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2668 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002669 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2670 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002671 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002672 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2673 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2674 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2675 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2676 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2677 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2678 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2679 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2680 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2681 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2682 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2683 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2684 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2685 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2686 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2687 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2688 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2689 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2690 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2691 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2692 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2693 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2694 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2695 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002696 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2697 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2698 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2699 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2700 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2701 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2702 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2703 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002704 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2705 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2706 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2707 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2708 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2709 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2710 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2711 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002712 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2713 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2714 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2715 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2716 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2717 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002718 "src/math/exp-avx512f-p5-scalef.c",
2719 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002720 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002721 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002722 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002723 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002724 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002725 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002726 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002727 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002728 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002729 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002730 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002731 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2732 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2733 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2734 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2735 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2736 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2737 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2738 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2739 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2740 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2741 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2742 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002743 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
2744 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2745 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
2746 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
2747 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2748 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002749 "src/math/sqrt-avx512f-nr1fma.c",
2750 "src/math/sqrt-avx512f-nr2fma.c",
2751 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752]
2753
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002754AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002755 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2756 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002757 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2758 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2759 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2760 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2761 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2762 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2763 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2764 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2765]
2766
Frank Barchardbcedc082020-08-17 18:00:51 -07002767WASM32_ASM_UKERNELS = [
2768 "src/f32-relu/wasm_shr_x1.S",
2769 "src/f32-relu/wasm_shr_x2.S",
2770 "src/f32-relu/wasm_shr_x4.S",
2771]
2772
Marat Dukhan08c4a432019-10-03 09:29:21 -07002773AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002774 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002775 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002776 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2777 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002778 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002779 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002780 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002781 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002782 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2783 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002784 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2785 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2786 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2787 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002788]
2789
2790AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002791 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002792 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2793 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002794 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002795 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2796 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002797 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2798 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2799 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002800 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002801 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2802 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2803 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002804 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002805 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2806 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002807 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002808 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2809 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2810 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2811 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2812 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2813 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2814 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2815 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2816 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2817 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2818 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2819 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2820 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2821 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2822 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2823 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2824 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2825 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002826 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2827 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002828 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002829 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2830 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2831 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2832 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2833 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2834 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2835 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2836 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2837 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2838 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2839 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2840 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2841 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2842 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2843 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2844 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2845 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2846 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002847 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2848 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2849 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2850 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2851 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2852 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2853 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2854 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2855 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2856 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2857 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2858 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2859 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2860 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2861 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2862 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2863 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2864 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002865 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002866 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002867 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002868]
2869
Marat Dukhan1b354632020-03-23 12:50:22 -07002870INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 "src/xnnpack/argmaxpool.h",
2872 "src/xnnpack/avgpool.h",
2873 "src/xnnpack/clamp.h",
2874 "src/xnnpack/common.h",
2875 "src/xnnpack/conv.h",
2876 "src/xnnpack/dwconv.h",
2877 "src/xnnpack/gavgpool.h",
2878 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002879 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002881 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002883 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002884 "src/xnnpack/lut.h",
2885 "src/xnnpack/math.h",
2886 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002887 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002888 "src/xnnpack/packx.h",
2889 "src/xnnpack/pad.h",
2890 "src/xnnpack/params.h",
2891 "src/xnnpack/pavgpool.h",
2892 "src/xnnpack/ppmm.h",
2893 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002894 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002895 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002896 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 "src/xnnpack/rmax.h",
2898 "src/xnnpack/scalar-utils.h",
2899 "src/xnnpack/spmm.h",
2900 "src/xnnpack/unpool.h",
2901 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002902 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002904 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002905 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002906 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002907 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002908 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002909]
2910
2911INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002912 "include/xnnpack.h",
2913 "src/xnnpack/allocator.h",
2914 "src/xnnpack/compute.h",
2915 "src/xnnpack/im2col.h",
2916 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002917 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002918 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 "src/xnnpack/operator.h",
2920 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002921 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002923 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002924 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002925]
2926
Marat Dukhan1b354632020-03-23 12:50:22 -07002927ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002928 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929]
2930
Marat Dukhan1b354632020-03-23 12:50:22 -07002931MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002932 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933 "include/xnnpack.h",
2934]
2935
Marat Dukhan1b354632020-03-23 12:50:22 -07002936MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002938 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 "src/xnnpack/requantization.h",
2940 "include/xnnpack.h",
2941]
2942
2943OPERATOR_TEST_PARAMS_HDRS = [
2944 "src/xnnpack/params.h",
2945 "src/xnnpack/common.h",
2946]
2947
2948WEIGHTS_PACK_HDRS = [
2949 "src/xnnpack/pack.h",
2950 "src/xnnpack/operator.h",
2951 "src/xnnpack/compute.h",
2952]
2953
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002954LOGGING_COPTS = select({
2955 # No logging in optimized mode
2956 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2957 # Full logging in debug mode
2958 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2959 # Error-only logging in default (fastbuild) mode
2960 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2961})
2962
Marat Dukhan3b59de22020-06-03 20:15:19 -07002963LOGGING_SRCS = select({
2964 # No logging in optimized mode
2965 ":optimized_build": [],
2966 "//conditions:default": [
2967 "src/operator-strings.c",
2968 "src/subgraph-strings.c",
2969 ],
2970})
2971
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002972LOGGING_HDRS = [
2973 "src/xnnpack/log.h",
2974]
2975
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002977 name = "tables",
2978 srcs = TABLE_SRCS,
2979 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002980 gcc_copts = xnnpack_gcc_std_copts(),
2981 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002982)
2983
2984xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 name = "scalar_ukernels",
2986 srcs = SCALAR_UKERNELS,
2987 hdrs = INTERNAL_HDRS,
2988 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002989 gcc_copts = xnnpack_gcc_std_copts(),
2990 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002992 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 "@FP16",
2994 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002995 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996 ],
2997)
2998
2999xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003000 name = "scalar_ukernels_test_mode",
3001 srcs = SCALAR_UKERNELS,
3002 hdrs = INTERNAL_HDRS,
3003 aarch32_copts = ["-marm"],
3004 copts = [
3005 "-UNDEBUG",
3006 "-DXNN_TEST_MODE=1",
3007 ],
3008 gcc_copts = xnnpack_gcc_std_copts(),
3009 msvc_copts = xnnpack_msvc_std_copts(),
3010 deps = [
3011 ":tables",
3012 "@FP16",
3013 "@FXdiv",
3014 "@pthreadpool",
3015 ],
3016)
3017
3018xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003019 name = "wasm_ukernels",
3020 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003021 gcc_copts = xnnpack_gcc_std_copts(),
3022 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003023 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003024 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003025 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003026 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003027 "@FP16",
3028 "@FXdiv",
3029 "@pthreadpool",
3030 ],
3031)
3032
3033xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003034 name = "wasm_ukernels_test_mode",
3035 hdrs = INTERNAL_HDRS,
3036 copts = [
3037 "-UNDEBUG",
3038 "-DXNN_TEST_MODE=1",
3039 ],
3040 gcc_copts = xnnpack_gcc_std_copts(),
3041 msvc_copts = xnnpack_msvc_std_copts(),
3042 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003043 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003044 deps = [
3045 ":tables",
3046 "@FP16",
3047 "@FXdiv",
3048 "@pthreadpool",
3049 ],
3050)
3051
3052xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003053 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 hdrs = INTERNAL_HDRS,
3055 aarch32_copts = [
3056 "-marm",
3057 "-mfpu=neon",
3058 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003059 gcc_copts = xnnpack_gcc_std_copts(),
3060 gcc_x86_copts = ["-msse2"],
3061 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062 optimized_copts = [
3063 "-O3",
3064 "-ffast-math",
3065 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003066 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003068 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 "@FP16",
3070 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003071 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003072 ],
3073)
3074
3075xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003076 name = "psimd_fastmath_ukernels_test_mode",
3077 hdrs = INTERNAL_HDRS,
3078 aarch32_copts = [
3079 "-marm",
3080 "-mfpu=neon",
3081 ],
3082 copts = [
3083 "-UNDEBUG",
3084 "-DXNN_TEST_MODE=1",
3085 ],
3086 gcc_copts = xnnpack_gcc_std_copts(),
3087 gcc_x86_copts = ["-msse2"],
3088 msvc_copts = xnnpack_msvc_std_copts(),
3089 optimized_copts = [
3090 "-O3",
3091 "-ffast-math",
3092 ],
3093 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3094 deps = [
3095 ":tables",
3096 "@FP16",
3097 "@psimd",
3098 "@pthreadpool",
3099 ],
3100)
3101
3102xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003103 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003104 hdrs = INTERNAL_HDRS,
3105 aarch32_copts = [
3106 "-marm",
3107 "-mfpu=neon",
3108 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003109 gcc_copts = xnnpack_gcc_std_copts(),
3110 gcc_x86_copts = ["-msse2"],
3111 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003112 optimized_copts = [
3113 "-O3",
3114 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003115 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003116 deps = [
3117 ":tables",
3118 "@FP16",
3119 "@psimd",
3120 "@pthreadpool",
3121 ],
3122)
3123
3124xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003125 name = "psimd_accmath_ukernels_test_mode",
3126 hdrs = INTERNAL_HDRS,
3127 aarch32_copts = [
3128 "-marm",
3129 "-mfpu=neon",
3130 ],
3131 copts = [
3132 "-UNDEBUG",
3133 "-DXNN_TEST_MODE=1",
3134 ],
3135 gcc_copts = xnnpack_gcc_std_copts(),
3136 gcc_x86_copts = ["-msse2"],
3137 msvc_copts = xnnpack_msvc_std_copts(),
3138 optimized_copts = [
3139 "-O3",
3140 ],
3141 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3142 deps = [
3143 ":tables",
3144 "@FP16",
3145 "@psimd",
3146 "@pthreadpool",
3147 ],
3148)
3149
3150xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 name = "neon_ukernels",
3152 hdrs = INTERNAL_HDRS,
3153 aarch32_copts = [
3154 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003155 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 "-mfpu=neon",
3157 ],
3158 aarch32_srcs = NEON_UKERNELS,
3159 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003160 gcc_copts = xnnpack_gcc_std_copts(),
3161 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003162 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003163 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003164 "@FP16",
3165 "@pthreadpool",
3166 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167)
3168
3169xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003170 name = "neon_ukernels_test_mode",
3171 hdrs = INTERNAL_HDRS,
3172 aarch32_copts = [
3173 "-marm",
3174 "-march=armv7-a",
3175 "-mfpu=neon",
3176 ],
3177 aarch32_srcs = NEON_UKERNELS,
3178 aarch64_srcs = NEON_UKERNELS,
3179 copts = [
3180 "-UNDEBUG",
3181 "-DXNN_TEST_MODE=1",
3182 ],
3183 gcc_copts = xnnpack_gcc_std_copts(),
3184 msvc_copts = xnnpack_msvc_std_copts(),
3185 deps = [
3186 ":tables",
3187 "@FP16",
3188 "@pthreadpool",
3189 ],
3190)
3191
3192xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 name = "neonfma_ukernels",
3194 hdrs = INTERNAL_HDRS,
3195 aarch32_copts = [
3196 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003197 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 "-mfpu=neon-vfpv4",
3199 ],
3200 aarch32_srcs = NEONFMA_UKERNELS,
3201 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003202 apple_aarch32_copts = [
3203 "-mcpu=swift",
3204 "-mtune=generic",
3205 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003206 gcc_copts = xnnpack_gcc_std_copts(),
3207 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003208 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003209 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003210 "@FP16",
3211 "@pthreadpool",
3212 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213)
3214
3215xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003216 name = "neonfma_ukernels_test_mode",
3217 hdrs = INTERNAL_HDRS,
3218 aarch32_copts = [
3219 "-marm",
3220 "-march=armv7-a",
3221 "-mfpu=neon-vfpv4",
3222 ],
3223 aarch32_srcs = NEONFMA_UKERNELS,
3224 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003225 apple_aarch32_copts = [
3226 "-mcpu=swift",
3227 "-mtune=generic",
3228 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003229 copts = [
3230 "-UNDEBUG",
3231 "-DXNN_TEST_MODE=1",
3232 ],
3233 gcc_copts = xnnpack_gcc_std_copts(),
3234 msvc_copts = xnnpack_msvc_std_copts(),
3235 deps = [
3236 ":tables",
3237 "@FP16",
3238 "@pthreadpool",
3239 ],
3240)
3241
3242xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003243 name = "neonv8_ukernels",
3244 hdrs = INTERNAL_HDRS,
3245 aarch32_copts = [
3246 "-marm",
3247 "-march=armv8-a",
3248 "-mfpu=neon-fp-armv8",
3249 ],
3250 aarch32_srcs = NEONV8_UKERNELS,
3251 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003252 apple_aarch32_copts = [
3253 "-mcpu=cyclone",
3254 "-mtune=generic",
3255 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003256 gcc_copts = xnnpack_gcc_std_copts(),
3257 msvc_copts = xnnpack_msvc_std_copts(),
3258 deps = [
3259 ":tables",
3260 "@FP16",
3261 "@pthreadpool",
3262 ],
3263)
3264
3265xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003266 name = "neonv8_ukernels_test_mode",
3267 hdrs = INTERNAL_HDRS,
3268 aarch32_copts = [
3269 "-marm",
3270 "-march=armv8-a",
3271 "-mfpu=neon-fp-armv8",
3272 ],
3273 aarch32_srcs = NEONV8_UKERNELS,
3274 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003275 apple_aarch32_copts = [
3276 "-mcpu=cyclone",
3277 "-mtune=generic",
3278 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003279 copts = [
3280 "-UNDEBUG",
3281 "-DXNN_TEST_MODE=1",
3282 ],
3283 gcc_copts = xnnpack_gcc_std_copts(),
3284 msvc_copts = xnnpack_msvc_std_copts(),
3285 deps = [
3286 ":tables",
3287 "@FP16",
3288 "@pthreadpool",
3289 ],
3290)
3291
3292xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003293 name = "neonfp16arith_ukernels",
3294 hdrs = INTERNAL_HDRS,
3295 aarch64_copts = ["-march=armv8.2-a+fp16"],
3296 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003297 gcc_copts = xnnpack_gcc_std_copts(),
3298 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003299 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003300 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003301 "@FP16",
3302 "@pthreadpool",
3303 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304)
3305
3306xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003307 name = "neonfp16arith_ukernels_test_mode",
3308 hdrs = INTERNAL_HDRS,
3309 aarch64_copts = ["-march=armv8.2-a+fp16"],
3310 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3311 copts = [
3312 "-UNDEBUG",
3313 "-DXNN_TEST_MODE=1",
3314 ],
3315 gcc_copts = xnnpack_gcc_std_copts(),
3316 msvc_copts = xnnpack_msvc_std_copts(),
3317 deps = [
3318 ":tables",
3319 "@FP16",
3320 "@pthreadpool",
3321 ],
3322)
3323
3324xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003325 name = "neondot_ukernels",
3326 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003327 aarch32_copts = [
3328 "-marm",
3329 "-march=armv8.2-a+dotprod",
3330 "-mfpu=neon-fp-armv8",
3331 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003332 aarch32_srcs = NEONDOT_UKERNELS,
3333 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3334 aarch64_srcs = NEONDOT_UKERNELS,
3335 gcc_copts = xnnpack_gcc_std_copts(),
3336 msvc_copts = xnnpack_msvc_std_copts(),
3337 deps = [
3338 ":tables",
3339 "@FP16",
3340 "@pthreadpool",
3341 ],
3342)
3343
3344xnnpack_cc_library(
3345 name = "neondot_ukernels_test_mode",
3346 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003347 aarch32_copts = [
3348 "-marm",
3349 "-march=armv8.2-a+dotprod",
3350 "-mfpu=neon-fp-armv8",
3351 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003352 aarch32_srcs = NEONDOT_UKERNELS,
3353 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3354 aarch64_srcs = NEONDOT_UKERNELS,
3355 copts = [
3356 "-UNDEBUG",
3357 "-DXNN_TEST_MODE=1",
3358 ],
3359 gcc_copts = xnnpack_gcc_std_copts(),
3360 msvc_copts = xnnpack_msvc_std_copts(),
3361 deps = [
3362 ":tables",
3363 "@FP16",
3364 "@pthreadpool",
3365 ],
3366)
3367
3368xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 name = "sse2_ukernels",
3370 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003371 gcc_copts = xnnpack_gcc_std_copts(),
3372 gcc_x86_copts = ["-msse2"],
3373 msvc_copts = xnnpack_msvc_std_copts(),
3374 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003376 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003377 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003378 "@FP16",
3379 "@pthreadpool",
3380 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381)
3382
3383xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003384 name = "sse2_ukernels_test_mode",
3385 hdrs = INTERNAL_HDRS,
3386 copts = [
3387 "-UNDEBUG",
3388 "-DXNN_TEST_MODE=1",
3389 ],
3390 gcc_copts = xnnpack_gcc_std_copts(),
3391 gcc_x86_copts = ["-msse2"],
3392 msvc_copts = xnnpack_msvc_std_copts(),
3393 msvc_x86_32_copts = ["/arch:SSE2"],
3394 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3395 deps = [
3396 ":tables",
3397 "@FP16",
3398 "@pthreadpool",
3399 ],
3400)
3401
3402xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003403 name = "ssse3_ukernels",
3404 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003405 gcc_copts = xnnpack_gcc_std_copts(),
3406 gcc_x86_copts = ["-mssse3"],
3407 msvc_copts = xnnpack_msvc_std_copts(),
3408 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003409 x86_srcs = SSSE3_UKERNELS,
3410 deps = [
3411 ":tables",
3412 "@FP16",
3413 "@pthreadpool",
3414 ],
3415)
3416
3417xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003418 name = "ssse3_ukernels_test_mode",
3419 hdrs = INTERNAL_HDRS,
3420 copts = [
3421 "-UNDEBUG",
3422 "-DXNN_TEST_MODE=1",
3423 ],
3424 gcc_copts = xnnpack_gcc_std_copts(),
3425 gcc_x86_copts = ["-mssse3"],
3426 msvc_copts = xnnpack_msvc_std_copts(),
3427 msvc_x86_32_copts = ["/arch:SSE2"],
3428 x86_srcs = SSSE3_UKERNELS,
3429 deps = [
3430 ":tables",
3431 "@FP16",
3432 "@pthreadpool",
3433 ],
3434)
3435
3436xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003437 name = "sse41_ukernels",
3438 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003439 gcc_copts = xnnpack_gcc_std_copts(),
3440 gcc_x86_copts = ["-msse4.1"],
3441 msvc_copts = xnnpack_msvc_std_copts(),
3442 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003443 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003444 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003445 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003446 "@FP16",
3447 "@pthreadpool",
3448 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003449)
3450
3451xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003452 name = "sse41_ukernels_test_mode",
3453 hdrs = INTERNAL_HDRS,
3454 copts = [
3455 "-UNDEBUG",
3456 "-DXNN_TEST_MODE=1",
3457 ],
3458 gcc_copts = xnnpack_gcc_std_copts(),
3459 gcc_x86_copts = ["-msse4.1"],
3460 msvc_copts = xnnpack_msvc_std_copts(),
3461 msvc_x86_32_copts = ["/arch:SSE2"],
3462 x86_srcs = SSE41_UKERNELS,
3463 deps = [
3464 ":tables",
3465 "@FP16",
3466 "@pthreadpool",
3467 ],
3468)
3469
3470xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003471 name = "avx_ukernels",
3472 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003473 gcc_copts = xnnpack_gcc_std_copts(),
3474 gcc_x86_copts = ["-mavx"],
3475 msvc_copts = xnnpack_msvc_std_copts(),
3476 msvc_x86_32_copts = ["/arch:AVX"],
3477 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003479 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003480 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003481 "@FP16",
3482 "@pthreadpool",
3483 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484)
3485
3486xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003487 name = "avx_ukernels_test_mode",
3488 hdrs = INTERNAL_HDRS,
3489 copts = [
3490 "-UNDEBUG",
3491 "-DXNN_TEST_MODE=1",
3492 ],
3493 gcc_copts = xnnpack_gcc_std_copts(),
3494 gcc_x86_copts = ["-mavx"],
3495 msvc_copts = xnnpack_msvc_std_copts(),
3496 msvc_x86_32_copts = ["/arch:AVX"],
3497 msvc_x86_64_copts = ["/arch:AVX"],
3498 x86_srcs = AVX_UKERNELS,
3499 deps = [
3500 ":tables",
3501 "@FP16",
3502 "@pthreadpool",
3503 ],
3504)
3505
3506xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003507 name = "xop_ukernels",
3508 hdrs = INTERNAL_HDRS,
3509 gcc_copts = xnnpack_gcc_std_copts(),
3510 gcc_x86_copts = ["-mxop"],
3511 msvc_copts = xnnpack_msvc_std_copts(),
3512 msvc_x86_32_copts = ["/arch:AVX"],
3513 msvc_x86_64_copts = ["/arch:AVX"],
3514 x86_srcs = XOP_UKERNELS,
3515 deps = [
3516 ":tables",
3517 "@FP16",
3518 "@pthreadpool",
3519 ],
3520)
3521
3522xnnpack_cc_library(
3523 name = "xop_ukernels_test_mode",
3524 hdrs = INTERNAL_HDRS,
3525 copts = [
3526 "-UNDEBUG",
3527 "-DXNN_TEST_MODE=1",
3528 ],
3529 gcc_copts = xnnpack_gcc_std_copts(),
3530 gcc_x86_copts = ["-mxop"],
3531 msvc_copts = xnnpack_msvc_std_copts(),
3532 msvc_x86_32_copts = ["/arch:AVX"],
3533 msvc_x86_64_copts = ["/arch:AVX"],
3534 x86_srcs = XOP_UKERNELS,
3535 deps = [
3536 ":tables",
3537 "@FP16",
3538 "@pthreadpool",
3539 ],
3540)
3541
3542xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003543 name = "fma3_ukernels",
3544 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003545 gcc_copts = xnnpack_gcc_std_copts(),
3546 gcc_x86_copts = ["-mfma"],
3547 msvc_copts = xnnpack_msvc_std_copts(),
3548 msvc_x86_32_copts = ["/arch:AVX"],
3549 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003550 x86_srcs = FMA3_UKERNELS,
3551 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003552 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003553 "@FP16",
3554 "@pthreadpool",
3555 ],
3556)
3557
3558xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003559 name = "fma3_ukernels_test_mode",
3560 hdrs = INTERNAL_HDRS,
3561 copts = [
3562 "-UNDEBUG",
3563 "-DXNN_TEST_MODE=1",
3564 ],
3565 gcc_copts = xnnpack_gcc_std_copts(),
3566 gcc_x86_copts = ["-mfma"],
3567 msvc_copts = xnnpack_msvc_std_copts(),
3568 msvc_x86_32_copts = ["/arch:AVX"],
3569 msvc_x86_64_copts = ["/arch:AVX"],
3570 x86_srcs = FMA3_UKERNELS,
3571 deps = [
3572 ":tables",
3573 "@FP16",
3574 "@pthreadpool",
3575 ],
3576)
3577
3578xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003579 name = "avx2_ukernels",
3580 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003581 gcc_copts = xnnpack_gcc_std_copts(),
3582 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003583 "-mfma",
3584 "-mavx2",
3585 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003586 msvc_copts = xnnpack_msvc_std_copts(),
3587 msvc_x86_32_copts = ["/arch:AVX2"],
3588 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003589 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003590 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003591 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003592 "@FP16",
3593 "@pthreadpool",
3594 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003595)
3596
3597xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003598 name = "avx2_ukernels_test_mode",
3599 hdrs = INTERNAL_HDRS,
3600 copts = [
3601 "-UNDEBUG",
3602 "-DXNN_TEST_MODE=1",
3603 ],
3604 gcc_copts = xnnpack_gcc_std_copts(),
3605 gcc_x86_copts = [
3606 "-mfma",
3607 "-mavx2",
3608 ],
3609 msvc_copts = xnnpack_msvc_std_copts(),
3610 msvc_x86_32_copts = ["/arch:AVX2"],
3611 msvc_x86_64_copts = ["/arch:AVX2"],
3612 x86_srcs = AVX2_UKERNELS,
3613 deps = [
3614 ":tables",
3615 "@FP16",
3616 "@pthreadpool",
3617 ],
3618)
3619
3620xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621 name = "avx512f_ukernels",
3622 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003623 gcc_copts = xnnpack_gcc_std_copts(),
3624 gcc_x86_copts = ["-mavx512f"],
3625 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3626 msvc_copts = xnnpack_msvc_std_copts(),
3627 msvc_x86_32_copts = ["/arch:AVX512"],
3628 msvc_x86_64_copts = ["/arch:AVX512"],
3629 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003631 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003632 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003633 "@FP16",
3634 "@pthreadpool",
3635 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636)
3637
3638xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003639 name = "avx512f_ukernels_test_mode",
3640 hdrs = INTERNAL_HDRS,
3641 copts = [
3642 "-UNDEBUG",
3643 "-DXNN_TEST_MODE=1",
3644 ],
3645 gcc_copts = xnnpack_gcc_std_copts(),
3646 gcc_x86_copts = ["-mavx512f"],
3647 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3648 msvc_copts = xnnpack_msvc_std_copts(),
3649 msvc_x86_32_copts = ["/arch:AVX512"],
3650 msvc_x86_64_copts = ["/arch:AVX512"],
3651 msys_copts = ["-fno-asynchronous-unwind-tables"],
3652 x86_srcs = AVX512F_UKERNELS,
3653 deps = [
3654 ":tables",
3655 "@FP16",
3656 "@pthreadpool",
3657 ],
3658)
3659
3660xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003661 name = "avx512skx_ukernels",
3662 hdrs = INTERNAL_HDRS,
3663 gcc_copts = xnnpack_gcc_std_copts(),
3664 gcc_x86_copts = [
3665 "-mavx512f",
3666 "-mavx512cd",
3667 "-mavx512bw",
3668 "-mavx512dq",
3669 "-mavx512vl",
3670 ],
3671 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3672 msvc_copts = xnnpack_msvc_std_copts(),
3673 msvc_x86_32_copts = ["/arch:AVX512"],
3674 msvc_x86_64_copts = ["/arch:AVX512"],
3675 msys_copts = ["-fno-asynchronous-unwind-tables"],
3676 x86_srcs = AVX512SKX_UKERNELS,
3677 deps = [
3678 ":tables",
3679 "@FP16",
3680 "@pthreadpool",
3681 ],
3682)
3683
3684xnnpack_cc_library(
3685 name = "avx512skx_ukernels_test_mode",
3686 hdrs = INTERNAL_HDRS,
3687 copts = [
3688 "-UNDEBUG",
3689 "-DXNN_TEST_MODE=1",
3690 ],
3691 gcc_copts = xnnpack_gcc_std_copts(),
3692 gcc_x86_copts = [
3693 "-mavx512f",
3694 "-mavx512cd",
3695 "-mavx512bw",
3696 "-mavx512dq",
3697 "-mavx512vl",
3698 ],
3699 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3700 msvc_copts = xnnpack_msvc_std_copts(),
3701 msvc_x86_32_copts = ["/arch:AVX512"],
3702 msvc_x86_64_copts = ["/arch:AVX512"],
3703 msys_copts = ["-fno-asynchronous-unwind-tables"],
3704 x86_srcs = AVX512SKX_UKERNELS,
3705 deps = [
3706 ":tables",
3707 "@FP16",
3708 "@pthreadpool",
3709 ],
3710)
3711
3712xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003713 name = "asm_ukernels",
3714 hdrs = ["src/xnnpack/assembly.h"],
3715 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003716 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003717 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003718 wasm_srcs = WASM32_ASM_UKERNELS,
3719 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003720)
3721
Marat Dukhan3b59de22020-06-03 20:15:19 -07003722xnnpack_cc_library(
3723 name = "logging_utils",
3724 srcs = LOGGING_SRCS,
3725 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3726 copts = LOGGING_COPTS + [
3727 "-Isrc",
3728 "-Iinclude",
3729 ] + select({
3730 ":debug_build": [],
3731 "//conditions:default": xnnpack_min_size_copts(),
3732 }),
3733 gcc_copts = xnnpack_gcc_std_copts(),
3734 msvc_copts = xnnpack_msvc_std_copts(),
3735 visibility = xnnpack_visibility(),
3736 deps = [
3737 "@FP16",
3738 "@clog",
3739 "@pthreadpool",
3740 ],
3741)
3742
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743xnnpack_aggregate_library(
3744 name = "ukernels",
3745 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746 ":neon_ukernels",
3747 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003748 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003749 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 ":asm_ukernels",
3751 ],
3752 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 ":neon_ukernels",
3754 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003755 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003757 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 ":asm_ukernels",
3759 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003760 generic_deps = [
3761 ":scalar_ukernels",
3762 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003763 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003764 ":psimd_fastmath_ukernels",
3765 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003767 wasm_deps = [
3768 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003769 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003770 ],
3771 wasmsimd_deps = [
3772 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003773 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003774 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003777 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003778 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003780 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003781 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003782 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003784 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003785 ],
3786)
3787
Marat Dukhan33fcf782020-05-24 14:27:15 -07003788xnnpack_aggregate_library(
3789 name = "ukernels_test_mode",
3790 aarch32_deps = [
3791 ":neon_ukernels_test_mode",
3792 ":neonfma_ukernels_test_mode",
3793 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003794 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003795 ":asm_ukernels",
3796 ],
3797 aarch64_deps = [
3798 ":neon_ukernels_test_mode",
3799 ":neonfma_ukernels_test_mode",
3800 ":neonv8_ukernels_test_mode",
3801 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003802 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003803 ":asm_ukernels",
3804 ],
3805 generic_deps = [
3806 ":scalar_ukernels_test_mode",
3807 ],
3808 psimd_deps = [
3809 ":psimd_fastmath_ukernels_test_mode",
3810 ":psimd_accmath_ukernels_test_mode",
3811 ],
3812 wasm_deps = [
3813 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003814 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003815 ],
3816 wasmsimd_deps = [
3817 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003818 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003819 ],
3820 x86_deps = [
3821 ":sse2_ukernels_test_mode",
3822 ":ssse3_ukernels_test_mode",
3823 ":sse41_ukernels_test_mode",
3824 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003825 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003826 ":fma3_ukernels_test_mode",
3827 ":avx2_ukernels_test_mode",
3828 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003829 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003830 ],
3831)
3832
Marat Dukhan86df3842020-06-17 14:03:31 -07003833# This target must be compatible with Android NDK r17 and compile with Clang 6.
3834# It is similar to :ukernels target, but excludes microkernels which require
3835# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003836xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003837 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003838 aarch32_deps = [
3839 ":neon_ukernels",
3840 ":neonfma_ukernels",
3841 ":neonv8_ukernels",
3842 ":asm_ukernels",
3843 ],
3844 aarch64_deps = [
3845 ":neon_ukernels",
3846 ":neonfma_ukernels",
3847 ":neonv8_ukernels",
3848 ":asm_ukernels",
3849 ],
3850 generic_deps = [
3851 ":scalar_ukernels",
3852 ],
3853 psimd_deps = [
3854 ":psimd_fastmath_ukernels",
3855 ":psimd_accmath_ukernels",
3856 ],
3857 wasm_deps = [
3858 ":wasm_ukernels",
3859 ],
3860 wasmsimd_deps = [
3861 ":wasm_ukernels",
3862 ],
3863 x86_deps = [
3864 ":sse2_ukernels",
3865 ":ssse3_ukernels",
3866 ":sse41_ukernels",
3867 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003868 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003869 ":fma3_ukernels",
3870 ":avx2_ukernels",
3871 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003872 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003873 ],
3874)
3875
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876xnnpack_cc_library(
3877 name = "im2col",
3878 srcs = ["src/im2col.c"],
3879 hdrs = [
3880 "src/xnnpack/common.h",
3881 "src/xnnpack/im2col.h",
3882 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003883 gcc_copts = xnnpack_gcc_std_copts(),
3884 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003885)
3886
3887xnnpack_cc_library(
3888 name = "indirection",
3889 srcs = ["src/indirection.c"],
3890 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003891 gcc_copts = xnnpack_gcc_std_copts(),
3892 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003893 deps = [
3894 "@FP16",
3895 "@FXdiv",
3896 "@pthreadpool",
3897 ],
3898)
3899
3900xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003901 name = "indirection_test_mode",
3902 srcs = ["src/indirection.c"],
3903 hdrs = INTERNAL_HDRS,
3904 copts = [
3905 "-UNDEBUG",
3906 "-DXNN_TEST_MODE=1",
3907 ],
3908 gcc_copts = xnnpack_gcc_std_copts(),
3909 msvc_copts = xnnpack_msvc_std_copts(),
3910 deps = [
3911 "@FP16",
3912 "@FXdiv",
3913 "@pthreadpool",
3914 ],
3915)
3916
3917xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003918 name = "packing",
3919 srcs = ["src/packing.c"],
3920 hdrs = INTERNAL_HDRS,
3921 gcc_copts = xnnpack_gcc_std_copts(),
3922 msvc_copts = xnnpack_msvc_std_copts(),
3923 deps = [
3924 "@FP16",
3925 "@FXdiv",
3926 "@pthreadpool",
3927 ],
3928)
3929
3930xnnpack_cc_library(
3931 name = "packing_test_mode",
3932 srcs = ["src/packing.c"],
3933 hdrs = INTERNAL_HDRS,
3934 copts = [
3935 "-UNDEBUG",
3936 "-DXNN_TEST_MODE=1",
3937 ],
3938 gcc_copts = xnnpack_gcc_std_copts(),
3939 msvc_copts = xnnpack_msvc_std_copts(),
3940 deps = [
3941 "@FP16",
3942 "@FXdiv",
3943 "@pthreadpool",
3944 ],
3945)
3946
3947xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003948 name = "operator_run",
3949 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003950 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003951 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003952 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3953 "//conditions:default": [],
3954 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003955 gcc_copts = xnnpack_gcc_std_copts(),
3956 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003957 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003958 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959 "@FP16",
3960 "@FXdiv",
3961 "@clog",
3962 "@pthreadpool",
3963 ],
3964)
3965
Chao Mei6ddfc602020-05-13 22:29:36 -07003966xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003967 name = "operator_run_test_mode",
3968 srcs = ["src/operator-run.c"],
3969 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3970 copts = LOGGING_COPTS + [
3971 "-UNDEBUG",
3972 "-DXNN_TEST_MODE=1",
3973 ] + select({
3974 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3975 "//conditions:default": [],
3976 }),
3977 gcc_copts = xnnpack_gcc_std_copts(),
3978 msvc_copts = xnnpack_msvc_std_copts(),
3979 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003980 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003981 "@FP16",
3982 "@FXdiv",
3983 "@clog",
3984 "@pthreadpool",
3985 ],
3986)
3987
3988xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003989 name = "memory_planner",
3990 srcs = ["src/memory-planner.c"],
3991 hdrs = INTERNAL_HDRS,
3992 defines = select({
3993 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3994 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3995 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3996 }),
3997 gcc_copts = xnnpack_gcc_std_copts(),
3998 msvc_copts = xnnpack_msvc_std_copts(),
3999 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004000 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004001 "@pthreadpool",
4002 ],
4003)
4004
Marat Dukhan33fcf782020-05-24 14:27:15 -07004005xnnpack_cc_library(
4006 name = "memory_planner_test_mode",
4007 srcs = ["src/memory-planner.c"],
4008 hdrs = INTERNAL_HDRS,
4009 copts = [
4010 "-UNDEBUG",
4011 "-DXNN_TEST_MODE=1",
4012 ],
4013 defines = select({
4014 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4015 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4016 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4017 }),
4018 gcc_copts = xnnpack_gcc_std_copts(),
4019 msvc_copts = xnnpack_msvc_std_copts(),
4020 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004021 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004022 "@pthreadpool",
4023 ],
4024)
4025
Marat Dukhan08c4a432019-10-03 09:29:21 -07004026cc_library(
4027 name = "enable_assembly",
4028 defines = select({
4029 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4030 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004031 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004032 }),
4033)
4034
Marat Dukhan9de90e02020-06-18 16:04:12 -07004035cc_library(
4036 name = "enable_sparse",
4037 defines = select({
4038 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4039 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004040 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004041 }),
4042)
4043
Marat Dukhancf056b22019-10-07 10:26:29 -07004044xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004045 name = "operators",
4046 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004047 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004048 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004049 ],
4050 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004051 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004052 "-Isrc",
4053 "-Iinclude",
4054 ] + select({
4055 ":debug_build": [],
4056 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004057 }) + select({
4058 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4059 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004061 gcc_copts = xnnpack_gcc_std_copts(),
4062 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004063 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004064 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004065 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004066 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004067 "@FP16",
4068 "@FXdiv",
4069 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004070 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004071 ],
4072)
4073
Marat Dukhan10a38082020-04-17 03:58:35 -07004074xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004075 name = "operators_test_mode",
4076 srcs = OPERATOR_SRCS + [
4077 "src/memory.c",
4078 "src/operator-delete.c",
4079 ],
4080 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4081 copts = LOGGING_COPTS + [
4082 "-Isrc",
4083 "-Iinclude",
4084 "-UNDEBUG",
4085 "-DXNN_TEST_MODE=1",
4086 ] + select({
4087 ":debug_build": [],
4088 "//conditions:default": xnnpack_min_size_copts(),
4089 }) + select({
4090 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4091 "//conditions:default": [],
4092 }),
4093 gcc_copts = xnnpack_gcc_std_copts(),
4094 msvc_copts = xnnpack_msvc_std_copts(),
4095 deps = [
4096 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004097 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004098 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004099 "@FP16",
4100 "@FXdiv",
4101 "@clog",
4102 "@pthreadpool",
4103 ],
4104)
4105
4106xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004107 name = "XNNPACK",
4108 srcs = [
4109 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004110 "src/runtime.c",
4111 "src/subgraph.c",
4112 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004113 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004114 hdrs = ["include/xnnpack.h"],
4115 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004116 "-Isrc",
4117 "-Iinclude",
4118 ] + select({
4119 ":debug_build": [],
4120 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004121 }) + select({
4122 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4123 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004124 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004125 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004126 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004127 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004128 visibility = xnnpack_visibility(),
4129 deps = [
4130 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004131 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004132 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004133 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004134 ":operator_run",
4135 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004136 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004137 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004138 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004139 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004140 ] + select({
4141 ":emscripten": [],
4142 "//conditions:default": ["@cpuinfo"],
4143 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144)
4145
Marat Dukhan10a38082020-04-17 03:58:35 -07004146xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004147 name = "XNNPACK_test_mode",
4148 srcs = [
4149 "src/init.c",
4150 "src/runtime.c",
4151 "src/subgraph.c",
4152 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004153 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004154 hdrs = ["include/xnnpack.h"],
4155 copts = LOGGING_COPTS + [
4156 "-Isrc",
4157 "-Iinclude",
4158 "-UNDEBUG",
4159 "-DXNN_TEST_MODE=1",
4160 ] + select({
4161 ":debug_build": [],
4162 "//conditions:default": xnnpack_min_size_copts(),
4163 }) + select({
4164 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4165 "//conditions:default": [],
4166 }),
4167 gcc_copts = xnnpack_gcc_std_copts(),
4168 includes = ["include"],
4169 msvc_copts = xnnpack_msvc_std_copts(),
4170 visibility = xnnpack_visibility(),
4171 deps = [
4172 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004173 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004174 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004175 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004176 ":operator_run_test_mode",
4177 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004178 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004179 "@clog",
4180 "@FP16",
4181 "@pthreadpool",
4182 ] + select({
4183 ":emscripten": [],
4184 "//conditions:default": ["@cpuinfo"],
4185 }),
4186)
4187
4188xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004189 name = "xnnpack_f32",
4190 srcs = [
4191 "src/init.c",
4192 "src/runtime.c",
4193 "src/subgraph.c",
4194 "src/tensor.c",
4195 ] + SUBGRAPH_SRCS,
4196 hdrs = ["include/xnnpack.h"],
4197 copts = LOGGING_COPTS + [
4198 "-Isrc",
4199 "-Iinclude",
4200 ] + select({
4201 ":debug_build": [],
4202 "//conditions:default": xnnpack_min_size_copts(),
4203 }) + select({
4204 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4205 "//conditions:default": [],
4206 }),
4207 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004208 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004209 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004210 "XNN_NO_U8_OPERATORS",
4211 "XNN_NO_X8_OPERATORS",
4212 "XNN_NO_F16_OPERATORS",
4213 "XNN_NO_X16_OPERATORS",
4214 ],
4215 gcc_copts = xnnpack_gcc_std_copts(),
4216 includes = ["include"],
4217 msvc_copts = xnnpack_msvc_std_copts(),
4218 visibility = xnnpack_visibility(),
4219 deps = [
4220 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004221 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004222 ":logging_utils",
4223 ":memory_planner",
4224 ":operator_run",
4225 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004226 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004227 "@clog",
4228 "@FP16",
4229 "@pthreadpool",
4230 ] + select({
4231 ":emscripten": [],
4232 "//conditions:default": ["@cpuinfo"],
4233 }),
4234)
4235
4236xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004237 name = "xnnpack_operators_nhwc_f32",
4238 srcs = [
4239 "src/init.c",
4240 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004241 hdrs = ["include/xnnpack.h"],
4242 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004243 "-Isrc",
4244 "-Iinclude",
4245 ] + select({
4246 ":debug_build": [],
4247 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004248 }) + select({
4249 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4250 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004251 }),
4252 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004253 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004254 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004255 "XNN_NO_U8_OPERATORS",
4256 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004257 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004258 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004259 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004261 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262 visibility = xnnpack_visibility(),
4263 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004264 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004265 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004266 ":operator_run",
4267 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004268 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004269 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004270 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004271 ] + select({
4272 ":emscripten": [],
4273 "//conditions:default": ["@cpuinfo"],
4274 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275)
4276
Marat Dukhancf056b22019-10-07 10:26:29 -07004277xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 name = "bench_utils",
4279 srcs = ["bench/utils.cc"],
4280 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004281 deps = [
4282 "@com_google_benchmark//:benchmark",
4283 "@cpuinfo",
4284 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285)
4286
Frank Barchard7e955972019-10-11 10:34:25 -07004287######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004288
4289xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004290 name = "qs8_gemm_bench",
4291 srcs = [
4292 "bench/gemm.h",
4293 "bench/qs8-gemm.cc",
4294 "src/xnnpack/AlignedAllocator.h",
4295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004296 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4297 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004298)
4299
4300xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004301 name = "qs8_requantization_bench",
4302 srcs = [
4303 "bench/qs8-requantization.cc",
4304 "src/xnnpack/requantization-stubs.h",
4305 "src/xnnpack/AlignedAllocator.h",
4306 ] + MICROKERNEL_BENCHMARK_HDRS,
4307 deps = MICROKERNEL_BENCHMARK_DEPS,
4308)
4309
4310xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004311 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 srcs = [
4313 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004314 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004315 "src/xnnpack/AlignedAllocator.h",
4316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004317 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004318 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319)
4320
4321xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004322 name = "qu8_requantization_bench",
4323 srcs = [
4324 "bench/qu8-requantization.cc",
4325 "src/xnnpack/requantization-stubs.h",
4326 "src/xnnpack/AlignedAllocator.h",
4327 ] + MICROKERNEL_BENCHMARK_HDRS,
4328 deps = MICROKERNEL_BENCHMARK_DEPS,
4329)
4330
4331xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004332 name = "f16_igemm_bench",
4333 srcs = [
4334 "bench/f16-igemm.cc",
4335 "bench/conv.h",
4336 "bench/google/conv.h",
4337 "src/xnnpack/AlignedAllocator.h",
4338 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004339 deps = MICROKERNEL_BENCHMARK_DEPS + [
4340 ":indirection",
4341 ":packing",
4342 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004343)
4344
4345xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004346 name = "f16_gemm_bench",
4347 srcs = [
4348 "bench/f16-gemm.cc",
4349 "bench/gemm.h",
4350 "src/xnnpack/AlignedAllocator.h",
4351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004352 deps = MICROKERNEL_BENCHMARK_DEPS + [
4353 ":packing",
4354 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355)
4356
4357xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004358 name = "f16_spmm_bench",
4359 srcs = [
4360 "bench/f16-spmm.cc",
4361 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004362 "src/xnnpack/AlignedAllocator.h",
4363 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004364 deps = MICROKERNEL_BENCHMARK_DEPS,
4365)
4366
4367xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 name = "f32_igemm_bench",
4369 srcs = [
4370 "bench/f32-igemm.cc",
4371 "bench/conv.h",
4372 "src/xnnpack/AlignedAllocator.h",
4373 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004374 deps = MICROKERNEL_BENCHMARK_DEPS + [
4375 ":indirection",
4376 ":packing",
4377 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378)
4379
4380xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004381 name = "f16_relu_bench",
4382 srcs = [
4383 "bench/f16-relu.cc",
4384 "src/xnnpack/AlignedAllocator.h",
4385 ] + MICROKERNEL_BENCHMARK_HDRS,
4386 deps = MICROKERNEL_BENCHMARK_DEPS,
4387)
4388
4389xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 name = "f32_conv_hwc_bench",
4391 srcs = [
4392 "bench/f32-conv-hwc.cc",
4393 "bench/dconv.h",
4394 "src/xnnpack/AlignedAllocator.h",
4395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004396 deps = MICROKERNEL_BENCHMARK_DEPS + [
4397 ":packing",
4398 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399)
4400
4401xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004402 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004403 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004404 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004405 "bench/dconv.h",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004408 deps = MICROKERNEL_BENCHMARK_DEPS + [
4409 ":packing",
4410 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004411)
4412
4413xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004414 name = "f16_dwconv_bench",
4415 srcs = [
4416 "bench/f16-dwconv.cc",
4417 "bench/dwconv.h",
4418 "bench/google/dwconv.h",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004421 deps = MICROKERNEL_BENCHMARK_DEPS + [
4422 ":indirection",
4423 ":packing",
4424 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004425)
4426
4427xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004428 name = "f32_dwconv_bench",
4429 srcs = [
4430 "bench/f32-dwconv.cc",
4431 "bench/dwconv.h",
4432 "src/xnnpack/AlignedAllocator.h",
4433 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004434 deps = MICROKERNEL_BENCHMARK_DEPS + [
4435 ":indirection",
4436 ":packing",
4437 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438)
4439
4440xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004441 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004443 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 "bench/dwconv.h",
4445 "src/xnnpack/AlignedAllocator.h",
4446 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004447 deps = MICROKERNEL_BENCHMARK_DEPS + [
4448 ":indirection",
4449 ":packing",
4450 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451)
4452
4453xnnpack_benchmark(
4454 name = "f32_gemm_bench",
4455 srcs = [
4456 "bench/f32-gemm.cc",
4457 "bench/gemm.h",
4458 "src/xnnpack/AlignedAllocator.h",
4459 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004460 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004461 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462)
4463
4464xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004465 name = "f32_hswish_bench",
4466 srcs = [
4467 "bench/f32-hswish.cc",
4468 "src/xnnpack/AlignedAllocator.h",
4469 ] + MICROKERNEL_BENCHMARK_HDRS,
4470 deps = MICROKERNEL_BENCHMARK_DEPS,
4471)
4472
4473xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004474 name = "f32_raddexpminusmax_bench",
4475 srcs = [
4476 "bench/f32-raddexpminusmax.cc",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + MICROKERNEL_BENCHMARK_HDRS,
4479 deps = MICROKERNEL_BENCHMARK_DEPS,
4480)
4481
4482xnnpack_benchmark(
4483 name = "f32_raddextexp_bench",
4484 srcs = [
4485 "bench/f32-raddextexp.cc",
4486 "src/xnnpack/AlignedAllocator.h",
4487 ] + MICROKERNEL_BENCHMARK_HDRS,
4488 deps = MICROKERNEL_BENCHMARK_DEPS,
4489)
4490
4491xnnpack_benchmark(
4492 name = "f32_raddstoreexpminusmax_bench",
4493 srcs = [
4494 "bench/f32-raddstoreexpminusmax.cc",
4495 "src/xnnpack/AlignedAllocator.h",
4496 ] + MICROKERNEL_BENCHMARK_HDRS,
4497 deps = MICROKERNEL_BENCHMARK_DEPS,
4498)
4499
4500xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004501 name = "f32_relu_bench",
4502 srcs = [
4503 "bench/f32-relu.cc",
4504 "src/xnnpack/AlignedAllocator.h",
4505 ] + MICROKERNEL_BENCHMARK_HDRS,
4506 deps = MICROKERNEL_BENCHMARK_DEPS,
4507)
4508
4509xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510 name = "f32_rmax_bench",
4511 srcs = [
4512 "bench/f32-rmax.cc",
4513 "src/xnnpack/AlignedAllocator.h",
4514 ] + MICROKERNEL_BENCHMARK_HDRS,
4515 deps = MICROKERNEL_BENCHMARK_DEPS,
4516)
4517
4518xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004519 name = "f32_sigmoid_bench",
4520 srcs = [
4521 "bench/f32-sigmoid.cc",
4522 "src/xnnpack/AlignedAllocator.h",
4523 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004524 deps = MICROKERNEL_BENCHMARK_DEPS,
4525)
4526
4527xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 name = "f32_spmm_bench",
4529 srcs = [
4530 "bench/f32-spmm.cc",
4531 "bench/gemm.h",
4532 "src/xnnpack/AlignedAllocator.h",
4533 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 deps = MICROKERNEL_BENCHMARK_DEPS,
4535)
4536
4537xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004538 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004539 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004540 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004541 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004542 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004543 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004544)
4545
4546xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004547 name = "f32_vscaleexpminusmax_bench",
4548 srcs = [
4549 "bench/f32-vscaleexpminusmax.cc",
4550 "src/xnnpack/AlignedAllocator.h",
4551 ] + MICROKERNEL_BENCHMARK_HDRS,
4552 deps = MICROKERNEL_BENCHMARK_DEPS,
4553)
4554
4555xnnpack_benchmark(
4556 name = "f32_vscaleextexp_bench",
4557 srcs = [
4558 "bench/f32-vscaleextexp.cc",
4559 "src/xnnpack/AlignedAllocator.h",
4560 ] + MICROKERNEL_BENCHMARK_HDRS,
4561 deps = MICROKERNEL_BENCHMARK_DEPS,
4562)
4563
4564xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004565 name = "f32_vsqrt_bench",
4566 srcs = [
4567 "bench/f32-vsqrt.cc",
4568 "src/xnnpack/AlignedAllocator.h",
4569 ] + MICROKERNEL_BENCHMARK_HDRS,
4570 deps = MICROKERNEL_BENCHMARK_DEPS,
4571)
4572
4573xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 name = "f32_im2col_gemm_bench",
4575 srcs = [
4576 "bench/f32-im2col-gemm.cc",
4577 "bench/conv.h",
4578 "src/xnnpack/AlignedAllocator.h",
4579 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004580 deps = MICROKERNEL_BENCHMARK_DEPS + [
4581 ":im2col",
4582 ":packing",
4583 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004584)
4585
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004586xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004587 name = "rounding_bench",
4588 srcs = [
4589 "bench/rounding.cc",
4590 "src/xnnpack/math-stubs.h",
4591 "src/xnnpack/AlignedAllocator.h",
4592 ] + MICROKERNEL_BENCHMARK_HDRS,
4593 deps = MICROKERNEL_BENCHMARK_DEPS,
4594)
4595
Marat Dukhan08c4a432019-10-03 09:29:21 -07004596########################### Benchmarks for operators ###########################
4597
4598xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004599 name = "average_pooling_bench",
4600 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004601 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004602 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004603 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604)
4605
4606xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004607 name = "bankers_rounding_bench",
4608 srcs = ["bench/bankers-rounding.cc"],
4609 copts = xnnpack_optional_tflite_copts(),
4610 tags = ["nowin32"],
4611 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4612)
4613
4614xnnpack_benchmark(
4615 name = "ceiling_bench",
4616 srcs = ["bench/ceiling.cc"],
4617 copts = xnnpack_optional_tflite_copts(),
4618 tags = ["nowin32"],
4619 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4620)
4621
4622xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623 name = "channel_shuffle_bench",
4624 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004625 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626)
4627
4628xnnpack_benchmark(
4629 name = "convolution_bench",
4630 srcs = ["bench/convolution.cc"],
4631 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004632 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004633 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004634)
4635
4636xnnpack_benchmark(
4637 name = "deconvolution_bench",
4638 srcs = ["bench/deconvolution.cc"],
4639 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004640 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004641 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004642)
4643
4644xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004645 name = "floor_bench",
4646 srcs = ["bench/floor.cc"],
4647 copts = xnnpack_optional_tflite_copts(),
4648 tags = ["nowin32"],
4649 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4650)
4651
4652xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004653 name = "global_average_pooling_bench",
4654 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004655 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656)
4657
4658xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004659 name = "hardswish_bench",
4660 srcs = ["bench/hardswish.cc"],
4661 copts = xnnpack_optional_tflite_copts(),
4662 tags = ["nowin32"],
4663 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4664)
4665
4666xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 name = "max_pooling_bench",
4668 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004669 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670)
4671
4672xnnpack_benchmark(
4673 name = "sigmoid_bench",
4674 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004675 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004676 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004677 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004678)
4679
4680xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004681 name = "prelu_bench",
4682 srcs = ["bench/prelu.cc"],
4683 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004684 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004685 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004686)
4687
4688xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004689 name = "softmax_bench",
4690 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004691 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004692 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004693 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004694)
4695
Marat Dukhan87727142020-06-24 15:24:10 -07004696xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004697 name = "square_root_bench",
4698 srcs = ["bench/square-root.cc"],
4699 copts = xnnpack_optional_tflite_copts(),
4700 tags = ["nowin32"],
4701 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4702)
4703
4704xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004705 name = "truncation_bench",
4706 srcs = ["bench/truncation.cc"],
4707 deps = OPERATOR_BENCHMARK_DEPS,
4708)
4709
Marat Dukhanc068bb62019-10-04 13:24:39 -07004710############################# End-to-end benchmarks ############################
4711
4712cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004713 name = "fp32_mobilenet_v1",
4714 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004715 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004716 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004717 linkstatic = True,
4718 deps = [
4719 ":XNNPACK",
4720 "@pthreadpool",
4721 ],
4722)
4723
4724cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004725 name = "fp16_mobilenet_v1",
4726 srcs = ["models/fp16-mobilenet-v1.cc"],
4727 hdrs = ["models/models.h"],
4728 copts = xnnpack_std_cxxopts(),
4729 linkstatic = True,
4730 deps = [
4731 ":XNNPACK",
4732 "@FP16",
4733 "@pthreadpool",
4734 ],
4735)
4736
4737cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004738 name = "qs8_mobilenet_v1",
4739 srcs = ["models/qs8-mobilenet-v1.cc"],
4740 hdrs = ["models/models.h"],
4741 copts = xnnpack_std_cxxopts(),
4742 linkstatic = True,
4743 deps = [
4744 ":XNNPACK",
4745 "@pthreadpool",
4746 ],
4747)
4748
4749cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004750 name = "qs8_mobilenet_v2",
4751 srcs = ["models/qs8-mobilenet-v2.cc"],
4752 hdrs = ["models/models.h"],
4753 copts = xnnpack_std_cxxopts(),
4754 linkstatic = True,
4755 deps = [
4756 ":XNNPACK",
4757 "@pthreadpool",
4758 ],
4759)
4760
4761cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004762 name = "fp32_mobilenet_v2",
4763 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004764 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004765 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004766 linkstatic = True,
4767 deps = [
4768 ":XNNPACK",
4769 "@pthreadpool",
4770 ],
4771)
4772
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004773cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004774 name = "fp16_mobilenet_v2",
4775 srcs = ["models/fp16-mobilenet-v2.cc"],
4776 hdrs = ["models/models.h"],
4777 copts = xnnpack_std_cxxopts(),
4778 linkstatic = True,
4779 deps = [
4780 ":XNNPACK",
4781 "@FP16",
4782 "@pthreadpool",
4783 ],
4784)
4785
4786cc_library(
4787 name = "fp32_mobilenet_v3_large",
4788 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004789 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004790 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004791 linkstatic = True,
4792 deps = [
4793 ":XNNPACK",
4794 "@pthreadpool",
4795 ],
4796)
4797
4798cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004799 name = "fp16_mobilenet_v3_large",
4800 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4801 hdrs = ["models/models.h"],
4802 copts = xnnpack_std_cxxopts(),
4803 linkstatic = True,
4804 deps = [
4805 ":XNNPACK",
4806 "@FP16",
4807 "@pthreadpool",
4808 ],
4809)
4810
4811cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004812 name = "fp32_mobilenet_v3_small",
4813 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004814 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004815 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004816 linkstatic = True,
4817 deps = [
4818 ":XNNPACK",
4819 "@pthreadpool",
4820 ],
4821)
4822
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004823cc_library(
4824 name = "fp16_mobilenet_v3_small",
4825 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4826 hdrs = ["models/models.h"],
4827 copts = xnnpack_std_cxxopts(),
4828 linkstatic = True,
4829 deps = [
4830 ":XNNPACK",
4831 "@FP16",
4832 "@pthreadpool",
4833 ],
4834)
4835
Marat Dukhanc068bb62019-10-04 13:24:39 -07004836xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004837 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004838 srcs = [
4839 "bench/f32-dwconv-e2e.cc",
4840 "bench/end2end.h",
4841 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004842 deps = MICROKERNEL_BENCHMARK_DEPS + [
4843 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004844 ":fp32_mobilenet_v1",
4845 ":fp32_mobilenet_v2",
4846 ":fp32_mobilenet_v3_large",
4847 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004848 ],
4849)
4850
4851xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004852 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004853 srcs = [
4854 "bench/f32-gemm-e2e.cc",
4855 "bench/end2end.h",
4856 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004857 deps = MICROKERNEL_BENCHMARK_DEPS + [
4858 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004859 ":fp32_mobilenet_v1",
4860 ":fp32_mobilenet_v2",
4861 ":fp32_mobilenet_v3_large",
4862 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004863 ],
4864)
4865
4866xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004867 name = "end2end_bench",
4868 srcs = ["bench/end2end.cc"],
4869 deps = [
4870 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004871 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004872 ":fp16_mobilenet_v1",
4873 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004874 ":fp16_mobilenet_v3_large",
4875 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004876 ":fp32_mobilenet_v1",
4877 ":fp32_mobilenet_v2",
4878 ":fp32_mobilenet_v3_large",
4879 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004880 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004881 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004882 "@pthreadpool",
4883 ],
4884)
4885
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004886#################### Accuracy evaluation for math functions ####################
4887
4888xnnpack_benchmark(
4889 name = "f32_exp_eval",
4890 srcs = [
4891 "eval/f32-exp.cc",
4892 "src/xnnpack/AlignedAllocator.h",
4893 ] + ACCURACY_EVAL_HDRS,
4894 deps = ACCURACY_EVAL_DEPS,
4895)
4896
Marat Dukhan515c9772019-10-17 18:07:57 -07004897xnnpack_benchmark(
4898 name = "f32_expminus_eval",
4899 srcs = [
4900 "eval/f32-expminus.cc",
4901 "src/xnnpack/AlignedAllocator.h",
4902 ] + ACCURACY_EVAL_HDRS,
4903 deps = ACCURACY_EVAL_DEPS,
4904)
4905
Marat Dukhan98ba4412019-10-23 02:14:28 -07004906xnnpack_benchmark(
4907 name = "f32_extexp_eval",
4908 srcs = [
4909 "eval/f32-extexp.cc",
4910 "src/xnnpack/AlignedAllocator.h",
4911 ] + ACCURACY_EVAL_HDRS,
4912 deps = ACCURACY_EVAL_DEPS,
4913)
4914
Marat Dukhan8853b822020-05-07 12:19:01 -07004915xnnpack_unit_test(
4916 name = "f32_roundne_eval",
4917 srcs = [
4918 "eval/f32-roundne.cc",
4919 "src/xnnpack/AlignedAllocator.h",
4920 "src/xnnpack/math-stubs.h",
4921 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004922 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004923 deps = MICROKERNEL_TEST_DEPS,
4924)
4925
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004926xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004927 name = "f32_roundd_eval",
4928 srcs = [
4929 "eval/f32-roundd.cc",
4930 "src/xnnpack/AlignedAllocator.h",
4931 "src/xnnpack/math-stubs.h",
4932 ] + MICROKERNEL_TEST_HDRS,
4933 automatic = False,
4934 deps = MICROKERNEL_TEST_DEPS,
4935)
4936
4937xnnpack_unit_test(
4938 name = "f32_roundu_eval",
4939 srcs = [
4940 "eval/f32-roundu.cc",
4941 "src/xnnpack/AlignedAllocator.h",
4942 "src/xnnpack/math-stubs.h",
4943 ] + MICROKERNEL_TEST_HDRS,
4944 automatic = False,
4945 deps = MICROKERNEL_TEST_DEPS,
4946)
4947
4948xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004949 name = "f32_roundz_eval",
4950 srcs = [
4951 "eval/f32-roundz.cc",
4952 "src/xnnpack/AlignedAllocator.h",
4953 "src/xnnpack/math-stubs.h",
4954 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004955 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004956 deps = MICROKERNEL_TEST_DEPS,
4957)
4958
Marat Dukhan346a9e52019-11-15 09:06:30 -08004959xnnpack_benchmark(
4960 name = "f32_sigmoid_eval",
4961 srcs = [
4962 "eval/f32-sigmoid.cc",
4963 "src/xnnpack/AlignedAllocator.h",
4964 ] + ACCURACY_EVAL_HDRS,
4965 deps = ACCURACY_EVAL_DEPS,
4966)
4967
Marat Dukhan84000762020-06-29 18:38:43 -07004968xnnpack_benchmark(
4969 name = "f32_sqrt_eval",
4970 srcs = [
4971 "eval/f32-sqrt.cc",
4972 "src/xnnpack/AlignedAllocator.h",
4973 ] + ACCURACY_EVAL_HDRS,
4974 deps = ACCURACY_EVAL_DEPS + [
4975 ":bench_utils",
4976 ],
4977)
4978
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979######################### Unit tests for micro-kernels #########################
4980
4981xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004982 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004984 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985 "test/gemm-microkernel-tester.h",
4986 "src/xnnpack/AlignedAllocator.h",
4987 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004988 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989)
4990
4991xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004992 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004993 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004994 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004995 "test/spmm-microkernel-tester.h",
4996 "src/xnnpack/AlignedAllocator.h",
4997 ] + MICROKERNEL_TEST_HDRS,
4998 deps = MICROKERNEL_TEST_DEPS,
4999)
5000
5001xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002 name = "f32_argmaxpool_test",
5003 srcs = [
5004 "test/f32-argmaxpool.cc",
5005 "test/argmaxpool-microkernel-tester.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + MICROKERNEL_TEST_HDRS,
5008 deps = MICROKERNEL_TEST_DEPS,
5009)
5010
5011xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005012 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005013 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005014 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 "test/avgpool-microkernel-tester.h",
5016 "src/xnnpack/AlignedAllocator.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005022 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005023 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005024 "test/f32-ibilinear.cc",
5025 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005026 "src/xnnpack/AlignedAllocator.h",
5027 ] + MICROKERNEL_TEST_HDRS,
5028 deps = MICROKERNEL_TEST_DEPS,
5029)
5030
5031xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005032 name = "f32_ibilinear_chw_test",
5033 srcs = [
5034 "test/f32-ibilinear-chw.cc",
5035 "test/ibilinear-chw-microkernel-tester.h",
5036 "src/xnnpack/AlignedAllocator.h",
5037 ] + MICROKERNEL_TEST_HDRS,
5038 deps = MICROKERNEL_TEST_DEPS,
5039)
5040
5041xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005042 name = "f16_clamp_test",
5043 srcs = [
5044 "test/f16-clamp.cc",
5045 "test/clamp-microkernel-tester.h",
5046 ] + MICROKERNEL_TEST_HDRS,
5047 deps = MICROKERNEL_TEST_DEPS,
5048)
5049
5050xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051 name = "f32_clamp_test",
5052 srcs = [
5053 "test/f32-clamp.cc",
5054 "test/clamp-microkernel-tester.h",
5055 ] + MICROKERNEL_TEST_HDRS,
5056 deps = MICROKERNEL_TEST_DEPS,
5057)
5058
5059xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005060 name = "f32_igemm_test",
5061 srcs = [
5062 "test/f32-igemm.cc",
5063 "test/gemm-microkernel-tester.h",
5064 "src/xnnpack/AlignedAllocator.h",
5065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005066 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005067)
5068
5069xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005070 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005072 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073 "test/gemm-microkernel-tester.h",
5074 "src/xnnpack/AlignedAllocator.h",
5075 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005076 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005077)
5078
5079xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005080 name = "f32_igemm_minmax_test",
5081 srcs = [
5082 "test/f32-igemm-minmax.cc",
5083 "test/gemm-microkernel-tester.h",
5084 "src/xnnpack/AlignedAllocator.h",
5085 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005086 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005087)
5088
5089xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005090 name = "f16_igemm_minmax_test",
5091 srcs = [
5092 "test/f16-igemm-minmax.cc",
5093 "test/gemm-microkernel-tester.h",
5094 "src/xnnpack/AlignedAllocator.h",
5095 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005096 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005097)
5098
5099xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005100 name = "f32_conv_hwc_test",
5101 srcs = [
5102 "test/f32-conv-hwc.cc",
5103 "test/conv-hwc-microkernel-tester.h",
5104 "src/xnnpack/AlignedAllocator.h",
5105 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005106 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005107)
5108
5109xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005110 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005111 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005112 "test/f32-conv-hwc2chw.cc",
5113 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 "src/xnnpack/AlignedAllocator.h",
5115 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005116 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005117)
5118
5119xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005120 name = "f32_dwconv_test",
5121 srcs = [
5122 "test/f32-dwconv.cc",
5123 "test/dwconv-microkernel-tester.h",
5124 "src/xnnpack/AlignedAllocator.h",
5125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005126 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005127)
5128
5129xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005130 name = "f16_dwconv_minmax_test",
5131 srcs = [
5132 "test/f16-dwconv-minmax.cc",
5133 "test/dwconv-microkernel-tester.h",
5134 "src/xnnpack/AlignedAllocator.h",
5135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005136 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005137)
5138
5139xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005140 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005141 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005142 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143 "test/dwconv-microkernel-tester.h",
5144 "src/xnnpack/AlignedAllocator.h",
5145 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005146 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005147)
5148
5149xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005150 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005152 "test/f32-dwconv-chw.cc",
5153 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154 "src/xnnpack/AlignedAllocator.h",
5155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005156 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157)
5158
5159xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005160 name = "f16_gavgpool_minmax_test",
5161 srcs = [
5162 "test/f16-gavgpool-minmax.cc",
5163 "test/gavgpool-microkernel-tester.h",
5164 "src/xnnpack/AlignedAllocator.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005170 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005172 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173 "test/gavgpool-microkernel-tester.h",
5174 "src/xnnpack/AlignedAllocator.h",
5175 ] + MICROKERNEL_TEST_HDRS,
5176 deps = MICROKERNEL_TEST_DEPS,
5177)
5178
5179xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005180 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005182 "test/f32-gavgpool-cw.cc",
5183 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005184 "src/xnnpack/AlignedAllocator.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005190 name = "f32_gemm_test",
5191 srcs = [
5192 "test/f32-gemm.cc",
5193 "test/gemm-microkernel-tester.h",
5194 "src/xnnpack/AlignedAllocator.h",
5195 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005196 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005197)
5198
5199xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005200 name = "f32_gemm_relu_test",
5201 srcs = [
5202 "test/f32-gemm-relu.cc",
5203 "test/gemm-microkernel-tester.h",
5204 "src/xnnpack/AlignedAllocator.h",
5205 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005206 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005207)
5208
5209xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005210 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005211 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005212 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213 "test/gemm-microkernel-tester.h",
5214 "src/xnnpack/AlignedAllocator.h",
5215 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005216 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217)
5218
5219xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005220 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005221 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005222 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 "test/gemm-microkernel-tester.h",
5224 "src/xnnpack/AlignedAllocator.h",
5225 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005226 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227)
5228
5229xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005230 name = "f16_hswish_test",
5231 srcs = [
5232 "test/f16-hswish.cc",
5233 "test/hswish-microkernel-tester.h",
5234 ] + MICROKERNEL_TEST_HDRS,
5235 deps = MICROKERNEL_TEST_DEPS,
5236)
5237
5238xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 name = "f32_hswish_test",
5240 srcs = [
5241 "test/f32-hswish.cc",
5242 "test/hswish-microkernel-tester.h",
5243 ] + MICROKERNEL_TEST_HDRS,
5244 deps = MICROKERNEL_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005248 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005250 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251 "test/maxpool-microkernel-tester.h",
5252 ] + MICROKERNEL_TEST_HDRS,
5253 deps = MICROKERNEL_TEST_DEPS,
5254)
5255
5256xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005257 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005259 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 "test/avgpool-microkernel-tester.h",
5261 "src/xnnpack/AlignedAllocator.h",
5262 ] + MICROKERNEL_TEST_HDRS,
5263 deps = MICROKERNEL_TEST_DEPS,
5264)
5265
5266xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005267 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005268 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005269 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005270 "test/gemm-microkernel-tester.h",
5271 "src/xnnpack/AlignedAllocator.h",
5272 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005273 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274)
5275
5276xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005277 name = "f16_prelu_test",
5278 srcs = [
5279 "test/f16-prelu.cc",
5280 "test/prelu-microkernel-tester.h",
5281 "src/xnnpack/AlignedAllocator.h",
5282 ] + MICROKERNEL_TEST_HDRS,
5283 deps = MICROKERNEL_TEST_DEPS,
5284)
5285
5286xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 name = "f32_prelu_test",
5288 srcs = [
5289 "test/f32-prelu.cc",
5290 "test/prelu-microkernel-tester.h",
5291 "src/xnnpack/AlignedAllocator.h",
5292 ] + MICROKERNEL_TEST_HDRS,
5293 deps = MICROKERNEL_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005297 name = "f32_raddexpminusmax_test",
5298 srcs = [
5299 "test/f32-raddexpminusmax.cc",
5300 "test/raddexpminusmax-microkernel-tester.h",
5301 ] + MICROKERNEL_TEST_HDRS,
5302 deps = MICROKERNEL_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005306 name = "f32_raddextexp_test",
5307 srcs = [
5308 "test/f32-raddextexp.cc",
5309 "test/raddextexp-microkernel-tester.h",
5310 ] + MICROKERNEL_TEST_HDRS,
5311 deps = MICROKERNEL_TEST_DEPS,
5312)
5313
5314xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005315 name = "f32_raddstoreexpminusmax_test",
5316 srcs = [
5317 "test/f32-raddstoreexpminusmax.cc",
5318 "test/raddstoreexpminusmax-microkernel-tester.h",
5319 ] + MICROKERNEL_TEST_HDRS,
5320 deps = MICROKERNEL_TEST_DEPS,
5321)
5322
5323xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005324 name = "f32_relu_test",
5325 srcs = [
5326 "test/f32-relu.cc",
5327 "test/vunary-microkernel-tester.h",
5328 ] + MICROKERNEL_TEST_HDRS,
5329 deps = MICROKERNEL_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 name = "f32_rmax_test",
5334 srcs = [
5335 "test/f32-rmax.cc",
5336 "test/rmax-microkernel-tester.h",
5337 ] + MICROKERNEL_TEST_HDRS,
5338 deps = MICROKERNEL_TEST_DEPS,
5339)
5340
5341xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005342 name = "f32_sigmoid_test",
5343 srcs = [
5344 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005345 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005346 ] + MICROKERNEL_TEST_HDRS,
5347 deps = MICROKERNEL_TEST_DEPS,
5348)
5349
5350xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005351 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005353 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354 "test/spmm-microkernel-tester.h",
5355 "src/xnnpack/AlignedAllocator.h",
5356 ] + MICROKERNEL_TEST_HDRS,
5357 deps = MICROKERNEL_TEST_DEPS,
5358)
5359
5360xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005361 name = "f16_vadd_minmax_test",
5362 srcs = [
5363 "test/f16-vadd-minmax.cc",
5364 "test/vbinary-microkernel-tester.h",
5365 ] + MICROKERNEL_TEST_HDRS,
5366 deps = MICROKERNEL_TEST_DEPS,
5367)
5368
5369xnnpack_unit_test(
5370 name = "f16_vaddc_minmax_test",
5371 srcs = [
5372 "test/f16-vaddc-minmax.cc",
5373 "test/vbinaryc-microkernel-tester.h",
5374 ] + MICROKERNEL_TEST_HDRS,
5375 deps = MICROKERNEL_TEST_DEPS,
5376)
5377
5378xnnpack_unit_test(
5379 name = "f16_vdiv_minmax_test",
5380 srcs = [
5381 "test/f16-vdiv-minmax.cc",
5382 "test/vbinary-microkernel-tester.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
5388 name = "f16_vdivc_minmax_test",
5389 srcs = [
5390 "test/f16-vdivc-minmax.cc",
5391 "test/vbinaryc-microkernel-tester.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 deps = MICROKERNEL_TEST_DEPS,
5394)
5395
5396xnnpack_unit_test(
5397 name = "f16_vrdivc_minmax_test",
5398 srcs = [
5399 "test/f16-vrdivc-minmax.cc",
5400 "test/vbinaryc-microkernel-tester.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
5406 name = "f16_vmax_test",
5407 srcs = [
5408 "test/f16-vmax.cc",
5409 "test/vbinary-microkernel-tester.h",
5410 ] + MICROKERNEL_TEST_HDRS,
5411 deps = MICROKERNEL_TEST_DEPS,
5412)
5413
5414xnnpack_unit_test(
5415 name = "f16_vmaxc_test",
5416 srcs = [
5417 "test/f16-vmaxc.cc",
5418 "test/vbinaryc-microkernel-tester.h",
5419 ] + MICROKERNEL_TEST_HDRS,
5420 deps = MICROKERNEL_TEST_DEPS,
5421)
5422
5423xnnpack_unit_test(
5424 name = "f16_vmin_test",
5425 srcs = [
5426 "test/f16-vmin.cc",
5427 "test/vbinary-microkernel-tester.h",
5428 ] + MICROKERNEL_TEST_HDRS,
5429 deps = MICROKERNEL_TEST_DEPS,
5430)
5431
5432xnnpack_unit_test(
5433 name = "f16_vminc_test",
5434 srcs = [
5435 "test/f16-vminc.cc",
5436 "test/vbinaryc-microkernel-tester.h",
5437 ] + MICROKERNEL_TEST_HDRS,
5438 deps = MICROKERNEL_TEST_DEPS,
5439)
5440
5441xnnpack_unit_test(
5442 name = "f16_vmul_minmax_test",
5443 srcs = [
5444 "test/f16-vmul-minmax.cc",
5445 "test/vbinary-microkernel-tester.h",
5446 ] + MICROKERNEL_TEST_HDRS,
5447 deps = MICROKERNEL_TEST_DEPS,
5448)
5449
5450xnnpack_unit_test(
5451 name = "f16_vmulc_minmax_test",
5452 srcs = [
5453 "test/f16-vmulc-minmax.cc",
5454 "test/vbinaryc-microkernel-tester.h",
5455 ] + MICROKERNEL_TEST_HDRS,
5456 deps = MICROKERNEL_TEST_DEPS,
5457)
5458
5459xnnpack_unit_test(
5460 name = "f16_vsub_minmax_test",
5461 srcs = [
5462 "test/f16-vsub-minmax.cc",
5463 "test/vbinary-microkernel-tester.h",
5464 ] + MICROKERNEL_TEST_HDRS,
5465 deps = MICROKERNEL_TEST_DEPS,
5466)
5467
5468xnnpack_unit_test(
5469 name = "f16_vsubc_minmax_test",
5470 srcs = [
5471 "test/f16-vsubc-minmax.cc",
5472 "test/vbinaryc-microkernel-tester.h",
5473 ] + MICROKERNEL_TEST_HDRS,
5474 deps = MICROKERNEL_TEST_DEPS,
5475)
5476
5477xnnpack_unit_test(
5478 name = "f16_vrsubc_minmax_test",
5479 srcs = [
5480 "test/f16-vrsubc-minmax.cc",
5481 "test/vbinaryc-microkernel-tester.h",
5482 ] + MICROKERNEL_TEST_HDRS,
5483 deps = MICROKERNEL_TEST_DEPS,
5484)
5485
5486xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005487 name = "f32_vabs_test",
5488 srcs = [
5489 "test/f32-vabs.cc",
5490 "test/vunary-microkernel-tester.h",
5491 ] + MICROKERNEL_TEST_HDRS,
5492 deps = MICROKERNEL_TEST_DEPS,
5493)
5494
5495xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005496 name = "f32_vadd_test",
5497 srcs = [
5498 "test/f32-vadd.cc",
5499 "test/vbinary-microkernel-tester.h",
5500 ] + MICROKERNEL_TEST_HDRS,
5501 deps = MICROKERNEL_TEST_DEPS,
5502)
5503
5504xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005505 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005506 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005507 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005508 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005509 ] + MICROKERNEL_TEST_HDRS,
5510 deps = MICROKERNEL_TEST_DEPS,
5511)
5512
5513xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005514 name = "f32_vadd_relu_test",
5515 srcs = [
5516 "test/f32-vadd-relu.cc",
5517 "test/vbinary-microkernel-tester.h",
5518 ] + MICROKERNEL_TEST_HDRS,
5519 deps = MICROKERNEL_TEST_DEPS,
5520)
5521
5522xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005523 name = "f32_vaddc_test",
5524 srcs = [
5525 "test/f32-vaddc.cc",
5526 "test/vbinaryc-microkernel-tester.h",
5527 ] + MICROKERNEL_TEST_HDRS,
5528 deps = MICROKERNEL_TEST_DEPS,
5529)
5530
5531xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005532 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005533 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005534 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005535 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005536 ] + MICROKERNEL_TEST_HDRS,
5537 deps = MICROKERNEL_TEST_DEPS,
5538)
5539
5540xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005541 name = "f32_vaddc_relu_test",
5542 srcs = [
5543 "test/f32-vaddc-relu.cc",
5544 "test/vbinaryc-microkernel-tester.h",
5545 ] + MICROKERNEL_TEST_HDRS,
5546 deps = MICROKERNEL_TEST_DEPS,
5547)
5548
5549xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005550 name = "f32_vdiv_test",
5551 srcs = [
5552 "test/f32-vdiv.cc",
5553 "test/vbinary-microkernel-tester.h",
5554 ] + MICROKERNEL_TEST_HDRS,
5555 deps = MICROKERNEL_TEST_DEPS,
5556)
5557
5558xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005559 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005560 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005561 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005562 "test/vbinary-microkernel-tester.h",
5563 ] + MICROKERNEL_TEST_HDRS,
5564 deps = MICROKERNEL_TEST_DEPS,
5565)
5566
5567xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005568 name = "f32_vdiv_relu_test",
5569 srcs = [
5570 "test/f32-vdiv-relu.cc",
5571 "test/vbinary-microkernel-tester.h",
5572 ] + MICROKERNEL_TEST_HDRS,
5573 deps = MICROKERNEL_TEST_DEPS,
5574)
5575
5576xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005577 name = "f32_vdivc_test",
5578 srcs = [
5579 "test/f32-vdivc.cc",
5580 "test/vbinaryc-microkernel-tester.h",
5581 ] + MICROKERNEL_TEST_HDRS,
5582 deps = MICROKERNEL_TEST_DEPS,
5583)
5584
5585xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005586 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005587 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005588 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005589 "test/vbinaryc-microkernel-tester.h",
5590 ] + MICROKERNEL_TEST_HDRS,
5591 deps = MICROKERNEL_TEST_DEPS,
5592)
5593
5594xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005595 name = "f32_vdivc_relu_test",
5596 srcs = [
5597 "test/f32-vdivc-relu.cc",
5598 "test/vbinaryc-microkernel-tester.h",
5599 ] + MICROKERNEL_TEST_HDRS,
5600 deps = MICROKERNEL_TEST_DEPS,
5601)
5602
5603xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005604 name = "f32_vrdivc_test",
5605 srcs = [
5606 "test/f32-vrdivc.cc",
5607 "test/vbinaryc-microkernel-tester.h",
5608 ] + MICROKERNEL_TEST_HDRS,
5609 deps = MICROKERNEL_TEST_DEPS,
5610)
5611
5612xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005613 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005614 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005615 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005616 "test/vbinaryc-microkernel-tester.h",
5617 ] + MICROKERNEL_TEST_HDRS,
5618 deps = MICROKERNEL_TEST_DEPS,
5619)
5620
5621xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005622 name = "f32_vrdivc_relu_test",
5623 srcs = [
5624 "test/f32-vrdivc-relu.cc",
5625 "test/vbinaryc-microkernel-tester.h",
5626 ] + MICROKERNEL_TEST_HDRS,
5627 deps = MICROKERNEL_TEST_DEPS,
5628)
5629
5630xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005631 name = "f32_vmax_test",
5632 srcs = [
5633 "test/f32-vmax.cc",
5634 "test/vbinary-microkernel-tester.h",
5635 ] + MICROKERNEL_TEST_HDRS,
5636 deps = MICROKERNEL_TEST_DEPS,
5637)
5638
5639xnnpack_unit_test(
5640 name = "f32_vmaxc_test",
5641 srcs = [
5642 "test/f32-vmaxc.cc",
5643 "test/vbinaryc-microkernel-tester.h",
5644 ] + MICROKERNEL_TEST_HDRS,
5645 deps = MICROKERNEL_TEST_DEPS,
5646)
5647
5648xnnpack_unit_test(
5649 name = "f32_vmin_test",
5650 srcs = [
5651 "test/f32-vmin.cc",
5652 "test/vbinary-microkernel-tester.h",
5653 ] + MICROKERNEL_TEST_HDRS,
5654 deps = MICROKERNEL_TEST_DEPS,
5655)
5656
5657xnnpack_unit_test(
5658 name = "f32_vminc_test",
5659 srcs = [
5660 "test/f32-vminc.cc",
5661 "test/vbinaryc-microkernel-tester.h",
5662 ] + MICROKERNEL_TEST_HDRS,
5663 deps = MICROKERNEL_TEST_DEPS,
5664)
5665
5666xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005667 name = "f32_vmul_test",
5668 srcs = [
5669 "test/f32-vmul.cc",
5670 "test/vbinary-microkernel-tester.h",
5671 ] + MICROKERNEL_TEST_HDRS,
5672 deps = MICROKERNEL_TEST_DEPS,
5673)
5674
5675xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005676 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005678 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005679 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005680 ] + MICROKERNEL_TEST_HDRS,
5681 deps = MICROKERNEL_TEST_DEPS,
5682)
5683
5684xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005685 name = "f32_vmul_relu_test",
5686 srcs = [
5687 "test/f32-vmul-relu.cc",
5688 "test/vbinary-microkernel-tester.h",
5689 ] + MICROKERNEL_TEST_HDRS,
5690 deps = MICROKERNEL_TEST_DEPS,
5691)
5692
5693xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005694 name = "f32_vmulc_test",
5695 srcs = [
5696 "test/f32-vmulc.cc",
5697 "test/vbinaryc-microkernel-tester.h",
5698 ] + MICROKERNEL_TEST_HDRS,
5699 deps = MICROKERNEL_TEST_DEPS,
5700)
5701
5702xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005703 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005704 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005705 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005706 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005707 ] + MICROKERNEL_TEST_HDRS,
5708 deps = MICROKERNEL_TEST_DEPS,
5709)
5710
5711xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005712 name = "f32_vmulc_relu_test",
5713 srcs = [
5714 "test/f32-vmulc-relu.cc",
5715 "test/vbinaryc-microkernel-tester.h",
5716 ] + MICROKERNEL_TEST_HDRS,
5717 deps = MICROKERNEL_TEST_DEPS,
5718)
5719
5720xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005721 name = "f16_vmulcaddc_minmax_test",
5722 srcs = [
5723 "test/f16-vmulcaddc-minmax.cc",
5724 "test/vmulcaddc-microkernel-tester.h",
5725 "src/xnnpack/AlignedAllocator.h",
5726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005727 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005728)
5729
5730xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005731 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005732 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005733 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734 "test/vmulcaddc-microkernel-tester.h",
5735 "src/xnnpack/AlignedAllocator.h",
5736 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005737 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005738)
5739
5740xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005741 name = "f32_vlrelu_test",
5742 srcs = [
5743 "test/f32-vlrelu.cc",
5744 "test/vunary-microkernel-tester.h",
5745 ] + MICROKERNEL_TEST_HDRS,
5746 deps = MICROKERNEL_TEST_DEPS,
5747)
5748
5749xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005750 name = "f32_vneg_test",
5751 srcs = [
5752 "test/f32-vneg.cc",
5753 "test/vunary-microkernel-tester.h",
5754 ] + MICROKERNEL_TEST_HDRS,
5755 deps = MICROKERNEL_TEST_DEPS,
5756)
5757
5758xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005759 name = "f32_vrndne_test",
5760 srcs = [
5761 "test/f32-vrndne.cc",
5762 "test/vunary-microkernel-tester.h",
5763 ] + MICROKERNEL_TEST_HDRS,
5764 deps = MICROKERNEL_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
5768 name = "f32_vrndz_test",
5769 srcs = [
5770 "test/f32-vrndz.cc",
5771 "test/vunary-microkernel-tester.h",
5772 ] + MICROKERNEL_TEST_HDRS,
5773 deps = MICROKERNEL_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
5777 name = "f32_vrndu_test",
5778 srcs = [
5779 "test/f32-vrndu.cc",
5780 "test/vunary-microkernel-tester.h",
5781 ] + MICROKERNEL_TEST_HDRS,
5782 deps = MICROKERNEL_TEST_DEPS,
5783)
5784
5785xnnpack_unit_test(
5786 name = "f32_vrndd_test",
5787 srcs = [
5788 "test/f32-vrndd.cc",
5789 "test/vunary-microkernel-tester.h",
5790 ] + MICROKERNEL_TEST_HDRS,
5791 deps = MICROKERNEL_TEST_DEPS,
5792)
5793
5794xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005795 name = "f32_vscale_test",
5796 srcs = [
5797 "test/f32-vscale.cc",
5798 "test/vscale-microkernel-tester.h",
5799 ] + MICROKERNEL_TEST_HDRS,
5800 deps = MICROKERNEL_TEST_DEPS,
5801)
5802
5803xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005804 name = "f32_vscaleexpminusmax_test",
5805 srcs = [
5806 "test/f32-vscaleexpminusmax.cc",
5807 "test/vscaleexpminusmax-microkernel-tester.h",
5808 ] + MICROKERNEL_TEST_HDRS,
5809 deps = MICROKERNEL_TEST_DEPS,
5810)
5811
5812xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005813 name = "f32_vscaleextexp_test",
5814 srcs = [
5815 "test/f32-vscaleextexp.cc",
5816 "test/vscaleextexp-microkernel-tester.h",
5817 ] + MICROKERNEL_TEST_HDRS,
5818 deps = MICROKERNEL_TEST_DEPS,
5819)
5820
5821xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005822 name = "f32_vsqr_test",
5823 srcs = [
5824 "test/f32-vsqr.cc",
5825 "test/vunary-microkernel-tester.h",
5826 ] + MICROKERNEL_TEST_HDRS,
5827 deps = MICROKERNEL_TEST_DEPS,
5828)
5829
5830xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005831 name = "f32_vsqrdiff_test",
5832 srcs = [
5833 "test/f32-vsqrdiff.cc",
5834 "test/vbinary-microkernel-tester.h",
5835 ] + MICROKERNEL_TEST_HDRS,
5836 deps = MICROKERNEL_TEST_DEPS,
5837)
5838
5839xnnpack_unit_test(
5840 name = "f32_vsqrdiffc_test",
5841 srcs = [
5842 "test/f32-vsqrdiffc.cc",
5843 "test/vbinaryc-microkernel-tester.h",
5844 ] + MICROKERNEL_TEST_HDRS,
5845 deps = MICROKERNEL_TEST_DEPS,
5846)
5847
5848xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005849 name = "f32_vsqrt_test",
5850 srcs = [
5851 "test/f32-vsqrt.cc",
5852 "test/vunary-microkernel-tester.h",
5853 ] + MICROKERNEL_TEST_HDRS,
5854 deps = MICROKERNEL_TEST_DEPS,
5855)
5856
5857xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005858 name = "f32_vsub_test",
5859 srcs = [
5860 "test/f32-vsub.cc",
5861 "test/vbinary-microkernel-tester.h",
5862 ] + MICROKERNEL_TEST_HDRS,
5863 deps = MICROKERNEL_TEST_DEPS,
5864)
5865
5866xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005867 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005868 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005869 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005870 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005871 ] + MICROKERNEL_TEST_HDRS,
5872 deps = MICROKERNEL_TEST_DEPS,
5873)
5874
5875xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005876 name = "f32_vsub_relu_test",
5877 srcs = [
5878 "test/f32-vsub-relu.cc",
5879 "test/vbinary-microkernel-tester.h",
5880 ] + MICROKERNEL_TEST_HDRS,
5881 deps = MICROKERNEL_TEST_DEPS,
5882)
5883
5884xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005885 name = "f32_vsubc_test",
5886 srcs = [
5887 "test/f32-vsubc.cc",
5888 "test/vbinaryc-microkernel-tester.h",
5889 ] + MICROKERNEL_TEST_HDRS,
5890 deps = MICROKERNEL_TEST_DEPS,
5891)
5892
5893xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005894 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005895 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005896 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005897 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005898 ] + MICROKERNEL_TEST_HDRS,
5899 deps = MICROKERNEL_TEST_DEPS,
5900)
5901
5902xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005903 name = "f32_vsubc_relu_test",
5904 srcs = [
5905 "test/f32-vsubc-relu.cc",
5906 "test/vbinaryc-microkernel-tester.h",
5907 ] + MICROKERNEL_TEST_HDRS,
5908 deps = MICROKERNEL_TEST_DEPS,
5909)
5910
5911xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005912 name = "f32_vrsubc_test",
5913 srcs = [
5914 "test/f32-vrsubc.cc",
5915 "test/vbinaryc-microkernel-tester.h",
5916 ] + MICROKERNEL_TEST_HDRS,
5917 deps = MICROKERNEL_TEST_DEPS,
5918)
5919
5920xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005921 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005922 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005923 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005924 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005925 ] + MICROKERNEL_TEST_HDRS,
5926 deps = MICROKERNEL_TEST_DEPS,
5927)
5928
5929xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005930 name = "f32_vrsubc_relu_test",
5931 srcs = [
5932 "test/f32-vrsubc-relu.cc",
5933 "test/vbinaryc-microkernel-tester.h",
5934 ] + MICROKERNEL_TEST_HDRS,
5935 deps = MICROKERNEL_TEST_DEPS,
5936)
5937
5938xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005939 name = "qs8_dwconv_minmax_test",
5940 srcs = [
5941 "test/qs8-dwconv-minmax.cc",
5942 "test/dwconv-microkernel-tester.h",
5943 "src/xnnpack/AlignedAllocator.h",
5944 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5946)
5947
5948xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005949 name = "qs8_gavgpool_minmax_test",
5950 srcs = [
5951 "test/qs8-gavgpool-minmax.cc",
5952 "test/gavgpool-microkernel-tester.h",
5953 "src/xnnpack/AlignedAllocator.h",
5954 ] + MICROKERNEL_TEST_HDRS,
5955 deps = MICROKERNEL_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005959 name = "qs8_gemm_minmax_test",
5960 srcs = [
5961 "test/qs8-gemm-minmax.cc",
5962 "test/gemm-microkernel-tester.h",
5963 "src/xnnpack/AlignedAllocator.h",
5964 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5965 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5966)
5967
5968xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005969 name = "qs8_igemm_minmax_test",
5970 srcs = [
5971 "test/qs8-igemm-minmax.cc",
5972 "test/gemm-microkernel-tester.h",
5973 "src/xnnpack/AlignedAllocator.h",
5974 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5975 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5976)
5977
5978xnnpack_unit_test(
5979 name = "qs8_requantization_test",
5980 srcs = [
5981 "src/xnnpack/requantization-stubs.h",
5982 "test/qs8-requantization.cc",
5983 "test/requantization-tester.h",
5984 ] + MICROKERNEL_TEST_HDRS,
5985 deps = MICROKERNEL_TEST_DEPS,
5986)
5987
5988xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005989 name = "qs8_vadd_minmax_test",
5990 srcs = [
5991 "test/qs8-vadd-minmax.cc",
5992 "test/vadd-microkernel-tester.h",
5993 ] + MICROKERNEL_TEST_HDRS,
5994 deps = MICROKERNEL_TEST_DEPS,
5995)
5996
5997xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005998 name = "qs8_vaddc_minmax_test",
5999 srcs = [
6000 "test/qs8-vaddc-minmax.cc",
6001 "test/vaddc-microkernel-tester.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006007 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006008 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006009 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006010 "test/avgpool-microkernel-tester.h",
6011 "src/xnnpack/AlignedAllocator.h",
6012 ] + MICROKERNEL_TEST_HDRS,
6013 deps = MICROKERNEL_TEST_DEPS,
6014)
6015
6016xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006017 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006019 "test/qu8-dwconv-minmax.cc",
6020 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006021 "src/xnnpack/AlignedAllocator.h",
6022 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006023 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006024)
6025
6026xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006027 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006028 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006029 "test/qu8-igemm-minmax.cc",
6030 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006031 "src/xnnpack/AlignedAllocator.h",
6032 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006033 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034)
6035
6036xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006037 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006039 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006040 "test/gavgpool-microkernel-tester.h",
6041 "src/xnnpack/AlignedAllocator.h",
6042 ] + MICROKERNEL_TEST_HDRS,
6043 deps = MICROKERNEL_TEST_DEPS,
6044)
6045
6046xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006047 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006048 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006049 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006050 "test/gemm-microkernel-tester.h",
6051 "src/xnnpack/AlignedAllocator.h",
6052 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006053 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054)
6055
6056xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006057 name = "qu8_requantization_test",
6058 srcs = [
6059 "src/xnnpack/requantization-stubs.h",
6060 "test/qu8-requantization.cc",
6061 "test/requantization-tester.h",
6062 ] + MICROKERNEL_TEST_HDRS,
6063 deps = MICROKERNEL_TEST_DEPS,
6064)
6065
6066xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006067 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006069 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006070 "test/vadd-microkernel-tester.h",
6071 ] + MICROKERNEL_TEST_HDRS,
6072 deps = MICROKERNEL_TEST_DEPS,
6073)
6074
6075xnnpack_unit_test(
6076 name = "u8_clamp_test",
6077 srcs = [
6078 "test/u8-clamp.cc",
6079 "test/clamp-microkernel-tester.h",
6080 ] + MICROKERNEL_TEST_HDRS,
6081 deps = MICROKERNEL_TEST_DEPS,
6082)
6083
6084xnnpack_unit_test(
6085 name = "u8_lut32norm_test",
6086 srcs = [
6087 "test/u8-lut32norm.cc",
6088 "test/lut-norm-microkernel-tester.h",
6089 ] + MICROKERNEL_TEST_HDRS,
6090 deps = MICROKERNEL_TEST_DEPS,
6091)
6092
6093xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006094 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006096 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006097 "test/maxpool-microkernel-tester.h",
6098 ] + MICROKERNEL_TEST_HDRS,
6099 deps = MICROKERNEL_TEST_DEPS,
6100)
6101
6102xnnpack_unit_test(
6103 name = "u8_rmax_test",
6104 srcs = [
6105 "test/u8-rmax.cc",
6106 "test/rmax-microkernel-tester.h",
6107 ] + MICROKERNEL_TEST_HDRS,
6108 deps = MICROKERNEL_TEST_DEPS,
6109)
6110
6111xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006112 name = "x32_fill_test",
6113 srcs = [
6114 "test/x32-fill.cc",
6115 "test/fill-microkernel-tester.h",
6116 ] + MICROKERNEL_TEST_HDRS,
6117 deps = MICROKERNEL_TEST_DEPS,
6118)
6119
6120xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006121 name = "x32_packx_test",
6122 srcs = [
6123 "test/x32-packx.cc",
6124 "test/pack-microkernel-tester.h",
6125 "src/xnnpack/AlignedAllocator.h",
6126 ] + MICROKERNEL_TEST_HDRS,
6127 deps = MICROKERNEL_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
6131 name = "x32_pad_test",
6132 srcs = [
6133 "test/x32-pad.cc",
6134 "test/pad-microkernel-tester.h",
6135 ] + MICROKERNEL_TEST_HDRS,
6136 deps = MICROKERNEL_TEST_DEPS,
6137)
6138
6139xnnpack_unit_test(
6140 name = "x32_unpool_test",
6141 srcs = [
6142 "test/x32-unpool.cc",
6143 "test/unpool-microkernel-tester.h",
6144 ] + MICROKERNEL_TEST_HDRS,
6145 deps = MICROKERNEL_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
6149 name = "x32_zip_test",
6150 srcs = [
6151 "test/x32-zip.cc",
6152 "test/zip-microkernel-tester.h",
6153 ] + MICROKERNEL_TEST_HDRS,
6154 deps = MICROKERNEL_TEST_DEPS,
6155)
6156
6157xnnpack_unit_test(
6158 name = "x8_lut_test",
6159 srcs = [
6160 "test/x8-lut.cc",
6161 "test/lut-microkernel-tester.h",
6162 ] + MICROKERNEL_TEST_HDRS,
6163 deps = MICROKERNEL_TEST_DEPS,
6164)
6165
6166xnnpack_unit_test(
6167 name = "x8_zip_test",
6168 srcs = [
6169 "test/x8-zip.cc",
6170 "test/zip-microkernel-tester.h",
6171 ] + MICROKERNEL_TEST_HDRS,
6172 deps = MICROKERNEL_TEST_DEPS,
6173)
6174
Marat Dukhan20c3b922020-03-10 03:45:06 -07006175########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006176
6177xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006178 name = "operator_size_test",
6179 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006180 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006181)
6182
Marat Dukhan20c3b922020-03-10 03:45:06 -07006183xnnpack_binary(
6184 name = "subgraph_size_test",
6185 srcs = ["test/subgraph-size.c"],
6186 deps = [":XNNPACK"],
6187)
6188
6189########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006190
6191xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006192 name = "abs_nc_test",
6193 srcs = [
6194 "test/abs-nc.cc",
6195 "test/abs-operator-tester.h",
6196 ],
6197 deps = OPERATOR_TEST_DEPS,
6198)
6199
6200xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006201 name = "add_nd_test",
6202 srcs = [
6203 "test/add-nd.cc",
6204 "test/binary-elementwise-operator-tester.h",
6205 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006206 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006207)
6208
6209xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006210 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006211 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006212 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 "test/argmax-pooling-operator-tester.h",
6214 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006215 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216)
6217
6218xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006219 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006220 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006221 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222 "test/average-pooling-operator-tester.h",
6223 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006224 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225)
6226
6227xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006228 name = "bankers_rounding_nc_test",
6229 srcs = [
6230 "test/bankers-rounding-nc.cc",
6231 "test/bankers-rounding-operator-tester.h",
6232 ],
6233 deps = OPERATOR_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
6237 name = "ceiling_nc_test",
6238 srcs = [
6239 "test/ceiling-nc.cc",
6240 "test/ceiling-operator-tester.h",
6241 ],
6242 deps = OPERATOR_TEST_DEPS,
6243)
6244
6245xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006246 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006247 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006248 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006249 "test/channel-shuffle-operator-tester.h",
6250 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006251 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006252)
6253
6254xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006255 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006257 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 "test/clamp-operator-tester.h",
6259 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006260 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261)
6262
6263xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006264 name = "constant_pad_nd_test",
6265 srcs = [
6266 "test/constant-pad-nd.cc",
6267 "test/constant-pad-operator-tester.h",
6268 ],
6269 deps = OPERATOR_TEST_DEPS,
6270)
6271
6272xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006273 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006275 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006276 "test/convolution-operator-tester.h",
6277 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006278 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006279)
6280
6281xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006282 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006284 "test/convolution-nchw.cc",
6285 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006287 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288)
6289
6290xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006291 name = "copy_nc_test",
6292 srcs = [
6293 "test/copy-nc.cc",
6294 "test/copy-operator-tester.h",
6295 ],
6296 deps = OPERATOR_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006300 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006301 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006302 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 "test/deconvolution-operator-tester.h",
6304 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006305 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306)
6307
6308xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006309 name = "divide_nd_test",
6310 srcs = [
6311 "test/binary-elementwise-operator-tester.h",
6312 "test/divide-nd.cc",
6313 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006314 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006315)
6316
6317xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006318 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006319 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006320 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006321 "test/fully-connected-operator-tester.h",
6322 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006323 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324)
6325
6326xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006327 name = "floor_nc_test",
6328 srcs = [
6329 "test/floor-nc.cc",
6330 "test/floor-operator-tester.h",
6331 ],
6332 deps = OPERATOR_TEST_DEPS,
6333)
6334
6335xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006336 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006337 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006338 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006339 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006340 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006341 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006342)
6343
6344xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006345 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006346 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006347 "test/global-average-pooling-ncw.cc",
6348 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006349 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006350 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006351)
6352
6353xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006354 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006355 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006356 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 "test/hardswish-operator-tester.h",
6358 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006359 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006360)
6361
6362xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006363 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006364 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006365 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006366 "test/leaky-relu-operator-tester.h",
6367 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006368 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006369)
6370
6371xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006372 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006373 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006374 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006375 "test/max-pooling-operator-tester.h",
6376 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006377 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006378)
6379
6380xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006381 name = "maximum_nd_test",
6382 srcs = [
6383 "test/binary-elementwise-operator-tester.h",
6384 "test/maximum-nd.cc",
6385 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006386 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006387)
6388
6389xnnpack_unit_test(
6390 name = "minimum_nd_test",
6391 srcs = [
6392 "test/binary-elementwise-operator-tester.h",
6393 "test/minimum-nd.cc",
6394 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006395 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006396)
6397
6398xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006399 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006400 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006401 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006402 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006403 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006404 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006405)
6406
6407xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006408 name = "negate_nc_test",
6409 srcs = [
6410 "test/negate-nc.cc",
6411 "test/negate-operator-tester.h",
6412 ],
6413 deps = OPERATOR_TEST_DEPS,
6414)
6415
6416xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006417 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006418 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006419 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006420 "test/prelu-operator-tester.h",
6421 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006422 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006423)
6424
6425xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006426 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006427 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006428 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006429 "test/resize-bilinear-operator-tester.h",
6430 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006431 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006432)
6433
6434xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006435 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006436 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006437 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006438 "test/sigmoid-operator-tester.h",
6439 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006440 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006441)
6442
6443xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006444 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006445 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006446 "test/softmax-nc.cc",
6447 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006448 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006449 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450)
6451
6452xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006453 name = "square_nc_test",
6454 srcs = [
6455 "test/square-nc.cc",
6456 "test/square-operator-tester.h",
6457 ],
6458 deps = OPERATOR_TEST_DEPS,
6459)
6460
6461xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006462 name = "square_root_nc_test",
6463 srcs = [
6464 "test/square-root-nc.cc",
6465 "test/square-root-operator-tester.h",
6466 ],
6467 deps = OPERATOR_TEST_DEPS,
6468)
6469
6470xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006471 name = "squared_difference_nd_test",
6472 srcs = [
6473 "test/binary-elementwise-operator-tester.h",
6474 "test/squared-difference-nd.cc",
6475 ],
6476 deps = OPERATOR_TEST_DEPS,
6477)
6478
6479xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006480 name = "subtract_nd_test",
6481 srcs = [
6482 "test/binary-elementwise-operator-tester.h",
6483 "test/subtract-nd.cc",
6484 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006485 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006486)
6487
6488xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006489 name = "truncation_nc_test",
6490 srcs = [
6491 "test/truncation-nc.cc",
6492 "test/truncation-operator-tester.h",
6493 ],
6494 deps = OPERATOR_TEST_DEPS,
6495)
6496
6497xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006498 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006499 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006500 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 "test/unpooling-operator-tester.h",
6502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006503 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006504)
6505
Chao Mei6ddfc602020-05-13 22:29:36 -07006506############################### Misc unit tests ###############################
6507
6508xnnpack_unit_test(
6509 name = "memory_planner_test",
6510 srcs = [
6511 "test/memory-planner-test.cc",
6512 ],
6513 deps = [
6514 ":XNNPACK",
6515 ":memory_planner",
6516 ],
6517)
6518
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006519xnnpack_unit_test(
6520 name = "subgraph_nchw_test",
6521 srcs = [
6522 "src/xnnpack/subgraph.h",
6523 "test/subgraph-nchw.cc",
6524 "test/subgraph-tester.h",
6525 ],
6526 deps = [
6527 ":XNNPACK",
6528 ],
6529)
6530
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531############################# Build configurations #############################
6532
Marat Dukhanb8642352019-10-30 15:43:02 -07006533# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006535 name = "xnn_enable_assembly_explicit_true",
6536 define_values = {"xnn_enable_assembly": "true"},
6537)
6538
6539# Disables usage of assembly kernels.
6540config_setting(
6541 name = "xnn_enable_assembly_explicit_false",
6542 define_values = {"xnn_enable_assembly": "false"},
6543)
6544
Marat Dukhan9de90e02020-06-18 16:04:12 -07006545# Enables usage of sparse inference.
6546config_setting(
6547 name = "xnn_enable_sparse_explicit_true",
6548 define_values = {"xnn_enable_sparse": "true"},
6549)
6550
6551# Disables usage of sparse inference.
6552config_setting(
6553 name = "xnn_enable_sparse_explicit_false",
6554 define_values = {"xnn_enable_sparse": "false"},
6555)
6556
Marat Dukhan05702cf2020-03-26 15:41:33 -07006557# Disables usage of HMP-aware optimizations.
6558config_setting(
6559 name = "xnn_enable_hmp_explicit_false",
6560 define_values = {"xnn_enable_hmp": "false"},
6561)
6562
Chao Mei6ddfc602020-05-13 22:29:36 -07006563# Enable usage of optimized memory allocation
6564config_setting(
6565 name = "xnn_enable_memopt_explicit_true",
6566 define_values = {"xnnpack_enable_memopt": "true"},
6567)
6568
6569# Disable usage of optimized memory allocation
6570config_setting(
6571 name = "xnn_enable_memopt_explicit_false",
6572 define_values = {"xnnpack_enable_memopt": "false"},
6573)
6574
Marat Dukhanb8642352019-10-30 15:43:02 -07006575# Builds with -c dbg
6576config_setting(
6577 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006578 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006579 "compilation_mode": "dbg",
6580 },
6581)
6582
6583# Builds with -c opt
6584config_setting(
6585 name = "optimized_build",
6586 values = {
6587 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006588 },
6589)
6590
6591config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006592 name = "linux_k8",
6593 values = {"cpu": "k8"},
6594)
6595
6596config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006597 name = "linux_arm",
6598 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006599)
6600
6601config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006602 name = "linux_armeabi",
6603 values = {"cpu": "armeabi"},
6604)
6605
6606config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006607 name = "linux_armhf",
6608 values = {"cpu": "armhf"},
6609)
6610
6611config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006612 name = "linux_armv7a",
6613 values = {"cpu": "armv7a"},
6614)
6615
6616config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006617 name = "linux_aarch64",
6618 values = {"cpu": "aarch64"},
6619)
6620
6621config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006622 name = "android",
6623 values = {"crosstool_top": "//external:android/crosstool"},
6624)
6625
6626config_setting(
6627 name = "android_armv7",
6628 values = {
6629 "crosstool_top": "//external:android/crosstool",
6630 "cpu": "armeabi-v7a",
6631 },
6632)
6633
6634config_setting(
6635 name = "android_arm64",
6636 values = {
6637 "crosstool_top": "//external:android/crosstool",
6638 "cpu": "arm64-v8a",
6639 },
6640)
6641
6642config_setting(
6643 name = "android_x86",
6644 values = {
6645 "crosstool_top": "//external:android/crosstool",
6646 "cpu": "x86",
6647 },
6648)
6649
6650config_setting(
6651 name = "android_x86_64",
6652 values = {
6653 "crosstool_top": "//external:android/crosstool",
6654 "cpu": "x86_64",
6655 },
6656)
6657
6658config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006659 name = "windows_x86_64",
6660 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006661)
6662
6663config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006664 name = "windows_x86_64_clang",
6665 values = {
6666 "compiler": "clang-cl",
6667 "cpu": "x64_windows",
6668 },
6669)
6670
6671config_setting(
6672 name = "windows_x86_64_mingw",
6673 values = {
6674 "compiler": "mingw-gcc",
6675 "cpu": "x64_windows",
6676 },
6677)
6678
6679config_setting(
6680 name = "windows_x86_64_msys",
6681 values = {
6682 "compiler": "msys-gcc",
6683 "cpu": "x64_windows",
6684 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006685)
6686
6687config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006688 name = "macos_x86_64",
6689 values = {
6690 "apple_platform_type": "macos",
6691 "cpu": "darwin",
6692 },
6693)
6694
6695config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006696 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006697 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698)
6699
6700config_setting(
6701 name = "emscripten_wasm",
6702 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006703 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006704 "cpu": "wasm",
6705 },
6706)
6707
6708config_setting(
6709 name = "emscripten_wasmsimd",
6710 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006711 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006712 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006713 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006714 },
6715)
6716
6717config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006718 name = "ios_armv7",
6719 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006720 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006721 "cpu": "ios_armv7",
6722 },
6723)
6724
6725config_setting(
6726 name = "ios_arm64",
6727 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006728 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006729 "cpu": "ios_arm64",
6730 },
6731)
6732
6733config_setting(
6734 name = "ios_arm64e",
6735 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006736 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006737 "cpu": "ios_arm64e",
6738 },
6739)
6740
6741config_setting(
6742 name = "ios_x86",
6743 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006744 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006745 "cpu": "ios_i386",
6746 },
6747)
6748
6749config_setting(
6750 name = "ios_x86_64",
6751 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006752 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006753 "cpu": "ios_x86_64",
6754 },
6755)
6756
6757config_setting(
6758 name = "watchos_armv7k",
6759 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006760 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006761 "cpu": "watchos_armv7k",
6762 },
6763)
6764
6765config_setting(
6766 name = "watchos_arm64_32",
6767 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006768 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006769 "cpu": "watchos_arm64_32",
6770 },
6771)
6772
6773config_setting(
6774 name = "watchos_x86",
6775 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006776 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006777 "cpu": "watchos_i386",
6778 },
6779)
6780
6781config_setting(
6782 name = "watchos_x86_64",
6783 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006784 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006785 "cpu": "watchos_x86_64",
6786 },
6787)
6788
6789config_setting(
6790 name = "tvos_arm64",
6791 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006792 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006793 "cpu": "tvos_arm64",
6794 },
6795)
6796
6797config_setting(
6798 name = "tvos_x86_64",
6799 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006800 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006801 "cpu": "tvos_x86_64",
6802 },
6803)