blob: f42cb19f60a31f030e99028208457367bfa4bda9 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800425 "src/math/sigmoid-scalar-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001074 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1075 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1076 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1078 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1079 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001082 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001083 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001084 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001085 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001086 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001087 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001088 "src/x32-zip/x2-wasmsimd.c",
1089 "src/x32-zip/x3-wasmsimd.c",
1090 "src/x32-zip/x4-wasmsimd.c",
1091 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001092 "src/math/roundne-wasmsimd-addsub.c",
1093 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001094 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001095 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001099 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1100 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001101]
1102
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001103PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001104 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001105 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001106 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001107 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001108 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001109 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1110 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1111 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/precise-psimd.c",
1116 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001117 "src/qu8-requantization/precise-psimd.c",
1118 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121# ISA-specific micro-kernels
1122NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001123 "src/f32-argmaxpool/4x-neon-c4.c",
1124 "src/f32-argmaxpool/9p8x-neon-c4.c",
1125 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001126 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1127 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001128 "src/f32-clamp/gen/neon-x4.c",
1129 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001130 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1131 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001134 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1135 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1144 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1152 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1154 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1160 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1161 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1162 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1164 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1165 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1166 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1167 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1170 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1171 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1173 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1181 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1182 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001184 "src/f32-hswish/gen/hswish-neon-x4.c",
1185 "src/f32-hswish/gen/hswish-neon-x8.c",
1186 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001187 "src/f32-ibilinear/gen/neon-c4.c",
1188 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1190 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1191 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1193 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1195 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1196 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1197 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1198 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1200 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1201 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1202 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1203 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001205 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1206 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1207 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001210 "src/f32-prelu/gen/neon-1x4.c",
1211 "src/f32-prelu/gen/neon-1x8.c",
1212 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001213 "src/f32-prelu/gen/neon-2x4.c",
1214 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001215 "src/f32-prelu/gen/neon-2x16.c",
1216 "src/f32-prelu/gen/neon-4x4.c",
1217 "src/f32-prelu/gen/neon-4x8.c",
1218 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001243 "src/f32-relu/gen/neon-x4.c",
1244 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001246 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001247 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1248 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1267 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001269 "src/f32-vbinary/gen/vmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1272 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1273 "src/f32-vbinary/gen/vmin-neon-x4.c",
1274 "src/f32-vbinary/gen/vmin-neon-x8.c",
1275 "src/f32-vbinary/gen/vminc-neon-x4.c",
1276 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001283 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1285 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001291 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1292 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001293 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1294 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001295 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1296 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1297 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001303 "src/f32-vunary/gen/vabs-neon-x4.c",
1304 "src/f32-vunary/gen/vabs-neon-x8.c",
1305 "src/f32-vunary/gen/vneg-neon-x4.c",
1306 "src/f32-vunary/gen/vneg-neon-x8.c",
1307 "src/f32-vunary/gen/vsqr-neon-x4.c",
1308 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001309 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1310 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1311 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001313 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1314 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1315 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1317 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1318 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001321 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1322 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1323 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1325 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1326 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001329 "src/qs8-requantization/precise-neon.c",
1330 "src/qs8-requantization/fp32-neon.c",
1331 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001332 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1333 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1334 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1336 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1337 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001340 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1341 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1342 "src/qu8-dwconv/up8x9-minmax-neon.c",
1343 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1344 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1345 "src/qu8-gemm/4x8-minmax-neon.c",
1346 "src/qu8-gemm/8x8-minmax-neon.c",
1347 "src/qu8-igemm/4x8-minmax-neon.c",
1348 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001349 "src/qu8-requantization/precise-neon.c",
1350 "src/qu8-requantization/fp32-neon.c",
1351 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001352 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001355 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001356 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001358 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001359 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360 "src/x32-zip/x2-neon.c",
1361 "src/x32-zip/x3-neon.c",
1362 "src/x32-zip/x4-neon.c",
1363 "src/x32-zip/xm-neon.c",
1364 "src/x8-zip/x2-neon.c",
1365 "src/x8-zip/x3-neon.c",
1366 "src/x8-zip/x4-neon.c",
1367 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001368 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-neon-addsub.c",
1370 "src/math/roundd-neon-cvt.c",
1371 "src/math/roundu-neon-addsub.c",
1372 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neon-addsub.c",
1374 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001375 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001376 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1377 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1378 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1379 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1380 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001382 "src/math/sqrt-neon-nr1rsqrts.c",
1383 "src/math/sqrt-neon-nr2rsqrts.c",
1384 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001385]
1386
1387NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001388 "src/f32-ibilinear/gen/neonfma-c4.c",
1389 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001390 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1391 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1392 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1394 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1395 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1396 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1397 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1405 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1412 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1413 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1416 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1417 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1418 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1424 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1425 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1430 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001509 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1510 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001531 "src/math/exp-neonfma-lut64-p2.c",
1532 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001533 "src/math/expminus-neonfma-lut2048-p1.c",
1534 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001535 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001536 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1537 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1539 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1542 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1543 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1545 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1546 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1548 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1551 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1552 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001554 "src/math/sqrt-neonfma-nr1fma.c",
1555 "src/math/sqrt-neonfma-nr2fma.c",
1556 "src/math/sqrt-neonfma-nr3fma.c",
1557 "src/math/sqrt-neonfma-nr2fma1adj.c",
1558 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001559]
1560
1561AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1569 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1573 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1575 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1579 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1581 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1583 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1585 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1587 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1589 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001592 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1593 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001596 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1598 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1599 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001601 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001619 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1620 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1621 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1622 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1623 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1625 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1626 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1627 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1628 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1630 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1631 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1632 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1633 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1635 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1636 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001637 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1638 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1640 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1641 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1643 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1644 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645]
1646
Marat Dukhan8853b822020-05-07 12:19:01 -07001647NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001648 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1649 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1650 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001656 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001657 "src/math/roundd-neonv8.c",
1658 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001659 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001660]
1661
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001663 "src/f16-clamp/gen/neonfp16arith-x16.c",
1664 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001665 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001677 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1678 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001679 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1680 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001681 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001682 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001683 "src/f16-relu/gen/neonfp16arith-x16.c",
1684 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001685 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001713 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1714 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001715 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001723 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001727 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001739 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1740 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1741 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Benoit Jacoba9644732020-08-13 12:48:55 -07001749NEONDOT_UKERNELS = [
1750 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1751 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001753 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1754 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1755 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001756]
1757
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001759 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1760 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001761 "src/f32-clamp/gen/sse-x4.c",
1762 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001763 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1764 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001765 "src/f32-dwconv-chw/3x3p1-sse.c",
1766 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001767 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1768 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1769 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1770 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1771 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1772 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1773 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1775 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1777 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001779 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001780 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1781 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001782 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1783 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1784 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1785 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1786 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1787 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1788 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1789 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1790 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1791 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1792 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1793 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1794 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001795 "src/f32-hswish/gen/hswish-sse-x4.c",
1796 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001797 "src/f32-ibilinear/gen/sse-c4.c",
1798 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001799 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1800 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1801 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1802 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1803 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1804 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1805 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001806 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1807 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1808 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001809 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001810 "src/f32-prelu/gen/sse-2x4.c",
1811 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001812 "src/f32-relu/gen/sse-x4.c",
1813 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001815 "src/f32-spmm/gen/4x1-minmax-sse.c",
1816 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001817 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001818 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1819 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1820 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1821 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1822 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001826 "src/f32-vbinary/gen/vmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1829 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1830 "src/f32-vbinary/gen/vmin-sse-x4.c",
1831 "src/f32-vbinary/gen/vmin-sse-x8.c",
1832 "src/f32-vbinary/gen/vminc-sse-x4.c",
1833 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001834 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1835 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1836 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1837 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1838 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001842 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1843 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1844 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1845 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001846 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1848 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1849 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001850 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1851 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001852 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1853 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001854 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1855 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001856 "src/f32-vunary/gen/vabs-sse-x4.c",
1857 "src/f32-vunary/gen/vabs-sse-x8.c",
1858 "src/f32-vunary/gen/vneg-sse-x4.c",
1859 "src/f32-vunary/gen/vneg-sse-x8.c",
1860 "src/f32-vunary/gen/vsqr-sse-x4.c",
1861 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001862 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001864 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001865 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001866 "src/math/roundd-sse-addsub.c",
1867 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001868 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001869 "src/math/sqrt-sse-hh1mac.c",
1870 "src/math/sqrt-sse-nr1mac.c",
1871 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872]
1873
1874SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001875 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1876 "src/f32-argmaxpool/4x-sse2-c4.c",
1877 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001878 "src/f32-prelu/gen/sse2-2x4.c",
1879 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001880 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1881 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1882 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1883 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1884 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1885 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001892 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001893 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001894 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001895 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001896 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1897 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001898 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1899 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001900 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1901 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1902 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1903 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1904 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1905 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1906 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1907 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001908 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1909 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1910 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001911 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1912 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1913 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001914 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1915 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1916 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001917 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1918 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001919 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1920 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001921 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1922 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001923 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1924 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001925 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001926 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1927 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001928 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001929 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1930 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1931 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001932 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1933 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001934 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1935 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001936 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1937 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001938 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001939 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1940 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001941 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001942 "src/qs8-requantization/precise-sse2.c",
1943 "src/qs8-requantization/fp32-sse2.c",
1944 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001945 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1946 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1947 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1948 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001949 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1950 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1951 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1952 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001953 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1954 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1955 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1956 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1957 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1958 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1959 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1960 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001961 "src/qu8-requantization/precise-sse2.c",
1962 "src/qu8-requantization/fp32-sse2.c",
1963 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001964 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001965 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001966 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001968 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "src/x32-zip/x2-sse2.c",
1970 "src/x32-zip/x3-sse2.c",
1971 "src/x32-zip/x4-sse2.c",
1972 "src/x32-zip/xm-sse2.c",
1973 "src/x8-zip/x2-sse2.c",
1974 "src/x8-zip/x3-sse2.c",
1975 "src/x8-zip/x4-sse2.c",
1976 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001977 "src/math/exp-sse2-p5.c",
1978 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001979 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001980 "src/math/roundd-sse2-cvt.c",
1981 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001982 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001983 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001984]
1985
1986SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001987 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1988 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1989 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001990 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1991 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1992 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001993 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1994 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1995 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001996 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1997 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001998 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1999 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002000 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2001 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002002 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2003 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002004 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002005 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2006 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002007 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002008 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2009 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2010 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002011 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2012 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002013 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2014 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002015 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2016 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002017 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002018 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2019 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002020 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002021 "src/qs8-requantization/precise-ssse3.c",
2022 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002023 "src/qu8-requantization/precise-ssse3.c",
2024 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025]
2026
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002027SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002028 "src/f32-prelu/gen/sse41-2x4.c",
2029 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002030 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002031 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002032 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002033 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002034 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2035 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002036 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2037 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002038 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2039 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2040 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2041 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2042 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2043 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2044 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2045 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002046 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2047 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2048 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002049 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2050 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2051 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002052 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2053 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2054 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002055 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2056 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002057 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2058 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002059 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2060 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002061 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2062 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002063 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002064 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2065 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002066 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002067 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2068 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2069 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002070 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2071 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002072 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2073 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002074 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2075 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002076 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002077 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2078 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002079 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002080 "src/qs8-requantization/fp32-sse4.c",
2081 "src/qs8-requantization/precise-sse4.c",
2082 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002083 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2084 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2085 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2086 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002087 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2088 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2089 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2090 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002091 "src/qu8-requantization/precise-sse4.c",
2092 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002093 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002094 "src/math/roundd-sse41.c",
2095 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002096 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002097]
2098
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002100 "src/f32-clamp/gen/avx-x8.c",
2101 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002102 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2103 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2104 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2105 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2106 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2107 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2108 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2109 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2110 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2111 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2112 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2113 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2114 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2115 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2116 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2117 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2118 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2119 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2120 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2121 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2122 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2123 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2124 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2125 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2126 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2127 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2128 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2129 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2130 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2131 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002132 "src/f32-hswish/gen/hswish-avx-x8.c",
2133 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002134 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2135 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2136 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2137 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2138 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2139 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2140 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2141 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2142 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002143 "src/f32-prelu/gen/avx-2x8.c",
2144 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002145 "src/f32-relu/gen/avx-x16.c",
2146 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002148 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2149 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2150 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2151 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2152 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002156 "src/f32-vbinary/gen/vmax-avx-x8.c",
2157 "src/f32-vbinary/gen/vmax-avx-x16.c",
2158 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2159 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2160 "src/f32-vbinary/gen/vmin-avx-x8.c",
2161 "src/f32-vbinary/gen/vmin-avx-x16.c",
2162 "src/f32-vbinary/gen/vminc-avx-x8.c",
2163 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002164 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2165 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2166 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2167 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2168 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2169 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2170 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2171 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002172 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2173 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2174 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2175 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002176 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2177 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2178 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2179 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002180 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2181 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002182 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002183 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2184 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2185 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2186 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2187 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2188 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2189 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2190 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002191 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2192 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002193 "src/f32-vunary/gen/vabs-avx-x8.c",
2194 "src/f32-vunary/gen/vabs-avx-x16.c",
2195 "src/f32-vunary/gen/vneg-avx-x8.c",
2196 "src/f32-vunary/gen/vneg-avx-x16.c",
2197 "src/f32-vunary/gen/vsqr-avx-x8.c",
2198 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002199]
2200
Marat Dukhan1566fee2020-08-02 21:55:41 -07002201XOP_UKERNELS = [
2202 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2203 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2204 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2205 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002206 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2207 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002208 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2209 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002210 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002211 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2212 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002213 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002214 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2215 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2216 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002217 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2218 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2219 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2220 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2221 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2222 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002223 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002224 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2225 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002226 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002227]
2228
Marat Dukhanfda12b82019-11-21 12:27:59 -08002229FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002230 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2231 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2232 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2233 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2234 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2235 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2236 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2237 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2238 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2239 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2240 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2241 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2242 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2243 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2244 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2245 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2246 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2247 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2248 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2249 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2250 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2251 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2252 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2253 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2254 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2255 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2256 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2257 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2258 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2259 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2260 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2261 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2262 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2263 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2264 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2265 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2266 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2267 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2268 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2269 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002270 "src/f32-hswish/gen/hswish-fma3-x8.c",
2271 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2273 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2274 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2275 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2276 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2277 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2278 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2279 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2280 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2281 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2282 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2283 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2284 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2285 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002286 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2287 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2288 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2289 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2290 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2291 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2292 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2293 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002294 "src/math/sqrt-fma3-nr1fma.c",
2295 "src/math/sqrt-fma3-nr2fma.c",
2296 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002297]
2298
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002299AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002300 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2301 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2302 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2303 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2304 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2305 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2306 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2307 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2308 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2309 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2310 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2311 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2312 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2313 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2314 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2315 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2316 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2317 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2318 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2319 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2320 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2321 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2322 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2323 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2324 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2325 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2326 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2327 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2328 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2329 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2330 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2331 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2332 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2333 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2334 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2335 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002336 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2337 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2338 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2339 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2340 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2341 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2342 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2343 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2344 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2345 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2350 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2351 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2352 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2353 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2354 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2355 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2356 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2357 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2358 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2359 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2360 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2361 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2362 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2363 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2364 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2365 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002366 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2367 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2368 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2369 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2370 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2371 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2372 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2373 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2374 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2375 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2376 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2377 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2378 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2379 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2380 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2381 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2382 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2383 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2384 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2385 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2386 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2387 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2388 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2389 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002390 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2391 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002392 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2393 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2394 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2395 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002396 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2397 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2398 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002399 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2400 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2401 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002402 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2403 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2404 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002405 "src/math/exp-avx2-p5.c",
2406 "src/math/exp-avx2-perm-p3.c",
2407 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002408 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002409 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002410 "src/math/sigmoid-avx2-rr2-p5-div.c",
2411 "src/math/sigmoid-avx2-rr1-p5-div.c",
2412 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2413 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2414 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2415 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002416]
2417
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002419 "src/f32-clamp/gen/avx512f-x16.c",
2420 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002421 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2422 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2423 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2424 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2425 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2426 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2427 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2428 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2429 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2430 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2431 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2432 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2433 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2434 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2435 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2436 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2437 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2438 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2439 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2440 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2441 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2442 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2443 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2444 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002445 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2446 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002447 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2448 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2449 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2450 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2451 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2452 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002453 "src/f32-prelu/gen/avx512f-2x16.c",
2454 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002455 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2456 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2457 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2458 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2459 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2460 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2461 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2462 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2463 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2464 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2465 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2466 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2467 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2468 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2469 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2470 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2471 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2472 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2473 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2474 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2475 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2476 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2477 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2478 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2479 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2480 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2481 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2482 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2483 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2484 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2485 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2486 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2487 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2488 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2489 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2490 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002491 "src/f32-relu/gen/avx512f-x16.c",
2492 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002493 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002494 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2495 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2496 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2497 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2498 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2499 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2500 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2501 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002502 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2503 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2504 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2505 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2506 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2507 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2508 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2509 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002510 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2511 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2512 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2513 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2514 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2515 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2516 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2517 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002518 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2519 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2520 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2521 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002522 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2523 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2524 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2525 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002526 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2527 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002528 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002529 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2530 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2531 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2532 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2533 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2534 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2535 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2536 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2537 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2538 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2539 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2540 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2541 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2542 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2543 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2544 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2545 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2546 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2547 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2548 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2549 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2550 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2551 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2552 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002553 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2554 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2555 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2556 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2557 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2558 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2559 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2560 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002561 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2562 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2563 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2564 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2565 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2566 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2567 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2568 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002569 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2570 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2571 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2572 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2573 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2574 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002575 "src/math/exp-avx512f-p5-scalef.c",
2576 "src/math/exp-avx512f-p5.c",
2577 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002578 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002579 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002580 "src/math/sqrt-avx512f-nr1fma.c",
2581 "src/math/sqrt-avx512f-nr2fma.c",
2582 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583]
2584
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002585AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002586 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2587 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002588 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2589 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2590 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2591 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2592 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2593 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2594 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2595 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2596]
2597
Frank Barchardbcedc082020-08-17 18:00:51 -07002598WASM32_ASM_UKERNELS = [
2599 "src/f32-relu/wasm_shr_x1.S",
2600 "src/f32-relu/wasm_shr_x2.S",
2601 "src/f32-relu/wasm_shr_x4.S",
2602]
2603
Marat Dukhan08c4a432019-10-03 09:29:21 -07002604AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002605 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002606 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002607 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2608 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002609 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002610 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002611 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002612 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002613 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2614 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002615 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2616 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2617 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2618 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619]
2620
2621AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002622 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002623 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2624 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002625 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002626 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2627 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002628 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2629 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2630 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002631 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002632 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2633 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2634 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002635 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002636 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2637 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002638 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002639 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2640 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2641 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2642 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2643 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2644 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2645 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2646 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2647 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2648 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2649 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2650 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2651 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2652 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2653 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2654 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2655 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2656 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002657 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2658 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002659 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002660 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2661 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2662 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2663 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2664 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2665 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2666 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2667 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2668 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2669 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2670 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2671 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2672 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2673 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2674 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2675 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2676 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2677 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002678 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2679 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2680 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2681 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2682 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2683 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2684 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2685 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2686 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2687 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2688 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2689 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2690 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2691 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2692 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2693 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2694 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2695 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002696]
2697
Marat Dukhan1b354632020-03-23 12:50:22 -07002698INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002699 "src/xnnpack/argmaxpool.h",
2700 "src/xnnpack/avgpool.h",
2701 "src/xnnpack/clamp.h",
2702 "src/xnnpack/common.h",
2703 "src/xnnpack/conv.h",
2704 "src/xnnpack/dwconv.h",
2705 "src/xnnpack/gavgpool.h",
2706 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002707 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002708 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002709 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002710 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002711 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712 "src/xnnpack/lut.h",
2713 "src/xnnpack/math.h",
2714 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002715 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002716 "src/xnnpack/packx.h",
2717 "src/xnnpack/pad.h",
2718 "src/xnnpack/params.h",
2719 "src/xnnpack/pavgpool.h",
2720 "src/xnnpack/ppmm.h",
2721 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002722 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002723 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002724 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002725 "src/xnnpack/rmax.h",
2726 "src/xnnpack/scalar-utils.h",
2727 "src/xnnpack/spmm.h",
2728 "src/xnnpack/unpool.h",
2729 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002730 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002731 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002732 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002733 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002734 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002735 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002736 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002737]
2738
2739INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 "include/xnnpack.h",
2741 "src/xnnpack/allocator.h",
2742 "src/xnnpack/compute.h",
2743 "src/xnnpack/im2col.h",
2744 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002745 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002746 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 "src/xnnpack/operator.h",
2748 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002749 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002750 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002751 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002752 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002753]
2754
Marat Dukhan1b354632020-03-23 12:50:22 -07002755ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002756 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757]
2758
Marat Dukhan1b354632020-03-23 12:50:22 -07002759MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002760 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002761 "include/xnnpack.h",
2762]
2763
Marat Dukhan1b354632020-03-23 12:50:22 -07002764MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002766 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 "src/xnnpack/requantization.h",
2768 "include/xnnpack.h",
2769]
2770
2771OPERATOR_TEST_PARAMS_HDRS = [
2772 "src/xnnpack/params.h",
2773 "src/xnnpack/common.h",
2774]
2775
2776WEIGHTS_PACK_HDRS = [
2777 "src/xnnpack/pack.h",
2778 "src/xnnpack/operator.h",
2779 "src/xnnpack/compute.h",
2780]
2781
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002782LOGGING_COPTS = select({
2783 # No logging in optimized mode
2784 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2785 # Full logging in debug mode
2786 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2787 # Error-only logging in default (fastbuild) mode
2788 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2789})
2790
Marat Dukhan3b59de22020-06-03 20:15:19 -07002791LOGGING_SRCS = select({
2792 # No logging in optimized mode
2793 ":optimized_build": [],
2794 "//conditions:default": [
2795 "src/operator-strings.c",
2796 "src/subgraph-strings.c",
2797 ],
2798})
2799
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002800LOGGING_HDRS = [
2801 "src/xnnpack/log.h",
2802]
2803
Marat Dukhan08c4a432019-10-03 09:29:21 -07002804xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002805 name = "tables",
2806 srcs = TABLE_SRCS,
2807 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002808 gcc_copts = xnnpack_gcc_std_copts(),
2809 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002810)
2811
2812xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002813 name = "scalar_ukernels",
2814 srcs = SCALAR_UKERNELS,
2815 hdrs = INTERNAL_HDRS,
2816 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002817 gcc_copts = xnnpack_gcc_std_copts(),
2818 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002820 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821 "@FP16",
2822 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002823 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 ],
2825)
2826
2827xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002828 name = "scalar_ukernels_test_mode",
2829 srcs = SCALAR_UKERNELS,
2830 hdrs = INTERNAL_HDRS,
2831 aarch32_copts = ["-marm"],
2832 copts = [
2833 "-UNDEBUG",
2834 "-DXNN_TEST_MODE=1",
2835 ],
2836 gcc_copts = xnnpack_gcc_std_copts(),
2837 msvc_copts = xnnpack_msvc_std_copts(),
2838 deps = [
2839 ":tables",
2840 "@FP16",
2841 "@FXdiv",
2842 "@pthreadpool",
2843 ],
2844)
2845
2846xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002847 name = "wasm_ukernels",
2848 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002849 gcc_copts = xnnpack_gcc_std_copts(),
2850 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002851 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002852 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002853 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002854 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002855 "@FP16",
2856 "@FXdiv",
2857 "@pthreadpool",
2858 ],
2859)
2860
2861xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002862 name = "wasm_ukernels_test_mode",
2863 hdrs = INTERNAL_HDRS,
2864 copts = [
2865 "-UNDEBUG",
2866 "-DXNN_TEST_MODE=1",
2867 ],
2868 gcc_copts = xnnpack_gcc_std_copts(),
2869 msvc_copts = xnnpack_msvc_std_copts(),
2870 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002871 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002872 deps = [
2873 ":tables",
2874 "@FP16",
2875 "@FXdiv",
2876 "@pthreadpool",
2877 ],
2878)
2879
2880xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002881 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 hdrs = INTERNAL_HDRS,
2883 aarch32_copts = [
2884 "-marm",
2885 "-mfpu=neon",
2886 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002887 gcc_copts = xnnpack_gcc_std_copts(),
2888 gcc_x86_copts = ["-msse2"],
2889 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 optimized_copts = [
2891 "-O3",
2892 "-ffast-math",
2893 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002894 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002895 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002896 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 "@FP16",
2898 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002899 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 ],
2901)
2902
2903xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002904 name = "psimd_fastmath_ukernels_test_mode",
2905 hdrs = INTERNAL_HDRS,
2906 aarch32_copts = [
2907 "-marm",
2908 "-mfpu=neon",
2909 ],
2910 copts = [
2911 "-UNDEBUG",
2912 "-DXNN_TEST_MODE=1",
2913 ],
2914 gcc_copts = xnnpack_gcc_std_copts(),
2915 gcc_x86_copts = ["-msse2"],
2916 msvc_copts = xnnpack_msvc_std_copts(),
2917 optimized_copts = [
2918 "-O3",
2919 "-ffast-math",
2920 ],
2921 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2922 deps = [
2923 ":tables",
2924 "@FP16",
2925 "@psimd",
2926 "@pthreadpool",
2927 ],
2928)
2929
2930xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002931 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002932 hdrs = INTERNAL_HDRS,
2933 aarch32_copts = [
2934 "-marm",
2935 "-mfpu=neon",
2936 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002937 gcc_copts = xnnpack_gcc_std_copts(),
2938 gcc_x86_copts = ["-msse2"],
2939 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002940 optimized_copts = [
2941 "-O3",
2942 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002943 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002944 deps = [
2945 ":tables",
2946 "@FP16",
2947 "@psimd",
2948 "@pthreadpool",
2949 ],
2950)
2951
2952xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002953 name = "psimd_accmath_ukernels_test_mode",
2954 hdrs = INTERNAL_HDRS,
2955 aarch32_copts = [
2956 "-marm",
2957 "-mfpu=neon",
2958 ],
2959 copts = [
2960 "-UNDEBUG",
2961 "-DXNN_TEST_MODE=1",
2962 ],
2963 gcc_copts = xnnpack_gcc_std_copts(),
2964 gcc_x86_copts = ["-msse2"],
2965 msvc_copts = xnnpack_msvc_std_copts(),
2966 optimized_copts = [
2967 "-O3",
2968 ],
2969 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2970 deps = [
2971 ":tables",
2972 "@FP16",
2973 "@psimd",
2974 "@pthreadpool",
2975 ],
2976)
2977
2978xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 name = "neon_ukernels",
2980 hdrs = INTERNAL_HDRS,
2981 aarch32_copts = [
2982 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002983 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002984 "-mfpu=neon",
2985 ],
2986 aarch32_srcs = NEON_UKERNELS,
2987 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002988 gcc_copts = xnnpack_gcc_std_copts(),
2989 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002990 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002991 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002992 "@FP16",
2993 "@pthreadpool",
2994 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995)
2996
2997xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002998 name = "neon_ukernels_test_mode",
2999 hdrs = INTERNAL_HDRS,
3000 aarch32_copts = [
3001 "-marm",
3002 "-march=armv7-a",
3003 "-mfpu=neon",
3004 ],
3005 aarch32_srcs = NEON_UKERNELS,
3006 aarch64_srcs = NEON_UKERNELS,
3007 copts = [
3008 "-UNDEBUG",
3009 "-DXNN_TEST_MODE=1",
3010 ],
3011 gcc_copts = xnnpack_gcc_std_copts(),
3012 msvc_copts = xnnpack_msvc_std_copts(),
3013 deps = [
3014 ":tables",
3015 "@FP16",
3016 "@pthreadpool",
3017 ],
3018)
3019
3020xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 name = "neonfma_ukernels",
3022 hdrs = INTERNAL_HDRS,
3023 aarch32_copts = [
3024 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003025 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 "-mfpu=neon-vfpv4",
3027 ],
3028 aarch32_srcs = NEONFMA_UKERNELS,
3029 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003030 apple_aarch32_copts = [
3031 "-mcpu=swift",
3032 "-mtune=generic",
3033 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003034 gcc_copts = xnnpack_gcc_std_copts(),
3035 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003036 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003037 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003038 "@FP16",
3039 "@pthreadpool",
3040 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003041)
3042
3043xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003044 name = "neonfma_ukernels_test_mode",
3045 hdrs = INTERNAL_HDRS,
3046 aarch32_copts = [
3047 "-marm",
3048 "-march=armv7-a",
3049 "-mfpu=neon-vfpv4",
3050 ],
3051 aarch32_srcs = NEONFMA_UKERNELS,
3052 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003053 apple_aarch32_copts = [
3054 "-mcpu=swift",
3055 "-mtune=generic",
3056 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003057 copts = [
3058 "-UNDEBUG",
3059 "-DXNN_TEST_MODE=1",
3060 ],
3061 gcc_copts = xnnpack_gcc_std_copts(),
3062 msvc_copts = xnnpack_msvc_std_copts(),
3063 deps = [
3064 ":tables",
3065 "@FP16",
3066 "@pthreadpool",
3067 ],
3068)
3069
3070xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003071 name = "neonv8_ukernels",
3072 hdrs = INTERNAL_HDRS,
3073 aarch32_copts = [
3074 "-marm",
3075 "-march=armv8-a",
3076 "-mfpu=neon-fp-armv8",
3077 ],
3078 aarch32_srcs = NEONV8_UKERNELS,
3079 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003080 apple_aarch32_copts = [
3081 "-mcpu=cyclone",
3082 "-mtune=generic",
3083 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003084 gcc_copts = xnnpack_gcc_std_copts(),
3085 msvc_copts = xnnpack_msvc_std_copts(),
3086 deps = [
3087 ":tables",
3088 "@FP16",
3089 "@pthreadpool",
3090 ],
3091)
3092
3093xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003094 name = "neonv8_ukernels_test_mode",
3095 hdrs = INTERNAL_HDRS,
3096 aarch32_copts = [
3097 "-marm",
3098 "-march=armv8-a",
3099 "-mfpu=neon-fp-armv8",
3100 ],
3101 aarch32_srcs = NEONV8_UKERNELS,
3102 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003103 apple_aarch32_copts = [
3104 "-mcpu=cyclone",
3105 "-mtune=generic",
3106 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003107 copts = [
3108 "-UNDEBUG",
3109 "-DXNN_TEST_MODE=1",
3110 ],
3111 gcc_copts = xnnpack_gcc_std_copts(),
3112 msvc_copts = xnnpack_msvc_std_copts(),
3113 deps = [
3114 ":tables",
3115 "@FP16",
3116 "@pthreadpool",
3117 ],
3118)
3119
3120xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 name = "neonfp16arith_ukernels",
3122 hdrs = INTERNAL_HDRS,
3123 aarch64_copts = ["-march=armv8.2-a+fp16"],
3124 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003125 gcc_copts = xnnpack_gcc_std_copts(),
3126 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003127 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003128 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003129 "@FP16",
3130 "@pthreadpool",
3131 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132)
3133
3134xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003135 name = "neonfp16arith_ukernels_test_mode",
3136 hdrs = INTERNAL_HDRS,
3137 aarch64_copts = ["-march=armv8.2-a+fp16"],
3138 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3139 copts = [
3140 "-UNDEBUG",
3141 "-DXNN_TEST_MODE=1",
3142 ],
3143 gcc_copts = xnnpack_gcc_std_copts(),
3144 msvc_copts = xnnpack_msvc_std_copts(),
3145 deps = [
3146 ":tables",
3147 "@FP16",
3148 "@pthreadpool",
3149 ],
3150)
3151
3152xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003153 name = "neondot_ukernels",
3154 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003155 aarch32_copts = [
3156 "-marm",
3157 "-march=armv8.2-a+dotprod",
3158 "-mfpu=neon-fp-armv8",
3159 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003160 aarch32_srcs = NEONDOT_UKERNELS,
3161 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3162 aarch64_srcs = NEONDOT_UKERNELS,
3163 gcc_copts = xnnpack_gcc_std_copts(),
3164 msvc_copts = xnnpack_msvc_std_copts(),
3165 deps = [
3166 ":tables",
3167 "@FP16",
3168 "@pthreadpool",
3169 ],
3170)
3171
3172xnnpack_cc_library(
3173 name = "neondot_ukernels_test_mode",
3174 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003175 aarch32_copts = [
3176 "-marm",
3177 "-march=armv8.2-a+dotprod",
3178 "-mfpu=neon-fp-armv8",
3179 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003180 aarch32_srcs = NEONDOT_UKERNELS,
3181 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3182 aarch64_srcs = NEONDOT_UKERNELS,
3183 copts = [
3184 "-UNDEBUG",
3185 "-DXNN_TEST_MODE=1",
3186 ],
3187 gcc_copts = xnnpack_gcc_std_copts(),
3188 msvc_copts = xnnpack_msvc_std_copts(),
3189 deps = [
3190 ":tables",
3191 "@FP16",
3192 "@pthreadpool",
3193 ],
3194)
3195
3196xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197 name = "sse2_ukernels",
3198 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003199 gcc_copts = xnnpack_gcc_std_copts(),
3200 gcc_x86_copts = ["-msse2"],
3201 msvc_copts = xnnpack_msvc_std_copts(),
3202 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003204 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003205 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003206 "@FP16",
3207 "@pthreadpool",
3208 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003209)
3210
3211xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003212 name = "sse2_ukernels_test_mode",
3213 hdrs = INTERNAL_HDRS,
3214 copts = [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ],
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 gcc_x86_copts = ["-msse2"],
3220 msvc_copts = xnnpack_msvc_std_copts(),
3221 msvc_x86_32_copts = ["/arch:SSE2"],
3222 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3223 deps = [
3224 ":tables",
3225 "@FP16",
3226 "@pthreadpool",
3227 ],
3228)
3229
3230xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003231 name = "ssse3_ukernels",
3232 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003233 gcc_copts = xnnpack_gcc_std_copts(),
3234 gcc_x86_copts = ["-mssse3"],
3235 msvc_copts = xnnpack_msvc_std_copts(),
3236 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003237 x86_srcs = SSSE3_UKERNELS,
3238 deps = [
3239 ":tables",
3240 "@FP16",
3241 "@pthreadpool",
3242 ],
3243)
3244
3245xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003246 name = "ssse3_ukernels_test_mode",
3247 hdrs = INTERNAL_HDRS,
3248 copts = [
3249 "-UNDEBUG",
3250 "-DXNN_TEST_MODE=1",
3251 ],
3252 gcc_copts = xnnpack_gcc_std_copts(),
3253 gcc_x86_copts = ["-mssse3"],
3254 msvc_copts = xnnpack_msvc_std_copts(),
3255 msvc_x86_32_copts = ["/arch:SSE2"],
3256 x86_srcs = SSSE3_UKERNELS,
3257 deps = [
3258 ":tables",
3259 "@FP16",
3260 "@pthreadpool",
3261 ],
3262)
3263
3264xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003265 name = "sse41_ukernels",
3266 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003267 gcc_copts = xnnpack_gcc_std_copts(),
3268 gcc_x86_copts = ["-msse4.1"],
3269 msvc_copts = xnnpack_msvc_std_copts(),
3270 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003271 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003272 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003273 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003274 "@FP16",
3275 "@pthreadpool",
3276 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003277)
3278
3279xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003280 name = "sse41_ukernels_test_mode",
3281 hdrs = INTERNAL_HDRS,
3282 copts = [
3283 "-UNDEBUG",
3284 "-DXNN_TEST_MODE=1",
3285 ],
3286 gcc_copts = xnnpack_gcc_std_copts(),
3287 gcc_x86_copts = ["-msse4.1"],
3288 msvc_copts = xnnpack_msvc_std_copts(),
3289 msvc_x86_32_copts = ["/arch:SSE2"],
3290 x86_srcs = SSE41_UKERNELS,
3291 deps = [
3292 ":tables",
3293 "@FP16",
3294 "@pthreadpool",
3295 ],
3296)
3297
3298xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299 name = "avx_ukernels",
3300 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003301 gcc_copts = xnnpack_gcc_std_copts(),
3302 gcc_x86_copts = ["-mavx"],
3303 msvc_copts = xnnpack_msvc_std_copts(),
3304 msvc_x86_32_copts = ["/arch:AVX"],
3305 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003306 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003307 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003308 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003309 "@FP16",
3310 "@pthreadpool",
3311 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003312)
3313
3314xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003315 name = "avx_ukernels_test_mode",
3316 hdrs = INTERNAL_HDRS,
3317 copts = [
3318 "-UNDEBUG",
3319 "-DXNN_TEST_MODE=1",
3320 ],
3321 gcc_copts = xnnpack_gcc_std_copts(),
3322 gcc_x86_copts = ["-mavx"],
3323 msvc_copts = xnnpack_msvc_std_copts(),
3324 msvc_x86_32_copts = ["/arch:AVX"],
3325 msvc_x86_64_copts = ["/arch:AVX"],
3326 x86_srcs = AVX_UKERNELS,
3327 deps = [
3328 ":tables",
3329 "@FP16",
3330 "@pthreadpool",
3331 ],
3332)
3333
3334xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003335 name = "xop_ukernels",
3336 hdrs = INTERNAL_HDRS,
3337 gcc_copts = xnnpack_gcc_std_copts(),
3338 gcc_x86_copts = ["-mxop"],
3339 msvc_copts = xnnpack_msvc_std_copts(),
3340 msvc_x86_32_copts = ["/arch:AVX"],
3341 msvc_x86_64_copts = ["/arch:AVX"],
3342 x86_srcs = XOP_UKERNELS,
3343 deps = [
3344 ":tables",
3345 "@FP16",
3346 "@pthreadpool",
3347 ],
3348)
3349
3350xnnpack_cc_library(
3351 name = "xop_ukernels_test_mode",
3352 hdrs = INTERNAL_HDRS,
3353 copts = [
3354 "-UNDEBUG",
3355 "-DXNN_TEST_MODE=1",
3356 ],
3357 gcc_copts = xnnpack_gcc_std_copts(),
3358 gcc_x86_copts = ["-mxop"],
3359 msvc_copts = xnnpack_msvc_std_copts(),
3360 msvc_x86_32_copts = ["/arch:AVX"],
3361 msvc_x86_64_copts = ["/arch:AVX"],
3362 x86_srcs = XOP_UKERNELS,
3363 deps = [
3364 ":tables",
3365 "@FP16",
3366 "@pthreadpool",
3367 ],
3368)
3369
3370xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003371 name = "fma3_ukernels",
3372 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003373 gcc_copts = xnnpack_gcc_std_copts(),
3374 gcc_x86_copts = ["-mfma"],
3375 msvc_copts = xnnpack_msvc_std_copts(),
3376 msvc_x86_32_copts = ["/arch:AVX"],
3377 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003378 x86_srcs = FMA3_UKERNELS,
3379 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003380 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003381 "@FP16",
3382 "@pthreadpool",
3383 ],
3384)
3385
3386xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003387 name = "fma3_ukernels_test_mode",
3388 hdrs = INTERNAL_HDRS,
3389 copts = [
3390 "-UNDEBUG",
3391 "-DXNN_TEST_MODE=1",
3392 ],
3393 gcc_copts = xnnpack_gcc_std_copts(),
3394 gcc_x86_copts = ["-mfma"],
3395 msvc_copts = xnnpack_msvc_std_copts(),
3396 msvc_x86_32_copts = ["/arch:AVX"],
3397 msvc_x86_64_copts = ["/arch:AVX"],
3398 x86_srcs = FMA3_UKERNELS,
3399 deps = [
3400 ":tables",
3401 "@FP16",
3402 "@pthreadpool",
3403 ],
3404)
3405
3406xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003407 name = "avx2_ukernels",
3408 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003409 gcc_copts = xnnpack_gcc_std_copts(),
3410 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003411 "-mfma",
3412 "-mavx2",
3413 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003414 msvc_copts = xnnpack_msvc_std_copts(),
3415 msvc_x86_32_copts = ["/arch:AVX2"],
3416 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003417 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003418 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003419 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003420 "@FP16",
3421 "@pthreadpool",
3422 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003423)
3424
3425xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003426 name = "avx2_ukernels_test_mode",
3427 hdrs = INTERNAL_HDRS,
3428 copts = [
3429 "-UNDEBUG",
3430 "-DXNN_TEST_MODE=1",
3431 ],
3432 gcc_copts = xnnpack_gcc_std_copts(),
3433 gcc_x86_copts = [
3434 "-mfma",
3435 "-mavx2",
3436 ],
3437 msvc_copts = xnnpack_msvc_std_copts(),
3438 msvc_x86_32_copts = ["/arch:AVX2"],
3439 msvc_x86_64_copts = ["/arch:AVX2"],
3440 x86_srcs = AVX2_UKERNELS,
3441 deps = [
3442 ":tables",
3443 "@FP16",
3444 "@pthreadpool",
3445 ],
3446)
3447
3448xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003449 name = "avx512f_ukernels",
3450 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003451 gcc_copts = xnnpack_gcc_std_copts(),
3452 gcc_x86_copts = ["-mavx512f"],
3453 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3454 msvc_copts = xnnpack_msvc_std_copts(),
3455 msvc_x86_32_copts = ["/arch:AVX512"],
3456 msvc_x86_64_copts = ["/arch:AVX512"],
3457 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003458 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003459 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003460 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003461 "@FP16",
3462 "@pthreadpool",
3463 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464)
3465
3466xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 name = "avx512f_ukernels_test_mode",
3468 hdrs = INTERNAL_HDRS,
3469 copts = [
3470 "-UNDEBUG",
3471 "-DXNN_TEST_MODE=1",
3472 ],
3473 gcc_copts = xnnpack_gcc_std_copts(),
3474 gcc_x86_copts = ["-mavx512f"],
3475 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3476 msvc_copts = xnnpack_msvc_std_copts(),
3477 msvc_x86_32_copts = ["/arch:AVX512"],
3478 msvc_x86_64_copts = ["/arch:AVX512"],
3479 msys_copts = ["-fno-asynchronous-unwind-tables"],
3480 x86_srcs = AVX512F_UKERNELS,
3481 deps = [
3482 ":tables",
3483 "@FP16",
3484 "@pthreadpool",
3485 ],
3486)
3487
3488xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003489 name = "avx512skx_ukernels",
3490 hdrs = INTERNAL_HDRS,
3491 gcc_copts = xnnpack_gcc_std_copts(),
3492 gcc_x86_copts = [
3493 "-mavx512f",
3494 "-mavx512cd",
3495 "-mavx512bw",
3496 "-mavx512dq",
3497 "-mavx512vl",
3498 ],
3499 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3500 msvc_copts = xnnpack_msvc_std_copts(),
3501 msvc_x86_32_copts = ["/arch:AVX512"],
3502 msvc_x86_64_copts = ["/arch:AVX512"],
3503 msys_copts = ["-fno-asynchronous-unwind-tables"],
3504 x86_srcs = AVX512SKX_UKERNELS,
3505 deps = [
3506 ":tables",
3507 "@FP16",
3508 "@pthreadpool",
3509 ],
3510)
3511
3512xnnpack_cc_library(
3513 name = "avx512skx_ukernels_test_mode",
3514 hdrs = INTERNAL_HDRS,
3515 copts = [
3516 "-UNDEBUG",
3517 "-DXNN_TEST_MODE=1",
3518 ],
3519 gcc_copts = xnnpack_gcc_std_copts(),
3520 gcc_x86_copts = [
3521 "-mavx512f",
3522 "-mavx512cd",
3523 "-mavx512bw",
3524 "-mavx512dq",
3525 "-mavx512vl",
3526 ],
3527 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3528 msvc_copts = xnnpack_msvc_std_copts(),
3529 msvc_x86_32_copts = ["/arch:AVX512"],
3530 msvc_x86_64_copts = ["/arch:AVX512"],
3531 msys_copts = ["-fno-asynchronous-unwind-tables"],
3532 x86_srcs = AVX512SKX_UKERNELS,
3533 deps = [
3534 ":tables",
3535 "@FP16",
3536 "@pthreadpool",
3537 ],
3538)
3539
3540xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003541 name = "asm_ukernels",
3542 hdrs = ["src/xnnpack/assembly.h"],
3543 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003544 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003545 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003546 wasm_srcs = WASM32_ASM_UKERNELS,
3547 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003548)
3549
Marat Dukhan3b59de22020-06-03 20:15:19 -07003550xnnpack_cc_library(
3551 name = "logging_utils",
3552 srcs = LOGGING_SRCS,
3553 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3554 copts = LOGGING_COPTS + [
3555 "-Isrc",
3556 "-Iinclude",
3557 ] + select({
3558 ":debug_build": [],
3559 "//conditions:default": xnnpack_min_size_copts(),
3560 }),
3561 gcc_copts = xnnpack_gcc_std_copts(),
3562 msvc_copts = xnnpack_msvc_std_copts(),
3563 visibility = xnnpack_visibility(),
3564 deps = [
3565 "@FP16",
3566 "@clog",
3567 "@pthreadpool",
3568 ],
3569)
3570
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571xnnpack_aggregate_library(
3572 name = "ukernels",
3573 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 ":neon_ukernels",
3575 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003576 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003577 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578 ":asm_ukernels",
3579 ],
3580 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 ":neon_ukernels",
3582 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003583 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003585 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 ":asm_ukernels",
3587 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003588 generic_deps = [
3589 ":scalar_ukernels",
3590 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003591 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003592 ":psimd_fastmath_ukernels",
3593 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003595 wasm_deps = [
3596 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003597 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003598 ],
3599 wasmsimd_deps = [
3600 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003601 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003602 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003605 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003606 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003608 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003609 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003610 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003612 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613 ],
3614)
3615
Marat Dukhan33fcf782020-05-24 14:27:15 -07003616xnnpack_aggregate_library(
3617 name = "ukernels_test_mode",
3618 aarch32_deps = [
3619 ":neon_ukernels_test_mode",
3620 ":neonfma_ukernels_test_mode",
3621 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003622 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003623 ":asm_ukernels",
3624 ],
3625 aarch64_deps = [
3626 ":neon_ukernels_test_mode",
3627 ":neonfma_ukernels_test_mode",
3628 ":neonv8_ukernels_test_mode",
3629 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003630 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003631 ":asm_ukernels",
3632 ],
3633 generic_deps = [
3634 ":scalar_ukernels_test_mode",
3635 ],
3636 psimd_deps = [
3637 ":psimd_fastmath_ukernels_test_mode",
3638 ":psimd_accmath_ukernels_test_mode",
3639 ],
3640 wasm_deps = [
3641 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003642 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003643 ],
3644 wasmsimd_deps = [
3645 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003646 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003647 ],
3648 x86_deps = [
3649 ":sse2_ukernels_test_mode",
3650 ":ssse3_ukernels_test_mode",
3651 ":sse41_ukernels_test_mode",
3652 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003653 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003654 ":fma3_ukernels_test_mode",
3655 ":avx2_ukernels_test_mode",
3656 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003657 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003658 ],
3659)
3660
Marat Dukhan86df3842020-06-17 14:03:31 -07003661# This target must be compatible with Android NDK r17 and compile with Clang 6.
3662# It is similar to :ukernels target, but excludes microkernels which require
3663# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003664xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003665 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003666 aarch32_deps = [
3667 ":neon_ukernels",
3668 ":neonfma_ukernels",
3669 ":neonv8_ukernels",
3670 ":asm_ukernels",
3671 ],
3672 aarch64_deps = [
3673 ":neon_ukernels",
3674 ":neonfma_ukernels",
3675 ":neonv8_ukernels",
3676 ":asm_ukernels",
3677 ],
3678 generic_deps = [
3679 ":scalar_ukernels",
3680 ],
3681 psimd_deps = [
3682 ":psimd_fastmath_ukernels",
3683 ":psimd_accmath_ukernels",
3684 ],
3685 wasm_deps = [
3686 ":wasm_ukernels",
3687 ],
3688 wasmsimd_deps = [
3689 ":wasm_ukernels",
3690 ],
3691 x86_deps = [
3692 ":sse2_ukernels",
3693 ":ssse3_ukernels",
3694 ":sse41_ukernels",
3695 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003696 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003697 ":fma3_ukernels",
3698 ":avx2_ukernels",
3699 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003700 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003701 ],
3702)
3703
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704xnnpack_cc_library(
3705 name = "im2col",
3706 srcs = ["src/im2col.c"],
3707 hdrs = [
3708 "src/xnnpack/common.h",
3709 "src/xnnpack/im2col.h",
3710 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003711 gcc_copts = xnnpack_gcc_std_copts(),
3712 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003713)
3714
3715xnnpack_cc_library(
3716 name = "indirection",
3717 srcs = ["src/indirection.c"],
3718 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003719 gcc_copts = xnnpack_gcc_std_copts(),
3720 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003721 deps = [
3722 "@FP16",
3723 "@FXdiv",
3724 "@pthreadpool",
3725 ],
3726)
3727
3728xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003729 name = "indirection_test_mode",
3730 srcs = ["src/indirection.c"],
3731 hdrs = INTERNAL_HDRS,
3732 copts = [
3733 "-UNDEBUG",
3734 "-DXNN_TEST_MODE=1",
3735 ],
3736 gcc_copts = xnnpack_gcc_std_copts(),
3737 msvc_copts = xnnpack_msvc_std_copts(),
3738 deps = [
3739 "@FP16",
3740 "@FXdiv",
3741 "@pthreadpool",
3742 ],
3743)
3744
3745xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003746 name = "packing",
3747 srcs = ["src/packing.c"],
3748 hdrs = INTERNAL_HDRS,
3749 gcc_copts = xnnpack_gcc_std_copts(),
3750 msvc_copts = xnnpack_msvc_std_copts(),
3751 deps = [
3752 "@FP16",
3753 "@FXdiv",
3754 "@pthreadpool",
3755 ],
3756)
3757
3758xnnpack_cc_library(
3759 name = "packing_test_mode",
3760 srcs = ["src/packing.c"],
3761 hdrs = INTERNAL_HDRS,
3762 copts = [
3763 "-UNDEBUG",
3764 "-DXNN_TEST_MODE=1",
3765 ],
3766 gcc_copts = xnnpack_gcc_std_copts(),
3767 msvc_copts = xnnpack_msvc_std_copts(),
3768 deps = [
3769 "@FP16",
3770 "@FXdiv",
3771 "@pthreadpool",
3772 ],
3773)
3774
3775xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 name = "operator_run",
3777 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003778 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003779 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003780 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3781 "//conditions:default": [],
3782 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003783 gcc_copts = xnnpack_gcc_std_copts(),
3784 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003785 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003786 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003787 "@FP16",
3788 "@FXdiv",
3789 "@clog",
3790 "@pthreadpool",
3791 ],
3792)
3793
Chao Mei6ddfc602020-05-13 22:29:36 -07003794xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003795 name = "operator_run_test_mode",
3796 srcs = ["src/operator-run.c"],
3797 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3798 copts = LOGGING_COPTS + [
3799 "-UNDEBUG",
3800 "-DXNN_TEST_MODE=1",
3801 ] + select({
3802 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3803 "//conditions:default": [],
3804 }),
3805 gcc_copts = xnnpack_gcc_std_copts(),
3806 msvc_copts = xnnpack_msvc_std_copts(),
3807 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003808 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003809 "@FP16",
3810 "@FXdiv",
3811 "@clog",
3812 "@pthreadpool",
3813 ],
3814)
3815
3816xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003817 name = "memory_planner",
3818 srcs = ["src/memory-planner.c"],
3819 hdrs = INTERNAL_HDRS,
3820 defines = select({
3821 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3822 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3823 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3824 }),
3825 gcc_copts = xnnpack_gcc_std_copts(),
3826 msvc_copts = xnnpack_msvc_std_copts(),
3827 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003828 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003829 "@pthreadpool",
3830 ],
3831)
3832
Marat Dukhan33fcf782020-05-24 14:27:15 -07003833xnnpack_cc_library(
3834 name = "memory_planner_test_mode",
3835 srcs = ["src/memory-planner.c"],
3836 hdrs = INTERNAL_HDRS,
3837 copts = [
3838 "-UNDEBUG",
3839 "-DXNN_TEST_MODE=1",
3840 ],
3841 defines = select({
3842 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3843 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3844 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3845 }),
3846 gcc_copts = xnnpack_gcc_std_copts(),
3847 msvc_copts = xnnpack_msvc_std_copts(),
3848 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003849 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003850 "@pthreadpool",
3851 ],
3852)
3853
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854cc_library(
3855 name = "enable_assembly",
3856 defines = select({
3857 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3858 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003859 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860 }),
3861)
3862
Marat Dukhan9de90e02020-06-18 16:04:12 -07003863cc_library(
3864 name = "enable_sparse",
3865 defines = select({
3866 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3867 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003868 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003869 }),
3870)
3871
Marat Dukhancf056b22019-10-07 10:26:29 -07003872xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873 name = "operators",
3874 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003875 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003877 ],
3878 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003879 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003880 "-Isrc",
3881 "-Iinclude",
3882 ] + select({
3883 ":debug_build": [],
3884 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003885 }) + select({
3886 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3887 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003888 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003889 gcc_copts = xnnpack_gcc_std_copts(),
3890 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003891 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003893 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003894 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003895 "@FP16",
3896 "@FXdiv",
3897 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003899 ],
3900)
3901
Marat Dukhan10a38082020-04-17 03:58:35 -07003902xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003903 name = "operators_test_mode",
3904 srcs = OPERATOR_SRCS + [
3905 "src/memory.c",
3906 "src/operator-delete.c",
3907 ],
3908 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3909 copts = LOGGING_COPTS + [
3910 "-Isrc",
3911 "-Iinclude",
3912 "-UNDEBUG",
3913 "-DXNN_TEST_MODE=1",
3914 ] + select({
3915 ":debug_build": [],
3916 "//conditions:default": xnnpack_min_size_copts(),
3917 }) + select({
3918 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3919 "//conditions:default": [],
3920 }),
3921 gcc_copts = xnnpack_gcc_std_copts(),
3922 msvc_copts = xnnpack_msvc_std_copts(),
3923 deps = [
3924 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003925 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003926 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003927 "@FP16",
3928 "@FXdiv",
3929 "@clog",
3930 "@pthreadpool",
3931 ],
3932)
3933
3934xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003935 name = "XNNPACK",
3936 srcs = [
3937 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003938 "src/runtime.c",
3939 "src/subgraph.c",
3940 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003941 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003942 hdrs = ["include/xnnpack.h"],
3943 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003944 "-Isrc",
3945 "-Iinclude",
3946 ] + select({
3947 ":debug_build": [],
3948 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003949 }) + select({
3950 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3951 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003952 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003953 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003954 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003955 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003956 visibility = xnnpack_visibility(),
3957 deps = [
3958 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003959 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003960 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003961 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003962 ":operator_run",
3963 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003964 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003965 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003966 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003967 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003968 ] + select({
3969 ":emscripten": [],
3970 "//conditions:default": ["@cpuinfo"],
3971 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003972)
3973
Marat Dukhan10a38082020-04-17 03:58:35 -07003974xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003975 name = "XNNPACK_test_mode",
3976 srcs = [
3977 "src/init.c",
3978 "src/runtime.c",
3979 "src/subgraph.c",
3980 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003981 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003982 hdrs = ["include/xnnpack.h"],
3983 copts = LOGGING_COPTS + [
3984 "-Isrc",
3985 "-Iinclude",
3986 "-UNDEBUG",
3987 "-DXNN_TEST_MODE=1",
3988 ] + select({
3989 ":debug_build": [],
3990 "//conditions:default": xnnpack_min_size_copts(),
3991 }) + select({
3992 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3993 "//conditions:default": [],
3994 }),
3995 gcc_copts = xnnpack_gcc_std_copts(),
3996 includes = ["include"],
3997 msvc_copts = xnnpack_msvc_std_copts(),
3998 visibility = xnnpack_visibility(),
3999 deps = [
4000 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004001 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004002 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004003 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004004 ":operator_run_test_mode",
4005 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004006 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004007 "@clog",
4008 "@FP16",
4009 "@pthreadpool",
4010 ] + select({
4011 ":emscripten": [],
4012 "//conditions:default": ["@cpuinfo"],
4013 }),
4014)
4015
4016xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004017 name = "xnnpack_f32",
4018 srcs = [
4019 "src/init.c",
4020 "src/runtime.c",
4021 "src/subgraph.c",
4022 "src/tensor.c",
4023 ] + SUBGRAPH_SRCS,
4024 hdrs = ["include/xnnpack.h"],
4025 copts = LOGGING_COPTS + [
4026 "-Isrc",
4027 "-Iinclude",
4028 ] + select({
4029 ":debug_build": [],
4030 "//conditions:default": xnnpack_min_size_copts(),
4031 }) + select({
4032 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4033 "//conditions:default": [],
4034 }),
4035 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004036 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004037 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004038 "XNN_NO_U8_OPERATORS",
4039 "XNN_NO_X8_OPERATORS",
4040 "XNN_NO_F16_OPERATORS",
4041 "XNN_NO_X16_OPERATORS",
4042 ],
4043 gcc_copts = xnnpack_gcc_std_copts(),
4044 includes = ["include"],
4045 msvc_copts = xnnpack_msvc_std_copts(),
4046 visibility = xnnpack_visibility(),
4047 deps = [
4048 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004049 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004050 ":logging_utils",
4051 ":memory_planner",
4052 ":operator_run",
4053 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004054 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004055 "@clog",
4056 "@FP16",
4057 "@pthreadpool",
4058 ] + select({
4059 ":emscripten": [],
4060 "//conditions:default": ["@cpuinfo"],
4061 }),
4062)
4063
4064xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004065 name = "xnnpack_operators_nhwc_f32",
4066 srcs = [
4067 "src/init.c",
4068 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004069 hdrs = ["include/xnnpack.h"],
4070 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004071 "-Isrc",
4072 "-Iinclude",
4073 ] + select({
4074 ":debug_build": [],
4075 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004076 }) + select({
4077 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4078 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004079 }),
4080 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004081 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004082 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004083 "XNN_NO_U8_OPERATORS",
4084 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004085 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004086 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004087 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004088 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004089 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090 visibility = xnnpack_visibility(),
4091 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004092 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004093 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004094 ":operator_run",
4095 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004096 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004097 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004099 ] + select({
4100 ":emscripten": [],
4101 "//conditions:default": ["@cpuinfo"],
4102 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004103)
4104
Marat Dukhancf056b22019-10-07 10:26:29 -07004105xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106 name = "bench_utils",
4107 srcs = ["bench/utils.cc"],
4108 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004109 deps = [
4110 "@com_google_benchmark//:benchmark",
4111 "@cpuinfo",
4112 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004113)
4114
Frank Barchard7e955972019-10-11 10:34:25 -07004115######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004116
4117xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004118 name = "qs8_gemm_bench",
4119 srcs = [
4120 "bench/gemm.h",
4121 "bench/qs8-gemm.cc",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4124 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4125)
4126
4127xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004128 name = "qs8_requantization_bench",
4129 srcs = [
4130 "bench/qs8-requantization.cc",
4131 "src/xnnpack/requantization-stubs.h",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + MICROKERNEL_BENCHMARK_HDRS,
4134 deps = MICROKERNEL_BENCHMARK_DEPS,
4135)
4136
4137xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004138 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004139 srcs = [
4140 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004141 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004142 "src/xnnpack/AlignedAllocator.h",
4143 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004144 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004145 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004146)
4147
4148xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004149 name = "qu8_requantization_bench",
4150 srcs = [
4151 "bench/qu8-requantization.cc",
4152 "src/xnnpack/requantization-stubs.h",
4153 "src/xnnpack/AlignedAllocator.h",
4154 ] + MICROKERNEL_BENCHMARK_HDRS,
4155 deps = MICROKERNEL_BENCHMARK_DEPS,
4156)
4157
4158xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004159 name = "f16_igemm_bench",
4160 srcs = [
4161 "bench/f16-igemm.cc",
4162 "bench/conv.h",
4163 "bench/google/conv.h",
4164 "src/xnnpack/AlignedAllocator.h",
4165 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004166 deps = MICROKERNEL_BENCHMARK_DEPS + [
4167 ":indirection",
4168 ":packing",
4169 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004170)
4171
4172xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004173 name = "f16_gemm_bench",
4174 srcs = [
4175 "bench/f16-gemm.cc",
4176 "bench/gemm.h",
4177 "src/xnnpack/AlignedAllocator.h",
4178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004179 deps = MICROKERNEL_BENCHMARK_DEPS + [
4180 ":packing",
4181 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004182)
4183
4184xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004185 name = "f16_spmm_bench",
4186 srcs = [
4187 "bench/f16-spmm.cc",
4188 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004189 "src/xnnpack/AlignedAllocator.h",
4190 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004191 deps = MICROKERNEL_BENCHMARK_DEPS,
4192)
4193
4194xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195 name = "f32_igemm_bench",
4196 srcs = [
4197 "bench/f32-igemm.cc",
4198 "bench/conv.h",
4199 "src/xnnpack/AlignedAllocator.h",
4200 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004201 deps = MICROKERNEL_BENCHMARK_DEPS + [
4202 ":indirection",
4203 ":packing",
4204 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205)
4206
4207xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004208 name = "f16_relu_bench",
4209 srcs = [
4210 "bench/f16-relu.cc",
4211 "src/xnnpack/AlignedAllocator.h",
4212 ] + MICROKERNEL_BENCHMARK_HDRS,
4213 deps = MICROKERNEL_BENCHMARK_DEPS,
4214)
4215
4216xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 name = "f32_conv_hwc_bench",
4218 srcs = [
4219 "bench/f32-conv-hwc.cc",
4220 "bench/dconv.h",
4221 "src/xnnpack/AlignedAllocator.h",
4222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004223 deps = MICROKERNEL_BENCHMARK_DEPS + [
4224 ":packing",
4225 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226)
4227
4228xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004229 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004230 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004231 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004232 "bench/dconv.h",
4233 "src/xnnpack/AlignedAllocator.h",
4234 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004235 deps = MICROKERNEL_BENCHMARK_DEPS + [
4236 ":packing",
4237 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004238)
4239
4240xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004241 name = "f16_dwconv_bench",
4242 srcs = [
4243 "bench/f16-dwconv.cc",
4244 "bench/dwconv.h",
4245 "bench/google/dwconv.h",
4246 "src/xnnpack/AlignedAllocator.h",
4247 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004248 deps = MICROKERNEL_BENCHMARK_DEPS + [
4249 ":indirection",
4250 ":packing",
4251 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004252)
4253
4254xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255 name = "f32_dwconv_bench",
4256 srcs = [
4257 "bench/f32-dwconv.cc",
4258 "bench/dwconv.h",
4259 "src/xnnpack/AlignedAllocator.h",
4260 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004261 deps = MICROKERNEL_BENCHMARK_DEPS + [
4262 ":indirection",
4263 ":packing",
4264 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265)
4266
4267xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004268 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004270 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271 "bench/dwconv.h",
4272 "src/xnnpack/AlignedAllocator.h",
4273 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004274 deps = MICROKERNEL_BENCHMARK_DEPS + [
4275 ":indirection",
4276 ":packing",
4277 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278)
4279
4280xnnpack_benchmark(
4281 name = "f32_gemm_bench",
4282 srcs = [
4283 "bench/f32-gemm.cc",
4284 "bench/gemm.h",
4285 "src/xnnpack/AlignedAllocator.h",
4286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004287 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004288 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004292 name = "f32_hswish_bench",
4293 srcs = [
4294 "bench/f32-hswish.cc",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + MICROKERNEL_BENCHMARK_HDRS,
4297 deps = MICROKERNEL_BENCHMARK_DEPS,
4298)
4299
4300xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004301 name = "f32_raddexpminusmax_bench",
4302 srcs = [
4303 "bench/f32-raddexpminusmax.cc",
4304 "src/xnnpack/AlignedAllocator.h",
4305 ] + MICROKERNEL_BENCHMARK_HDRS,
4306 deps = MICROKERNEL_BENCHMARK_DEPS,
4307)
4308
4309xnnpack_benchmark(
4310 name = "f32_raddextexp_bench",
4311 srcs = [
4312 "bench/f32-raddextexp.cc",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + MICROKERNEL_BENCHMARK_HDRS,
4315 deps = MICROKERNEL_BENCHMARK_DEPS,
4316)
4317
4318xnnpack_benchmark(
4319 name = "f32_raddstoreexpminusmax_bench",
4320 srcs = [
4321 "bench/f32-raddstoreexpminusmax.cc",
4322 "src/xnnpack/AlignedAllocator.h",
4323 ] + MICROKERNEL_BENCHMARK_HDRS,
4324 deps = MICROKERNEL_BENCHMARK_DEPS,
4325)
4326
4327xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004328 name = "f32_relu_bench",
4329 srcs = [
4330 "bench/f32-relu.cc",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + MICROKERNEL_BENCHMARK_HDRS,
4333 deps = MICROKERNEL_BENCHMARK_DEPS,
4334)
4335
4336xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337 name = "f32_rmax_bench",
4338 srcs = [
4339 "bench/f32-rmax.cc",
4340 "src/xnnpack/AlignedAllocator.h",
4341 ] + MICROKERNEL_BENCHMARK_HDRS,
4342 deps = MICROKERNEL_BENCHMARK_DEPS,
4343)
4344
4345xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004346 name = "f32_sigmoid_bench",
4347 srcs = [
4348 "bench/f32-sigmoid.cc",
4349 "src/xnnpack/AlignedAllocator.h",
4350 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004351 deps = MICROKERNEL_BENCHMARK_DEPS,
4352)
4353
4354xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355 name = "f32_spmm_bench",
4356 srcs = [
4357 "bench/f32-spmm.cc",
4358 "bench/gemm.h",
4359 "src/xnnpack/AlignedAllocator.h",
4360 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004361 deps = MICROKERNEL_BENCHMARK_DEPS,
4362)
4363
4364xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004365 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004366 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004367 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004368 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004369 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004370 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004371)
4372
4373xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004374 name = "f32_vscaleexpminusmax_bench",
4375 srcs = [
4376 "bench/f32-vscaleexpminusmax.cc",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + MICROKERNEL_BENCHMARK_HDRS,
4379 deps = MICROKERNEL_BENCHMARK_DEPS,
4380)
4381
4382xnnpack_benchmark(
4383 name = "f32_vscaleextexp_bench",
4384 srcs = [
4385 "bench/f32-vscaleextexp.cc",
4386 "src/xnnpack/AlignedAllocator.h",
4387 ] + MICROKERNEL_BENCHMARK_HDRS,
4388 deps = MICROKERNEL_BENCHMARK_DEPS,
4389)
4390
4391xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004392 name = "f32_vsqrt_bench",
4393 srcs = [
4394 "bench/f32-vsqrt.cc",
4395 "src/xnnpack/AlignedAllocator.h",
4396 ] + MICROKERNEL_BENCHMARK_HDRS,
4397 deps = MICROKERNEL_BENCHMARK_DEPS,
4398)
4399
4400xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004401 name = "f32_im2col_gemm_bench",
4402 srcs = [
4403 "bench/f32-im2col-gemm.cc",
4404 "bench/conv.h",
4405 "src/xnnpack/AlignedAllocator.h",
4406 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004407 deps = MICROKERNEL_BENCHMARK_DEPS + [
4408 ":im2col",
4409 ":packing",
4410 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004411)
4412
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004413xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004414 name = "rounding_bench",
4415 srcs = [
4416 "bench/rounding.cc",
4417 "src/xnnpack/math-stubs.h",
4418 "src/xnnpack/AlignedAllocator.h",
4419 ] + MICROKERNEL_BENCHMARK_HDRS,
4420 deps = MICROKERNEL_BENCHMARK_DEPS,
4421)
4422
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423########################### Benchmarks for operators ###########################
4424
4425xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426 name = "average_pooling_bench",
4427 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004428 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004429 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004430 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004431)
4432
4433xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004434 name = "bankers_rounding_bench",
4435 srcs = ["bench/bankers-rounding.cc"],
4436 copts = xnnpack_optional_tflite_copts(),
4437 tags = ["nowin32"],
4438 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4439)
4440
4441xnnpack_benchmark(
4442 name = "ceiling_bench",
4443 srcs = ["bench/ceiling.cc"],
4444 copts = xnnpack_optional_tflite_copts(),
4445 tags = ["nowin32"],
4446 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4447)
4448
4449xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450 name = "channel_shuffle_bench",
4451 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004452 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004453)
4454
4455xnnpack_benchmark(
4456 name = "convolution_bench",
4457 srcs = ["bench/convolution.cc"],
4458 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004459 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004460 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461)
4462
4463xnnpack_benchmark(
4464 name = "deconvolution_bench",
4465 srcs = ["bench/deconvolution.cc"],
4466 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004467 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004468 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004469)
4470
4471xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004472 name = "floor_bench",
4473 srcs = ["bench/floor.cc"],
4474 copts = xnnpack_optional_tflite_copts(),
4475 tags = ["nowin32"],
4476 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4477)
4478
4479xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 name = "global_average_pooling_bench",
4481 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004482 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483)
4484
4485xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004486 name = "hardswish_bench",
4487 srcs = ["bench/hardswish.cc"],
4488 copts = xnnpack_optional_tflite_copts(),
4489 tags = ["nowin32"],
4490 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4491)
4492
4493xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494 name = "max_pooling_bench",
4495 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004496 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497)
4498
4499xnnpack_benchmark(
4500 name = "sigmoid_bench",
4501 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004502 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004503 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004504 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505)
4506
4507xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004508 name = "prelu_bench",
4509 srcs = ["bench/prelu.cc"],
4510 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004511 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004512 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004513)
4514
4515xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004516 name = "softmax_bench",
4517 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004518 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004519 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004520 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521)
4522
Marat Dukhan87727142020-06-24 15:24:10 -07004523xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004524 name = "square_root_bench",
4525 srcs = ["bench/square-root.cc"],
4526 copts = xnnpack_optional_tflite_copts(),
4527 tags = ["nowin32"],
4528 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4529)
4530
4531xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004532 name = "truncation_bench",
4533 srcs = ["bench/truncation.cc"],
4534 deps = OPERATOR_BENCHMARK_DEPS,
4535)
4536
Marat Dukhanc068bb62019-10-04 13:24:39 -07004537############################# End-to-end benchmarks ############################
4538
4539cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004540 name = "fp32_mobilenet_v1",
4541 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004542 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004543 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004544 linkstatic = True,
4545 deps = [
4546 ":XNNPACK",
4547 "@pthreadpool",
4548 ],
4549)
4550
4551cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004552 name = "fp16_mobilenet_v1",
4553 srcs = ["models/fp16-mobilenet-v1.cc"],
4554 hdrs = ["models/models.h"],
4555 copts = xnnpack_std_cxxopts(),
4556 linkstatic = True,
4557 deps = [
4558 ":XNNPACK",
4559 "@FP16",
4560 "@pthreadpool",
4561 ],
4562)
4563
4564cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004565 name = "qs8_mobilenet_v1",
4566 srcs = ["models/qs8-mobilenet-v1.cc"],
4567 hdrs = ["models/models.h"],
4568 copts = xnnpack_std_cxxopts(),
4569 linkstatic = True,
4570 deps = [
4571 ":XNNPACK",
4572 "@pthreadpool",
4573 ],
4574)
4575
4576cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004577 name = "qs8_mobilenet_v2",
4578 srcs = ["models/qs8-mobilenet-v2.cc"],
4579 hdrs = ["models/models.h"],
4580 copts = xnnpack_std_cxxopts(),
4581 linkstatic = True,
4582 deps = [
4583 ":XNNPACK",
4584 "@pthreadpool",
4585 ],
4586)
4587
4588cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004589 name = "fp32_mobilenet_v2",
4590 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004591 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004592 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004593 linkstatic = True,
4594 deps = [
4595 ":XNNPACK",
4596 "@pthreadpool",
4597 ],
4598)
4599
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004600cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004601 name = "fp16_mobilenet_v2",
4602 srcs = ["models/fp16-mobilenet-v2.cc"],
4603 hdrs = ["models/models.h"],
4604 copts = xnnpack_std_cxxopts(),
4605 linkstatic = True,
4606 deps = [
4607 ":XNNPACK",
4608 "@FP16",
4609 "@pthreadpool",
4610 ],
4611)
4612
4613cc_library(
4614 name = "fp32_mobilenet_v3_large",
4615 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004616 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004617 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004618 linkstatic = True,
4619 deps = [
4620 ":XNNPACK",
4621 "@pthreadpool",
4622 ],
4623)
4624
4625cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004626 name = "fp32_mobilenet_v3_small",
4627 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004628 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004629 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004630 linkstatic = True,
4631 deps = [
4632 ":XNNPACK",
4633 "@pthreadpool",
4634 ],
4635)
4636
Marat Dukhanc068bb62019-10-04 13:24:39 -07004637xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004638 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004639 srcs = [
4640 "bench/f32-dwconv-e2e.cc",
4641 "bench/end2end.h",
4642 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004643 deps = MICROKERNEL_BENCHMARK_DEPS + [
4644 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004645 ":fp32_mobilenet_v1",
4646 ":fp32_mobilenet_v2",
4647 ":fp32_mobilenet_v3_large",
4648 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004649 ],
4650)
4651
4652xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004653 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004654 srcs = [
4655 "bench/f32-gemm-e2e.cc",
4656 "bench/end2end.h",
4657 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004658 deps = MICROKERNEL_BENCHMARK_DEPS + [
4659 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004660 ":fp32_mobilenet_v1",
4661 ":fp32_mobilenet_v2",
4662 ":fp32_mobilenet_v3_large",
4663 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004664 ],
4665)
4666
4667xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004668 name = "end2end_bench",
4669 srcs = ["bench/end2end.cc"],
4670 deps = [
4671 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004672 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004673 ":fp16_mobilenet_v1",
4674 ":fp16_mobilenet_v2",
4675 ":fp32_mobilenet_v1",
4676 ":fp32_mobilenet_v2",
4677 ":fp32_mobilenet_v3_large",
4678 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004679 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004680 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004681 "@pthreadpool",
4682 ],
4683)
4684
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004685#################### Accuracy evaluation for math functions ####################
4686
4687xnnpack_benchmark(
4688 name = "f32_exp_eval",
4689 srcs = [
4690 "eval/f32-exp.cc",
4691 "src/xnnpack/AlignedAllocator.h",
4692 ] + ACCURACY_EVAL_HDRS,
4693 deps = ACCURACY_EVAL_DEPS,
4694)
4695
Marat Dukhan515c9772019-10-17 18:07:57 -07004696xnnpack_benchmark(
4697 name = "f32_expminus_eval",
4698 srcs = [
4699 "eval/f32-expminus.cc",
4700 "src/xnnpack/AlignedAllocator.h",
4701 ] + ACCURACY_EVAL_HDRS,
4702 deps = ACCURACY_EVAL_DEPS,
4703)
4704
Marat Dukhan98ba4412019-10-23 02:14:28 -07004705xnnpack_benchmark(
4706 name = "f32_extexp_eval",
4707 srcs = [
4708 "eval/f32-extexp.cc",
4709 "src/xnnpack/AlignedAllocator.h",
4710 ] + ACCURACY_EVAL_HDRS,
4711 deps = ACCURACY_EVAL_DEPS,
4712)
4713
Marat Dukhan8853b822020-05-07 12:19:01 -07004714xnnpack_unit_test(
4715 name = "f32_roundne_eval",
4716 srcs = [
4717 "eval/f32-roundne.cc",
4718 "src/xnnpack/AlignedAllocator.h",
4719 "src/xnnpack/math-stubs.h",
4720 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004721 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004725xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004726 name = "f32_roundd_eval",
4727 srcs = [
4728 "eval/f32-roundd.cc",
4729 "src/xnnpack/AlignedAllocator.h",
4730 "src/xnnpack/math-stubs.h",
4731 ] + MICROKERNEL_TEST_HDRS,
4732 automatic = False,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
4737 name = "f32_roundu_eval",
4738 srcs = [
4739 "eval/f32-roundu.cc",
4740 "src/xnnpack/AlignedAllocator.h",
4741 "src/xnnpack/math-stubs.h",
4742 ] + MICROKERNEL_TEST_HDRS,
4743 automatic = False,
4744 deps = MICROKERNEL_TEST_DEPS,
4745)
4746
4747xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004748 name = "f32_roundz_eval",
4749 srcs = [
4750 "eval/f32-roundz.cc",
4751 "src/xnnpack/AlignedAllocator.h",
4752 "src/xnnpack/math-stubs.h",
4753 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004754 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004755 deps = MICROKERNEL_TEST_DEPS,
4756)
4757
Marat Dukhan346a9e52019-11-15 09:06:30 -08004758xnnpack_benchmark(
4759 name = "f32_sigmoid_eval",
4760 srcs = [
4761 "eval/f32-sigmoid.cc",
4762 "src/xnnpack/AlignedAllocator.h",
4763 ] + ACCURACY_EVAL_HDRS,
4764 deps = ACCURACY_EVAL_DEPS,
4765)
4766
Marat Dukhan84000762020-06-29 18:38:43 -07004767xnnpack_benchmark(
4768 name = "f32_sqrt_eval",
4769 srcs = [
4770 "eval/f32-sqrt.cc",
4771 "src/xnnpack/AlignedAllocator.h",
4772 ] + ACCURACY_EVAL_HDRS,
4773 deps = ACCURACY_EVAL_DEPS + [
4774 ":bench_utils",
4775 ],
4776)
4777
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778######################### Unit tests for micro-kernels #########################
4779
4780xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004781 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004783 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004784 "test/gemm-microkernel-tester.h",
4785 "src/xnnpack/AlignedAllocator.h",
4786 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004787 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004788)
4789
4790xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004791 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004792 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004793 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004794 "test/spmm-microkernel-tester.h",
4795 "src/xnnpack/AlignedAllocator.h",
4796 ] + MICROKERNEL_TEST_HDRS,
4797 deps = MICROKERNEL_TEST_DEPS,
4798)
4799
4800xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801 name = "f32_argmaxpool_test",
4802 srcs = [
4803 "test/f32-argmaxpool.cc",
4804 "test/argmaxpool-microkernel-tester.h",
4805 "src/xnnpack/AlignedAllocator.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004811 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004813 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004814 "test/avgpool-microkernel-tester.h",
4815 "src/xnnpack/AlignedAllocator.h",
4816 ] + MICROKERNEL_TEST_HDRS,
4817 deps = MICROKERNEL_TEST_DEPS,
4818)
4819
4820xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004821 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004822 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004823 "test/f32-ibilinear.cc",
4824 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004825 "src/xnnpack/AlignedAllocator.h",
4826 ] + MICROKERNEL_TEST_HDRS,
4827 deps = MICROKERNEL_TEST_DEPS,
4828)
4829
4830xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004831 name = "f16_clamp_test",
4832 srcs = [
4833 "test/f16-clamp.cc",
4834 "test/clamp-microkernel-tester.h",
4835 ] + MICROKERNEL_TEST_HDRS,
4836 deps = MICROKERNEL_TEST_DEPS,
4837)
4838
4839xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004840 name = "f32_clamp_test",
4841 srcs = [
4842 "test/f32-clamp.cc",
4843 "test/clamp-microkernel-tester.h",
4844 ] + MICROKERNEL_TEST_HDRS,
4845 deps = MICROKERNEL_TEST_DEPS,
4846)
4847
4848xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004849 name = "f32_igemm_test",
4850 srcs = [
4851 "test/f32-igemm.cc",
4852 "test/gemm-microkernel-tester.h",
4853 "src/xnnpack/AlignedAllocator.h",
4854 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004855 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004856)
4857
4858xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004859 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004861 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004862 "test/gemm-microkernel-tester.h",
4863 "src/xnnpack/AlignedAllocator.h",
4864 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004865 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866)
4867
4868xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004869 name = "f32_igemm_minmax_test",
4870 srcs = [
4871 "test/f32-igemm-minmax.cc",
4872 "test/gemm-microkernel-tester.h",
4873 "src/xnnpack/AlignedAllocator.h",
4874 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004875 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004876)
4877
4878xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004879 name = "f16_igemm_minmax_test",
4880 srcs = [
4881 "test/f16-igemm-minmax.cc",
4882 "test/gemm-microkernel-tester.h",
4883 "src/xnnpack/AlignedAllocator.h",
4884 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004885 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004886)
4887
4888xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004889 name = "f32_conv_hwc_test",
4890 srcs = [
4891 "test/f32-conv-hwc.cc",
4892 "test/conv-hwc-microkernel-tester.h",
4893 "src/xnnpack/AlignedAllocator.h",
4894 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004895 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004896)
4897
4898xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004899 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004900 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004901 "test/f32-conv-hwc2chw.cc",
4902 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004903 "src/xnnpack/AlignedAllocator.h",
4904 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004905 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004906)
4907
4908xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004909 name = "f32_dwconv_test",
4910 srcs = [
4911 "test/f32-dwconv.cc",
4912 "test/dwconv-microkernel-tester.h",
4913 "src/xnnpack/AlignedAllocator.h",
4914 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004915 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004916)
4917
4918xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004919 name = "f16_dwconv_minmax_test",
4920 srcs = [
4921 "test/f16-dwconv-minmax.cc",
4922 "test/dwconv-microkernel-tester.h",
4923 "src/xnnpack/AlignedAllocator.h",
4924 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004925 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004926)
4927
4928xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004929 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004930 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004931 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004932 "test/dwconv-microkernel-tester.h",
4933 "src/xnnpack/AlignedAllocator.h",
4934 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004935 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004936)
4937
4938xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004939 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004940 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004941 "test/f32-dwconv-chw.cc",
4942 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004943 "src/xnnpack/AlignedAllocator.h",
4944 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004945 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946)
4947
4948xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004949 name = "f16_gavgpool_minmax_test",
4950 srcs = [
4951 "test/f16-gavgpool-minmax.cc",
4952 "test/gavgpool-microkernel-tester.h",
4953 "src/xnnpack/AlignedAllocator.h",
4954 ] + MICROKERNEL_TEST_HDRS,
4955 deps = MICROKERNEL_TEST_DEPS,
4956)
4957
4958xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004959 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004961 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004962 "test/gavgpool-microkernel-tester.h",
4963 "src/xnnpack/AlignedAllocator.h",
4964 ] + MICROKERNEL_TEST_HDRS,
4965 deps = MICROKERNEL_TEST_DEPS,
4966)
4967
4968xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004969 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004970 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004971 "test/f32-gavgpool-cw.cc",
4972 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973 "src/xnnpack/AlignedAllocator.h",
4974 ] + MICROKERNEL_TEST_HDRS,
4975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
4978xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004979 name = "f32_gemm_test",
4980 srcs = [
4981 "test/f32-gemm.cc",
4982 "test/gemm-microkernel-tester.h",
4983 "src/xnnpack/AlignedAllocator.h",
4984 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004985 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004986)
4987
4988xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004989 name = "f32_gemm_relu_test",
4990 srcs = [
4991 "test/f32-gemm-relu.cc",
4992 "test/gemm-microkernel-tester.h",
4993 "src/xnnpack/AlignedAllocator.h",
4994 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004995 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004996)
4997
4998xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004999 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005001 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002 "test/gemm-microkernel-tester.h",
5003 "src/xnnpack/AlignedAllocator.h",
5004 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005005 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005006)
5007
5008xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005009 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005010 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005011 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012 "test/gemm-microkernel-tester.h",
5013 "src/xnnpack/AlignedAllocator.h",
5014 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005015 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016)
5017
5018xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005019 name = "f16_hswish_test",
5020 srcs = [
5021 "test/f16-hswish.cc",
5022 "test/hswish-microkernel-tester.h",
5023 ] + MICROKERNEL_TEST_HDRS,
5024 deps = MICROKERNEL_TEST_DEPS,
5025)
5026
5027xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028 name = "f32_hswish_test",
5029 srcs = [
5030 "test/f32-hswish.cc",
5031 "test/hswish-microkernel-tester.h",
5032 ] + MICROKERNEL_TEST_HDRS,
5033 deps = MICROKERNEL_TEST_DEPS,
5034)
5035
5036xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005037 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005039 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040 "test/maxpool-microkernel-tester.h",
5041 ] + MICROKERNEL_TEST_HDRS,
5042 deps = MICROKERNEL_TEST_DEPS,
5043)
5044
5045xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005046 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005048 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 "test/avgpool-microkernel-tester.h",
5050 "src/xnnpack/AlignedAllocator.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005056 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005058 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059 "test/gemm-microkernel-tester.h",
5060 "src/xnnpack/AlignedAllocator.h",
5061 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005062 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063)
5064
5065xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005066 name = "f16_prelu_test",
5067 srcs = [
5068 "test/f16-prelu.cc",
5069 "test/prelu-microkernel-tester.h",
5070 "src/xnnpack/AlignedAllocator.h",
5071 ] + MICROKERNEL_TEST_HDRS,
5072 deps = MICROKERNEL_TEST_DEPS,
5073)
5074
5075xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076 name = "f32_prelu_test",
5077 srcs = [
5078 "test/f32-prelu.cc",
5079 "test/prelu-microkernel-tester.h",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + MICROKERNEL_TEST_HDRS,
5082 deps = MICROKERNEL_TEST_DEPS,
5083)
5084
5085xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005086 name = "f32_raddexpminusmax_test",
5087 srcs = [
5088 "test/f32-raddexpminusmax.cc",
5089 "test/raddexpminusmax-microkernel-tester.h",
5090 ] + MICROKERNEL_TEST_HDRS,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005095 name = "f32_raddextexp_test",
5096 srcs = [
5097 "test/f32-raddextexp.cc",
5098 "test/raddextexp-microkernel-tester.h",
5099 ] + MICROKERNEL_TEST_HDRS,
5100 deps = MICROKERNEL_TEST_DEPS,
5101)
5102
5103xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005104 name = "f32_raddstoreexpminusmax_test",
5105 srcs = [
5106 "test/f32-raddstoreexpminusmax.cc",
5107 "test/raddstoreexpminusmax-microkernel-tester.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005113 name = "f32_relu_test",
5114 srcs = [
5115 "test/f32-relu.cc",
5116 "test/vunary-microkernel-tester.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005122 name = "f32_rmax_test",
5123 srcs = [
5124 "test/f32-rmax.cc",
5125 "test/rmax-microkernel-tester.h",
5126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005131 name = "f32_sigmoid_test",
5132 srcs = [
5133 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005134 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005135 ] + MICROKERNEL_TEST_HDRS,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005140 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005141 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005142 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143 "test/spmm-microkernel-tester.h",
5144 "src/xnnpack/AlignedAllocator.h",
5145 ] + MICROKERNEL_TEST_HDRS,
5146 deps = MICROKERNEL_TEST_DEPS,
5147)
5148
5149xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005150 name = "f16_vadd_minmax_test",
5151 srcs = [
5152 "test/f16-vadd-minmax.cc",
5153 "test/vbinary-microkernel-tester.h",
5154 ] + MICROKERNEL_TEST_HDRS,
5155 deps = MICROKERNEL_TEST_DEPS,
5156)
5157
5158xnnpack_unit_test(
5159 name = "f16_vaddc_minmax_test",
5160 srcs = [
5161 "test/f16-vaddc-minmax.cc",
5162 "test/vbinaryc-microkernel-tester.h",
5163 ] + MICROKERNEL_TEST_HDRS,
5164 deps = MICROKERNEL_TEST_DEPS,
5165)
5166
5167xnnpack_unit_test(
5168 name = "f16_vdiv_minmax_test",
5169 srcs = [
5170 "test/f16-vdiv-minmax.cc",
5171 "test/vbinary-microkernel-tester.h",
5172 ] + MICROKERNEL_TEST_HDRS,
5173 deps = MICROKERNEL_TEST_DEPS,
5174)
5175
5176xnnpack_unit_test(
5177 name = "f16_vdivc_minmax_test",
5178 srcs = [
5179 "test/f16-vdivc-minmax.cc",
5180 "test/vbinaryc-microkernel-tester.h",
5181 ] + MICROKERNEL_TEST_HDRS,
5182 deps = MICROKERNEL_TEST_DEPS,
5183)
5184
5185xnnpack_unit_test(
5186 name = "f16_vrdivc_minmax_test",
5187 srcs = [
5188 "test/f16-vrdivc-minmax.cc",
5189 "test/vbinaryc-microkernel-tester.h",
5190 ] + MICROKERNEL_TEST_HDRS,
5191 deps = MICROKERNEL_TEST_DEPS,
5192)
5193
5194xnnpack_unit_test(
5195 name = "f16_vmax_test",
5196 srcs = [
5197 "test/f16-vmax.cc",
5198 "test/vbinary-microkernel-tester.h",
5199 ] + MICROKERNEL_TEST_HDRS,
5200 deps = MICROKERNEL_TEST_DEPS,
5201)
5202
5203xnnpack_unit_test(
5204 name = "f16_vmaxc_test",
5205 srcs = [
5206 "test/f16-vmaxc.cc",
5207 "test/vbinaryc-microkernel-tester.h",
5208 ] + MICROKERNEL_TEST_HDRS,
5209 deps = MICROKERNEL_TEST_DEPS,
5210)
5211
5212xnnpack_unit_test(
5213 name = "f16_vmin_test",
5214 srcs = [
5215 "test/f16-vmin.cc",
5216 "test/vbinary-microkernel-tester.h",
5217 ] + MICROKERNEL_TEST_HDRS,
5218 deps = MICROKERNEL_TEST_DEPS,
5219)
5220
5221xnnpack_unit_test(
5222 name = "f16_vminc_test",
5223 srcs = [
5224 "test/f16-vminc.cc",
5225 "test/vbinaryc-microkernel-tester.h",
5226 ] + MICROKERNEL_TEST_HDRS,
5227 deps = MICROKERNEL_TEST_DEPS,
5228)
5229
5230xnnpack_unit_test(
5231 name = "f16_vmul_minmax_test",
5232 srcs = [
5233 "test/f16-vmul-minmax.cc",
5234 "test/vbinary-microkernel-tester.h",
5235 ] + MICROKERNEL_TEST_HDRS,
5236 deps = MICROKERNEL_TEST_DEPS,
5237)
5238
5239xnnpack_unit_test(
5240 name = "f16_vmulc_minmax_test",
5241 srcs = [
5242 "test/f16-vmulc-minmax.cc",
5243 "test/vbinaryc-microkernel-tester.h",
5244 ] + MICROKERNEL_TEST_HDRS,
5245 deps = MICROKERNEL_TEST_DEPS,
5246)
5247
5248xnnpack_unit_test(
5249 name = "f16_vsub_minmax_test",
5250 srcs = [
5251 "test/f16-vsub-minmax.cc",
5252 "test/vbinary-microkernel-tester.h",
5253 ] + MICROKERNEL_TEST_HDRS,
5254 deps = MICROKERNEL_TEST_DEPS,
5255)
5256
5257xnnpack_unit_test(
5258 name = "f16_vsubc_minmax_test",
5259 srcs = [
5260 "test/f16-vsubc-minmax.cc",
5261 "test/vbinaryc-microkernel-tester.h",
5262 ] + MICROKERNEL_TEST_HDRS,
5263 deps = MICROKERNEL_TEST_DEPS,
5264)
5265
5266xnnpack_unit_test(
5267 name = "f16_vrsubc_minmax_test",
5268 srcs = [
5269 "test/f16-vrsubc-minmax.cc",
5270 "test/vbinaryc-microkernel-tester.h",
5271 ] + MICROKERNEL_TEST_HDRS,
5272 deps = MICROKERNEL_TEST_DEPS,
5273)
5274
5275xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005276 name = "f32_vabs_test",
5277 srcs = [
5278 "test/f32-vabs.cc",
5279 "test/vunary-microkernel-tester.h",
5280 ] + MICROKERNEL_TEST_HDRS,
5281 deps = MICROKERNEL_TEST_DEPS,
5282)
5283
5284xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005285 name = "f32_vadd_test",
5286 srcs = [
5287 "test/f32-vadd.cc",
5288 "test/vbinary-microkernel-tester.h",
5289 ] + MICROKERNEL_TEST_HDRS,
5290 deps = MICROKERNEL_TEST_DEPS,
5291)
5292
5293xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005294 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005296 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005297 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005298 ] + MICROKERNEL_TEST_HDRS,
5299 deps = MICROKERNEL_TEST_DEPS,
5300)
5301
5302xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005303 name = "f32_vadd_relu_test",
5304 srcs = [
5305 "test/f32-vadd-relu.cc",
5306 "test/vbinary-microkernel-tester.h",
5307 ] + MICROKERNEL_TEST_HDRS,
5308 deps = MICROKERNEL_TEST_DEPS,
5309)
5310
5311xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005312 name = "f32_vaddc_test",
5313 srcs = [
5314 "test/f32-vaddc.cc",
5315 "test/vbinaryc-microkernel-tester.h",
5316 ] + MICROKERNEL_TEST_HDRS,
5317 deps = MICROKERNEL_TEST_DEPS,
5318)
5319
5320xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005321 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005322 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005323 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005324 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 ] + MICROKERNEL_TEST_HDRS,
5326 deps = MICROKERNEL_TEST_DEPS,
5327)
5328
5329xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005330 name = "f32_vaddc_relu_test",
5331 srcs = [
5332 "test/f32-vaddc-relu.cc",
5333 "test/vbinaryc-microkernel-tester.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005339 name = "f32_vdiv_test",
5340 srcs = [
5341 "test/f32-vdiv.cc",
5342 "test/vbinary-microkernel-tester.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005348 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005349 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005350 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005351 "test/vbinary-microkernel-tester.h",
5352 ] + MICROKERNEL_TEST_HDRS,
5353 deps = MICROKERNEL_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005357 name = "f32_vdiv_relu_test",
5358 srcs = [
5359 "test/f32-vdiv-relu.cc",
5360 "test/vbinary-microkernel-tester.h",
5361 ] + MICROKERNEL_TEST_HDRS,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
5365xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005366 name = "f32_vdivc_test",
5367 srcs = [
5368 "test/f32-vdivc.cc",
5369 "test/vbinaryc-microkernel-tester.h",
5370 ] + MICROKERNEL_TEST_HDRS,
5371 deps = MICROKERNEL_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005375 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005376 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005377 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005378 "test/vbinaryc-microkernel-tester.h",
5379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005384 name = "f32_vdivc_relu_test",
5385 srcs = [
5386 "test/f32-vdivc-relu.cc",
5387 "test/vbinaryc-microkernel-tester.h",
5388 ] + MICROKERNEL_TEST_HDRS,
5389 deps = MICROKERNEL_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005393 name = "f32_vrdivc_test",
5394 srcs = [
5395 "test/f32-vrdivc.cc",
5396 "test/vbinaryc-microkernel-tester.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005402 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005403 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005404 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005405 "test/vbinaryc-microkernel-tester.h",
5406 ] + MICROKERNEL_TEST_HDRS,
5407 deps = MICROKERNEL_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005411 name = "f32_vrdivc_relu_test",
5412 srcs = [
5413 "test/f32-vrdivc-relu.cc",
5414 "test/vbinaryc-microkernel-tester.h",
5415 ] + MICROKERNEL_TEST_HDRS,
5416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
5419xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005420 name = "f32_vmax_test",
5421 srcs = [
5422 "test/f32-vmax.cc",
5423 "test/vbinary-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
5429 name = "f32_vmaxc_test",
5430 srcs = [
5431 "test/f32-vmaxc.cc",
5432 "test/vbinaryc-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
5438 name = "f32_vmin_test",
5439 srcs = [
5440 "test/f32-vmin.cc",
5441 "test/vbinary-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
5447 name = "f32_vminc_test",
5448 srcs = [
5449 "test/f32-vminc.cc",
5450 "test/vbinaryc-microkernel-tester.h",
5451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005456 name = "f32_vmul_test",
5457 srcs = [
5458 "test/f32-vmul.cc",
5459 "test/vbinary-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005465 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005467 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005468 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005469 ] + MICROKERNEL_TEST_HDRS,
5470 deps = MICROKERNEL_TEST_DEPS,
5471)
5472
5473xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005474 name = "f32_vmul_relu_test",
5475 srcs = [
5476 "test/f32-vmul-relu.cc",
5477 "test/vbinary-microkernel-tester.h",
5478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005483 name = "f32_vmulc_test",
5484 srcs = [
5485 "test/f32-vmulc.cc",
5486 "test/vbinaryc-microkernel-tester.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005492 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005493 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005494 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005495 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 ] + MICROKERNEL_TEST_HDRS,
5497 deps = MICROKERNEL_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005501 name = "f32_vmulc_relu_test",
5502 srcs = [
5503 "test/f32-vmulc-relu.cc",
5504 "test/vbinaryc-microkernel-tester.h",
5505 ] + MICROKERNEL_TEST_HDRS,
5506 deps = MICROKERNEL_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005510 name = "f16_vmulcaddc_minmax_test",
5511 srcs = [
5512 "test/f16-vmulcaddc-minmax.cc",
5513 "test/vmulcaddc-microkernel-tester.h",
5514 "src/xnnpack/AlignedAllocator.h",
5515 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005516 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005517)
5518
5519xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005520 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005521 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005522 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005523 "test/vmulcaddc-microkernel-tester.h",
5524 "src/xnnpack/AlignedAllocator.h",
5525 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005526 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005527)
5528
5529xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005530 name = "f32_vlrelu_test",
5531 srcs = [
5532 "test/f32-vlrelu.cc",
5533 "test/vunary-microkernel-tester.h",
5534 ] + MICROKERNEL_TEST_HDRS,
5535 deps = MICROKERNEL_TEST_DEPS,
5536)
5537
5538xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005539 name = "f32_vneg_test",
5540 srcs = [
5541 "test/f32-vneg.cc",
5542 "test/vunary-microkernel-tester.h",
5543 ] + MICROKERNEL_TEST_HDRS,
5544 deps = MICROKERNEL_TEST_DEPS,
5545)
5546
5547xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005548 name = "f32_vrndne_test",
5549 srcs = [
5550 "test/f32-vrndne.cc",
5551 "test/vunary-microkernel-tester.h",
5552 ] + MICROKERNEL_TEST_HDRS,
5553 deps = MICROKERNEL_TEST_DEPS,
5554)
5555
5556xnnpack_unit_test(
5557 name = "f32_vrndz_test",
5558 srcs = [
5559 "test/f32-vrndz.cc",
5560 "test/vunary-microkernel-tester.h",
5561 ] + MICROKERNEL_TEST_HDRS,
5562 deps = MICROKERNEL_TEST_DEPS,
5563)
5564
5565xnnpack_unit_test(
5566 name = "f32_vrndu_test",
5567 srcs = [
5568 "test/f32-vrndu.cc",
5569 "test/vunary-microkernel-tester.h",
5570 ] + MICROKERNEL_TEST_HDRS,
5571 deps = MICROKERNEL_TEST_DEPS,
5572)
5573
5574xnnpack_unit_test(
5575 name = "f32_vrndd_test",
5576 srcs = [
5577 "test/f32-vrndd.cc",
5578 "test/vunary-microkernel-tester.h",
5579 ] + MICROKERNEL_TEST_HDRS,
5580 deps = MICROKERNEL_TEST_DEPS,
5581)
5582
5583xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005584 name = "f32_vscale_test",
5585 srcs = [
5586 "test/f32-vscale.cc",
5587 "test/vscale-microkernel-tester.h",
5588 ] + MICROKERNEL_TEST_HDRS,
5589 deps = MICROKERNEL_TEST_DEPS,
5590)
5591
5592xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005593 name = "f32_vscaleexpminusmax_test",
5594 srcs = [
5595 "test/f32-vscaleexpminusmax.cc",
5596 "test/vscaleexpminusmax-microkernel-tester.h",
5597 ] + MICROKERNEL_TEST_HDRS,
5598 deps = MICROKERNEL_TEST_DEPS,
5599)
5600
5601xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005602 name = "f32_vscaleextexp_test",
5603 srcs = [
5604 "test/f32-vscaleextexp.cc",
5605 "test/vscaleextexp-microkernel-tester.h",
5606 ] + MICROKERNEL_TEST_HDRS,
5607 deps = MICROKERNEL_TEST_DEPS,
5608)
5609
5610xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005611 name = "f32_vsqr_test",
5612 srcs = [
5613 "test/f32-vsqr.cc",
5614 "test/vunary-microkernel-tester.h",
5615 ] + MICROKERNEL_TEST_HDRS,
5616 deps = MICROKERNEL_TEST_DEPS,
5617)
5618
5619xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005620 name = "f32_vsqrdiff_test",
5621 srcs = [
5622 "test/f32-vsqrdiff.cc",
5623 "test/vbinary-microkernel-tester.h",
5624 ] + MICROKERNEL_TEST_HDRS,
5625 deps = MICROKERNEL_TEST_DEPS,
5626)
5627
5628xnnpack_unit_test(
5629 name = "f32_vsqrdiffc_test",
5630 srcs = [
5631 "test/f32-vsqrdiffc.cc",
5632 "test/vbinaryc-microkernel-tester.h",
5633 ] + MICROKERNEL_TEST_HDRS,
5634 deps = MICROKERNEL_TEST_DEPS,
5635)
5636
5637xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005638 name = "f32_vsqrt_test",
5639 srcs = [
5640 "test/f32-vsqrt.cc",
5641 "test/vunary-microkernel-tester.h",
5642 ] + MICROKERNEL_TEST_HDRS,
5643 deps = MICROKERNEL_TEST_DEPS,
5644)
5645
5646xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005647 name = "f32_vsub_test",
5648 srcs = [
5649 "test/f32-vsub.cc",
5650 "test/vbinary-microkernel-tester.h",
5651 ] + MICROKERNEL_TEST_HDRS,
5652 deps = MICROKERNEL_TEST_DEPS,
5653)
5654
5655xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005656 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005657 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005658 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005659 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005660 ] + MICROKERNEL_TEST_HDRS,
5661 deps = MICROKERNEL_TEST_DEPS,
5662)
5663
5664xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005665 name = "f32_vsub_relu_test",
5666 srcs = [
5667 "test/f32-vsub-relu.cc",
5668 "test/vbinary-microkernel-tester.h",
5669 ] + MICROKERNEL_TEST_HDRS,
5670 deps = MICROKERNEL_TEST_DEPS,
5671)
5672
5673xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005674 name = "f32_vsubc_test",
5675 srcs = [
5676 "test/f32-vsubc.cc",
5677 "test/vbinaryc-microkernel-tester.h",
5678 ] + MICROKERNEL_TEST_HDRS,
5679 deps = MICROKERNEL_TEST_DEPS,
5680)
5681
5682xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005683 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005684 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005685 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005686 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005687 ] + MICROKERNEL_TEST_HDRS,
5688 deps = MICROKERNEL_TEST_DEPS,
5689)
5690
5691xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005692 name = "f32_vsubc_relu_test",
5693 srcs = [
5694 "test/f32-vsubc-relu.cc",
5695 "test/vbinaryc-microkernel-tester.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 deps = MICROKERNEL_TEST_DEPS,
5698)
5699
5700xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005701 name = "f32_vrsubc_test",
5702 srcs = [
5703 "test/f32-vrsubc.cc",
5704 "test/vbinaryc-microkernel-tester.h",
5705 ] + MICROKERNEL_TEST_HDRS,
5706 deps = MICROKERNEL_TEST_DEPS,
5707)
5708
5709xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005710 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005711 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005712 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005713 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005714 ] + MICROKERNEL_TEST_HDRS,
5715 deps = MICROKERNEL_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005719 name = "f32_vrsubc_relu_test",
5720 srcs = [
5721 "test/f32-vrsubc-relu.cc",
5722 "test/vbinaryc-microkernel-tester.h",
5723 ] + MICROKERNEL_TEST_HDRS,
5724 deps = MICROKERNEL_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005728 name = "qs8_dwconv_minmax_test",
5729 srcs = [
5730 "test/qs8-dwconv-minmax.cc",
5731 "test/dwconv-microkernel-tester.h",
5732 "src/xnnpack/AlignedAllocator.h",
5733 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5734 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5735)
5736
5737xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005738 name = "qs8_gavgpool_minmax_test",
5739 srcs = [
5740 "test/qs8-gavgpool-minmax.cc",
5741 "test/gavgpool-microkernel-tester.h",
5742 "src/xnnpack/AlignedAllocator.h",
5743 ] + MICROKERNEL_TEST_HDRS,
5744 deps = MICROKERNEL_TEST_DEPS,
5745)
5746
5747xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005748 name = "qs8_gemm_minmax_test",
5749 srcs = [
5750 "test/qs8-gemm-minmax.cc",
5751 "test/gemm-microkernel-tester.h",
5752 "src/xnnpack/AlignedAllocator.h",
5753 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5754 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5755)
5756
5757xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005758 name = "qs8_igemm_minmax_test",
5759 srcs = [
5760 "test/qs8-igemm-minmax.cc",
5761 "test/gemm-microkernel-tester.h",
5762 "src/xnnpack/AlignedAllocator.h",
5763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5765)
5766
5767xnnpack_unit_test(
5768 name = "qs8_requantization_test",
5769 srcs = [
5770 "src/xnnpack/requantization-stubs.h",
5771 "test/qs8-requantization.cc",
5772 "test/requantization-tester.h",
5773 ] + MICROKERNEL_TEST_HDRS,
5774 deps = MICROKERNEL_TEST_DEPS,
5775)
5776
5777xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005778 name = "qs8_vadd_minmax_test",
5779 srcs = [
5780 "test/qs8-vadd-minmax.cc",
5781 "test/vadd-microkernel-tester.h",
5782 ] + MICROKERNEL_TEST_HDRS,
5783 deps = MICROKERNEL_TEST_DEPS,
5784)
5785
5786xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005787 name = "qs8_vaddc_minmax_test",
5788 srcs = [
5789 "test/qs8-vaddc-minmax.cc",
5790 "test/vaddc-microkernel-tester.h",
5791 ] + MICROKERNEL_TEST_HDRS,
5792 deps = MICROKERNEL_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005796 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005797 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005798 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005799 "test/avgpool-microkernel-tester.h",
5800 "src/xnnpack/AlignedAllocator.h",
5801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005806 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005807 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005808 "test/qu8-dwconv-minmax.cc",
5809 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005810 "src/xnnpack/AlignedAllocator.h",
5811 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005812 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005813)
5814
5815xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005816 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005817 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005818 "test/qu8-igemm-minmax.cc",
5819 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005820 "src/xnnpack/AlignedAllocator.h",
5821 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005822 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823)
5824
5825xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005826 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005827 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005828 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005829 "test/gavgpool-microkernel-tester.h",
5830 "src/xnnpack/AlignedAllocator.h",
5831 ] + MICROKERNEL_TEST_HDRS,
5832 deps = MICROKERNEL_TEST_DEPS,
5833)
5834
5835xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005836 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005837 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005838 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005839 "test/gemm-microkernel-tester.h",
5840 "src/xnnpack/AlignedAllocator.h",
5841 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005842 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005843)
5844
5845xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005846 name = "qu8_requantization_test",
5847 srcs = [
5848 "src/xnnpack/requantization-stubs.h",
5849 "test/qu8-requantization.cc",
5850 "test/requantization-tester.h",
5851 ] + MICROKERNEL_TEST_HDRS,
5852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
5855xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005856 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005857 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005858 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005859 "test/vadd-microkernel-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
5865 name = "u8_clamp_test",
5866 srcs = [
5867 "test/u8-clamp.cc",
5868 "test/clamp-microkernel-tester.h",
5869 ] + MICROKERNEL_TEST_HDRS,
5870 deps = MICROKERNEL_TEST_DEPS,
5871)
5872
5873xnnpack_unit_test(
5874 name = "u8_lut32norm_test",
5875 srcs = [
5876 "test/u8-lut32norm.cc",
5877 "test/lut-norm-microkernel-tester.h",
5878 ] + MICROKERNEL_TEST_HDRS,
5879 deps = MICROKERNEL_TEST_DEPS,
5880)
5881
5882xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005883 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005885 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005886 "test/maxpool-microkernel-tester.h",
5887 ] + MICROKERNEL_TEST_HDRS,
5888 deps = MICROKERNEL_TEST_DEPS,
5889)
5890
5891xnnpack_unit_test(
5892 name = "u8_rmax_test",
5893 srcs = [
5894 "test/u8-rmax.cc",
5895 "test/rmax-microkernel-tester.h",
5896 ] + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS,
5898)
5899
5900xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005901 name = "x32_fill_test",
5902 srcs = [
5903 "test/x32-fill.cc",
5904 "test/fill-microkernel-tester.h",
5905 ] + MICROKERNEL_TEST_HDRS,
5906 deps = MICROKERNEL_TEST_DEPS,
5907)
5908
5909xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910 name = "x32_packx_test",
5911 srcs = [
5912 "test/x32-packx.cc",
5913 "test/pack-microkernel-tester.h",
5914 "src/xnnpack/AlignedAllocator.h",
5915 ] + MICROKERNEL_TEST_HDRS,
5916 deps = MICROKERNEL_TEST_DEPS,
5917)
5918
5919xnnpack_unit_test(
5920 name = "x32_pad_test",
5921 srcs = [
5922 "test/x32-pad.cc",
5923 "test/pad-microkernel-tester.h",
5924 ] + MICROKERNEL_TEST_HDRS,
5925 deps = MICROKERNEL_TEST_DEPS,
5926)
5927
5928xnnpack_unit_test(
5929 name = "x32_unpool_test",
5930 srcs = [
5931 "test/x32-unpool.cc",
5932 "test/unpool-microkernel-tester.h",
5933 ] + MICROKERNEL_TEST_HDRS,
5934 deps = MICROKERNEL_TEST_DEPS,
5935)
5936
5937xnnpack_unit_test(
5938 name = "x32_zip_test",
5939 srcs = [
5940 "test/x32-zip.cc",
5941 "test/zip-microkernel-tester.h",
5942 ] + MICROKERNEL_TEST_HDRS,
5943 deps = MICROKERNEL_TEST_DEPS,
5944)
5945
5946xnnpack_unit_test(
5947 name = "x8_lut_test",
5948 srcs = [
5949 "test/x8-lut.cc",
5950 "test/lut-microkernel-tester.h",
5951 ] + MICROKERNEL_TEST_HDRS,
5952 deps = MICROKERNEL_TEST_DEPS,
5953)
5954
5955xnnpack_unit_test(
5956 name = "x8_zip_test",
5957 srcs = [
5958 "test/x8-zip.cc",
5959 "test/zip-microkernel-tester.h",
5960 ] + MICROKERNEL_TEST_HDRS,
5961 deps = MICROKERNEL_TEST_DEPS,
5962)
5963
Marat Dukhan20c3b922020-03-10 03:45:06 -07005964########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005965
5966xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005967 name = "operator_size_test",
5968 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005969 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970)
5971
Marat Dukhan20c3b922020-03-10 03:45:06 -07005972xnnpack_binary(
5973 name = "subgraph_size_test",
5974 srcs = ["test/subgraph-size.c"],
5975 deps = [":XNNPACK"],
5976)
5977
5978########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005979
5980xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005981 name = "abs_nc_test",
5982 srcs = [
5983 "test/abs-nc.cc",
5984 "test/abs-operator-tester.h",
5985 ],
5986 deps = OPERATOR_TEST_DEPS,
5987)
5988
5989xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005990 name = "add_nd_test",
5991 srcs = [
5992 "test/add-nd.cc",
5993 "test/binary-elementwise-operator-tester.h",
5994 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005995 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005996)
5997
5998xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005999 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006001 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006002 "test/argmax-pooling-operator-tester.h",
6003 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006004 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006005)
6006
6007xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006008 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006010 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006011 "test/average-pooling-operator-tester.h",
6012 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006013 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006014)
6015
6016xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006017 name = "bankers_rounding_nc_test",
6018 srcs = [
6019 "test/bankers-rounding-nc.cc",
6020 "test/bankers-rounding-operator-tester.h",
6021 ],
6022 deps = OPERATOR_TEST_DEPS,
6023)
6024
6025xnnpack_unit_test(
6026 name = "ceiling_nc_test",
6027 srcs = [
6028 "test/ceiling-nc.cc",
6029 "test/ceiling-operator-tester.h",
6030 ],
6031 deps = OPERATOR_TEST_DEPS,
6032)
6033
6034xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006035 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006036 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006037 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038 "test/channel-shuffle-operator-tester.h",
6039 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006040 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006041)
6042
6043xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006044 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006045 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006046 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006047 "test/clamp-operator-tester.h",
6048 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006049 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006050)
6051
6052xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006053 name = "constant_pad_nd_test",
6054 srcs = [
6055 "test/constant-pad-nd.cc",
6056 "test/constant-pad-operator-tester.h",
6057 ],
6058 deps = OPERATOR_TEST_DEPS,
6059)
6060
6061xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006062 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006064 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006065 "test/convolution-operator-tester.h",
6066 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006067 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068)
6069
6070xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006071 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006072 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006073 "test/convolution-nchw.cc",
6074 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006075 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006076 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006077)
6078
6079xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006080 name = "copy_nc_test",
6081 srcs = [
6082 "test/copy-nc.cc",
6083 "test/copy-operator-tester.h",
6084 ],
6085 deps = OPERATOR_TEST_DEPS,
6086)
6087
6088xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006089 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006090 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006091 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006092 "test/deconvolution-operator-tester.h",
6093 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006094 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095)
6096
6097xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006098 name = "divide_nd_test",
6099 srcs = [
6100 "test/binary-elementwise-operator-tester.h",
6101 "test/divide-nd.cc",
6102 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006103 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006104)
6105
6106xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006107 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006108 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006109 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006110 "test/fully-connected-operator-tester.h",
6111 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006112 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006113)
6114
6115xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006116 name = "floor_nc_test",
6117 srcs = [
6118 "test/floor-nc.cc",
6119 "test/floor-operator-tester.h",
6120 ],
6121 deps = OPERATOR_TEST_DEPS,
6122)
6123
6124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006125 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006127 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006128 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006129 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006130 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131)
6132
6133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006134 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006136 "test/global-average-pooling-ncw.cc",
6137 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006138 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006139 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006140)
6141
6142xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006143 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006145 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006146 "test/hardswish-operator-tester.h",
6147 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006148 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149)
6150
6151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006152 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006154 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006155 "test/leaky-relu-operator-tester.h",
6156 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006157 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006158)
6159
6160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006161 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006163 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164 "test/max-pooling-operator-tester.h",
6165 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006167)
6168
6169xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006170 name = "maximum_nd_test",
6171 srcs = [
6172 "test/binary-elementwise-operator-tester.h",
6173 "test/maximum-nd.cc",
6174 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006175 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006176)
6177
6178xnnpack_unit_test(
6179 name = "minimum_nd_test",
6180 srcs = [
6181 "test/binary-elementwise-operator-tester.h",
6182 "test/minimum-nd.cc",
6183 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006184 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006185)
6186
6187xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006188 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006189 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006190 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006191 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006192 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006193 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006194)
6195
6196xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006197 name = "negate_nc_test",
6198 srcs = [
6199 "test/negate-nc.cc",
6200 "test/negate-operator-tester.h",
6201 ],
6202 deps = OPERATOR_TEST_DEPS,
6203)
6204
6205xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006206 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006208 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209 "test/prelu-operator-tester.h",
6210 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006211 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006212)
6213
6214xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006215 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006216 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006217 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006218 "test/resize-bilinear-operator-tester.h",
6219 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006220 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006221)
6222
6223xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006224 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006226 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227 "test/sigmoid-operator-tester.h",
6228 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006229 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230)
6231
6232xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006233 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006234 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006235 "test/softmax-nc.cc",
6236 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006237 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006238 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239)
6240
6241xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006242 name = "square_nc_test",
6243 srcs = [
6244 "test/square-nc.cc",
6245 "test/square-operator-tester.h",
6246 ],
6247 deps = OPERATOR_TEST_DEPS,
6248)
6249
6250xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006251 name = "square_root_nc_test",
6252 srcs = [
6253 "test/square-root-nc.cc",
6254 "test/square-root-operator-tester.h",
6255 ],
6256 deps = OPERATOR_TEST_DEPS,
6257)
6258
6259xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006260 name = "squared_difference_nd_test",
6261 srcs = [
6262 "test/binary-elementwise-operator-tester.h",
6263 "test/squared-difference-nd.cc",
6264 ],
6265 deps = OPERATOR_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006269 name = "subtract_nd_test",
6270 srcs = [
6271 "test/binary-elementwise-operator-tester.h",
6272 "test/subtract-nd.cc",
6273 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006274 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006275)
6276
6277xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006278 name = "truncation_nc_test",
6279 srcs = [
6280 "test/truncation-nc.cc",
6281 "test/truncation-operator-tester.h",
6282 ],
6283 deps = OPERATOR_TEST_DEPS,
6284)
6285
6286xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006287 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006289 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290 "test/unpooling-operator-tester.h",
6291 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006292 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293)
6294
Chao Mei6ddfc602020-05-13 22:29:36 -07006295############################### Misc unit tests ###############################
6296
6297xnnpack_unit_test(
6298 name = "memory_planner_test",
6299 srcs = [
6300 "test/memory-planner-test.cc",
6301 ],
6302 deps = [
6303 ":XNNPACK",
6304 ":memory_planner",
6305 ],
6306)
6307
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308############################# Build configurations #############################
6309
Marat Dukhanb8642352019-10-30 15:43:02 -07006310# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006311config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006312 name = "xnn_enable_assembly_explicit_true",
6313 define_values = {"xnn_enable_assembly": "true"},
6314)
6315
6316# Disables usage of assembly kernels.
6317config_setting(
6318 name = "xnn_enable_assembly_explicit_false",
6319 define_values = {"xnn_enable_assembly": "false"},
6320)
6321
Marat Dukhan9de90e02020-06-18 16:04:12 -07006322# Enables usage of sparse inference.
6323config_setting(
6324 name = "xnn_enable_sparse_explicit_true",
6325 define_values = {"xnn_enable_sparse": "true"},
6326)
6327
6328# Disables usage of sparse inference.
6329config_setting(
6330 name = "xnn_enable_sparse_explicit_false",
6331 define_values = {"xnn_enable_sparse": "false"},
6332)
6333
Marat Dukhan05702cf2020-03-26 15:41:33 -07006334# Disables usage of HMP-aware optimizations.
6335config_setting(
6336 name = "xnn_enable_hmp_explicit_false",
6337 define_values = {"xnn_enable_hmp": "false"},
6338)
6339
Chao Mei6ddfc602020-05-13 22:29:36 -07006340# Enable usage of optimized memory allocation
6341config_setting(
6342 name = "xnn_enable_memopt_explicit_true",
6343 define_values = {"xnnpack_enable_memopt": "true"},
6344)
6345
6346# Disable usage of optimized memory allocation
6347config_setting(
6348 name = "xnn_enable_memopt_explicit_false",
6349 define_values = {"xnnpack_enable_memopt": "false"},
6350)
6351
Marat Dukhanb8642352019-10-30 15:43:02 -07006352# Builds with -c dbg
6353config_setting(
6354 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006355 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006356 "compilation_mode": "dbg",
6357 },
6358)
6359
6360# Builds with -c opt
6361config_setting(
6362 name = "optimized_build",
6363 values = {
6364 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006365 },
6366)
6367
6368config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006369 name = "linux_k8",
6370 values = {"cpu": "k8"},
6371)
6372
6373config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006374 name = "linux_arm",
6375 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006376)
6377
6378config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006379 name = "linux_armeabi",
6380 values = {"cpu": "armeabi"},
6381)
6382
6383config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006384 name = "linux_armhf",
6385 values = {"cpu": "armhf"},
6386)
6387
6388config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006389 name = "linux_armv7a",
6390 values = {"cpu": "armv7a"},
6391)
6392
6393config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006394 name = "linux_aarch64",
6395 values = {"cpu": "aarch64"},
6396)
6397
6398config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006399 name = "android",
6400 values = {"crosstool_top": "//external:android/crosstool"},
6401)
6402
6403config_setting(
6404 name = "android_armv7",
6405 values = {
6406 "crosstool_top": "//external:android/crosstool",
6407 "cpu": "armeabi-v7a",
6408 },
6409)
6410
6411config_setting(
6412 name = "android_arm64",
6413 values = {
6414 "crosstool_top": "//external:android/crosstool",
6415 "cpu": "arm64-v8a",
6416 },
6417)
6418
6419config_setting(
6420 name = "android_x86",
6421 values = {
6422 "crosstool_top": "//external:android/crosstool",
6423 "cpu": "x86",
6424 },
6425)
6426
6427config_setting(
6428 name = "android_x86_64",
6429 values = {
6430 "crosstool_top": "//external:android/crosstool",
6431 "cpu": "x86_64",
6432 },
6433)
6434
6435config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006436 name = "windows_x86_64",
6437 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006438)
6439
6440config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006441 name = "windows_x86_64_clang",
6442 values = {
6443 "compiler": "clang-cl",
6444 "cpu": "x64_windows",
6445 },
6446)
6447
6448config_setting(
6449 name = "windows_x86_64_mingw",
6450 values = {
6451 "compiler": "mingw-gcc",
6452 "cpu": "x64_windows",
6453 },
6454)
6455
6456config_setting(
6457 name = "windows_x86_64_msys",
6458 values = {
6459 "compiler": "msys-gcc",
6460 "cpu": "x64_windows",
6461 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006462)
6463
6464config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006465 name = "macos_x86_64",
6466 values = {
6467 "apple_platform_type": "macos",
6468 "cpu": "darwin",
6469 },
6470)
6471
6472config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006473 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006474 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475)
6476
6477config_setting(
6478 name = "emscripten_wasm",
6479 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006480 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006481 "cpu": "wasm",
6482 },
6483)
6484
6485config_setting(
6486 name = "emscripten_wasmsimd",
6487 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006488 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006490 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006491 },
6492)
6493
6494config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006495 name = "ios_armv7",
6496 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006497 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006498 "cpu": "ios_armv7",
6499 },
6500)
6501
6502config_setting(
6503 name = "ios_arm64",
6504 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006505 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006506 "cpu": "ios_arm64",
6507 },
6508)
6509
6510config_setting(
6511 name = "ios_arm64e",
6512 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006513 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006514 "cpu": "ios_arm64e",
6515 },
6516)
6517
6518config_setting(
6519 name = "ios_x86",
6520 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006521 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006522 "cpu": "ios_i386",
6523 },
6524)
6525
6526config_setting(
6527 name = "ios_x86_64",
6528 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006529 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006530 "cpu": "ios_x86_64",
6531 },
6532)
6533
6534config_setting(
6535 name = "watchos_armv7k",
6536 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006537 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006538 "cpu": "watchos_armv7k",
6539 },
6540)
6541
6542config_setting(
6543 name = "watchos_arm64_32",
6544 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006545 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006546 "cpu": "watchos_arm64_32",
6547 },
6548)
6549
6550config_setting(
6551 name = "watchos_x86",
6552 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006553 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006554 "cpu": "watchos_i386",
6555 },
6556)
6557
6558config_setting(
6559 name = "watchos_x86_64",
6560 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006561 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006562 "cpu": "watchos_x86_64",
6563 },
6564)
6565
6566config_setting(
6567 name = "tvos_arm64",
6568 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006569 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006570 "cpu": "tvos_arm64",
6571 },
6572)
6573
6574config_setting(
6575 name = "tvos_x86_64",
6576 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006577 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006578 "cpu": "tvos_x86_64",
6579 },
6580)