blob: ced71c0a53a23cd5bb5ec64b05d84f606c222351 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700425 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001074 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1075 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1076 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1078 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1079 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001082 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001083 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001084 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001085 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001086 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001087 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001088 "src/x32-zip/x2-wasmsimd.c",
1089 "src/x32-zip/x3-wasmsimd.c",
1090 "src/x32-zip/x4-wasmsimd.c",
1091 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001092 "src/math/roundne-wasmsimd-addsub.c",
1093 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001094 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001095 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001099 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1100 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001101]
1102
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001103PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001104 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001105 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001106 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001107 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001108 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001109 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1110 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1111 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/precise-psimd.c",
1116 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001117 "src/qu8-requantization/precise-psimd.c",
1118 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121# ISA-specific micro-kernels
1122NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001123 "src/f32-argmaxpool/4x-neon-c4.c",
1124 "src/f32-argmaxpool/9p8x-neon-c4.c",
1125 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001126 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1127 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001128 "src/f32-clamp/gen/neon-x4.c",
1129 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001130 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1131 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001134 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1135 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1144 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1152 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1154 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1160 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1161 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1162 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1164 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1165 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1166 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1167 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1170 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1171 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1173 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1181 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1182 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001184 "src/f32-hswish/gen/hswish-neon-x4.c",
1185 "src/f32-hswish/gen/hswish-neon-x8.c",
1186 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001187 "src/f32-ibilinear/gen/neon-c4.c",
1188 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1190 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1191 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1193 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1195 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1196 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1197 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1198 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1200 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1201 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1202 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1203 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001205 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1206 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1207 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001210 "src/f32-prelu/gen/neon-1x4.c",
1211 "src/f32-prelu/gen/neon-1x8.c",
1212 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001213 "src/f32-prelu/gen/neon-2x4.c",
1214 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001215 "src/f32-prelu/gen/neon-2x16.c",
1216 "src/f32-prelu/gen/neon-4x4.c",
1217 "src/f32-prelu/gen/neon-4x8.c",
1218 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001243 "src/f32-relu/gen/neon-x4.c",
1244 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001246 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001247 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1248 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1267 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001269 "src/f32-vbinary/gen/vmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1272 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1273 "src/f32-vbinary/gen/vmin-neon-x4.c",
1274 "src/f32-vbinary/gen/vmin-neon-x8.c",
1275 "src/f32-vbinary/gen/vminc-neon-x4.c",
1276 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001283 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1285 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001291 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1292 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001293 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1294 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001295 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1296 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1297 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001303 "src/f32-vunary/gen/vabs-neon-x4.c",
1304 "src/f32-vunary/gen/vabs-neon-x8.c",
1305 "src/f32-vunary/gen/vneg-neon-x4.c",
1306 "src/f32-vunary/gen/vneg-neon-x8.c",
1307 "src/f32-vunary/gen/vsqr-neon-x4.c",
1308 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001309 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1310 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1311 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001313 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1314 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1315 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1317 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1318 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001321 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1322 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1323 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1325 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1326 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001329 "src/qs8-requantization/precise-neon.c",
1330 "src/qs8-requantization/fp32-neon.c",
1331 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001332 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1333 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1334 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1336 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1337 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001340 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1341 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1342 "src/qu8-dwconv/up8x9-minmax-neon.c",
1343 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1344 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1345 "src/qu8-gemm/4x8-minmax-neon.c",
1346 "src/qu8-gemm/8x8-minmax-neon.c",
1347 "src/qu8-igemm/4x8-minmax-neon.c",
1348 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001349 "src/qu8-requantization/precise-neon.c",
1350 "src/qu8-requantization/fp32-neon.c",
1351 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001352 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001355 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001356 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001358 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001359 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360 "src/x32-zip/x2-neon.c",
1361 "src/x32-zip/x3-neon.c",
1362 "src/x32-zip/x4-neon.c",
1363 "src/x32-zip/xm-neon.c",
1364 "src/x8-zip/x2-neon.c",
1365 "src/x8-zip/x3-neon.c",
1366 "src/x8-zip/x4-neon.c",
1367 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001368 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-neon-addsub.c",
1370 "src/math/roundd-neon-cvt.c",
1371 "src/math/roundu-neon-addsub.c",
1372 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neon-addsub.c",
1374 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001375 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001376 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1377 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1378 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1379 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1380 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001382 "src/math/sqrt-neon-nr1rsqrts.c",
1383 "src/math/sqrt-neon-nr2rsqrts.c",
1384 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001385]
1386
1387NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001388 "src/f32-ibilinear/gen/neonfma-c4.c",
1389 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001390 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1391 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1392 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1394 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1395 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1396 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1397 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1405 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1412 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1413 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1416 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1417 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1418 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1424 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1425 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1430 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001509 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1510 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001531 "src/math/exp-neonfma-lut64-p2.c",
1532 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001533 "src/math/expminus-neonfma-lut2048-p1.c",
1534 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001535 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001536 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1537 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1539 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1542 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1543 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1545 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1546 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1548 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1551 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1552 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001554 "src/math/sqrt-neonfma-nr1fma.c",
1555 "src/math/sqrt-neonfma-nr2fma.c",
1556 "src/math/sqrt-neonfma-nr3fma.c",
1557 "src/math/sqrt-neonfma-nr2fma1adj.c",
1558 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001559]
1560
1561AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1569 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1573 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1575 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1579 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1581 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1583 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1585 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1587 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1589 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001592 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1593 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001596 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1598 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1599 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001601 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001619 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1620 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1621 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1622 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1623 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1625 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1626 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1627 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1628 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1630 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1631 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1632 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1633 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1635 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1636 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001637 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1638 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1640 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1641 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1643 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1644 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645]
1646
Marat Dukhan8853b822020-05-07 12:19:01 -07001647NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001648 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1649 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1650 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001656 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001657 "src/math/roundd-neonv8.c",
1658 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001659 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001660]
1661
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001663 "src/f16-clamp/gen/neonfp16arith-x16.c",
1664 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001665 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001677 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1678 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001679 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1680 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001681 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001682 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001683 "src/f16-relu/gen/neonfp16arith-x16.c",
1684 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001685 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001713 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1714 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001715 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001723 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001727 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001739 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1740 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1741 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Benoit Jacoba9644732020-08-13 12:48:55 -07001749NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001750 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001751 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001753 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001754 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001755 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001756 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001757 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1758 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001759 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001760 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001761 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001762]
1763
Marat Dukhan08c4a432019-10-03 09:29:21 -07001764SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001765 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1766 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001767 "src/f32-clamp/gen/sse-x4.c",
1768 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001769 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1770 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001771 "src/f32-dwconv-chw/3x3p1-sse.c",
1772 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001773 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1775 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1777 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1779 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1780 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1781 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1783 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1784 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001785 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001786 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1787 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001788 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1789 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1790 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1791 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1792 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1793 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1794 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1795 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1796 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1797 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1798 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1799 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1800 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001801 "src/f32-hswish/gen/hswish-sse-x4.c",
1802 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001803 "src/f32-ibilinear/gen/sse-c4.c",
1804 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001805 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1806 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1807 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1808 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1809 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1810 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1811 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001812 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1813 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1814 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001815 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001816 "src/f32-prelu/gen/sse-2x4.c",
1817 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001818 "src/f32-relu/gen/sse-x4.c",
1819 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001821 "src/f32-spmm/gen/4x1-minmax-sse.c",
1822 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001823 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001824 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1826 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1829 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1830 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001832 "src/f32-vbinary/gen/vmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vmax-sse-x8.c",
1834 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1835 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1836 "src/f32-vbinary/gen/vmin-sse-x4.c",
1837 "src/f32-vbinary/gen/vmin-sse-x8.c",
1838 "src/f32-vbinary/gen/vminc-sse-x4.c",
1839 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001840 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1843 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1844 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1845 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1846 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001848 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1849 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1850 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1851 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001852 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1854 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1855 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001856 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1857 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001858 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1859 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001860 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1861 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001862 "src/f32-vunary/gen/vabs-sse-x4.c",
1863 "src/f32-vunary/gen/vabs-sse-x8.c",
1864 "src/f32-vunary/gen/vneg-sse-x4.c",
1865 "src/f32-vunary/gen/vneg-sse-x8.c",
1866 "src/f32-vunary/gen/vsqr-sse-x4.c",
1867 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001868 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001869 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001870 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001871 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001872 "src/math/roundd-sse-addsub.c",
1873 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001874 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001875 "src/math/sqrt-sse-hh1mac.c",
1876 "src/math/sqrt-sse-nr1mac.c",
1877 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878]
1879
1880SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001881 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1882 "src/f32-argmaxpool/4x-sse2-c4.c",
1883 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001884 "src/f32-prelu/gen/sse2-2x4.c",
1885 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1896 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1897 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001898 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1899 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1900 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1901 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1902 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1903 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001904 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001905 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001906 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001907 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001908 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1909 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001910 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1911 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001912 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1913 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1914 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1915 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1916 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1917 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1918 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1919 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001920 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1921 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1922 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001923 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1924 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1925 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001926 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1927 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1928 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001929 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001931 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1932 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001933 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1934 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001935 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1936 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001937 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001938 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1939 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001940 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001941 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1942 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1943 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001944 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1945 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001946 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1947 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001948 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1949 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001950 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001951 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1952 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001953 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001954 "src/qs8-requantization/precise-sse2.c",
1955 "src/qs8-requantization/fp32-sse2.c",
1956 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001957 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1958 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1959 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1960 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001961 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1962 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1963 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1964 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001965 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1966 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1967 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1968 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1969 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1970 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1971 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1972 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001973 "src/qu8-requantization/precise-sse2.c",
1974 "src/qu8-requantization/fp32-sse2.c",
1975 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001976 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001977 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001978 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001980 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 "src/x32-zip/x2-sse2.c",
1982 "src/x32-zip/x3-sse2.c",
1983 "src/x32-zip/x4-sse2.c",
1984 "src/x32-zip/xm-sse2.c",
1985 "src/x8-zip/x2-sse2.c",
1986 "src/x8-zip/x3-sse2.c",
1987 "src/x8-zip/x4-sse2.c",
1988 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001989 "src/math/exp-sse2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07001990 "src/math/exp-sse2-lut64-p2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001991 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001992 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001993 "src/math/roundd-sse2-cvt.c",
1994 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001995 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07001996 "src/math/sigmoid-sse2-rr2-p5-div.c",
1997 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001998]
1999
2000SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002001 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2002 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2003 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002004 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2005 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2006 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002007 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2008 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2009 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002010 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2011 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002012 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2013 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002014 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2015 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002016 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2017 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002018 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002019 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2020 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002021 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002022 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2023 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2024 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002025 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2026 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002027 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2028 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002029 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2030 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002031 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002032 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2033 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002034 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002035 "src/qs8-requantization/precise-ssse3.c",
2036 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002037 "src/qu8-requantization/precise-ssse3.c",
2038 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039]
2040
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002041SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002042 "src/f32-prelu/gen/sse41-2x4.c",
2043 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002044 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2045 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2046 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2047 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2048 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2049 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002050 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002051 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002052 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002053 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002054 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2055 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002056 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2057 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002058 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2059 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2060 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2061 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2062 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2063 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2064 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2065 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002066 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2067 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2068 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002069 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2070 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2071 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002072 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2073 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2074 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002075 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2076 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002077 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2078 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002079 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2080 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002081 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2082 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002083 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002084 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2085 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002086 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002087 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2088 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2089 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002090 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2091 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002092 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2093 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002094 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2095 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002096 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002097 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2098 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002099 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002100 "src/qs8-requantization/fp32-sse4.c",
2101 "src/qs8-requantization/precise-sse4.c",
2102 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002103 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2104 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2105 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2106 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002107 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2108 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2109 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2110 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002111 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2112 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2113 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2114 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002115 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2116 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2117 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2118 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002119 "src/qu8-requantization/precise-sse4.c",
2120 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002121 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002122 "src/math/roundd-sse41.c",
2123 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002124 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002125]
2126
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002128 "src/f32-clamp/gen/avx-x8.c",
2129 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002130 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2131 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2132 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2133 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2134 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2135 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2136 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2137 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2138 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2139 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2140 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2141 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2142 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2143 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2144 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2145 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2146 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2147 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2148 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2149 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2150 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2151 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2152 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2153 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2154 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2157 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2158 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2159 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002160 "src/f32-hswish/gen/hswish-avx-x8.c",
2161 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002162 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2163 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2164 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2165 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2166 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2167 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2168 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2169 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2170 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002171 "src/f32-prelu/gen/avx-2x8.c",
2172 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002173 "src/f32-relu/gen/avx-x16.c",
2174 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002176 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2177 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2178 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2179 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2180 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2181 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2182 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2183 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002184 "src/f32-vbinary/gen/vmax-avx-x8.c",
2185 "src/f32-vbinary/gen/vmax-avx-x16.c",
2186 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2187 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2188 "src/f32-vbinary/gen/vmin-avx-x8.c",
2189 "src/f32-vbinary/gen/vmin-avx-x16.c",
2190 "src/f32-vbinary/gen/vminc-avx-x8.c",
2191 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002192 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2193 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2194 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2195 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2196 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2197 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2198 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2199 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002200 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2201 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2202 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2203 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002204 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2205 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2206 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2207 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002208 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2209 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002210 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002211 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2212 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2213 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2214 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2215 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2216 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2217 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2218 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002219 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2220 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002221 "src/f32-vunary/gen/vabs-avx-x8.c",
2222 "src/f32-vunary/gen/vabs-avx-x16.c",
2223 "src/f32-vunary/gen/vneg-avx-x8.c",
2224 "src/f32-vunary/gen/vneg-avx-x16.c",
2225 "src/f32-vunary/gen/vsqr-avx-x8.c",
2226 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002227 "src/math/exp-avx-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002228]
2229
Marat Dukhan1566fee2020-08-02 21:55:41 -07002230XOP_UKERNELS = [
2231 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2232 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2233 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2234 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002235 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2236 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002237 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2238 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002239 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002240 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2241 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002242 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002243 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2244 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2245 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002246 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2247 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2248 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2249 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2250 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2251 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002252 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002253 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2254 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002255 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002256 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2257 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2258 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2259 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2260 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2261 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2262 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2263 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002264]
2265
Marat Dukhanfda12b82019-11-21 12:27:59 -08002266FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002267 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2268 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2269 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2270 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2271 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2272 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2273 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2274 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2275 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2276 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2277 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2278 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2279 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2280 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2281 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2282 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2283 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2284 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2285 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2286 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2287 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2288 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2289 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2291 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2292 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2293 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2294 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2295 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2296 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2297 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2298 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2299 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2300 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2301 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2302 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2303 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2304 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2305 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2306 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002307 "src/f32-hswish/gen/hswish-fma3-x8.c",
2308 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002309 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2310 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2311 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2312 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2313 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2314 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2315 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2316 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2317 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2318 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2319 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2320 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2321 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2322 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002323 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2324 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2325 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2326 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2327 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2328 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2329 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2330 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002331 "src/math/sqrt-fma3-nr1fma.c",
2332 "src/math/sqrt-fma3-nr2fma.c",
2333 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002334]
2335
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002336AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002337 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2338 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2339 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2340 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2341 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2342 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2343 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2344 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2345 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2346 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2347 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2348 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2349 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2350 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2351 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2352 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2353 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2354 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2355 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2356 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2357 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2358 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2359 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2360 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2361 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2362 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2363 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2364 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2365 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2366 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2367 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2368 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2369 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002373 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2374 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2375 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2376 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2377 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2378 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2379 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2380 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2381 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2383 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002403 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2404 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2405 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2406 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2407 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2408 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2409 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2410 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2411 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2415 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2416 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2417 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2418 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2419 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2420 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2421 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2422 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2423 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2424 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2425 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2426 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002427 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2428 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002429 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2430 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2431 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2432 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002433 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2434 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2435 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002436 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2437 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2438 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002439 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2440 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2441 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002442 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2443 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2444 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2445 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2446 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2447 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2448 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2449 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002450 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002451 "src/math/exp-avx2-lut8-p3-perm.c",
2452 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002453 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002454 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002455 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002456 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002457 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002458 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002459 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002460 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002461]
2462
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002464 "src/f32-clamp/gen/avx512f-x16.c",
2465 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002466 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2467 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2468 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2469 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2470 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2471 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2472 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2473 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2474 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2475 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2476 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2477 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2478 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2479 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2480 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2481 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2482 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2483 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2484 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2485 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2486 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2487 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2488 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2489 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002490 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2491 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002492 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2493 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2494 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2495 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2496 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2497 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002498 "src/f32-prelu/gen/avx512f-2x16.c",
2499 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002500 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2501 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2502 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2503 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2504 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2505 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2506 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2507 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2508 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2509 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2510 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2511 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2512 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2513 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2514 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2515 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2516 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2517 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2518 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2519 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2520 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2521 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2522 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2523 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2524 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2525 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2526 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2527 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2528 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2529 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2530 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2531 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2532 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2533 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2534 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2535 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002536 "src/f32-relu/gen/avx512f-x16.c",
2537 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002539 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2540 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2541 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2542 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2543 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2544 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2545 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2546 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002547 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2548 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2549 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2550 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2551 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2552 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2553 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2554 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002555 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2556 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2557 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2558 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2559 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2560 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2561 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2562 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002563 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2564 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2565 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2566 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002567 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2568 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2569 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2570 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002571 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2572 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002573 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002574 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2575 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2576 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2577 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2578 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2579 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2580 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2581 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2582 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2583 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2584 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2585 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2586 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2587 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2588 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2589 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2590 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2591 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2592 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2593 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2594 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2595 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2596 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2597 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002598 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2599 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2600 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2601 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2602 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2603 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2604 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2605 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002606 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2607 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2608 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2609 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2610 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2611 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2612 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2613 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002614 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2615 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2616 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2617 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2618 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2619 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002620 "src/math/exp-avx512f-p5-scalef.c",
2621 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002622 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002623 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002624 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002625 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002626 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002627 "src/math/sigmoid-avx512f-rr1-p5-div.c",
2628 "src/math/sigmoid-avx512f-rr1-p5-nr1fma.c",
2629 "src/math/sigmoid-avx512f-rr1-p5-nr2fma.c",
2630 "src/math/sigmoid-avx512f-rr2-p5-div.c",
2631 "src/math/sigmoid-avx512f-rr2-p5-nr1fma.c",
2632 "src/math/sigmoid-avx512f-rr2-p5-nr2fma.c",
2633 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2634 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2635 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr2fma.c",
2636 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2637 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2638 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr2fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002639 "src/math/sqrt-avx512f-nr1fma.c",
2640 "src/math/sqrt-avx512f-nr2fma.c",
2641 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002642]
2643
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002644AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002645 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2646 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002647 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2648 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2649 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2650 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2651 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2652 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2653 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2654 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2655]
2656
Frank Barchardbcedc082020-08-17 18:00:51 -07002657WASM32_ASM_UKERNELS = [
2658 "src/f32-relu/wasm_shr_x1.S",
2659 "src/f32-relu/wasm_shr_x2.S",
2660 "src/f32-relu/wasm_shr_x4.S",
2661]
2662
Marat Dukhan08c4a432019-10-03 09:29:21 -07002663AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002664 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002665 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002666 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2667 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002668 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002669 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002670 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002671 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002672 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2673 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002674 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2675 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2676 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2677 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678]
2679
2680AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002681 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002682 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2683 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002684 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002685 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2686 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002687 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2688 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2689 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002690 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002691 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2692 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2693 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002694 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002695 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2696 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002697 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002698 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2699 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2700 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2701 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2702 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2703 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2704 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2705 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2706 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2707 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2708 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2709 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2710 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2711 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2712 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2713 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2714 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2715 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002716 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2717 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002718 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002719 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2720 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2721 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2722 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2723 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2724 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2725 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2726 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2727 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2728 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2729 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2730 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2731 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2732 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2733 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2734 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2735 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2736 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002737 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2738 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2739 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2740 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2741 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2742 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2743 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2744 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2745 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2746 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2747 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2748 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2749 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2750 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2751 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2752 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2753 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2754 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755]
2756
Marat Dukhan1b354632020-03-23 12:50:22 -07002757INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 "src/xnnpack/argmaxpool.h",
2759 "src/xnnpack/avgpool.h",
2760 "src/xnnpack/clamp.h",
2761 "src/xnnpack/common.h",
2762 "src/xnnpack/conv.h",
2763 "src/xnnpack/dwconv.h",
2764 "src/xnnpack/gavgpool.h",
2765 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002766 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002768 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002770 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002771 "src/xnnpack/lut.h",
2772 "src/xnnpack/math.h",
2773 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002774 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002775 "src/xnnpack/packx.h",
2776 "src/xnnpack/pad.h",
2777 "src/xnnpack/params.h",
2778 "src/xnnpack/pavgpool.h",
2779 "src/xnnpack/ppmm.h",
2780 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002781 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002782 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002783 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002784 "src/xnnpack/rmax.h",
2785 "src/xnnpack/scalar-utils.h",
2786 "src/xnnpack/spmm.h",
2787 "src/xnnpack/unpool.h",
2788 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002789 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002791 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002792 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002793 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002794 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002796]
2797
2798INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002799 "include/xnnpack.h",
2800 "src/xnnpack/allocator.h",
2801 "src/xnnpack/compute.h",
2802 "src/xnnpack/im2col.h",
2803 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002804 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002805 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 "src/xnnpack/operator.h",
2807 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002808 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002809 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002810 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002811 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002812]
2813
Marat Dukhan1b354632020-03-23 12:50:22 -07002814ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002815 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002816]
2817
Marat Dukhan1b354632020-03-23 12:50:22 -07002818MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002819 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 "include/xnnpack.h",
2821]
2822
Marat Dukhan1b354632020-03-23 12:50:22 -07002823MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002825 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002826 "src/xnnpack/requantization.h",
2827 "include/xnnpack.h",
2828]
2829
2830OPERATOR_TEST_PARAMS_HDRS = [
2831 "src/xnnpack/params.h",
2832 "src/xnnpack/common.h",
2833]
2834
2835WEIGHTS_PACK_HDRS = [
2836 "src/xnnpack/pack.h",
2837 "src/xnnpack/operator.h",
2838 "src/xnnpack/compute.h",
2839]
2840
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002841LOGGING_COPTS = select({
2842 # No logging in optimized mode
2843 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2844 # Full logging in debug mode
2845 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2846 # Error-only logging in default (fastbuild) mode
2847 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2848})
2849
Marat Dukhan3b59de22020-06-03 20:15:19 -07002850LOGGING_SRCS = select({
2851 # No logging in optimized mode
2852 ":optimized_build": [],
2853 "//conditions:default": [
2854 "src/operator-strings.c",
2855 "src/subgraph-strings.c",
2856 ],
2857})
2858
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002859LOGGING_HDRS = [
2860 "src/xnnpack/log.h",
2861]
2862
Marat Dukhan08c4a432019-10-03 09:29:21 -07002863xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002864 name = "tables",
2865 srcs = TABLE_SRCS,
2866 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002867 gcc_copts = xnnpack_gcc_std_copts(),
2868 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002869)
2870
2871xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 name = "scalar_ukernels",
2873 srcs = SCALAR_UKERNELS,
2874 hdrs = INTERNAL_HDRS,
2875 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002876 gcc_copts = xnnpack_gcc_std_copts(),
2877 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002879 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 "@FP16",
2881 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002882 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002883 ],
2884)
2885
2886xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002887 name = "scalar_ukernels_test_mode",
2888 srcs = SCALAR_UKERNELS,
2889 hdrs = INTERNAL_HDRS,
2890 aarch32_copts = ["-marm"],
2891 copts = [
2892 "-UNDEBUG",
2893 "-DXNN_TEST_MODE=1",
2894 ],
2895 gcc_copts = xnnpack_gcc_std_copts(),
2896 msvc_copts = xnnpack_msvc_std_copts(),
2897 deps = [
2898 ":tables",
2899 "@FP16",
2900 "@FXdiv",
2901 "@pthreadpool",
2902 ],
2903)
2904
2905xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002906 name = "wasm_ukernels",
2907 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002908 gcc_copts = xnnpack_gcc_std_copts(),
2909 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002910 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002911 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002912 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002913 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002914 "@FP16",
2915 "@FXdiv",
2916 "@pthreadpool",
2917 ],
2918)
2919
2920xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002921 name = "wasm_ukernels_test_mode",
2922 hdrs = INTERNAL_HDRS,
2923 copts = [
2924 "-UNDEBUG",
2925 "-DXNN_TEST_MODE=1",
2926 ],
2927 gcc_copts = xnnpack_gcc_std_copts(),
2928 msvc_copts = xnnpack_msvc_std_copts(),
2929 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002930 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002931 deps = [
2932 ":tables",
2933 "@FP16",
2934 "@FXdiv",
2935 "@pthreadpool",
2936 ],
2937)
2938
2939xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002940 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 hdrs = INTERNAL_HDRS,
2942 aarch32_copts = [
2943 "-marm",
2944 "-mfpu=neon",
2945 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002946 gcc_copts = xnnpack_gcc_std_copts(),
2947 gcc_x86_copts = ["-msse2"],
2948 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 optimized_copts = [
2950 "-O3",
2951 "-ffast-math",
2952 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002953 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002955 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 "@FP16",
2957 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002958 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002959 ],
2960)
2961
2962xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002963 name = "psimd_fastmath_ukernels_test_mode",
2964 hdrs = INTERNAL_HDRS,
2965 aarch32_copts = [
2966 "-marm",
2967 "-mfpu=neon",
2968 ],
2969 copts = [
2970 "-UNDEBUG",
2971 "-DXNN_TEST_MODE=1",
2972 ],
2973 gcc_copts = xnnpack_gcc_std_copts(),
2974 gcc_x86_copts = ["-msse2"],
2975 msvc_copts = xnnpack_msvc_std_copts(),
2976 optimized_copts = [
2977 "-O3",
2978 "-ffast-math",
2979 ],
2980 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2981 deps = [
2982 ":tables",
2983 "@FP16",
2984 "@psimd",
2985 "@pthreadpool",
2986 ],
2987)
2988
2989xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002990 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002991 hdrs = INTERNAL_HDRS,
2992 aarch32_copts = [
2993 "-marm",
2994 "-mfpu=neon",
2995 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002996 gcc_copts = xnnpack_gcc_std_copts(),
2997 gcc_x86_copts = ["-msse2"],
2998 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002999 optimized_copts = [
3000 "-O3",
3001 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003002 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003003 deps = [
3004 ":tables",
3005 "@FP16",
3006 "@psimd",
3007 "@pthreadpool",
3008 ],
3009)
3010
3011xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003012 name = "psimd_accmath_ukernels_test_mode",
3013 hdrs = INTERNAL_HDRS,
3014 aarch32_copts = [
3015 "-marm",
3016 "-mfpu=neon",
3017 ],
3018 copts = [
3019 "-UNDEBUG",
3020 "-DXNN_TEST_MODE=1",
3021 ],
3022 gcc_copts = xnnpack_gcc_std_copts(),
3023 gcc_x86_copts = ["-msse2"],
3024 msvc_copts = xnnpack_msvc_std_copts(),
3025 optimized_copts = [
3026 "-O3",
3027 ],
3028 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3029 deps = [
3030 ":tables",
3031 "@FP16",
3032 "@psimd",
3033 "@pthreadpool",
3034 ],
3035)
3036
3037xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003038 name = "neon_ukernels",
3039 hdrs = INTERNAL_HDRS,
3040 aarch32_copts = [
3041 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003042 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 "-mfpu=neon",
3044 ],
3045 aarch32_srcs = NEON_UKERNELS,
3046 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003047 gcc_copts = xnnpack_gcc_std_copts(),
3048 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003049 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003050 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003051 "@FP16",
3052 "@pthreadpool",
3053 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054)
3055
3056xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003057 name = "neon_ukernels_test_mode",
3058 hdrs = INTERNAL_HDRS,
3059 aarch32_copts = [
3060 "-marm",
3061 "-march=armv7-a",
3062 "-mfpu=neon",
3063 ],
3064 aarch32_srcs = NEON_UKERNELS,
3065 aarch64_srcs = NEON_UKERNELS,
3066 copts = [
3067 "-UNDEBUG",
3068 "-DXNN_TEST_MODE=1",
3069 ],
3070 gcc_copts = xnnpack_gcc_std_copts(),
3071 msvc_copts = xnnpack_msvc_std_copts(),
3072 deps = [
3073 ":tables",
3074 "@FP16",
3075 "@pthreadpool",
3076 ],
3077)
3078
3079xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080 name = "neonfma_ukernels",
3081 hdrs = INTERNAL_HDRS,
3082 aarch32_copts = [
3083 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003084 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085 "-mfpu=neon-vfpv4",
3086 ],
3087 aarch32_srcs = NEONFMA_UKERNELS,
3088 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003089 apple_aarch32_copts = [
3090 "-mcpu=swift",
3091 "-mtune=generic",
3092 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003093 gcc_copts = xnnpack_gcc_std_copts(),
3094 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003095 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003096 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003097 "@FP16",
3098 "@pthreadpool",
3099 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003100)
3101
3102xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003103 name = "neonfma_ukernels_test_mode",
3104 hdrs = INTERNAL_HDRS,
3105 aarch32_copts = [
3106 "-marm",
3107 "-march=armv7-a",
3108 "-mfpu=neon-vfpv4",
3109 ],
3110 aarch32_srcs = NEONFMA_UKERNELS,
3111 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003112 apple_aarch32_copts = [
3113 "-mcpu=swift",
3114 "-mtune=generic",
3115 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003116 copts = [
3117 "-UNDEBUG",
3118 "-DXNN_TEST_MODE=1",
3119 ],
3120 gcc_copts = xnnpack_gcc_std_copts(),
3121 msvc_copts = xnnpack_msvc_std_copts(),
3122 deps = [
3123 ":tables",
3124 "@FP16",
3125 "@pthreadpool",
3126 ],
3127)
3128
3129xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003130 name = "neonv8_ukernels",
3131 hdrs = INTERNAL_HDRS,
3132 aarch32_copts = [
3133 "-marm",
3134 "-march=armv8-a",
3135 "-mfpu=neon-fp-armv8",
3136 ],
3137 aarch32_srcs = NEONV8_UKERNELS,
3138 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003139 apple_aarch32_copts = [
3140 "-mcpu=cyclone",
3141 "-mtune=generic",
3142 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003143 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(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003153 name = "neonv8_ukernels_test_mode",
3154 hdrs = INTERNAL_HDRS,
3155 aarch32_copts = [
3156 "-marm",
3157 "-march=armv8-a",
3158 "-mfpu=neon-fp-armv8",
3159 ],
3160 aarch32_srcs = NEONV8_UKERNELS,
3161 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003162 apple_aarch32_copts = [
3163 "-mcpu=cyclone",
3164 "-mtune=generic",
3165 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003166 copts = [
3167 "-UNDEBUG",
3168 "-DXNN_TEST_MODE=1",
3169 ],
3170 gcc_copts = xnnpack_gcc_std_copts(),
3171 msvc_copts = xnnpack_msvc_std_copts(),
3172 deps = [
3173 ":tables",
3174 "@FP16",
3175 "@pthreadpool",
3176 ],
3177)
3178
3179xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 name = "neonfp16arith_ukernels",
3181 hdrs = INTERNAL_HDRS,
3182 aarch64_copts = ["-march=armv8.2-a+fp16"],
3183 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003184 gcc_copts = xnnpack_gcc_std_copts(),
3185 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003186 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003187 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003188 "@FP16",
3189 "@pthreadpool",
3190 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191)
3192
3193xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003194 name = "neonfp16arith_ukernels_test_mode",
3195 hdrs = INTERNAL_HDRS,
3196 aarch64_copts = ["-march=armv8.2-a+fp16"],
3197 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3198 copts = [
3199 "-UNDEBUG",
3200 "-DXNN_TEST_MODE=1",
3201 ],
3202 gcc_copts = xnnpack_gcc_std_copts(),
3203 msvc_copts = xnnpack_msvc_std_copts(),
3204 deps = [
3205 ":tables",
3206 "@FP16",
3207 "@pthreadpool",
3208 ],
3209)
3210
3211xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003212 name = "neondot_ukernels",
3213 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003214 aarch32_copts = [
3215 "-marm",
3216 "-march=armv8.2-a+dotprod",
3217 "-mfpu=neon-fp-armv8",
3218 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003219 aarch32_srcs = NEONDOT_UKERNELS,
3220 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3221 aarch64_srcs = NEONDOT_UKERNELS,
3222 gcc_copts = xnnpack_gcc_std_copts(),
3223 msvc_copts = xnnpack_msvc_std_copts(),
3224 deps = [
3225 ":tables",
3226 "@FP16",
3227 "@pthreadpool",
3228 ],
3229)
3230
3231xnnpack_cc_library(
3232 name = "neondot_ukernels_test_mode",
3233 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003234 aarch32_copts = [
3235 "-marm",
3236 "-march=armv8.2-a+dotprod",
3237 "-mfpu=neon-fp-armv8",
3238 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003239 aarch32_srcs = NEONDOT_UKERNELS,
3240 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3241 aarch64_srcs = NEONDOT_UKERNELS,
3242 copts = [
3243 "-UNDEBUG",
3244 "-DXNN_TEST_MODE=1",
3245 ],
3246 gcc_copts = xnnpack_gcc_std_copts(),
3247 msvc_copts = xnnpack_msvc_std_copts(),
3248 deps = [
3249 ":tables",
3250 "@FP16",
3251 "@pthreadpool",
3252 ],
3253)
3254
3255xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256 name = "sse2_ukernels",
3257 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003258 gcc_copts = xnnpack_gcc_std_copts(),
3259 gcc_x86_copts = ["-msse2"],
3260 msvc_copts = xnnpack_msvc_std_copts(),
3261 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003263 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003264 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003265 "@FP16",
3266 "@pthreadpool",
3267 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268)
3269
3270xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003271 name = "sse2_ukernels_test_mode",
3272 hdrs = INTERNAL_HDRS,
3273 copts = [
3274 "-UNDEBUG",
3275 "-DXNN_TEST_MODE=1",
3276 ],
3277 gcc_copts = xnnpack_gcc_std_copts(),
3278 gcc_x86_copts = ["-msse2"],
3279 msvc_copts = xnnpack_msvc_std_copts(),
3280 msvc_x86_32_copts = ["/arch:SSE2"],
3281 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3282 deps = [
3283 ":tables",
3284 "@FP16",
3285 "@pthreadpool",
3286 ],
3287)
3288
3289xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003290 name = "ssse3_ukernels",
3291 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003292 gcc_copts = xnnpack_gcc_std_copts(),
3293 gcc_x86_copts = ["-mssse3"],
3294 msvc_copts = xnnpack_msvc_std_copts(),
3295 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003296 x86_srcs = SSSE3_UKERNELS,
3297 deps = [
3298 ":tables",
3299 "@FP16",
3300 "@pthreadpool",
3301 ],
3302)
3303
3304xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003305 name = "ssse3_ukernels_test_mode",
3306 hdrs = INTERNAL_HDRS,
3307 copts = [
3308 "-UNDEBUG",
3309 "-DXNN_TEST_MODE=1",
3310 ],
3311 gcc_copts = xnnpack_gcc_std_copts(),
3312 gcc_x86_copts = ["-mssse3"],
3313 msvc_copts = xnnpack_msvc_std_copts(),
3314 msvc_x86_32_copts = ["/arch:SSE2"],
3315 x86_srcs = SSSE3_UKERNELS,
3316 deps = [
3317 ":tables",
3318 "@FP16",
3319 "@pthreadpool",
3320 ],
3321)
3322
3323xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003324 name = "sse41_ukernels",
3325 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003326 gcc_copts = xnnpack_gcc_std_copts(),
3327 gcc_x86_copts = ["-msse4.1"],
3328 msvc_copts = xnnpack_msvc_std_copts(),
3329 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003330 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003331 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003332 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003333 "@FP16",
3334 "@pthreadpool",
3335 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003336)
3337
3338xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003339 name = "sse41_ukernels_test_mode",
3340 hdrs = INTERNAL_HDRS,
3341 copts = [
3342 "-UNDEBUG",
3343 "-DXNN_TEST_MODE=1",
3344 ],
3345 gcc_copts = xnnpack_gcc_std_copts(),
3346 gcc_x86_copts = ["-msse4.1"],
3347 msvc_copts = xnnpack_msvc_std_copts(),
3348 msvc_x86_32_copts = ["/arch:SSE2"],
3349 x86_srcs = SSE41_UKERNELS,
3350 deps = [
3351 ":tables",
3352 "@FP16",
3353 "@pthreadpool",
3354 ],
3355)
3356
3357xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003358 name = "avx_ukernels",
3359 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003360 gcc_copts = xnnpack_gcc_std_copts(),
3361 gcc_x86_copts = ["-mavx"],
3362 msvc_copts = xnnpack_msvc_std_copts(),
3363 msvc_x86_32_copts = ["/arch:AVX"],
3364 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003366 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003367 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003368 "@FP16",
3369 "@pthreadpool",
3370 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003371)
3372
3373xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003374 name = "avx_ukernels_test_mode",
3375 hdrs = INTERNAL_HDRS,
3376 copts = [
3377 "-UNDEBUG",
3378 "-DXNN_TEST_MODE=1",
3379 ],
3380 gcc_copts = xnnpack_gcc_std_copts(),
3381 gcc_x86_copts = ["-mavx"],
3382 msvc_copts = xnnpack_msvc_std_copts(),
3383 msvc_x86_32_copts = ["/arch:AVX"],
3384 msvc_x86_64_copts = ["/arch:AVX"],
3385 x86_srcs = AVX_UKERNELS,
3386 deps = [
3387 ":tables",
3388 "@FP16",
3389 "@pthreadpool",
3390 ],
3391)
3392
3393xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003394 name = "xop_ukernels",
3395 hdrs = INTERNAL_HDRS,
3396 gcc_copts = xnnpack_gcc_std_copts(),
3397 gcc_x86_copts = ["-mxop"],
3398 msvc_copts = xnnpack_msvc_std_copts(),
3399 msvc_x86_32_copts = ["/arch:AVX"],
3400 msvc_x86_64_copts = ["/arch:AVX"],
3401 x86_srcs = XOP_UKERNELS,
3402 deps = [
3403 ":tables",
3404 "@FP16",
3405 "@pthreadpool",
3406 ],
3407)
3408
3409xnnpack_cc_library(
3410 name = "xop_ukernels_test_mode",
3411 hdrs = INTERNAL_HDRS,
3412 copts = [
3413 "-UNDEBUG",
3414 "-DXNN_TEST_MODE=1",
3415 ],
3416 gcc_copts = xnnpack_gcc_std_copts(),
3417 gcc_x86_copts = ["-mxop"],
3418 msvc_copts = xnnpack_msvc_std_copts(),
3419 msvc_x86_32_copts = ["/arch:AVX"],
3420 msvc_x86_64_copts = ["/arch:AVX"],
3421 x86_srcs = XOP_UKERNELS,
3422 deps = [
3423 ":tables",
3424 "@FP16",
3425 "@pthreadpool",
3426 ],
3427)
3428
3429xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003430 name = "fma3_ukernels",
3431 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003432 gcc_copts = xnnpack_gcc_std_copts(),
3433 gcc_x86_copts = ["-mfma"],
3434 msvc_copts = xnnpack_msvc_std_copts(),
3435 msvc_x86_32_copts = ["/arch:AVX"],
3436 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003437 x86_srcs = FMA3_UKERNELS,
3438 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003439 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003440 "@FP16",
3441 "@pthreadpool",
3442 ],
3443)
3444
3445xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003446 name = "fma3_ukernels_test_mode",
3447 hdrs = INTERNAL_HDRS,
3448 copts = [
3449 "-UNDEBUG",
3450 "-DXNN_TEST_MODE=1",
3451 ],
3452 gcc_copts = xnnpack_gcc_std_copts(),
3453 gcc_x86_copts = ["-mfma"],
3454 msvc_copts = xnnpack_msvc_std_copts(),
3455 msvc_x86_32_copts = ["/arch:AVX"],
3456 msvc_x86_64_copts = ["/arch:AVX"],
3457 x86_srcs = FMA3_UKERNELS,
3458 deps = [
3459 ":tables",
3460 "@FP16",
3461 "@pthreadpool",
3462 ],
3463)
3464
3465xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003466 name = "avx2_ukernels",
3467 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003468 gcc_copts = xnnpack_gcc_std_copts(),
3469 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003470 "-mfma",
3471 "-mavx2",
3472 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003473 msvc_copts = xnnpack_msvc_std_copts(),
3474 msvc_x86_32_copts = ["/arch:AVX2"],
3475 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003476 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003477 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003478 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003479 "@FP16",
3480 "@pthreadpool",
3481 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003482)
3483
3484xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003485 name = "avx2_ukernels_test_mode",
3486 hdrs = INTERNAL_HDRS,
3487 copts = [
3488 "-UNDEBUG",
3489 "-DXNN_TEST_MODE=1",
3490 ],
3491 gcc_copts = xnnpack_gcc_std_copts(),
3492 gcc_x86_copts = [
3493 "-mfma",
3494 "-mavx2",
3495 ],
3496 msvc_copts = xnnpack_msvc_std_copts(),
3497 msvc_x86_32_copts = ["/arch:AVX2"],
3498 msvc_x86_64_copts = ["/arch:AVX2"],
3499 x86_srcs = AVX2_UKERNELS,
3500 deps = [
3501 ":tables",
3502 "@FP16",
3503 "@pthreadpool",
3504 ],
3505)
3506
3507xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 name = "avx512f_ukernels",
3509 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003510 gcc_copts = xnnpack_gcc_std_copts(),
3511 gcc_x86_copts = ["-mavx512f"],
3512 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3513 msvc_copts = xnnpack_msvc_std_copts(),
3514 msvc_x86_32_copts = ["/arch:AVX512"],
3515 msvc_x86_64_copts = ["/arch:AVX512"],
3516 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003518 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003519 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003520 "@FP16",
3521 "@pthreadpool",
3522 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523)
3524
3525xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003526 name = "avx512f_ukernels_test_mode",
3527 hdrs = INTERNAL_HDRS,
3528 copts = [
3529 "-UNDEBUG",
3530 "-DXNN_TEST_MODE=1",
3531 ],
3532 gcc_copts = xnnpack_gcc_std_copts(),
3533 gcc_x86_copts = ["-mavx512f"],
3534 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3535 msvc_copts = xnnpack_msvc_std_copts(),
3536 msvc_x86_32_copts = ["/arch:AVX512"],
3537 msvc_x86_64_copts = ["/arch:AVX512"],
3538 msys_copts = ["-fno-asynchronous-unwind-tables"],
3539 x86_srcs = AVX512F_UKERNELS,
3540 deps = [
3541 ":tables",
3542 "@FP16",
3543 "@pthreadpool",
3544 ],
3545)
3546
3547xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003548 name = "avx512skx_ukernels",
3549 hdrs = INTERNAL_HDRS,
3550 gcc_copts = xnnpack_gcc_std_copts(),
3551 gcc_x86_copts = [
3552 "-mavx512f",
3553 "-mavx512cd",
3554 "-mavx512bw",
3555 "-mavx512dq",
3556 "-mavx512vl",
3557 ],
3558 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3559 msvc_copts = xnnpack_msvc_std_copts(),
3560 msvc_x86_32_copts = ["/arch:AVX512"],
3561 msvc_x86_64_copts = ["/arch:AVX512"],
3562 msys_copts = ["-fno-asynchronous-unwind-tables"],
3563 x86_srcs = AVX512SKX_UKERNELS,
3564 deps = [
3565 ":tables",
3566 "@FP16",
3567 "@pthreadpool",
3568 ],
3569)
3570
3571xnnpack_cc_library(
3572 name = "avx512skx_ukernels_test_mode",
3573 hdrs = INTERNAL_HDRS,
3574 copts = [
3575 "-UNDEBUG",
3576 "-DXNN_TEST_MODE=1",
3577 ],
3578 gcc_copts = xnnpack_gcc_std_copts(),
3579 gcc_x86_copts = [
3580 "-mavx512f",
3581 "-mavx512cd",
3582 "-mavx512bw",
3583 "-mavx512dq",
3584 "-mavx512vl",
3585 ],
3586 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3587 msvc_copts = xnnpack_msvc_std_copts(),
3588 msvc_x86_32_copts = ["/arch:AVX512"],
3589 msvc_x86_64_copts = ["/arch:AVX512"],
3590 msys_copts = ["-fno-asynchronous-unwind-tables"],
3591 x86_srcs = AVX512SKX_UKERNELS,
3592 deps = [
3593 ":tables",
3594 "@FP16",
3595 "@pthreadpool",
3596 ],
3597)
3598
3599xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 name = "asm_ukernels",
3601 hdrs = ["src/xnnpack/assembly.h"],
3602 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003603 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003604 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003605 wasm_srcs = WASM32_ASM_UKERNELS,
3606 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607)
3608
Marat Dukhan3b59de22020-06-03 20:15:19 -07003609xnnpack_cc_library(
3610 name = "logging_utils",
3611 srcs = LOGGING_SRCS,
3612 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3613 copts = LOGGING_COPTS + [
3614 "-Isrc",
3615 "-Iinclude",
3616 ] + select({
3617 ":debug_build": [],
3618 "//conditions:default": xnnpack_min_size_copts(),
3619 }),
3620 gcc_copts = xnnpack_gcc_std_copts(),
3621 msvc_copts = xnnpack_msvc_std_copts(),
3622 visibility = xnnpack_visibility(),
3623 deps = [
3624 "@FP16",
3625 "@clog",
3626 "@pthreadpool",
3627 ],
3628)
3629
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630xnnpack_aggregate_library(
3631 name = "ukernels",
3632 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 ":neon_ukernels",
3634 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003635 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003636 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637 ":asm_ukernels",
3638 ],
3639 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003640 ":neon_ukernels",
3641 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003642 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003644 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 ":asm_ukernels",
3646 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003647 generic_deps = [
3648 ":scalar_ukernels",
3649 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003650 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003651 ":psimd_fastmath_ukernels",
3652 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003654 wasm_deps = [
3655 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003656 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003657 ],
3658 wasmsimd_deps = [
3659 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003660 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003661 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003664 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003665 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003667 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003668 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003669 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003671 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672 ],
3673)
3674
Marat Dukhan33fcf782020-05-24 14:27:15 -07003675xnnpack_aggregate_library(
3676 name = "ukernels_test_mode",
3677 aarch32_deps = [
3678 ":neon_ukernels_test_mode",
3679 ":neonfma_ukernels_test_mode",
3680 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003681 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003682 ":asm_ukernels",
3683 ],
3684 aarch64_deps = [
3685 ":neon_ukernels_test_mode",
3686 ":neonfma_ukernels_test_mode",
3687 ":neonv8_ukernels_test_mode",
3688 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003689 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003690 ":asm_ukernels",
3691 ],
3692 generic_deps = [
3693 ":scalar_ukernels_test_mode",
3694 ],
3695 psimd_deps = [
3696 ":psimd_fastmath_ukernels_test_mode",
3697 ":psimd_accmath_ukernels_test_mode",
3698 ],
3699 wasm_deps = [
3700 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003701 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003702 ],
3703 wasmsimd_deps = [
3704 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003705 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003706 ],
3707 x86_deps = [
3708 ":sse2_ukernels_test_mode",
3709 ":ssse3_ukernels_test_mode",
3710 ":sse41_ukernels_test_mode",
3711 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003712 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003713 ":fma3_ukernels_test_mode",
3714 ":avx2_ukernels_test_mode",
3715 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003716 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003717 ],
3718)
3719
Marat Dukhan86df3842020-06-17 14:03:31 -07003720# This target must be compatible with Android NDK r17 and compile with Clang 6.
3721# It is similar to :ukernels target, but excludes microkernels which require
3722# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003723xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003724 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003725 aarch32_deps = [
3726 ":neon_ukernels",
3727 ":neonfma_ukernels",
3728 ":neonv8_ukernels",
3729 ":asm_ukernels",
3730 ],
3731 aarch64_deps = [
3732 ":neon_ukernels",
3733 ":neonfma_ukernels",
3734 ":neonv8_ukernels",
3735 ":asm_ukernels",
3736 ],
3737 generic_deps = [
3738 ":scalar_ukernels",
3739 ],
3740 psimd_deps = [
3741 ":psimd_fastmath_ukernels",
3742 ":psimd_accmath_ukernels",
3743 ],
3744 wasm_deps = [
3745 ":wasm_ukernels",
3746 ],
3747 wasmsimd_deps = [
3748 ":wasm_ukernels",
3749 ],
3750 x86_deps = [
3751 ":sse2_ukernels",
3752 ":ssse3_ukernels",
3753 ":sse41_ukernels",
3754 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003755 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003756 ":fma3_ukernels",
3757 ":avx2_ukernels",
3758 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003759 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003760 ],
3761)
3762
Marat Dukhan08c4a432019-10-03 09:29:21 -07003763xnnpack_cc_library(
3764 name = "im2col",
3765 srcs = ["src/im2col.c"],
3766 hdrs = [
3767 "src/xnnpack/common.h",
3768 "src/xnnpack/im2col.h",
3769 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003770 gcc_copts = xnnpack_gcc_std_copts(),
3771 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772)
3773
3774xnnpack_cc_library(
3775 name = "indirection",
3776 srcs = ["src/indirection.c"],
3777 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003778 gcc_copts = xnnpack_gcc_std_copts(),
3779 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003780 deps = [
3781 "@FP16",
3782 "@FXdiv",
3783 "@pthreadpool",
3784 ],
3785)
3786
3787xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003788 name = "indirection_test_mode",
3789 srcs = ["src/indirection.c"],
3790 hdrs = INTERNAL_HDRS,
3791 copts = [
3792 "-UNDEBUG",
3793 "-DXNN_TEST_MODE=1",
3794 ],
3795 gcc_copts = xnnpack_gcc_std_copts(),
3796 msvc_copts = xnnpack_msvc_std_copts(),
3797 deps = [
3798 "@FP16",
3799 "@FXdiv",
3800 "@pthreadpool",
3801 ],
3802)
3803
3804xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003805 name = "packing",
3806 srcs = ["src/packing.c"],
3807 hdrs = INTERNAL_HDRS,
3808 gcc_copts = xnnpack_gcc_std_copts(),
3809 msvc_copts = xnnpack_msvc_std_copts(),
3810 deps = [
3811 "@FP16",
3812 "@FXdiv",
3813 "@pthreadpool",
3814 ],
3815)
3816
3817xnnpack_cc_library(
3818 name = "packing_test_mode",
3819 srcs = ["src/packing.c"],
3820 hdrs = INTERNAL_HDRS,
3821 copts = [
3822 "-UNDEBUG",
3823 "-DXNN_TEST_MODE=1",
3824 ],
3825 gcc_copts = xnnpack_gcc_std_copts(),
3826 msvc_copts = xnnpack_msvc_std_copts(),
3827 deps = [
3828 "@FP16",
3829 "@FXdiv",
3830 "@pthreadpool",
3831 ],
3832)
3833
3834xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003835 name = "operator_run",
3836 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003837 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003838 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003839 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3840 "//conditions:default": [],
3841 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003842 gcc_copts = xnnpack_gcc_std_copts(),
3843 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003845 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846 "@FP16",
3847 "@FXdiv",
3848 "@clog",
3849 "@pthreadpool",
3850 ],
3851)
3852
Chao Mei6ddfc602020-05-13 22:29:36 -07003853xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003854 name = "operator_run_test_mode",
3855 srcs = ["src/operator-run.c"],
3856 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3857 copts = LOGGING_COPTS + [
3858 "-UNDEBUG",
3859 "-DXNN_TEST_MODE=1",
3860 ] + select({
3861 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3862 "//conditions:default": [],
3863 }),
3864 gcc_copts = xnnpack_gcc_std_copts(),
3865 msvc_copts = xnnpack_msvc_std_copts(),
3866 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003867 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003868 "@FP16",
3869 "@FXdiv",
3870 "@clog",
3871 "@pthreadpool",
3872 ],
3873)
3874
3875xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003876 name = "memory_planner",
3877 srcs = ["src/memory-planner.c"],
3878 hdrs = INTERNAL_HDRS,
3879 defines = select({
3880 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3881 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3882 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3883 }),
3884 gcc_copts = xnnpack_gcc_std_copts(),
3885 msvc_copts = xnnpack_msvc_std_copts(),
3886 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003887 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003888 "@pthreadpool",
3889 ],
3890)
3891
Marat Dukhan33fcf782020-05-24 14:27:15 -07003892xnnpack_cc_library(
3893 name = "memory_planner_test_mode",
3894 srcs = ["src/memory-planner.c"],
3895 hdrs = INTERNAL_HDRS,
3896 copts = [
3897 "-UNDEBUG",
3898 "-DXNN_TEST_MODE=1",
3899 ],
3900 defines = select({
3901 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3902 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3903 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3904 }),
3905 gcc_copts = xnnpack_gcc_std_copts(),
3906 msvc_copts = xnnpack_msvc_std_copts(),
3907 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003908 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003909 "@pthreadpool",
3910 ],
3911)
3912
Marat Dukhan08c4a432019-10-03 09:29:21 -07003913cc_library(
3914 name = "enable_assembly",
3915 defines = select({
3916 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3917 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003918 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003919 }),
3920)
3921
Marat Dukhan9de90e02020-06-18 16:04:12 -07003922cc_library(
3923 name = "enable_sparse",
3924 defines = select({
3925 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3926 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003927 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003928 }),
3929)
3930
Marat Dukhancf056b22019-10-07 10:26:29 -07003931xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932 name = "operators",
3933 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003934 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003935 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003936 ],
3937 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003938 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003939 "-Isrc",
3940 "-Iinclude",
3941 ] + select({
3942 ":debug_build": [],
3943 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003944 }) + select({
3945 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3946 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003947 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003948 gcc_copts = xnnpack_gcc_std_copts(),
3949 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003950 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003952 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003953 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 "@FP16",
3955 "@FXdiv",
3956 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003957 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003958 ],
3959)
3960
Marat Dukhan10a38082020-04-17 03:58:35 -07003961xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003962 name = "operators_test_mode",
3963 srcs = OPERATOR_SRCS + [
3964 "src/memory.c",
3965 "src/operator-delete.c",
3966 ],
3967 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3968 copts = LOGGING_COPTS + [
3969 "-Isrc",
3970 "-Iinclude",
3971 "-UNDEBUG",
3972 "-DXNN_TEST_MODE=1",
3973 ] + select({
3974 ":debug_build": [],
3975 "//conditions:default": xnnpack_min_size_copts(),
3976 }) + select({
3977 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3978 "//conditions:default": [],
3979 }),
3980 gcc_copts = xnnpack_gcc_std_copts(),
3981 msvc_copts = xnnpack_msvc_std_copts(),
3982 deps = [
3983 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003984 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003985 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003986 "@FP16",
3987 "@FXdiv",
3988 "@clog",
3989 "@pthreadpool",
3990 ],
3991)
3992
3993xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003994 name = "XNNPACK",
3995 srcs = [
3996 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003997 "src/runtime.c",
3998 "src/subgraph.c",
3999 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004000 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004001 hdrs = ["include/xnnpack.h"],
4002 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004003 "-Isrc",
4004 "-Iinclude",
4005 ] + select({
4006 ":debug_build": [],
4007 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004008 }) + select({
4009 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4010 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004011 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004012 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004013 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004015 visibility = xnnpack_visibility(),
4016 deps = [
4017 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004018 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004019 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004020 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004021 ":operator_run",
4022 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004023 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004024 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004025 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004026 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004027 ] + select({
4028 ":emscripten": [],
4029 "//conditions:default": ["@cpuinfo"],
4030 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004031)
4032
Marat Dukhan10a38082020-04-17 03:58:35 -07004033xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004034 name = "XNNPACK_test_mode",
4035 srcs = [
4036 "src/init.c",
4037 "src/runtime.c",
4038 "src/subgraph.c",
4039 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004040 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004041 hdrs = ["include/xnnpack.h"],
4042 copts = LOGGING_COPTS + [
4043 "-Isrc",
4044 "-Iinclude",
4045 "-UNDEBUG",
4046 "-DXNN_TEST_MODE=1",
4047 ] + select({
4048 ":debug_build": [],
4049 "//conditions:default": xnnpack_min_size_copts(),
4050 }) + select({
4051 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4052 "//conditions:default": [],
4053 }),
4054 gcc_copts = xnnpack_gcc_std_copts(),
4055 includes = ["include"],
4056 msvc_copts = xnnpack_msvc_std_copts(),
4057 visibility = xnnpack_visibility(),
4058 deps = [
4059 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004060 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004061 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004062 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004063 ":operator_run_test_mode",
4064 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004065 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004066 "@clog",
4067 "@FP16",
4068 "@pthreadpool",
4069 ] + select({
4070 ":emscripten": [],
4071 "//conditions:default": ["@cpuinfo"],
4072 }),
4073)
4074
4075xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004076 name = "xnnpack_f32",
4077 srcs = [
4078 "src/init.c",
4079 "src/runtime.c",
4080 "src/subgraph.c",
4081 "src/tensor.c",
4082 ] + SUBGRAPH_SRCS,
4083 hdrs = ["include/xnnpack.h"],
4084 copts = LOGGING_COPTS + [
4085 "-Isrc",
4086 "-Iinclude",
4087 ] + select({
4088 ":debug_build": [],
4089 "//conditions:default": xnnpack_min_size_copts(),
4090 }) + select({
4091 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4092 "//conditions:default": [],
4093 }),
4094 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004095 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004096 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004097 "XNN_NO_U8_OPERATORS",
4098 "XNN_NO_X8_OPERATORS",
4099 "XNN_NO_F16_OPERATORS",
4100 "XNN_NO_X16_OPERATORS",
4101 ],
4102 gcc_copts = xnnpack_gcc_std_copts(),
4103 includes = ["include"],
4104 msvc_copts = xnnpack_msvc_std_copts(),
4105 visibility = xnnpack_visibility(),
4106 deps = [
4107 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004108 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004109 ":logging_utils",
4110 ":memory_planner",
4111 ":operator_run",
4112 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004113 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004114 "@clog",
4115 "@FP16",
4116 "@pthreadpool",
4117 ] + select({
4118 ":emscripten": [],
4119 "//conditions:default": ["@cpuinfo"],
4120 }),
4121)
4122
4123xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004124 name = "xnnpack_operators_nhwc_f32",
4125 srcs = [
4126 "src/init.c",
4127 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004128 hdrs = ["include/xnnpack.h"],
4129 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004130 "-Isrc",
4131 "-Iinclude",
4132 ] + select({
4133 ":debug_build": [],
4134 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004135 }) + select({
4136 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4137 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004138 }),
4139 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004140 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004141 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004142 "XNN_NO_U8_OPERATORS",
4143 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004144 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004145 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004146 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004148 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004149 visibility = xnnpack_visibility(),
4150 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004151 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004152 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 ":operator_run",
4154 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004155 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004156 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004158 ] + select({
4159 ":emscripten": [],
4160 "//conditions:default": ["@cpuinfo"],
4161 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004162)
4163
Marat Dukhancf056b22019-10-07 10:26:29 -07004164xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004165 name = "bench_utils",
4166 srcs = ["bench/utils.cc"],
4167 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004168 deps = [
4169 "@com_google_benchmark//:benchmark",
4170 "@cpuinfo",
4171 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172)
4173
Frank Barchard7e955972019-10-11 10:34:25 -07004174######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004175
4176xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004177 name = "qs8_gemm_bench",
4178 srcs = [
4179 "bench/gemm.h",
4180 "bench/qs8-gemm.cc",
4181 "src/xnnpack/AlignedAllocator.h",
4182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4183 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4184)
4185
4186xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004187 name = "qs8_requantization_bench",
4188 srcs = [
4189 "bench/qs8-requantization.cc",
4190 "src/xnnpack/requantization-stubs.h",
4191 "src/xnnpack/AlignedAllocator.h",
4192 ] + MICROKERNEL_BENCHMARK_HDRS,
4193 deps = MICROKERNEL_BENCHMARK_DEPS,
4194)
4195
4196xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004197 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 srcs = [
4199 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004200 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004201 "src/xnnpack/AlignedAllocator.h",
4202 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004203 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004204 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205)
4206
4207xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004208 name = "qu8_requantization_bench",
4209 srcs = [
4210 "bench/qu8-requantization.cc",
4211 "src/xnnpack/requantization-stubs.h",
4212 "src/xnnpack/AlignedAllocator.h",
4213 ] + MICROKERNEL_BENCHMARK_HDRS,
4214 deps = MICROKERNEL_BENCHMARK_DEPS,
4215)
4216
4217xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004218 name = "f16_igemm_bench",
4219 srcs = [
4220 "bench/f16-igemm.cc",
4221 "bench/conv.h",
4222 "bench/google/conv.h",
4223 "src/xnnpack/AlignedAllocator.h",
4224 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004225 deps = MICROKERNEL_BENCHMARK_DEPS + [
4226 ":indirection",
4227 ":packing",
4228 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004229)
4230
4231xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004232 name = "f16_gemm_bench",
4233 srcs = [
4234 "bench/f16-gemm.cc",
4235 "bench/gemm.h",
4236 "src/xnnpack/AlignedAllocator.h",
4237 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004238 deps = MICROKERNEL_BENCHMARK_DEPS + [
4239 ":packing",
4240 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004241)
4242
4243xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004244 name = "f16_spmm_bench",
4245 srcs = [
4246 "bench/f16-spmm.cc",
4247 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004248 "src/xnnpack/AlignedAllocator.h",
4249 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004250 deps = MICROKERNEL_BENCHMARK_DEPS,
4251)
4252
4253xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004254 name = "f32_igemm_bench",
4255 srcs = [
4256 "bench/f32-igemm.cc",
4257 "bench/conv.h",
4258 "src/xnnpack/AlignedAllocator.h",
4259 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004260 deps = MICROKERNEL_BENCHMARK_DEPS + [
4261 ":indirection",
4262 ":packing",
4263 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004264)
4265
4266xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004267 name = "f16_relu_bench",
4268 srcs = [
4269 "bench/f16-relu.cc",
4270 "src/xnnpack/AlignedAllocator.h",
4271 ] + MICROKERNEL_BENCHMARK_HDRS,
4272 deps = MICROKERNEL_BENCHMARK_DEPS,
4273)
4274
4275xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 name = "f32_conv_hwc_bench",
4277 srcs = [
4278 "bench/f32-conv-hwc.cc",
4279 "bench/dconv.h",
4280 "src/xnnpack/AlignedAllocator.h",
4281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004282 deps = MICROKERNEL_BENCHMARK_DEPS + [
4283 ":packing",
4284 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285)
4286
4287xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004288 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004289 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004290 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004291 "bench/dconv.h",
4292 "src/xnnpack/AlignedAllocator.h",
4293 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004294 deps = MICROKERNEL_BENCHMARK_DEPS + [
4295 ":packing",
4296 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004297)
4298
4299xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004300 name = "f16_dwconv_bench",
4301 srcs = [
4302 "bench/f16-dwconv.cc",
4303 "bench/dwconv.h",
4304 "bench/google/dwconv.h",
4305 "src/xnnpack/AlignedAllocator.h",
4306 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004307 deps = MICROKERNEL_BENCHMARK_DEPS + [
4308 ":indirection",
4309 ":packing",
4310 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004311)
4312
4313xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314 name = "f32_dwconv_bench",
4315 srcs = [
4316 "bench/f32-dwconv.cc",
4317 "bench/dwconv.h",
4318 "src/xnnpack/AlignedAllocator.h",
4319 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004320 deps = MICROKERNEL_BENCHMARK_DEPS + [
4321 ":indirection",
4322 ":packing",
4323 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324)
4325
4326xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004327 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004328 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004329 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004330 "bench/dwconv.h",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004333 deps = MICROKERNEL_BENCHMARK_DEPS + [
4334 ":indirection",
4335 ":packing",
4336 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337)
4338
4339xnnpack_benchmark(
4340 name = "f32_gemm_bench",
4341 srcs = [
4342 "bench/f32-gemm.cc",
4343 "bench/gemm.h",
4344 "src/xnnpack/AlignedAllocator.h",
4345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004346 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004347 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004348)
4349
4350xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004351 name = "f32_hswish_bench",
4352 srcs = [
4353 "bench/f32-hswish.cc",
4354 "src/xnnpack/AlignedAllocator.h",
4355 ] + MICROKERNEL_BENCHMARK_HDRS,
4356 deps = MICROKERNEL_BENCHMARK_DEPS,
4357)
4358
4359xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004360 name = "f32_raddexpminusmax_bench",
4361 srcs = [
4362 "bench/f32-raddexpminusmax.cc",
4363 "src/xnnpack/AlignedAllocator.h",
4364 ] + MICROKERNEL_BENCHMARK_HDRS,
4365 deps = MICROKERNEL_BENCHMARK_DEPS,
4366)
4367
4368xnnpack_benchmark(
4369 name = "f32_raddextexp_bench",
4370 srcs = [
4371 "bench/f32-raddextexp.cc",
4372 "src/xnnpack/AlignedAllocator.h",
4373 ] + MICROKERNEL_BENCHMARK_HDRS,
4374 deps = MICROKERNEL_BENCHMARK_DEPS,
4375)
4376
4377xnnpack_benchmark(
4378 name = "f32_raddstoreexpminusmax_bench",
4379 srcs = [
4380 "bench/f32-raddstoreexpminusmax.cc",
4381 "src/xnnpack/AlignedAllocator.h",
4382 ] + MICROKERNEL_BENCHMARK_HDRS,
4383 deps = MICROKERNEL_BENCHMARK_DEPS,
4384)
4385
4386xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004387 name = "f32_relu_bench",
4388 srcs = [
4389 "bench/f32-relu.cc",
4390 "src/xnnpack/AlignedAllocator.h",
4391 ] + MICROKERNEL_BENCHMARK_HDRS,
4392 deps = MICROKERNEL_BENCHMARK_DEPS,
4393)
4394
4395xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 name = "f32_rmax_bench",
4397 srcs = [
4398 "bench/f32-rmax.cc",
4399 "src/xnnpack/AlignedAllocator.h",
4400 ] + MICROKERNEL_BENCHMARK_HDRS,
4401 deps = MICROKERNEL_BENCHMARK_DEPS,
4402)
4403
4404xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004405 name = "f32_sigmoid_bench",
4406 srcs = [
4407 "bench/f32-sigmoid.cc",
4408 "src/xnnpack/AlignedAllocator.h",
4409 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004410 deps = MICROKERNEL_BENCHMARK_DEPS,
4411)
4412
4413xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414 name = "f32_spmm_bench",
4415 srcs = [
4416 "bench/f32-spmm.cc",
4417 "bench/gemm.h",
4418 "src/xnnpack/AlignedAllocator.h",
4419 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004420 deps = MICROKERNEL_BENCHMARK_DEPS,
4421)
4422
4423xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004424 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004425 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004426 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004427 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004428 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004429 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004430)
4431
4432xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004433 name = "f32_vscaleexpminusmax_bench",
4434 srcs = [
4435 "bench/f32-vscaleexpminusmax.cc",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + MICROKERNEL_BENCHMARK_HDRS,
4438 deps = MICROKERNEL_BENCHMARK_DEPS,
4439)
4440
4441xnnpack_benchmark(
4442 name = "f32_vscaleextexp_bench",
4443 srcs = [
4444 "bench/f32-vscaleextexp.cc",
4445 "src/xnnpack/AlignedAllocator.h",
4446 ] + MICROKERNEL_BENCHMARK_HDRS,
4447 deps = MICROKERNEL_BENCHMARK_DEPS,
4448)
4449
4450xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004451 name = "f32_vsqrt_bench",
4452 srcs = [
4453 "bench/f32-vsqrt.cc",
4454 "src/xnnpack/AlignedAllocator.h",
4455 ] + MICROKERNEL_BENCHMARK_HDRS,
4456 deps = MICROKERNEL_BENCHMARK_DEPS,
4457)
4458
4459xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 name = "f32_im2col_gemm_bench",
4461 srcs = [
4462 "bench/f32-im2col-gemm.cc",
4463 "bench/conv.h",
4464 "src/xnnpack/AlignedAllocator.h",
4465 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004466 deps = MICROKERNEL_BENCHMARK_DEPS + [
4467 ":im2col",
4468 ":packing",
4469 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470)
4471
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004472xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004473 name = "rounding_bench",
4474 srcs = [
4475 "bench/rounding.cc",
4476 "src/xnnpack/math-stubs.h",
4477 "src/xnnpack/AlignedAllocator.h",
4478 ] + MICROKERNEL_BENCHMARK_HDRS,
4479 deps = MICROKERNEL_BENCHMARK_DEPS,
4480)
4481
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482########################### Benchmarks for operators ###########################
4483
4484xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485 name = "average_pooling_bench",
4486 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004487 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004488 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004489 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490)
4491
4492xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004493 name = "bankers_rounding_bench",
4494 srcs = ["bench/bankers-rounding.cc"],
4495 copts = xnnpack_optional_tflite_copts(),
4496 tags = ["nowin32"],
4497 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4498)
4499
4500xnnpack_benchmark(
4501 name = "ceiling_bench",
4502 srcs = ["bench/ceiling.cc"],
4503 copts = xnnpack_optional_tflite_copts(),
4504 tags = ["nowin32"],
4505 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4506)
4507
4508xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004509 name = "channel_shuffle_bench",
4510 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004511 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512)
4513
4514xnnpack_benchmark(
4515 name = "convolution_bench",
4516 srcs = ["bench/convolution.cc"],
4517 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004518 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004519 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004520)
4521
4522xnnpack_benchmark(
4523 name = "deconvolution_bench",
4524 srcs = ["bench/deconvolution.cc"],
4525 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004526 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004527 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528)
4529
4530xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004531 name = "floor_bench",
4532 srcs = ["bench/floor.cc"],
4533 copts = xnnpack_optional_tflite_copts(),
4534 tags = ["nowin32"],
4535 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4536)
4537
4538xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004539 name = "global_average_pooling_bench",
4540 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004541 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004542)
4543
4544xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004545 name = "hardswish_bench",
4546 srcs = ["bench/hardswish.cc"],
4547 copts = xnnpack_optional_tflite_copts(),
4548 tags = ["nowin32"],
4549 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4550)
4551
4552xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 name = "max_pooling_bench",
4554 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004555 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556)
4557
4558xnnpack_benchmark(
4559 name = "sigmoid_bench",
4560 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004561 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004562 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004563 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004564)
4565
4566xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004567 name = "prelu_bench",
4568 srcs = ["bench/prelu.cc"],
4569 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004570 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004571 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004572)
4573
4574xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004575 name = "softmax_bench",
4576 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004577 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004578 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004579 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580)
4581
Marat Dukhan87727142020-06-24 15:24:10 -07004582xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004583 name = "square_root_bench",
4584 srcs = ["bench/square-root.cc"],
4585 copts = xnnpack_optional_tflite_copts(),
4586 tags = ["nowin32"],
4587 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4588)
4589
4590xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004591 name = "truncation_bench",
4592 srcs = ["bench/truncation.cc"],
4593 deps = OPERATOR_BENCHMARK_DEPS,
4594)
4595
Marat Dukhanc068bb62019-10-04 13:24:39 -07004596############################# End-to-end benchmarks ############################
4597
4598cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004599 name = "fp32_mobilenet_v1",
4600 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004601 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004602 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004603 linkstatic = True,
4604 deps = [
4605 ":XNNPACK",
4606 "@pthreadpool",
4607 ],
4608)
4609
4610cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004611 name = "fp16_mobilenet_v1",
4612 srcs = ["models/fp16-mobilenet-v1.cc"],
4613 hdrs = ["models/models.h"],
4614 copts = xnnpack_std_cxxopts(),
4615 linkstatic = True,
4616 deps = [
4617 ":XNNPACK",
4618 "@FP16",
4619 "@pthreadpool",
4620 ],
4621)
4622
4623cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004624 name = "qs8_mobilenet_v1",
4625 srcs = ["models/qs8-mobilenet-v1.cc"],
4626 hdrs = ["models/models.h"],
4627 copts = xnnpack_std_cxxopts(),
4628 linkstatic = True,
4629 deps = [
4630 ":XNNPACK",
4631 "@pthreadpool",
4632 ],
4633)
4634
4635cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004636 name = "qs8_mobilenet_v2",
4637 srcs = ["models/qs8-mobilenet-v2.cc"],
4638 hdrs = ["models/models.h"],
4639 copts = xnnpack_std_cxxopts(),
4640 linkstatic = True,
4641 deps = [
4642 ":XNNPACK",
4643 "@pthreadpool",
4644 ],
4645)
4646
4647cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004648 name = "fp32_mobilenet_v2",
4649 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004650 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004651 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004652 linkstatic = True,
4653 deps = [
4654 ":XNNPACK",
4655 "@pthreadpool",
4656 ],
4657)
4658
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004659cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004660 name = "fp16_mobilenet_v2",
4661 srcs = ["models/fp16-mobilenet-v2.cc"],
4662 hdrs = ["models/models.h"],
4663 copts = xnnpack_std_cxxopts(),
4664 linkstatic = True,
4665 deps = [
4666 ":XNNPACK",
4667 "@FP16",
4668 "@pthreadpool",
4669 ],
4670)
4671
4672cc_library(
4673 name = "fp32_mobilenet_v3_large",
4674 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004675 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004676 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004677 linkstatic = True,
4678 deps = [
4679 ":XNNPACK",
4680 "@pthreadpool",
4681 ],
4682)
4683
4684cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004685 name = "fp16_mobilenet_v3_large",
4686 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4687 hdrs = ["models/models.h"],
4688 copts = xnnpack_std_cxxopts(),
4689 linkstatic = True,
4690 deps = [
4691 ":XNNPACK",
4692 "@FP16",
4693 "@pthreadpool",
4694 ],
4695)
4696
4697cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004698 name = "fp32_mobilenet_v3_small",
4699 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004700 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004701 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004702 linkstatic = True,
4703 deps = [
4704 ":XNNPACK",
4705 "@pthreadpool",
4706 ],
4707)
4708
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004709cc_library(
4710 name = "fp16_mobilenet_v3_small",
4711 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4712 hdrs = ["models/models.h"],
4713 copts = xnnpack_std_cxxopts(),
4714 linkstatic = True,
4715 deps = [
4716 ":XNNPACK",
4717 "@FP16",
4718 "@pthreadpool",
4719 ],
4720)
4721
Marat Dukhanc068bb62019-10-04 13:24:39 -07004722xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004723 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004724 srcs = [
4725 "bench/f32-dwconv-e2e.cc",
4726 "bench/end2end.h",
4727 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004728 deps = MICROKERNEL_BENCHMARK_DEPS + [
4729 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004730 ":fp32_mobilenet_v1",
4731 ":fp32_mobilenet_v2",
4732 ":fp32_mobilenet_v3_large",
4733 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004734 ],
4735)
4736
4737xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004738 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004739 srcs = [
4740 "bench/f32-gemm-e2e.cc",
4741 "bench/end2end.h",
4742 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004743 deps = MICROKERNEL_BENCHMARK_DEPS + [
4744 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004745 ":fp32_mobilenet_v1",
4746 ":fp32_mobilenet_v2",
4747 ":fp32_mobilenet_v3_large",
4748 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004749 ],
4750)
4751
4752xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004753 name = "end2end_bench",
4754 srcs = ["bench/end2end.cc"],
4755 deps = [
4756 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004757 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004758 ":fp16_mobilenet_v1",
4759 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004760 ":fp16_mobilenet_v3_large",
4761 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004762 ":fp32_mobilenet_v1",
4763 ":fp32_mobilenet_v2",
4764 ":fp32_mobilenet_v3_large",
4765 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004766 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004767 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004768 "@pthreadpool",
4769 ],
4770)
4771
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004772#################### Accuracy evaluation for math functions ####################
4773
4774xnnpack_benchmark(
4775 name = "f32_exp_eval",
4776 srcs = [
4777 "eval/f32-exp.cc",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + ACCURACY_EVAL_HDRS,
4780 deps = ACCURACY_EVAL_DEPS,
4781)
4782
Marat Dukhan515c9772019-10-17 18:07:57 -07004783xnnpack_benchmark(
4784 name = "f32_expminus_eval",
4785 srcs = [
4786 "eval/f32-expminus.cc",
4787 "src/xnnpack/AlignedAllocator.h",
4788 ] + ACCURACY_EVAL_HDRS,
4789 deps = ACCURACY_EVAL_DEPS,
4790)
4791
Marat Dukhan98ba4412019-10-23 02:14:28 -07004792xnnpack_benchmark(
4793 name = "f32_extexp_eval",
4794 srcs = [
4795 "eval/f32-extexp.cc",
4796 "src/xnnpack/AlignedAllocator.h",
4797 ] + ACCURACY_EVAL_HDRS,
4798 deps = ACCURACY_EVAL_DEPS,
4799)
4800
Marat Dukhan8853b822020-05-07 12:19:01 -07004801xnnpack_unit_test(
4802 name = "f32_roundne_eval",
4803 srcs = [
4804 "eval/f32-roundne.cc",
4805 "src/xnnpack/AlignedAllocator.h",
4806 "src/xnnpack/math-stubs.h",
4807 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004808 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004809 deps = MICROKERNEL_TEST_DEPS,
4810)
4811
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004812xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004813 name = "f32_roundd_eval",
4814 srcs = [
4815 "eval/f32-roundd.cc",
4816 "src/xnnpack/AlignedAllocator.h",
4817 "src/xnnpack/math-stubs.h",
4818 ] + MICROKERNEL_TEST_HDRS,
4819 automatic = False,
4820 deps = MICROKERNEL_TEST_DEPS,
4821)
4822
4823xnnpack_unit_test(
4824 name = "f32_roundu_eval",
4825 srcs = [
4826 "eval/f32-roundu.cc",
4827 "src/xnnpack/AlignedAllocator.h",
4828 "src/xnnpack/math-stubs.h",
4829 ] + MICROKERNEL_TEST_HDRS,
4830 automatic = False,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004835 name = "f32_roundz_eval",
4836 srcs = [
4837 "eval/f32-roundz.cc",
4838 "src/xnnpack/AlignedAllocator.h",
4839 "src/xnnpack/math-stubs.h",
4840 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004841 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004842 deps = MICROKERNEL_TEST_DEPS,
4843)
4844
Marat Dukhan346a9e52019-11-15 09:06:30 -08004845xnnpack_benchmark(
4846 name = "f32_sigmoid_eval",
4847 srcs = [
4848 "eval/f32-sigmoid.cc",
4849 "src/xnnpack/AlignedAllocator.h",
4850 ] + ACCURACY_EVAL_HDRS,
4851 deps = ACCURACY_EVAL_DEPS,
4852)
4853
Marat Dukhan84000762020-06-29 18:38:43 -07004854xnnpack_benchmark(
4855 name = "f32_sqrt_eval",
4856 srcs = [
4857 "eval/f32-sqrt.cc",
4858 "src/xnnpack/AlignedAllocator.h",
4859 ] + ACCURACY_EVAL_HDRS,
4860 deps = ACCURACY_EVAL_DEPS + [
4861 ":bench_utils",
4862 ],
4863)
4864
Marat Dukhan08c4a432019-10-03 09:29:21 -07004865######################### Unit tests for micro-kernels #########################
4866
4867xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004868 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004870 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871 "test/gemm-microkernel-tester.h",
4872 "src/xnnpack/AlignedAllocator.h",
4873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004874 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004878 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004879 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004880 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004881 "test/spmm-microkernel-tester.h",
4882 "src/xnnpack/AlignedAllocator.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 deps = MICROKERNEL_TEST_DEPS,
4885)
4886
4887xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004888 name = "f32_argmaxpool_test",
4889 srcs = [
4890 "test/f32-argmaxpool.cc",
4891 "test/argmaxpool-microkernel-tester.h",
4892 "src/xnnpack/AlignedAllocator.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004898 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004900 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 "test/avgpool-microkernel-tester.h",
4902 "src/xnnpack/AlignedAllocator.h",
4903 ] + MICROKERNEL_TEST_HDRS,
4904 deps = MICROKERNEL_TEST_DEPS,
4905)
4906
4907xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004908 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004909 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004910 "test/f32-ibilinear.cc",
4911 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004912 "src/xnnpack/AlignedAllocator.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004918 name = "f16_clamp_test",
4919 srcs = [
4920 "test/f16-clamp.cc",
4921 "test/clamp-microkernel-tester.h",
4922 ] + MICROKERNEL_TEST_HDRS,
4923 deps = MICROKERNEL_TEST_DEPS,
4924)
4925
4926xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004927 name = "f32_clamp_test",
4928 srcs = [
4929 "test/f32-clamp.cc",
4930 "test/clamp-microkernel-tester.h",
4931 ] + MICROKERNEL_TEST_HDRS,
4932 deps = MICROKERNEL_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004936 name = "f32_igemm_test",
4937 srcs = [
4938 "test/f32-igemm.cc",
4939 "test/gemm-microkernel-tester.h",
4940 "src/xnnpack/AlignedAllocator.h",
4941 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004942 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004943)
4944
4945xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004946 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004947 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004948 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004949 "test/gemm-microkernel-tester.h",
4950 "src/xnnpack/AlignedAllocator.h",
4951 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004952 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004953)
4954
4955xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004956 name = "f32_igemm_minmax_test",
4957 srcs = [
4958 "test/f32-igemm-minmax.cc",
4959 "test/gemm-microkernel-tester.h",
4960 "src/xnnpack/AlignedAllocator.h",
4961 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004962 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004963)
4964
4965xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004966 name = "f16_igemm_minmax_test",
4967 srcs = [
4968 "test/f16-igemm-minmax.cc",
4969 "test/gemm-microkernel-tester.h",
4970 "src/xnnpack/AlignedAllocator.h",
4971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004972 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004973)
4974
4975xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 name = "f32_conv_hwc_test",
4977 srcs = [
4978 "test/f32-conv-hwc.cc",
4979 "test/conv-hwc-microkernel-tester.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004982 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983)
4984
4985xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004986 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004988 "test/f32-conv-hwc2chw.cc",
4989 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004990 "src/xnnpack/AlignedAllocator.h",
4991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004992 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993)
4994
4995xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004996 name = "f32_dwconv_test",
4997 srcs = [
4998 "test/f32-dwconv.cc",
4999 "test/dwconv-microkernel-tester.h",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005002 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005003)
5004
5005xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005006 name = "f16_dwconv_minmax_test",
5007 srcs = [
5008 "test/f16-dwconv-minmax.cc",
5009 "test/dwconv-microkernel-tester.h",
5010 "src/xnnpack/AlignedAllocator.h",
5011 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005012 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005013)
5014
5015xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005016 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005018 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019 "test/dwconv-microkernel-tester.h",
5020 "src/xnnpack/AlignedAllocator.h",
5021 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005022 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005023)
5024
5025xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005026 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005028 "test/f32-dwconv-chw.cc",
5029 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005030 "src/xnnpack/AlignedAllocator.h",
5031 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005032 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005033)
5034
5035xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005036 name = "f16_gavgpool_minmax_test",
5037 srcs = [
5038 "test/f16-gavgpool-minmax.cc",
5039 "test/gavgpool-microkernel-tester.h",
5040 "src/xnnpack/AlignedAllocator.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_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005048 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049 "test/gavgpool-microkernel-tester.h",
5050 "src/xnnpack/AlignedAllocator.h",
5051 ] + MICROKERNEL_TEST_HDRS,
5052 deps = MICROKERNEL_TEST_DEPS,
5053)
5054
5055xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005056 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005058 "test/f32-gavgpool-cw.cc",
5059 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005060 "src/xnnpack/AlignedAllocator.h",
5061 ] + MICROKERNEL_TEST_HDRS,
5062 deps = MICROKERNEL_TEST_DEPS,
5063)
5064
5065xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005066 name = "f32_gemm_test",
5067 srcs = [
5068 "test/f32-gemm.cc",
5069 "test/gemm-microkernel-tester.h",
5070 "src/xnnpack/AlignedAllocator.h",
5071 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005072 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005073)
5074
5075xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005076 name = "f32_gemm_relu_test",
5077 srcs = [
5078 "test/f32-gemm-relu.cc",
5079 "test/gemm-microkernel-tester.h",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005082 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005083)
5084
5085xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005086 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005087 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005088 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089 "test/gemm-microkernel-tester.h",
5090 "src/xnnpack/AlignedAllocator.h",
5091 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005092 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093)
5094
5095xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005096 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005097 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005098 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005099 "test/gemm-microkernel-tester.h",
5100 "src/xnnpack/AlignedAllocator.h",
5101 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005102 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005103)
5104
5105xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005106 name = "f16_hswish_test",
5107 srcs = [
5108 "test/f16-hswish.cc",
5109 "test/hswish-microkernel-tester.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005115 name = "f32_hswish_test",
5116 srcs = [
5117 "test/f32-hswish.cc",
5118 "test/hswish-microkernel-tester.h",
5119 ] + MICROKERNEL_TEST_HDRS,
5120 deps = MICROKERNEL_TEST_DEPS,
5121)
5122
5123xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005124 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005126 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127 "test/maxpool-microkernel-tester.h",
5128 ] + MICROKERNEL_TEST_HDRS,
5129 deps = MICROKERNEL_TEST_DEPS,
5130)
5131
5132xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005133 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005135 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005136 "test/avgpool-microkernel-tester.h",
5137 "src/xnnpack/AlignedAllocator.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005143 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005145 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005146 "test/gemm-microkernel-tester.h",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005149 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005150)
5151
5152xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005153 name = "f16_prelu_test",
5154 srcs = [
5155 "test/f16-prelu.cc",
5156 "test/prelu-microkernel-tester.h",
5157 "src/xnnpack/AlignedAllocator.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163 name = "f32_prelu_test",
5164 srcs = [
5165 "test/f32-prelu.cc",
5166 "test/prelu-microkernel-tester.h",
5167 "src/xnnpack/AlignedAllocator.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005173 name = "f32_raddexpminusmax_test",
5174 srcs = [
5175 "test/f32-raddexpminusmax.cc",
5176 "test/raddexpminusmax-microkernel-tester.h",
5177 ] + MICROKERNEL_TEST_HDRS,
5178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005182 name = "f32_raddextexp_test",
5183 srcs = [
5184 "test/f32-raddextexp.cc",
5185 "test/raddextexp-microkernel-tester.h",
5186 ] + MICROKERNEL_TEST_HDRS,
5187 deps = MICROKERNEL_TEST_DEPS,
5188)
5189
5190xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005191 name = "f32_raddstoreexpminusmax_test",
5192 srcs = [
5193 "test/f32-raddstoreexpminusmax.cc",
5194 "test/raddstoreexpminusmax-microkernel-tester.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005200 name = "f32_relu_test",
5201 srcs = [
5202 "test/f32-relu.cc",
5203 "test/vunary-microkernel-tester.h",
5204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209 name = "f32_rmax_test",
5210 srcs = [
5211 "test/f32-rmax.cc",
5212 "test/rmax-microkernel-tester.h",
5213 ] + MICROKERNEL_TEST_HDRS,
5214 deps = MICROKERNEL_TEST_DEPS,
5215)
5216
5217xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005218 name = "f32_sigmoid_test",
5219 srcs = [
5220 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005221 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005222 ] + MICROKERNEL_TEST_HDRS,
5223 deps = MICROKERNEL_TEST_DEPS,
5224)
5225
5226xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005227 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005229 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 "test/spmm-microkernel-tester.h",
5231 "src/xnnpack/AlignedAllocator.h",
5232 ] + MICROKERNEL_TEST_HDRS,
5233 deps = MICROKERNEL_TEST_DEPS,
5234)
5235
5236xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005237 name = "f16_vadd_minmax_test",
5238 srcs = [
5239 "test/f16-vadd-minmax.cc",
5240 "test/vbinary-microkernel-tester.h",
5241 ] + MICROKERNEL_TEST_HDRS,
5242 deps = MICROKERNEL_TEST_DEPS,
5243)
5244
5245xnnpack_unit_test(
5246 name = "f16_vaddc_minmax_test",
5247 srcs = [
5248 "test/f16-vaddc-minmax.cc",
5249 "test/vbinaryc-microkernel-tester.h",
5250 ] + MICROKERNEL_TEST_HDRS,
5251 deps = MICROKERNEL_TEST_DEPS,
5252)
5253
5254xnnpack_unit_test(
5255 name = "f16_vdiv_minmax_test",
5256 srcs = [
5257 "test/f16-vdiv-minmax.cc",
5258 "test/vbinary-microkernel-tester.h",
5259 ] + MICROKERNEL_TEST_HDRS,
5260 deps = MICROKERNEL_TEST_DEPS,
5261)
5262
5263xnnpack_unit_test(
5264 name = "f16_vdivc_minmax_test",
5265 srcs = [
5266 "test/f16-vdivc-minmax.cc",
5267 "test/vbinaryc-microkernel-tester.h",
5268 ] + MICROKERNEL_TEST_HDRS,
5269 deps = MICROKERNEL_TEST_DEPS,
5270)
5271
5272xnnpack_unit_test(
5273 name = "f16_vrdivc_minmax_test",
5274 srcs = [
5275 "test/f16-vrdivc-minmax.cc",
5276 "test/vbinaryc-microkernel-tester.h",
5277 ] + MICROKERNEL_TEST_HDRS,
5278 deps = MICROKERNEL_TEST_DEPS,
5279)
5280
5281xnnpack_unit_test(
5282 name = "f16_vmax_test",
5283 srcs = [
5284 "test/f16-vmax.cc",
5285 "test/vbinary-microkernel-tester.h",
5286 ] + MICROKERNEL_TEST_HDRS,
5287 deps = MICROKERNEL_TEST_DEPS,
5288)
5289
5290xnnpack_unit_test(
5291 name = "f16_vmaxc_test",
5292 srcs = [
5293 "test/f16-vmaxc.cc",
5294 "test/vbinaryc-microkernel-tester.h",
5295 ] + MICROKERNEL_TEST_HDRS,
5296 deps = MICROKERNEL_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
5300 name = "f16_vmin_test",
5301 srcs = [
5302 "test/f16-vmin.cc",
5303 "test/vbinary-microkernel-tester.h",
5304 ] + MICROKERNEL_TEST_HDRS,
5305 deps = MICROKERNEL_TEST_DEPS,
5306)
5307
5308xnnpack_unit_test(
5309 name = "f16_vminc_test",
5310 srcs = [
5311 "test/f16-vminc.cc",
5312 "test/vbinaryc-microkernel-tester.h",
5313 ] + MICROKERNEL_TEST_HDRS,
5314 deps = MICROKERNEL_TEST_DEPS,
5315)
5316
5317xnnpack_unit_test(
5318 name = "f16_vmul_minmax_test",
5319 srcs = [
5320 "test/f16-vmul-minmax.cc",
5321 "test/vbinary-microkernel-tester.h",
5322 ] + MICROKERNEL_TEST_HDRS,
5323 deps = MICROKERNEL_TEST_DEPS,
5324)
5325
5326xnnpack_unit_test(
5327 name = "f16_vmulc_minmax_test",
5328 srcs = [
5329 "test/f16-vmulc-minmax.cc",
5330 "test/vbinaryc-microkernel-tester.h",
5331 ] + MICROKERNEL_TEST_HDRS,
5332 deps = MICROKERNEL_TEST_DEPS,
5333)
5334
5335xnnpack_unit_test(
5336 name = "f16_vsub_minmax_test",
5337 srcs = [
5338 "test/f16-vsub-minmax.cc",
5339 "test/vbinary-microkernel-tester.h",
5340 ] + MICROKERNEL_TEST_HDRS,
5341 deps = MICROKERNEL_TEST_DEPS,
5342)
5343
5344xnnpack_unit_test(
5345 name = "f16_vsubc_minmax_test",
5346 srcs = [
5347 "test/f16-vsubc-minmax.cc",
5348 "test/vbinaryc-microkernel-tester.h",
5349 ] + MICROKERNEL_TEST_HDRS,
5350 deps = MICROKERNEL_TEST_DEPS,
5351)
5352
5353xnnpack_unit_test(
5354 name = "f16_vrsubc_minmax_test",
5355 srcs = [
5356 "test/f16-vrsubc-minmax.cc",
5357 "test/vbinaryc-microkernel-tester.h",
5358 ] + MICROKERNEL_TEST_HDRS,
5359 deps = MICROKERNEL_TEST_DEPS,
5360)
5361
5362xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005363 name = "f32_vabs_test",
5364 srcs = [
5365 "test/f32-vabs.cc",
5366 "test/vunary-microkernel-tester.h",
5367 ] + MICROKERNEL_TEST_HDRS,
5368 deps = MICROKERNEL_TEST_DEPS,
5369)
5370
5371xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005372 name = "f32_vadd_test",
5373 srcs = [
5374 "test/f32-vadd.cc",
5375 "test/vbinary-microkernel-tester.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005381 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005382 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005383 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005384 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005385 ] + MICROKERNEL_TEST_HDRS,
5386 deps = MICROKERNEL_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005390 name = "f32_vadd_relu_test",
5391 srcs = [
5392 "test/f32-vadd-relu.cc",
5393 "test/vbinary-microkernel-tester.h",
5394 ] + MICROKERNEL_TEST_HDRS,
5395 deps = MICROKERNEL_TEST_DEPS,
5396)
5397
5398xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005399 name = "f32_vaddc_test",
5400 srcs = [
5401 "test/f32-vaddc.cc",
5402 "test/vbinaryc-microkernel-tester.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 deps = MICROKERNEL_TEST_DEPS,
5405)
5406
5407xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005408 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005409 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005410 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005411 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 ] + MICROKERNEL_TEST_HDRS,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005417 name = "f32_vaddc_relu_test",
5418 srcs = [
5419 "test/f32-vaddc-relu.cc",
5420 "test/vbinaryc-microkernel-tester.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005426 name = "f32_vdiv_test",
5427 srcs = [
5428 "test/f32-vdiv.cc",
5429 "test/vbinary-microkernel-tester.h",
5430 ] + MICROKERNEL_TEST_HDRS,
5431 deps = MICROKERNEL_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005435 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005436 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005437 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005438 "test/vbinary-microkernel-tester.h",
5439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005444 name = "f32_vdiv_relu_test",
5445 srcs = [
5446 "test/f32-vdiv-relu.cc",
5447 "test/vbinary-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005453 name = "f32_vdivc_test",
5454 srcs = [
5455 "test/f32-vdivc.cc",
5456 "test/vbinaryc-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005462 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005463 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005464 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005465 "test/vbinaryc-microkernel-tester.h",
5466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005471 name = "f32_vdivc_relu_test",
5472 srcs = [
5473 "test/f32-vdivc-relu.cc",
5474 "test/vbinaryc-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005480 name = "f32_vrdivc_test",
5481 srcs = [
5482 "test/f32-vrdivc.cc",
5483 "test/vbinaryc-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005489 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005490 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005491 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005492 "test/vbinaryc-microkernel-tester.h",
5493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005498 name = "f32_vrdivc_relu_test",
5499 srcs = [
5500 "test/f32-vrdivc-relu.cc",
5501 "test/vbinaryc-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005507 name = "f32_vmax_test",
5508 srcs = [
5509 "test/f32-vmax.cc",
5510 "test/vbinary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
5516 name = "f32_vmaxc_test",
5517 srcs = [
5518 "test/f32-vmaxc.cc",
5519 "test/vbinaryc-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
5525 name = "f32_vmin_test",
5526 srcs = [
5527 "test/f32-vmin.cc",
5528 "test/vbinary-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
5534 name = "f32_vminc_test",
5535 srcs = [
5536 "test/f32-vminc.cc",
5537 "test/vbinaryc-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005543 name = "f32_vmul_test",
5544 srcs = [
5545 "test/f32-vmul.cc",
5546 "test/vbinary-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005552 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005554 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005555 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005561 name = "f32_vmul_relu_test",
5562 srcs = [
5563 "test/f32-vmul-relu.cc",
5564 "test/vbinary-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005570 name = "f32_vmulc_test",
5571 srcs = [
5572 "test/f32-vmulc.cc",
5573 "test/vbinaryc-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005579 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005580 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005581 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005582 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005588 name = "f32_vmulc_relu_test",
5589 srcs = [
5590 "test/f32-vmulc-relu.cc",
5591 "test/vbinaryc-microkernel-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005597 name = "f16_vmulcaddc_minmax_test",
5598 srcs = [
5599 "test/f16-vmulcaddc-minmax.cc",
5600 "test/vmulcaddc-microkernel-tester.h",
5601 "src/xnnpack/AlignedAllocator.h",
5602 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005603 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005604)
5605
5606xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005607 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005608 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005609 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005610 "test/vmulcaddc-microkernel-tester.h",
5611 "src/xnnpack/AlignedAllocator.h",
5612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005613 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005614)
5615
5616xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005617 name = "f32_vlrelu_test",
5618 srcs = [
5619 "test/f32-vlrelu.cc",
5620 "test/vunary-microkernel-tester.h",
5621 ] + MICROKERNEL_TEST_HDRS,
5622 deps = MICROKERNEL_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005626 name = "f32_vneg_test",
5627 srcs = [
5628 "test/f32-vneg.cc",
5629 "test/vunary-microkernel-tester.h",
5630 ] + MICROKERNEL_TEST_HDRS,
5631 deps = MICROKERNEL_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005635 name = "f32_vrndne_test",
5636 srcs = [
5637 "test/f32-vrndne.cc",
5638 "test/vunary-microkernel-tester.h",
5639 ] + MICROKERNEL_TEST_HDRS,
5640 deps = MICROKERNEL_TEST_DEPS,
5641)
5642
5643xnnpack_unit_test(
5644 name = "f32_vrndz_test",
5645 srcs = [
5646 "test/f32-vrndz.cc",
5647 "test/vunary-microkernel-tester.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
5653 name = "f32_vrndu_test",
5654 srcs = [
5655 "test/f32-vrndu.cc",
5656 "test/vunary-microkernel-tester.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 deps = MICROKERNEL_TEST_DEPS,
5659)
5660
5661xnnpack_unit_test(
5662 name = "f32_vrndd_test",
5663 srcs = [
5664 "test/f32-vrndd.cc",
5665 "test/vunary-microkernel-tester.h",
5666 ] + MICROKERNEL_TEST_HDRS,
5667 deps = MICROKERNEL_TEST_DEPS,
5668)
5669
5670xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005671 name = "f32_vscale_test",
5672 srcs = [
5673 "test/f32-vscale.cc",
5674 "test/vscale-microkernel-tester.h",
5675 ] + MICROKERNEL_TEST_HDRS,
5676 deps = MICROKERNEL_TEST_DEPS,
5677)
5678
5679xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005680 name = "f32_vscaleexpminusmax_test",
5681 srcs = [
5682 "test/f32-vscaleexpminusmax.cc",
5683 "test/vscaleexpminusmax-microkernel-tester.h",
5684 ] + MICROKERNEL_TEST_HDRS,
5685 deps = MICROKERNEL_TEST_DEPS,
5686)
5687
5688xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005689 name = "f32_vscaleextexp_test",
5690 srcs = [
5691 "test/f32-vscaleextexp.cc",
5692 "test/vscaleextexp-microkernel-tester.h",
5693 ] + MICROKERNEL_TEST_HDRS,
5694 deps = MICROKERNEL_TEST_DEPS,
5695)
5696
5697xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005698 name = "f32_vsqr_test",
5699 srcs = [
5700 "test/f32-vsqr.cc",
5701 "test/vunary-microkernel-tester.h",
5702 ] + MICROKERNEL_TEST_HDRS,
5703 deps = MICROKERNEL_TEST_DEPS,
5704)
5705
5706xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005707 name = "f32_vsqrdiff_test",
5708 srcs = [
5709 "test/f32-vsqrdiff.cc",
5710 "test/vbinary-microkernel-tester.h",
5711 ] + MICROKERNEL_TEST_HDRS,
5712 deps = MICROKERNEL_TEST_DEPS,
5713)
5714
5715xnnpack_unit_test(
5716 name = "f32_vsqrdiffc_test",
5717 srcs = [
5718 "test/f32-vsqrdiffc.cc",
5719 "test/vbinaryc-microkernel-tester.h",
5720 ] + MICROKERNEL_TEST_HDRS,
5721 deps = MICROKERNEL_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005725 name = "f32_vsqrt_test",
5726 srcs = [
5727 "test/f32-vsqrt.cc",
5728 "test/vunary-microkernel-tester.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 deps = MICROKERNEL_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005734 name = "f32_vsub_test",
5735 srcs = [
5736 "test/f32-vsub.cc",
5737 "test/vbinary-microkernel-tester.h",
5738 ] + MICROKERNEL_TEST_HDRS,
5739 deps = MICROKERNEL_TEST_DEPS,
5740)
5741
5742xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005743 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005744 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005745 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005746 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005747 ] + MICROKERNEL_TEST_HDRS,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005752 name = "f32_vsub_relu_test",
5753 srcs = [
5754 "test/f32-vsub-relu.cc",
5755 "test/vbinary-microkernel-tester.h",
5756 ] + MICROKERNEL_TEST_HDRS,
5757 deps = MICROKERNEL_TEST_DEPS,
5758)
5759
5760xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005761 name = "f32_vsubc_test",
5762 srcs = [
5763 "test/f32-vsubc.cc",
5764 "test/vbinaryc-microkernel-tester.h",
5765 ] + MICROKERNEL_TEST_HDRS,
5766 deps = MICROKERNEL_TEST_DEPS,
5767)
5768
5769xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005770 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005771 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005772 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005773 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005779 name = "f32_vsubc_relu_test",
5780 srcs = [
5781 "test/f32-vsubc-relu.cc",
5782 "test/vbinaryc-microkernel-tester.h",
5783 ] + MICROKERNEL_TEST_HDRS,
5784 deps = MICROKERNEL_TEST_DEPS,
5785)
5786
5787xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005788 name = "f32_vrsubc_test",
5789 srcs = [
5790 "test/f32-vrsubc.cc",
5791 "test/vbinaryc-microkernel-tester.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005797 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005798 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005799 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005800 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005806 name = "f32_vrsubc_relu_test",
5807 srcs = [
5808 "test/f32-vrsubc-relu.cc",
5809 "test/vbinaryc-microkernel-tester.h",
5810 ] + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS,
5812)
5813
5814xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005815 name = "qs8_dwconv_minmax_test",
5816 srcs = [
5817 "test/qs8-dwconv-minmax.cc",
5818 "test/dwconv-microkernel-tester.h",
5819 "src/xnnpack/AlignedAllocator.h",
5820 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5821 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5822)
5823
5824xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005825 name = "qs8_gavgpool_minmax_test",
5826 srcs = [
5827 "test/qs8-gavgpool-minmax.cc",
5828 "test/gavgpool-microkernel-tester.h",
5829 "src/xnnpack/AlignedAllocator.h",
5830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005835 name = "qs8_gemm_minmax_test",
5836 srcs = [
5837 "test/qs8-gemm-minmax.cc",
5838 "test/gemm-microkernel-tester.h",
5839 "src/xnnpack/AlignedAllocator.h",
5840 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5841 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5842)
5843
5844xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005845 name = "qs8_igemm_minmax_test",
5846 srcs = [
5847 "test/qs8-igemm-minmax.cc",
5848 "test/gemm-microkernel-tester.h",
5849 "src/xnnpack/AlignedAllocator.h",
5850 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5851 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5852)
5853
5854xnnpack_unit_test(
5855 name = "qs8_requantization_test",
5856 srcs = [
5857 "src/xnnpack/requantization-stubs.h",
5858 "test/qs8-requantization.cc",
5859 "test/requantization-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005865 name = "qs8_vadd_minmax_test",
5866 srcs = [
5867 "test/qs8-vadd-minmax.cc",
5868 "test/vadd-microkernel-tester.h",
5869 ] + MICROKERNEL_TEST_HDRS,
5870 deps = MICROKERNEL_TEST_DEPS,
5871)
5872
5873xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005874 name = "qs8_vaddc_minmax_test",
5875 srcs = [
5876 "test/qs8-vaddc-minmax.cc",
5877 "test/vaddc-microkernel-tester.h",
5878 ] + MICROKERNEL_TEST_HDRS,
5879 deps = MICROKERNEL_TEST_DEPS,
5880)
5881
5882xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005883 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005885 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005886 "test/avgpool-microkernel-tester.h",
5887 "src/xnnpack/AlignedAllocator.h",
5888 ] + MICROKERNEL_TEST_HDRS,
5889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
5892xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005893 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005894 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005895 "test/qu8-dwconv-minmax.cc",
5896 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005897 "src/xnnpack/AlignedAllocator.h",
5898 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005899 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900)
5901
5902xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005903 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005904 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005905 "test/qu8-igemm-minmax.cc",
5906 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005907 "src/xnnpack/AlignedAllocator.h",
5908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005909 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910)
5911
5912xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005913 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005914 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005915 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005916 "test/gavgpool-microkernel-tester.h",
5917 "src/xnnpack/AlignedAllocator.h",
5918 ] + MICROKERNEL_TEST_HDRS,
5919 deps = MICROKERNEL_TEST_DEPS,
5920)
5921
5922xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005923 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005925 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926 "test/gemm-microkernel-tester.h",
5927 "src/xnnpack/AlignedAllocator.h",
5928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005929 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005930)
5931
5932xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005933 name = "qu8_requantization_test",
5934 srcs = [
5935 "src/xnnpack/requantization-stubs.h",
5936 "test/qu8-requantization.cc",
5937 "test/requantization-tester.h",
5938 ] + MICROKERNEL_TEST_HDRS,
5939 deps = MICROKERNEL_TEST_DEPS,
5940)
5941
5942xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005943 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005944 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005945 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005946 "test/vadd-microkernel-tester.h",
5947 ] + MICROKERNEL_TEST_HDRS,
5948 deps = MICROKERNEL_TEST_DEPS,
5949)
5950
5951xnnpack_unit_test(
5952 name = "u8_clamp_test",
5953 srcs = [
5954 "test/u8-clamp.cc",
5955 "test/clamp-microkernel-tester.h",
5956 ] + MICROKERNEL_TEST_HDRS,
5957 deps = MICROKERNEL_TEST_DEPS,
5958)
5959
5960xnnpack_unit_test(
5961 name = "u8_lut32norm_test",
5962 srcs = [
5963 "test/u8-lut32norm.cc",
5964 "test/lut-norm-microkernel-tester.h",
5965 ] + MICROKERNEL_TEST_HDRS,
5966 deps = MICROKERNEL_TEST_DEPS,
5967)
5968
5969xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005970 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005971 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005972 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005973 "test/maxpool-microkernel-tester.h",
5974 ] + MICROKERNEL_TEST_HDRS,
5975 deps = MICROKERNEL_TEST_DEPS,
5976)
5977
5978xnnpack_unit_test(
5979 name = "u8_rmax_test",
5980 srcs = [
5981 "test/u8-rmax.cc",
5982 "test/rmax-microkernel-tester.h",
5983 ] + MICROKERNEL_TEST_HDRS,
5984 deps = MICROKERNEL_TEST_DEPS,
5985)
5986
5987xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005988 name = "x32_fill_test",
5989 srcs = [
5990 "test/x32-fill.cc",
5991 "test/fill-microkernel-tester.h",
5992 ] + MICROKERNEL_TEST_HDRS,
5993 deps = MICROKERNEL_TEST_DEPS,
5994)
5995
5996xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005997 name = "x32_packx_test",
5998 srcs = [
5999 "test/x32-packx.cc",
6000 "test/pack-microkernel-tester.h",
6001 "src/xnnpack/AlignedAllocator.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
6007 name = "x32_pad_test",
6008 srcs = [
6009 "test/x32-pad.cc",
6010 "test/pad-microkernel-tester.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
6016 name = "x32_unpool_test",
6017 srcs = [
6018 "test/x32-unpool.cc",
6019 "test/unpool-microkernel-tester.h",
6020 ] + MICROKERNEL_TEST_HDRS,
6021 deps = MICROKERNEL_TEST_DEPS,
6022)
6023
6024xnnpack_unit_test(
6025 name = "x32_zip_test",
6026 srcs = [
6027 "test/x32-zip.cc",
6028 "test/zip-microkernel-tester.h",
6029 ] + MICROKERNEL_TEST_HDRS,
6030 deps = MICROKERNEL_TEST_DEPS,
6031)
6032
6033xnnpack_unit_test(
6034 name = "x8_lut_test",
6035 srcs = [
6036 "test/x8-lut.cc",
6037 "test/lut-microkernel-tester.h",
6038 ] + MICROKERNEL_TEST_HDRS,
6039 deps = MICROKERNEL_TEST_DEPS,
6040)
6041
6042xnnpack_unit_test(
6043 name = "x8_zip_test",
6044 srcs = [
6045 "test/x8-zip.cc",
6046 "test/zip-microkernel-tester.h",
6047 ] + MICROKERNEL_TEST_HDRS,
6048 deps = MICROKERNEL_TEST_DEPS,
6049)
6050
Marat Dukhan20c3b922020-03-10 03:45:06 -07006051########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052
6053xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006054 name = "operator_size_test",
6055 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006056 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006057)
6058
Marat Dukhan20c3b922020-03-10 03:45:06 -07006059xnnpack_binary(
6060 name = "subgraph_size_test",
6061 srcs = ["test/subgraph-size.c"],
6062 deps = [":XNNPACK"],
6063)
6064
6065########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066
6067xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006068 name = "abs_nc_test",
6069 srcs = [
6070 "test/abs-nc.cc",
6071 "test/abs-operator-tester.h",
6072 ],
6073 deps = OPERATOR_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006077 name = "add_nd_test",
6078 srcs = [
6079 "test/add-nd.cc",
6080 "test/binary-elementwise-operator-tester.h",
6081 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006082 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006083)
6084
6085xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006086 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006087 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006088 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006089 "test/argmax-pooling-operator-tester.h",
6090 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006091 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006092)
6093
6094xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006095 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006096 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006097 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006098 "test/average-pooling-operator-tester.h",
6099 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006100 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006101)
6102
6103xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006104 name = "bankers_rounding_nc_test",
6105 srcs = [
6106 "test/bankers-rounding-nc.cc",
6107 "test/bankers-rounding-operator-tester.h",
6108 ],
6109 deps = OPERATOR_TEST_DEPS,
6110)
6111
6112xnnpack_unit_test(
6113 name = "ceiling_nc_test",
6114 srcs = [
6115 "test/ceiling-nc.cc",
6116 "test/ceiling-operator-tester.h",
6117 ],
6118 deps = OPERATOR_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006122 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006123 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006124 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006125 "test/channel-shuffle-operator-tester.h",
6126 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006127 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006128)
6129
6130xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006131 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006133 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006134 "test/clamp-operator-tester.h",
6135 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006136 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006137)
6138
6139xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006140 name = "constant_pad_nd_test",
6141 srcs = [
6142 "test/constant-pad-nd.cc",
6143 "test/constant-pad-operator-tester.h",
6144 ],
6145 deps = OPERATOR_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006149 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006150 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006151 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152 "test/convolution-operator-tester.h",
6153 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006154 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006155)
6156
6157xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006158 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006159 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006160 "test/convolution-nchw.cc",
6161 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006163 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164)
6165
6166xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006167 name = "copy_nc_test",
6168 srcs = [
6169 "test/copy-nc.cc",
6170 "test/copy-operator-tester.h",
6171 ],
6172 deps = OPERATOR_TEST_DEPS,
6173)
6174
6175xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006176 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006177 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006178 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006179 "test/deconvolution-operator-tester.h",
6180 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006181 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006182)
6183
6184xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006185 name = "divide_nd_test",
6186 srcs = [
6187 "test/binary-elementwise-operator-tester.h",
6188 "test/divide-nd.cc",
6189 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006190 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006191)
6192
6193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006194 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006196 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197 "test/fully-connected-operator-tester.h",
6198 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200)
6201
6202xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006203 name = "floor_nc_test",
6204 srcs = [
6205 "test/floor-nc.cc",
6206 "test/floor-operator-tester.h",
6207 ],
6208 deps = OPERATOR_TEST_DEPS,
6209)
6210
6211xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006212 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006214 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006216 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006217 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218)
6219
6220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006221 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006223 "test/global-average-pooling-ncw.cc",
6224 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227)
6228
6229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006230 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006232 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 "test/hardswish-operator-tester.h",
6234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236)
6237
6238xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006239 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006241 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242 "test/leaky-relu-operator-tester.h",
6243 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245)
6246
6247xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006248 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006249 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006250 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006251 "test/max-pooling-operator-tester.h",
6252 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006254)
6255
6256xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006257 name = "maximum_nd_test",
6258 srcs = [
6259 "test/binary-elementwise-operator-tester.h",
6260 "test/maximum-nd.cc",
6261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006263)
6264
6265xnnpack_unit_test(
6266 name = "minimum_nd_test",
6267 srcs = [
6268 "test/binary-elementwise-operator-tester.h",
6269 "test/minimum-nd.cc",
6270 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006271 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006272)
6273
6274xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006275 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006276 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006277 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006278 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006279 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006280 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006281)
6282
6283xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006284 name = "negate_nc_test",
6285 srcs = [
6286 "test/negate-nc.cc",
6287 "test/negate-operator-tester.h",
6288 ],
6289 deps = OPERATOR_TEST_DEPS,
6290)
6291
6292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006293 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006295 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 "test/prelu-operator-tester.h",
6297 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299)
6300
6301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006302 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006304 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006305 "test/resize-bilinear-operator-tester.h",
6306 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006308)
6309
6310xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006311 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006312 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006313 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006314 "test/sigmoid-operator-tester.h",
6315 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006316 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317)
6318
6319xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006320 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006321 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006322 "test/softmax-nc.cc",
6323 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006326)
6327
6328xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006329 name = "square_nc_test",
6330 srcs = [
6331 "test/square-nc.cc",
6332 "test/square-operator-tester.h",
6333 ],
6334 deps = OPERATOR_TEST_DEPS,
6335)
6336
6337xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006338 name = "square_root_nc_test",
6339 srcs = [
6340 "test/square-root-nc.cc",
6341 "test/square-root-operator-tester.h",
6342 ],
6343 deps = OPERATOR_TEST_DEPS,
6344)
6345
6346xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006347 name = "squared_difference_nd_test",
6348 srcs = [
6349 "test/binary-elementwise-operator-tester.h",
6350 "test/squared-difference-nd.cc",
6351 ],
6352 deps = OPERATOR_TEST_DEPS,
6353)
6354
6355xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006356 name = "subtract_nd_test",
6357 srcs = [
6358 "test/binary-elementwise-operator-tester.h",
6359 "test/subtract-nd.cc",
6360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006362)
6363
6364xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006365 name = "truncation_nc_test",
6366 srcs = [
6367 "test/truncation-nc.cc",
6368 "test/truncation-operator-tester.h",
6369 ],
6370 deps = OPERATOR_TEST_DEPS,
6371)
6372
6373xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006374 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006375 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006376 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006377 "test/unpooling-operator-tester.h",
6378 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380)
6381
Chao Mei6ddfc602020-05-13 22:29:36 -07006382############################### Misc unit tests ###############################
6383
6384xnnpack_unit_test(
6385 name = "memory_planner_test",
6386 srcs = [
6387 "test/memory-planner-test.cc",
6388 ],
6389 deps = [
6390 ":XNNPACK",
6391 ":memory_planner",
6392 ],
6393)
6394
Marat Dukhan08c4a432019-10-03 09:29:21 -07006395############################# Build configurations #############################
6396
Marat Dukhanb8642352019-10-30 15:43:02 -07006397# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006399 name = "xnn_enable_assembly_explicit_true",
6400 define_values = {"xnn_enable_assembly": "true"},
6401)
6402
6403# Disables usage of assembly kernels.
6404config_setting(
6405 name = "xnn_enable_assembly_explicit_false",
6406 define_values = {"xnn_enable_assembly": "false"},
6407)
6408
Marat Dukhan9de90e02020-06-18 16:04:12 -07006409# Enables usage of sparse inference.
6410config_setting(
6411 name = "xnn_enable_sparse_explicit_true",
6412 define_values = {"xnn_enable_sparse": "true"},
6413)
6414
6415# Disables usage of sparse inference.
6416config_setting(
6417 name = "xnn_enable_sparse_explicit_false",
6418 define_values = {"xnn_enable_sparse": "false"},
6419)
6420
Marat Dukhan05702cf2020-03-26 15:41:33 -07006421# Disables usage of HMP-aware optimizations.
6422config_setting(
6423 name = "xnn_enable_hmp_explicit_false",
6424 define_values = {"xnn_enable_hmp": "false"},
6425)
6426
Chao Mei6ddfc602020-05-13 22:29:36 -07006427# Enable usage of optimized memory allocation
6428config_setting(
6429 name = "xnn_enable_memopt_explicit_true",
6430 define_values = {"xnnpack_enable_memopt": "true"},
6431)
6432
6433# Disable usage of optimized memory allocation
6434config_setting(
6435 name = "xnn_enable_memopt_explicit_false",
6436 define_values = {"xnnpack_enable_memopt": "false"},
6437)
6438
Marat Dukhanb8642352019-10-30 15:43:02 -07006439# Builds with -c dbg
6440config_setting(
6441 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006443 "compilation_mode": "dbg",
6444 },
6445)
6446
6447# Builds with -c opt
6448config_setting(
6449 name = "optimized_build",
6450 values = {
6451 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452 },
6453)
6454
6455config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006456 name = "linux_k8",
6457 values = {"cpu": "k8"},
6458)
6459
6460config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006461 name = "linux_arm",
6462 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006463)
6464
6465config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006466 name = "linux_armeabi",
6467 values = {"cpu": "armeabi"},
6468)
6469
6470config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006471 name = "linux_armhf",
6472 values = {"cpu": "armhf"},
6473)
6474
6475config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006476 name = "linux_armv7a",
6477 values = {"cpu": "armv7a"},
6478)
6479
6480config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006481 name = "linux_aarch64",
6482 values = {"cpu": "aarch64"},
6483)
6484
6485config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006486 name = "android",
6487 values = {"crosstool_top": "//external:android/crosstool"},
6488)
6489
6490config_setting(
6491 name = "android_armv7",
6492 values = {
6493 "crosstool_top": "//external:android/crosstool",
6494 "cpu": "armeabi-v7a",
6495 },
6496)
6497
6498config_setting(
6499 name = "android_arm64",
6500 values = {
6501 "crosstool_top": "//external:android/crosstool",
6502 "cpu": "arm64-v8a",
6503 },
6504)
6505
6506config_setting(
6507 name = "android_x86",
6508 values = {
6509 "crosstool_top": "//external:android/crosstool",
6510 "cpu": "x86",
6511 },
6512)
6513
6514config_setting(
6515 name = "android_x86_64",
6516 values = {
6517 "crosstool_top": "//external:android/crosstool",
6518 "cpu": "x86_64",
6519 },
6520)
6521
6522config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006523 name = "windows_x86_64",
6524 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006525)
6526
6527config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006528 name = "windows_x86_64_clang",
6529 values = {
6530 "compiler": "clang-cl",
6531 "cpu": "x64_windows",
6532 },
6533)
6534
6535config_setting(
6536 name = "windows_x86_64_mingw",
6537 values = {
6538 "compiler": "mingw-gcc",
6539 "cpu": "x64_windows",
6540 },
6541)
6542
6543config_setting(
6544 name = "windows_x86_64_msys",
6545 values = {
6546 "compiler": "msys-gcc",
6547 "cpu": "x64_windows",
6548 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006549)
6550
6551config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006552 name = "macos_x86_64",
6553 values = {
6554 "apple_platform_type": "macos",
6555 "cpu": "darwin",
6556 },
6557)
6558
6559config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006560 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006561 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006562)
6563
6564config_setting(
6565 name = "emscripten_wasm",
6566 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006567 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006568 "cpu": "wasm",
6569 },
6570)
6571
6572config_setting(
6573 name = "emscripten_wasmsimd",
6574 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006575 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006576 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006577 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006578 },
6579)
6580
6581config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006582 name = "ios_armv7",
6583 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006584 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006585 "cpu": "ios_armv7",
6586 },
6587)
6588
6589config_setting(
6590 name = "ios_arm64",
6591 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006592 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006593 "cpu": "ios_arm64",
6594 },
6595)
6596
6597config_setting(
6598 name = "ios_arm64e",
6599 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006600 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006601 "cpu": "ios_arm64e",
6602 },
6603)
6604
6605config_setting(
6606 name = "ios_x86",
6607 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006608 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006609 "cpu": "ios_i386",
6610 },
6611)
6612
6613config_setting(
6614 name = "ios_x86_64",
6615 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006616 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006617 "cpu": "ios_x86_64",
6618 },
6619)
6620
6621config_setting(
6622 name = "watchos_armv7k",
6623 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006624 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006625 "cpu": "watchos_armv7k",
6626 },
6627)
6628
6629config_setting(
6630 name = "watchos_arm64_32",
6631 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006632 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006633 "cpu": "watchos_arm64_32",
6634 },
6635)
6636
6637config_setting(
6638 name = "watchos_x86",
6639 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006640 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006641 "cpu": "watchos_i386",
6642 },
6643)
6644
6645config_setting(
6646 name = "watchos_x86_64",
6647 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006648 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006649 "cpu": "watchos_x86_64",
6650 },
6651)
6652
6653config_setting(
6654 name = "tvos_arm64",
6655 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006656 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006657 "cpu": "tvos_arm64",
6658 },
6659)
6660
6661config_setting(
6662 name = "tvos_x86_64",
6663 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006664 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006665 "cpu": "tvos_x86_64",
6666 },
6667)