blob: d4c1565200b876920e82c7feb6be86105e5138ce [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",
Marat Dukhand4c83032020-09-09 23:20:12 -07001755 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001756 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1757 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001758 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001759 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001760]
1761
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001763 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1764 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001765 "src/f32-clamp/gen/sse-x4.c",
1766 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001767 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1768 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001769 "src/f32-dwconv-chw/3x3p1-sse.c",
1770 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001771 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1772 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1773 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1775 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1777 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1779 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1780 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1781 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001783 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001784 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1785 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001786 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1787 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1788 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1789 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1790 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1791 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1792 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1793 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1794 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1795 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1796 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1797 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1798 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001799 "src/f32-hswish/gen/hswish-sse-x4.c",
1800 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001801 "src/f32-ibilinear/gen/sse-c4.c",
1802 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001803 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1804 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1805 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1806 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1807 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1808 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1809 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001810 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1811 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1812 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001813 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001814 "src/f32-prelu/gen/sse-2x4.c",
1815 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001816 "src/f32-relu/gen/sse-x4.c",
1817 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001819 "src/f32-spmm/gen/4x1-minmax-sse.c",
1820 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001821 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001822 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1826 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1829 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001830 "src/f32-vbinary/gen/vmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vmax-sse-x8.c",
1832 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1833 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1834 "src/f32-vbinary/gen/vmin-sse-x4.c",
1835 "src/f32-vbinary/gen/vmin-sse-x8.c",
1836 "src/f32-vbinary/gen/vminc-sse-x4.c",
1837 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1843 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1844 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1845 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001846 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1847 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1848 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1849 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001850 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1851 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1852 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001854 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1855 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001856 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1857 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001858 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1859 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001860 "src/f32-vunary/gen/vabs-sse-x4.c",
1861 "src/f32-vunary/gen/vabs-sse-x8.c",
1862 "src/f32-vunary/gen/vneg-sse-x4.c",
1863 "src/f32-vunary/gen/vneg-sse-x8.c",
1864 "src/f32-vunary/gen/vsqr-sse-x4.c",
1865 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001866 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001868 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001869 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001870 "src/math/roundd-sse-addsub.c",
1871 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001872 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001873 "src/math/sqrt-sse-hh1mac.c",
1874 "src/math/sqrt-sse-nr1mac.c",
1875 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876]
1877
1878SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001879 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1880 "src/f32-argmaxpool/4x-sse2-c4.c",
1881 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001882 "src/f32-prelu/gen/sse2-2x4.c",
1883 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001884 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1885 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001896 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001897 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001898 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001899 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001900 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1901 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001902 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1903 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001904 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1905 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1906 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1907 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1908 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1909 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1910 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1911 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001912 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1913 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1914 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001915 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1916 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1917 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001918 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1919 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1920 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001921 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1922 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001923 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1924 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001925 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1926 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001927 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1928 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001929 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001930 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1931 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001932 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001933 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1934 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1935 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001936 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1937 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001938 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1939 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001940 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1941 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001942 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001943 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1944 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001945 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001946 "src/qs8-requantization/precise-sse2.c",
1947 "src/qs8-requantization/fp32-sse2.c",
1948 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001949 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1950 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1951 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1952 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001953 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1954 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1955 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1956 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001957 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1958 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1959 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1960 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1961 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1962 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1963 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1964 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001965 "src/qu8-requantization/precise-sse2.c",
1966 "src/qu8-requantization/fp32-sse2.c",
1967 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001968 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001969 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001970 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001971 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001972 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973 "src/x32-zip/x2-sse2.c",
1974 "src/x32-zip/x3-sse2.c",
1975 "src/x32-zip/x4-sse2.c",
1976 "src/x32-zip/xm-sse2.c",
1977 "src/x8-zip/x2-sse2.c",
1978 "src/x8-zip/x3-sse2.c",
1979 "src/x8-zip/x4-sse2.c",
1980 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001981 "src/math/exp-sse2-p5.c",
1982 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001983 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001984 "src/math/roundd-sse2-cvt.c",
1985 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001986 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07001987 "src/math/sigmoid-sse2-rr2-p5-div.c",
1988 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001989]
1990
1991SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001992 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1993 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1994 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001995 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1996 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1997 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001998 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1999 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2000 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002001 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2002 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002003 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2004 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002005 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2006 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002007 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2008 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002009 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002010 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2011 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002012 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002013 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2014 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2015 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002016 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2017 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002018 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2019 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002020 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2021 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002022 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002023 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2024 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002025 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002026 "src/qs8-requantization/precise-ssse3.c",
2027 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002028 "src/qu8-requantization/precise-ssse3.c",
2029 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002030]
2031
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002032SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002033 "src/f32-prelu/gen/sse41-2x4.c",
2034 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002035 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002036 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002037 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002038 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002039 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2040 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002041 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2042 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002043 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2044 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2045 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2046 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2047 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2048 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2049 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2050 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002051 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2052 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2053 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002054 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2055 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2056 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002057 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2058 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2059 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002060 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2061 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002062 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2063 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002064 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2065 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002066 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2067 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002068 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002069 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2070 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002071 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002072 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2073 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2074 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002075 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2076 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002077 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2078 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002079 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2080 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002081 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002082 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2083 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002084 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002085 "src/qs8-requantization/fp32-sse4.c",
2086 "src/qs8-requantization/precise-sse4.c",
2087 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002088 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2089 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2090 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2091 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002092 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2093 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2094 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2095 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002096 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2097 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2098 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2099 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002100 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2101 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2102 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2103 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002104 "src/qu8-requantization/precise-sse4.c",
2105 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002106 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002107 "src/math/roundd-sse41.c",
2108 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002109 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002110]
2111
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002113 "src/f32-clamp/gen/avx-x8.c",
2114 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002115 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2116 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2117 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2118 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2119 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2120 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2121 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2122 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2123 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2124 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2125 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2126 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2127 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2128 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2129 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2130 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2131 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2132 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2133 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2134 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2135 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2136 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2137 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2138 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2139 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2140 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2141 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2142 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2143 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2144 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002145 "src/f32-hswish/gen/hswish-avx-x8.c",
2146 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002147 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2148 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2149 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2150 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2151 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2152 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2153 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2154 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2155 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002156 "src/f32-prelu/gen/avx-2x8.c",
2157 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002158 "src/f32-relu/gen/avx-x16.c",
2159 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002160 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002161 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2162 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2163 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2164 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2165 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2166 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2167 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2168 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002169 "src/f32-vbinary/gen/vmax-avx-x8.c",
2170 "src/f32-vbinary/gen/vmax-avx-x16.c",
2171 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2172 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2173 "src/f32-vbinary/gen/vmin-avx-x8.c",
2174 "src/f32-vbinary/gen/vmin-avx-x16.c",
2175 "src/f32-vbinary/gen/vminc-avx-x8.c",
2176 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002177 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2178 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2179 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2180 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2181 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2182 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2183 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2184 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002185 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2186 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2187 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2188 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002189 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2190 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2191 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2192 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002193 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2194 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002195 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002196 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2197 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2198 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2199 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2200 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2201 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2202 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2203 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002204 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2205 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002206 "src/f32-vunary/gen/vabs-avx-x8.c",
2207 "src/f32-vunary/gen/vabs-avx-x16.c",
2208 "src/f32-vunary/gen/vneg-avx-x8.c",
2209 "src/f32-vunary/gen/vneg-avx-x16.c",
2210 "src/f32-vunary/gen/vsqr-avx-x8.c",
2211 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002212]
2213
Marat Dukhan1566fee2020-08-02 21:55:41 -07002214XOP_UKERNELS = [
2215 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2216 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2217 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2218 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002219 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2220 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002221 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2222 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002223 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002224 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2225 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002226 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002227 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2228 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2229 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002230 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2231 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2232 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2233 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2234 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2235 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002236 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002237 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2238 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002239 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002240 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2241 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2242 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2243 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2244 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2245 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2246 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2247 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002248]
2249
Marat Dukhanfda12b82019-11-21 12:27:59 -08002250FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002251 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2252 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2253 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2254 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2255 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2256 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2257 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2258 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2259 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2260 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2261 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2262 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2263 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2264 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2265 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2266 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2267 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2268 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2269 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2270 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2271 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2272 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2273 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2274 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2275 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2276 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2277 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2278 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2279 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2280 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2281 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2282 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2283 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2284 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2285 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2286 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2287 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2288 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2289 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002291 "src/f32-hswish/gen/hswish-fma3-x8.c",
2292 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002293 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2294 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2295 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2296 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2297 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2298 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2299 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2300 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2301 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2302 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2303 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2304 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2305 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2306 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002307 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2308 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2309 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2310 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2311 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2312 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2313 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2314 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002315 "src/math/sqrt-fma3-nr1fma.c",
2316 "src/math/sqrt-fma3-nr2fma.c",
2317 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002318]
2319
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002320AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002321 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2322 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2323 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2324 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2325 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2326 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2327 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2328 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2329 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2330 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2331 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2332 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2333 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2334 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2335 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2336 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2337 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2338 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2339 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2340 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2341 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2342 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2343 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2344 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2345 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2346 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2347 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2348 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2349 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2350 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2351 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2352 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2353 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2354 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2355 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2356 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002357 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2358 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2359 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2360 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2361 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2362 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2363 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2364 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2365 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2366 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2367 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2368 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2369 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2370 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2371 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2372 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2373 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2374 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2375 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2376 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2377 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2378 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2379 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2380 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2381 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2382 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2383 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002387 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2388 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2389 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2390 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2391 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2392 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2393 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2394 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2395 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2396 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2397 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2398 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2399 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2400 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2401 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2402 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2403 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2404 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2405 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2406 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2407 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2408 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2409 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2410 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002411 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2412 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002413 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2414 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2415 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2416 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002417 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2418 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2419 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002420 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2421 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2422 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002423 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2424 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2425 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002426 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2427 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2428 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2429 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2430 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2431 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2432 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2433 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002434 "src/math/exp-avx2-p5.c",
2435 "src/math/exp-avx2-perm-p3.c",
2436 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002437 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002438 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002439 "src/math/sigmoid-avx2-rr2-p5-div.c",
2440 "src/math/sigmoid-avx2-rr1-p5-div.c",
2441 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2442 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2443 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2444 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002445]
2446
Marat Dukhan08c4a432019-10-03 09:29:21 -07002447AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002448 "src/f32-clamp/gen/avx512f-x16.c",
2449 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002450 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2451 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2452 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2453 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2454 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2455 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2456 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2457 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2458 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2459 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2460 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2461 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2462 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2463 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2464 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2465 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2466 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2467 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2468 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2469 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2470 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2471 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2472 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2473 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002474 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2475 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002476 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2477 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2478 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2479 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2480 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2481 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002482 "src/f32-prelu/gen/avx512f-2x16.c",
2483 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002484 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2485 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2486 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2487 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2488 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2489 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2490 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2491 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2492 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2493 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2494 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2495 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2496 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2497 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2498 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2499 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2500 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2501 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2502 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2503 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2504 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2505 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2506 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2507 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2508 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2509 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2510 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2511 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2512 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2513 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2514 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2515 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2516 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2517 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2518 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2519 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002520 "src/f32-relu/gen/avx512f-x16.c",
2521 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002522 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002523 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2524 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2525 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2526 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2527 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2528 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2529 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2530 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002531 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2532 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2533 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2534 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2535 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2536 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2537 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2538 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002539 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2540 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2541 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2542 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2543 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2544 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2545 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2546 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002547 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2548 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2549 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2550 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002551 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2552 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2553 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2554 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002555 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2556 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002557 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002558 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2559 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2560 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2561 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2562 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2563 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2564 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2565 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2566 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2567 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2568 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2569 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2570 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2571 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2572 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2573 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2574 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2575 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2576 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2577 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2578 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2579 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2580 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2581 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002582 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2583 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2584 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2585 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2586 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2587 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2588 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2589 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002590 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2591 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2592 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2593 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2594 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2595 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2596 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2597 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002598 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2599 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2600 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2601 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2602 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2603 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002604 "src/math/exp-avx512f-p5-scalef.c",
2605 "src/math/exp-avx512f-p5.c",
2606 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002607 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002608 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002609 "src/math/sqrt-avx512f-nr1fma.c",
2610 "src/math/sqrt-avx512f-nr2fma.c",
2611 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612]
2613
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002614AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002615 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2616 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002617 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2618 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2619 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2620 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2621 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2622 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2623 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2624 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2625]
2626
Frank Barchardbcedc082020-08-17 18:00:51 -07002627WASM32_ASM_UKERNELS = [
2628 "src/f32-relu/wasm_shr_x1.S",
2629 "src/f32-relu/wasm_shr_x2.S",
2630 "src/f32-relu/wasm_shr_x4.S",
2631]
2632
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002634 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002635 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002636 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2637 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002638 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002639 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002640 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002641 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002642 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2643 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002644 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2645 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2646 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2647 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002648]
2649
2650AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002651 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002652 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2653 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002654 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002655 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2656 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002657 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2658 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2659 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002660 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002661 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2662 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2663 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002664 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002665 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2666 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002667 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002668 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2669 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2670 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2671 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2672 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2673 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2674 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2675 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2676 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2677 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2678 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2679 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2680 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2681 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2682 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2683 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2684 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2685 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002686 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2687 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002688 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002689 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2690 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2691 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2692 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2693 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2694 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2695 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2696 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2697 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2698 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2699 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2700 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2701 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2702 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2703 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2704 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2705 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2706 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002707 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2708 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2709 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2710 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2711 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2712 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2713 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2714 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2715 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2716 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2717 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2718 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2719 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2720 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2721 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2722 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2723 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2724 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002725]
2726
Marat Dukhan1b354632020-03-23 12:50:22 -07002727INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002728 "src/xnnpack/argmaxpool.h",
2729 "src/xnnpack/avgpool.h",
2730 "src/xnnpack/clamp.h",
2731 "src/xnnpack/common.h",
2732 "src/xnnpack/conv.h",
2733 "src/xnnpack/dwconv.h",
2734 "src/xnnpack/gavgpool.h",
2735 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002736 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002738 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002740 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 "src/xnnpack/lut.h",
2742 "src/xnnpack/math.h",
2743 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002744 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002745 "src/xnnpack/packx.h",
2746 "src/xnnpack/pad.h",
2747 "src/xnnpack/params.h",
2748 "src/xnnpack/pavgpool.h",
2749 "src/xnnpack/ppmm.h",
2750 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002751 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002752 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002753 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002754 "src/xnnpack/rmax.h",
2755 "src/xnnpack/scalar-utils.h",
2756 "src/xnnpack/spmm.h",
2757 "src/xnnpack/unpool.h",
2758 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002759 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002760 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002761 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002762 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002763 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002764 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002766]
2767
2768INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 "include/xnnpack.h",
2770 "src/xnnpack/allocator.h",
2771 "src/xnnpack/compute.h",
2772 "src/xnnpack/im2col.h",
2773 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002774 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002775 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 "src/xnnpack/operator.h",
2777 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002778 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002779 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002780 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002781 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002782]
2783
Marat Dukhan1b354632020-03-23 12:50:22 -07002784ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002785 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002786]
2787
Marat Dukhan1b354632020-03-23 12:50:22 -07002788MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002789 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 "include/xnnpack.h",
2791]
2792
Marat Dukhan1b354632020-03-23 12:50:22 -07002793MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002794 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002795 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002796 "src/xnnpack/requantization.h",
2797 "include/xnnpack.h",
2798]
2799
2800OPERATOR_TEST_PARAMS_HDRS = [
2801 "src/xnnpack/params.h",
2802 "src/xnnpack/common.h",
2803]
2804
2805WEIGHTS_PACK_HDRS = [
2806 "src/xnnpack/pack.h",
2807 "src/xnnpack/operator.h",
2808 "src/xnnpack/compute.h",
2809]
2810
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002811LOGGING_COPTS = select({
2812 # No logging in optimized mode
2813 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2814 # Full logging in debug mode
2815 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2816 # Error-only logging in default (fastbuild) mode
2817 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2818})
2819
Marat Dukhan3b59de22020-06-03 20:15:19 -07002820LOGGING_SRCS = select({
2821 # No logging in optimized mode
2822 ":optimized_build": [],
2823 "//conditions:default": [
2824 "src/operator-strings.c",
2825 "src/subgraph-strings.c",
2826 ],
2827})
2828
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002829LOGGING_HDRS = [
2830 "src/xnnpack/log.h",
2831]
2832
Marat Dukhan08c4a432019-10-03 09:29:21 -07002833xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002834 name = "tables",
2835 srcs = TABLE_SRCS,
2836 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002837 gcc_copts = xnnpack_gcc_std_copts(),
2838 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002839)
2840
2841xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002842 name = "scalar_ukernels",
2843 srcs = SCALAR_UKERNELS,
2844 hdrs = INTERNAL_HDRS,
2845 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002846 gcc_copts = xnnpack_gcc_std_copts(),
2847 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002848 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002849 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002850 "@FP16",
2851 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002852 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002853 ],
2854)
2855
2856xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002857 name = "scalar_ukernels_test_mode",
2858 srcs = SCALAR_UKERNELS,
2859 hdrs = INTERNAL_HDRS,
2860 aarch32_copts = ["-marm"],
2861 copts = [
2862 "-UNDEBUG",
2863 "-DXNN_TEST_MODE=1",
2864 ],
2865 gcc_copts = xnnpack_gcc_std_copts(),
2866 msvc_copts = xnnpack_msvc_std_copts(),
2867 deps = [
2868 ":tables",
2869 "@FP16",
2870 "@FXdiv",
2871 "@pthreadpool",
2872 ],
2873)
2874
2875xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002876 name = "wasm_ukernels",
2877 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002878 gcc_copts = xnnpack_gcc_std_copts(),
2879 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002880 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002881 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002882 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002883 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002884 "@FP16",
2885 "@FXdiv",
2886 "@pthreadpool",
2887 ],
2888)
2889
2890xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002891 name = "wasm_ukernels_test_mode",
2892 hdrs = INTERNAL_HDRS,
2893 copts = [
2894 "-UNDEBUG",
2895 "-DXNN_TEST_MODE=1",
2896 ],
2897 gcc_copts = xnnpack_gcc_std_copts(),
2898 msvc_copts = xnnpack_msvc_std_copts(),
2899 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002900 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002901 deps = [
2902 ":tables",
2903 "@FP16",
2904 "@FXdiv",
2905 "@pthreadpool",
2906 ],
2907)
2908
2909xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002910 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 hdrs = INTERNAL_HDRS,
2912 aarch32_copts = [
2913 "-marm",
2914 "-mfpu=neon",
2915 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002916 gcc_copts = xnnpack_gcc_std_copts(),
2917 gcc_x86_copts = ["-msse2"],
2918 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 optimized_copts = [
2920 "-O3",
2921 "-ffast-math",
2922 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002923 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002924 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002925 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 "@FP16",
2927 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002928 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929 ],
2930)
2931
2932xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002933 name = "psimd_fastmath_ukernels_test_mode",
2934 hdrs = INTERNAL_HDRS,
2935 aarch32_copts = [
2936 "-marm",
2937 "-mfpu=neon",
2938 ],
2939 copts = [
2940 "-UNDEBUG",
2941 "-DXNN_TEST_MODE=1",
2942 ],
2943 gcc_copts = xnnpack_gcc_std_copts(),
2944 gcc_x86_copts = ["-msse2"],
2945 msvc_copts = xnnpack_msvc_std_copts(),
2946 optimized_copts = [
2947 "-O3",
2948 "-ffast-math",
2949 ],
2950 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2951 deps = [
2952 ":tables",
2953 "@FP16",
2954 "@psimd",
2955 "@pthreadpool",
2956 ],
2957)
2958
2959xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002960 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002961 hdrs = INTERNAL_HDRS,
2962 aarch32_copts = [
2963 "-marm",
2964 "-mfpu=neon",
2965 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002966 gcc_copts = xnnpack_gcc_std_copts(),
2967 gcc_x86_copts = ["-msse2"],
2968 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002969 optimized_copts = [
2970 "-O3",
2971 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002972 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002973 deps = [
2974 ":tables",
2975 "@FP16",
2976 "@psimd",
2977 "@pthreadpool",
2978 ],
2979)
2980
2981xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002982 name = "psimd_accmath_ukernels_test_mode",
2983 hdrs = INTERNAL_HDRS,
2984 aarch32_copts = [
2985 "-marm",
2986 "-mfpu=neon",
2987 ],
2988 copts = [
2989 "-UNDEBUG",
2990 "-DXNN_TEST_MODE=1",
2991 ],
2992 gcc_copts = xnnpack_gcc_std_copts(),
2993 gcc_x86_copts = ["-msse2"],
2994 msvc_copts = xnnpack_msvc_std_copts(),
2995 optimized_copts = [
2996 "-O3",
2997 ],
2998 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2999 deps = [
3000 ":tables",
3001 "@FP16",
3002 "@psimd",
3003 "@pthreadpool",
3004 ],
3005)
3006
3007xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 name = "neon_ukernels",
3009 hdrs = INTERNAL_HDRS,
3010 aarch32_copts = [
3011 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003012 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 "-mfpu=neon",
3014 ],
3015 aarch32_srcs = NEON_UKERNELS,
3016 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003017 gcc_copts = xnnpack_gcc_std_copts(),
3018 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003019 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003020 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003021 "@FP16",
3022 "@pthreadpool",
3023 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024)
3025
3026xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003027 name = "neon_ukernels_test_mode",
3028 hdrs = INTERNAL_HDRS,
3029 aarch32_copts = [
3030 "-marm",
3031 "-march=armv7-a",
3032 "-mfpu=neon",
3033 ],
3034 aarch32_srcs = NEON_UKERNELS,
3035 aarch64_srcs = NEON_UKERNELS,
3036 copts = [
3037 "-UNDEBUG",
3038 "-DXNN_TEST_MODE=1",
3039 ],
3040 gcc_copts = xnnpack_gcc_std_copts(),
3041 msvc_copts = xnnpack_msvc_std_copts(),
3042 deps = [
3043 ":tables",
3044 "@FP16",
3045 "@pthreadpool",
3046 ],
3047)
3048
3049xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003050 name = "neonfma_ukernels",
3051 hdrs = INTERNAL_HDRS,
3052 aarch32_copts = [
3053 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003054 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003055 "-mfpu=neon-vfpv4",
3056 ],
3057 aarch32_srcs = NEONFMA_UKERNELS,
3058 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003059 apple_aarch32_copts = [
3060 "-mcpu=swift",
3061 "-mtune=generic",
3062 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003063 gcc_copts = xnnpack_gcc_std_copts(),
3064 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003065 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003066 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003067 "@FP16",
3068 "@pthreadpool",
3069 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070)
3071
3072xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003073 name = "neonfma_ukernels_test_mode",
3074 hdrs = INTERNAL_HDRS,
3075 aarch32_copts = [
3076 "-marm",
3077 "-march=armv7-a",
3078 "-mfpu=neon-vfpv4",
3079 ],
3080 aarch32_srcs = NEONFMA_UKERNELS,
3081 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003082 apple_aarch32_copts = [
3083 "-mcpu=swift",
3084 "-mtune=generic",
3085 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003086 copts = [
3087 "-UNDEBUG",
3088 "-DXNN_TEST_MODE=1",
3089 ],
3090 gcc_copts = xnnpack_gcc_std_copts(),
3091 msvc_copts = xnnpack_msvc_std_copts(),
3092 deps = [
3093 ":tables",
3094 "@FP16",
3095 "@pthreadpool",
3096 ],
3097)
3098
3099xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003100 name = "neonv8_ukernels",
3101 hdrs = INTERNAL_HDRS,
3102 aarch32_copts = [
3103 "-marm",
3104 "-march=armv8-a",
3105 "-mfpu=neon-fp-armv8",
3106 ],
3107 aarch32_srcs = NEONV8_UKERNELS,
3108 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003109 apple_aarch32_copts = [
3110 "-mcpu=cyclone",
3111 "-mtune=generic",
3112 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003113 gcc_copts = xnnpack_gcc_std_copts(),
3114 msvc_copts = xnnpack_msvc_std_copts(),
3115 deps = [
3116 ":tables",
3117 "@FP16",
3118 "@pthreadpool",
3119 ],
3120)
3121
3122xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003123 name = "neonv8_ukernels_test_mode",
3124 hdrs = INTERNAL_HDRS,
3125 aarch32_copts = [
3126 "-marm",
3127 "-march=armv8-a",
3128 "-mfpu=neon-fp-armv8",
3129 ],
3130 aarch32_srcs = NEONV8_UKERNELS,
3131 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003132 apple_aarch32_copts = [
3133 "-mcpu=cyclone",
3134 "-mtune=generic",
3135 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003136 copts = [
3137 "-UNDEBUG",
3138 "-DXNN_TEST_MODE=1",
3139 ],
3140 gcc_copts = xnnpack_gcc_std_copts(),
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 deps = [
3143 ":tables",
3144 "@FP16",
3145 "@pthreadpool",
3146 ],
3147)
3148
3149xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 name = "neonfp16arith_ukernels",
3151 hdrs = INTERNAL_HDRS,
3152 aarch64_copts = ["-march=armv8.2-a+fp16"],
3153 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003154 gcc_copts = xnnpack_gcc_std_copts(),
3155 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003156 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003157 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003158 "@FP16",
3159 "@pthreadpool",
3160 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161)
3162
3163xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003164 name = "neonfp16arith_ukernels_test_mode",
3165 hdrs = INTERNAL_HDRS,
3166 aarch64_copts = ["-march=armv8.2-a+fp16"],
3167 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3168 copts = [
3169 "-UNDEBUG",
3170 "-DXNN_TEST_MODE=1",
3171 ],
3172 gcc_copts = xnnpack_gcc_std_copts(),
3173 msvc_copts = xnnpack_msvc_std_copts(),
3174 deps = [
3175 ":tables",
3176 "@FP16",
3177 "@pthreadpool",
3178 ],
3179)
3180
3181xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003182 name = "neondot_ukernels",
3183 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003184 aarch32_copts = [
3185 "-marm",
3186 "-march=armv8.2-a+dotprod",
3187 "-mfpu=neon-fp-armv8",
3188 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003189 aarch32_srcs = NEONDOT_UKERNELS,
3190 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3191 aarch64_srcs = NEONDOT_UKERNELS,
3192 gcc_copts = xnnpack_gcc_std_copts(),
3193 msvc_copts = xnnpack_msvc_std_copts(),
3194 deps = [
3195 ":tables",
3196 "@FP16",
3197 "@pthreadpool",
3198 ],
3199)
3200
3201xnnpack_cc_library(
3202 name = "neondot_ukernels_test_mode",
3203 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003204 aarch32_copts = [
3205 "-marm",
3206 "-march=armv8.2-a+dotprod",
3207 "-mfpu=neon-fp-armv8",
3208 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003209 aarch32_srcs = NEONDOT_UKERNELS,
3210 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3211 aarch64_srcs = NEONDOT_UKERNELS,
3212 copts = [
3213 "-UNDEBUG",
3214 "-DXNN_TEST_MODE=1",
3215 ],
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 deps = [
3219 ":tables",
3220 "@FP16",
3221 "@pthreadpool",
3222 ],
3223)
3224
3225xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226 name = "sse2_ukernels",
3227 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003228 gcc_copts = xnnpack_gcc_std_copts(),
3229 gcc_x86_copts = ["-msse2"],
3230 msvc_copts = xnnpack_msvc_std_copts(),
3231 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003233 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003234 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003235 "@FP16",
3236 "@pthreadpool",
3237 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238)
3239
3240xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003241 name = "sse2_ukernels_test_mode",
3242 hdrs = INTERNAL_HDRS,
3243 copts = [
3244 "-UNDEBUG",
3245 "-DXNN_TEST_MODE=1",
3246 ],
3247 gcc_copts = xnnpack_gcc_std_copts(),
3248 gcc_x86_copts = ["-msse2"],
3249 msvc_copts = xnnpack_msvc_std_copts(),
3250 msvc_x86_32_copts = ["/arch:SSE2"],
3251 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3252 deps = [
3253 ":tables",
3254 "@FP16",
3255 "@pthreadpool",
3256 ],
3257)
3258
3259xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003260 name = "ssse3_ukernels",
3261 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003262 gcc_copts = xnnpack_gcc_std_copts(),
3263 gcc_x86_copts = ["-mssse3"],
3264 msvc_copts = xnnpack_msvc_std_copts(),
3265 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003266 x86_srcs = SSSE3_UKERNELS,
3267 deps = [
3268 ":tables",
3269 "@FP16",
3270 "@pthreadpool",
3271 ],
3272)
3273
3274xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003275 name = "ssse3_ukernels_test_mode",
3276 hdrs = INTERNAL_HDRS,
3277 copts = [
3278 "-UNDEBUG",
3279 "-DXNN_TEST_MODE=1",
3280 ],
3281 gcc_copts = xnnpack_gcc_std_copts(),
3282 gcc_x86_copts = ["-mssse3"],
3283 msvc_copts = xnnpack_msvc_std_copts(),
3284 msvc_x86_32_copts = ["/arch:SSE2"],
3285 x86_srcs = SSSE3_UKERNELS,
3286 deps = [
3287 ":tables",
3288 "@FP16",
3289 "@pthreadpool",
3290 ],
3291)
3292
3293xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003294 name = "sse41_ukernels",
3295 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003296 gcc_copts = xnnpack_gcc_std_copts(),
3297 gcc_x86_copts = ["-msse4.1"],
3298 msvc_copts = xnnpack_msvc_std_copts(),
3299 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003300 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003301 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003302 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003303 "@FP16",
3304 "@pthreadpool",
3305 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003306)
3307
3308xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003309 name = "sse41_ukernels_test_mode",
3310 hdrs = INTERNAL_HDRS,
3311 copts = [
3312 "-UNDEBUG",
3313 "-DXNN_TEST_MODE=1",
3314 ],
3315 gcc_copts = xnnpack_gcc_std_copts(),
3316 gcc_x86_copts = ["-msse4.1"],
3317 msvc_copts = xnnpack_msvc_std_copts(),
3318 msvc_x86_32_copts = ["/arch:SSE2"],
3319 x86_srcs = SSE41_UKERNELS,
3320 deps = [
3321 ":tables",
3322 "@FP16",
3323 "@pthreadpool",
3324 ],
3325)
3326
3327xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328 name = "avx_ukernels",
3329 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003330 gcc_copts = xnnpack_gcc_std_copts(),
3331 gcc_x86_copts = ["-mavx"],
3332 msvc_copts = xnnpack_msvc_std_copts(),
3333 msvc_x86_32_copts = ["/arch:AVX"],
3334 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003336 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003337 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003338 "@FP16",
3339 "@pthreadpool",
3340 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341)
3342
3343xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003344 name = "avx_ukernels_test_mode",
3345 hdrs = INTERNAL_HDRS,
3346 copts = [
3347 "-UNDEBUG",
3348 "-DXNN_TEST_MODE=1",
3349 ],
3350 gcc_copts = xnnpack_gcc_std_copts(),
3351 gcc_x86_copts = ["-mavx"],
3352 msvc_copts = xnnpack_msvc_std_copts(),
3353 msvc_x86_32_copts = ["/arch:AVX"],
3354 msvc_x86_64_copts = ["/arch:AVX"],
3355 x86_srcs = AVX_UKERNELS,
3356 deps = [
3357 ":tables",
3358 "@FP16",
3359 "@pthreadpool",
3360 ],
3361)
3362
3363xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003364 name = "xop_ukernels",
3365 hdrs = INTERNAL_HDRS,
3366 gcc_copts = xnnpack_gcc_std_copts(),
3367 gcc_x86_copts = ["-mxop"],
3368 msvc_copts = xnnpack_msvc_std_copts(),
3369 msvc_x86_32_copts = ["/arch:AVX"],
3370 msvc_x86_64_copts = ["/arch:AVX"],
3371 x86_srcs = XOP_UKERNELS,
3372 deps = [
3373 ":tables",
3374 "@FP16",
3375 "@pthreadpool",
3376 ],
3377)
3378
3379xnnpack_cc_library(
3380 name = "xop_ukernels_test_mode",
3381 hdrs = INTERNAL_HDRS,
3382 copts = [
3383 "-UNDEBUG",
3384 "-DXNN_TEST_MODE=1",
3385 ],
3386 gcc_copts = xnnpack_gcc_std_copts(),
3387 gcc_x86_copts = ["-mxop"],
3388 msvc_copts = xnnpack_msvc_std_copts(),
3389 msvc_x86_32_copts = ["/arch:AVX"],
3390 msvc_x86_64_copts = ["/arch:AVX"],
3391 x86_srcs = XOP_UKERNELS,
3392 deps = [
3393 ":tables",
3394 "@FP16",
3395 "@pthreadpool",
3396 ],
3397)
3398
3399xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003400 name = "fma3_ukernels",
3401 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003402 gcc_copts = xnnpack_gcc_std_copts(),
3403 gcc_x86_copts = ["-mfma"],
3404 msvc_copts = xnnpack_msvc_std_copts(),
3405 msvc_x86_32_copts = ["/arch:AVX"],
3406 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003407 x86_srcs = FMA3_UKERNELS,
3408 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003409 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003410 "@FP16",
3411 "@pthreadpool",
3412 ],
3413)
3414
3415xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003416 name = "fma3_ukernels_test_mode",
3417 hdrs = INTERNAL_HDRS,
3418 copts = [
3419 "-UNDEBUG",
3420 "-DXNN_TEST_MODE=1",
3421 ],
3422 gcc_copts = xnnpack_gcc_std_copts(),
3423 gcc_x86_copts = ["-mfma"],
3424 msvc_copts = xnnpack_msvc_std_copts(),
3425 msvc_x86_32_copts = ["/arch:AVX"],
3426 msvc_x86_64_copts = ["/arch:AVX"],
3427 x86_srcs = FMA3_UKERNELS,
3428 deps = [
3429 ":tables",
3430 "@FP16",
3431 "@pthreadpool",
3432 ],
3433)
3434
3435xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003436 name = "avx2_ukernels",
3437 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003438 gcc_copts = xnnpack_gcc_std_copts(),
3439 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003440 "-mfma",
3441 "-mavx2",
3442 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003443 msvc_copts = xnnpack_msvc_std_copts(),
3444 msvc_x86_32_copts = ["/arch:AVX2"],
3445 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003446 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003447 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003448 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003449 "@FP16",
3450 "@pthreadpool",
3451 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003452)
3453
3454xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003455 name = "avx2_ukernels_test_mode",
3456 hdrs = INTERNAL_HDRS,
3457 copts = [
3458 "-UNDEBUG",
3459 "-DXNN_TEST_MODE=1",
3460 ],
3461 gcc_copts = xnnpack_gcc_std_copts(),
3462 gcc_x86_copts = [
3463 "-mfma",
3464 "-mavx2",
3465 ],
3466 msvc_copts = xnnpack_msvc_std_copts(),
3467 msvc_x86_32_copts = ["/arch:AVX2"],
3468 msvc_x86_64_copts = ["/arch:AVX2"],
3469 x86_srcs = AVX2_UKERNELS,
3470 deps = [
3471 ":tables",
3472 "@FP16",
3473 "@pthreadpool",
3474 ],
3475)
3476
3477xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 name = "avx512f_ukernels",
3479 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003480 gcc_copts = xnnpack_gcc_std_copts(),
3481 gcc_x86_copts = ["-mavx512f"],
3482 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3483 msvc_copts = xnnpack_msvc_std_copts(),
3484 msvc_x86_32_copts = ["/arch:AVX512"],
3485 msvc_x86_64_copts = ["/arch:AVX512"],
3486 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003488 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003489 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003490 "@FP16",
3491 "@pthreadpool",
3492 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493)
3494
3495xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003496 name = "avx512f_ukernels_test_mode",
3497 hdrs = INTERNAL_HDRS,
3498 copts = [
3499 "-UNDEBUG",
3500 "-DXNN_TEST_MODE=1",
3501 ],
3502 gcc_copts = xnnpack_gcc_std_copts(),
3503 gcc_x86_copts = ["-mavx512f"],
3504 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3505 msvc_copts = xnnpack_msvc_std_copts(),
3506 msvc_x86_32_copts = ["/arch:AVX512"],
3507 msvc_x86_64_copts = ["/arch:AVX512"],
3508 msys_copts = ["-fno-asynchronous-unwind-tables"],
3509 x86_srcs = AVX512F_UKERNELS,
3510 deps = [
3511 ":tables",
3512 "@FP16",
3513 "@pthreadpool",
3514 ],
3515)
3516
3517xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003518 name = "avx512skx_ukernels",
3519 hdrs = INTERNAL_HDRS,
3520 gcc_copts = xnnpack_gcc_std_copts(),
3521 gcc_x86_copts = [
3522 "-mavx512f",
3523 "-mavx512cd",
3524 "-mavx512bw",
3525 "-mavx512dq",
3526 "-mavx512vl",
3527 ],
3528 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3529 msvc_copts = xnnpack_msvc_std_copts(),
3530 msvc_x86_32_copts = ["/arch:AVX512"],
3531 msvc_x86_64_copts = ["/arch:AVX512"],
3532 msys_copts = ["-fno-asynchronous-unwind-tables"],
3533 x86_srcs = AVX512SKX_UKERNELS,
3534 deps = [
3535 ":tables",
3536 "@FP16",
3537 "@pthreadpool",
3538 ],
3539)
3540
3541xnnpack_cc_library(
3542 name = "avx512skx_ukernels_test_mode",
3543 hdrs = INTERNAL_HDRS,
3544 copts = [
3545 "-UNDEBUG",
3546 "-DXNN_TEST_MODE=1",
3547 ],
3548 gcc_copts = xnnpack_gcc_std_copts(),
3549 gcc_x86_copts = [
3550 "-mavx512f",
3551 "-mavx512cd",
3552 "-mavx512bw",
3553 "-mavx512dq",
3554 "-mavx512vl",
3555 ],
3556 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3557 msvc_copts = xnnpack_msvc_std_copts(),
3558 msvc_x86_32_copts = ["/arch:AVX512"],
3559 msvc_x86_64_copts = ["/arch:AVX512"],
3560 msys_copts = ["-fno-asynchronous-unwind-tables"],
3561 x86_srcs = AVX512SKX_UKERNELS,
3562 deps = [
3563 ":tables",
3564 "@FP16",
3565 "@pthreadpool",
3566 ],
3567)
3568
3569xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 name = "asm_ukernels",
3571 hdrs = ["src/xnnpack/assembly.h"],
3572 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003573 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003575 wasm_srcs = WASM32_ASM_UKERNELS,
3576 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577)
3578
Marat Dukhan3b59de22020-06-03 20:15:19 -07003579xnnpack_cc_library(
3580 name = "logging_utils",
3581 srcs = LOGGING_SRCS,
3582 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3583 copts = LOGGING_COPTS + [
3584 "-Isrc",
3585 "-Iinclude",
3586 ] + select({
3587 ":debug_build": [],
3588 "//conditions:default": xnnpack_min_size_copts(),
3589 }),
3590 gcc_copts = xnnpack_gcc_std_copts(),
3591 msvc_copts = xnnpack_msvc_std_copts(),
3592 visibility = xnnpack_visibility(),
3593 deps = [
3594 "@FP16",
3595 "@clog",
3596 "@pthreadpool",
3597 ],
3598)
3599
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600xnnpack_aggregate_library(
3601 name = "ukernels",
3602 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003603 ":neon_ukernels",
3604 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003605 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003606 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 ":asm_ukernels",
3608 ],
3609 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610 ":neon_ukernels",
3611 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003612 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003613 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003614 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 ":asm_ukernels",
3616 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003617 generic_deps = [
3618 ":scalar_ukernels",
3619 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003620 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003621 ":psimd_fastmath_ukernels",
3622 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003624 wasm_deps = [
3625 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003626 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003627 ],
3628 wasmsimd_deps = [
3629 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003630 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003631 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003634 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003635 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003637 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003638 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003639 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003640 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003641 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 ],
3643)
3644
Marat Dukhan33fcf782020-05-24 14:27:15 -07003645xnnpack_aggregate_library(
3646 name = "ukernels_test_mode",
3647 aarch32_deps = [
3648 ":neon_ukernels_test_mode",
3649 ":neonfma_ukernels_test_mode",
3650 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003651 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003652 ":asm_ukernels",
3653 ],
3654 aarch64_deps = [
3655 ":neon_ukernels_test_mode",
3656 ":neonfma_ukernels_test_mode",
3657 ":neonv8_ukernels_test_mode",
3658 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003659 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003660 ":asm_ukernels",
3661 ],
3662 generic_deps = [
3663 ":scalar_ukernels_test_mode",
3664 ],
3665 psimd_deps = [
3666 ":psimd_fastmath_ukernels_test_mode",
3667 ":psimd_accmath_ukernels_test_mode",
3668 ],
3669 wasm_deps = [
3670 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003671 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003672 ],
3673 wasmsimd_deps = [
3674 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003675 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003676 ],
3677 x86_deps = [
3678 ":sse2_ukernels_test_mode",
3679 ":ssse3_ukernels_test_mode",
3680 ":sse41_ukernels_test_mode",
3681 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003682 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003683 ":fma3_ukernels_test_mode",
3684 ":avx2_ukernels_test_mode",
3685 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003686 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003687 ],
3688)
3689
Marat Dukhan86df3842020-06-17 14:03:31 -07003690# This target must be compatible with Android NDK r17 and compile with Clang 6.
3691# It is similar to :ukernels target, but excludes microkernels which require
3692# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003693xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003694 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003695 aarch32_deps = [
3696 ":neon_ukernels",
3697 ":neonfma_ukernels",
3698 ":neonv8_ukernels",
3699 ":asm_ukernels",
3700 ],
3701 aarch64_deps = [
3702 ":neon_ukernels",
3703 ":neonfma_ukernels",
3704 ":neonv8_ukernels",
3705 ":asm_ukernels",
3706 ],
3707 generic_deps = [
3708 ":scalar_ukernels",
3709 ],
3710 psimd_deps = [
3711 ":psimd_fastmath_ukernels",
3712 ":psimd_accmath_ukernels",
3713 ],
3714 wasm_deps = [
3715 ":wasm_ukernels",
3716 ],
3717 wasmsimd_deps = [
3718 ":wasm_ukernels",
3719 ],
3720 x86_deps = [
3721 ":sse2_ukernels",
3722 ":ssse3_ukernels",
3723 ":sse41_ukernels",
3724 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003725 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003726 ":fma3_ukernels",
3727 ":avx2_ukernels",
3728 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003729 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003730 ],
3731)
3732
Marat Dukhan08c4a432019-10-03 09:29:21 -07003733xnnpack_cc_library(
3734 name = "im2col",
3735 srcs = ["src/im2col.c"],
3736 hdrs = [
3737 "src/xnnpack/common.h",
3738 "src/xnnpack/im2col.h",
3739 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003740 gcc_copts = xnnpack_gcc_std_copts(),
3741 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742)
3743
3744xnnpack_cc_library(
3745 name = "indirection",
3746 srcs = ["src/indirection.c"],
3747 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003748 gcc_copts = xnnpack_gcc_std_copts(),
3749 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 deps = [
3751 "@FP16",
3752 "@FXdiv",
3753 "@pthreadpool",
3754 ],
3755)
3756
3757xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003758 name = "indirection_test_mode",
3759 srcs = ["src/indirection.c"],
3760 hdrs = INTERNAL_HDRS,
3761 copts = [
3762 "-UNDEBUG",
3763 "-DXNN_TEST_MODE=1",
3764 ],
3765 gcc_copts = xnnpack_gcc_std_copts(),
3766 msvc_copts = xnnpack_msvc_std_copts(),
3767 deps = [
3768 "@FP16",
3769 "@FXdiv",
3770 "@pthreadpool",
3771 ],
3772)
3773
3774xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003775 name = "packing",
3776 srcs = ["src/packing.c"],
3777 hdrs = INTERNAL_HDRS,
3778 gcc_copts = xnnpack_gcc_std_copts(),
3779 msvc_copts = xnnpack_msvc_std_copts(),
3780 deps = [
3781 "@FP16",
3782 "@FXdiv",
3783 "@pthreadpool",
3784 ],
3785)
3786
3787xnnpack_cc_library(
3788 name = "packing_test_mode",
3789 srcs = ["src/packing.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 Dukhan08c4a432019-10-03 09:29:21 -07003805 name = "operator_run",
3806 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003807 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003809 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3810 "//conditions:default": [],
3811 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003812 gcc_copts = xnnpack_gcc_std_copts(),
3813 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003815 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816 "@FP16",
3817 "@FXdiv",
3818 "@clog",
3819 "@pthreadpool",
3820 ],
3821)
3822
Chao Mei6ddfc602020-05-13 22:29:36 -07003823xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003824 name = "operator_run_test_mode",
3825 srcs = ["src/operator-run.c"],
3826 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3827 copts = LOGGING_COPTS + [
3828 "-UNDEBUG",
3829 "-DXNN_TEST_MODE=1",
3830 ] + select({
3831 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3832 "//conditions:default": [],
3833 }),
3834 gcc_copts = xnnpack_gcc_std_copts(),
3835 msvc_copts = xnnpack_msvc_std_copts(),
3836 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003837 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003838 "@FP16",
3839 "@FXdiv",
3840 "@clog",
3841 "@pthreadpool",
3842 ],
3843)
3844
3845xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003846 name = "memory_planner",
3847 srcs = ["src/memory-planner.c"],
3848 hdrs = INTERNAL_HDRS,
3849 defines = select({
3850 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3851 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3852 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3853 }),
3854 gcc_copts = xnnpack_gcc_std_copts(),
3855 msvc_copts = xnnpack_msvc_std_copts(),
3856 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003857 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003858 "@pthreadpool",
3859 ],
3860)
3861
Marat Dukhan33fcf782020-05-24 14:27:15 -07003862xnnpack_cc_library(
3863 name = "memory_planner_test_mode",
3864 srcs = ["src/memory-planner.c"],
3865 hdrs = INTERNAL_HDRS,
3866 copts = [
3867 "-UNDEBUG",
3868 "-DXNN_TEST_MODE=1",
3869 ],
3870 defines = select({
3871 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3872 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3873 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3874 }),
3875 gcc_copts = xnnpack_gcc_std_copts(),
3876 msvc_copts = xnnpack_msvc_std_copts(),
3877 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003878 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003879 "@pthreadpool",
3880 ],
3881)
3882
Marat Dukhan08c4a432019-10-03 09:29:21 -07003883cc_library(
3884 name = "enable_assembly",
3885 defines = select({
3886 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3887 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003888 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889 }),
3890)
3891
Marat Dukhan9de90e02020-06-18 16:04:12 -07003892cc_library(
3893 name = "enable_sparse",
3894 defines = select({
3895 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3896 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003897 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003898 }),
3899)
3900
Marat Dukhancf056b22019-10-07 10:26:29 -07003901xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003902 name = "operators",
3903 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003904 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003906 ],
3907 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003908 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 "-Isrc",
3910 "-Iinclude",
3911 ] + select({
3912 ":debug_build": [],
3913 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003914 }) + select({
3915 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3916 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003917 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003918 gcc_copts = xnnpack_gcc_std_copts(),
3919 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003920 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003922 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003923 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003924 "@FP16",
3925 "@FXdiv",
3926 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003927 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003928 ],
3929)
3930
Marat Dukhan10a38082020-04-17 03:58:35 -07003931xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003932 name = "operators_test_mode",
3933 srcs = OPERATOR_SRCS + [
3934 "src/memory.c",
3935 "src/operator-delete.c",
3936 ],
3937 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3938 copts = LOGGING_COPTS + [
3939 "-Isrc",
3940 "-Iinclude",
3941 "-UNDEBUG",
3942 "-DXNN_TEST_MODE=1",
3943 ] + select({
3944 ":debug_build": [],
3945 "//conditions:default": xnnpack_min_size_copts(),
3946 }) + select({
3947 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3948 "//conditions:default": [],
3949 }),
3950 gcc_copts = xnnpack_gcc_std_copts(),
3951 msvc_copts = xnnpack_msvc_std_copts(),
3952 deps = [
3953 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003954 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003955 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003956 "@FP16",
3957 "@FXdiv",
3958 "@clog",
3959 "@pthreadpool",
3960 ],
3961)
3962
3963xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003964 name = "XNNPACK",
3965 srcs = [
3966 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003967 "src/runtime.c",
3968 "src/subgraph.c",
3969 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003970 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003971 hdrs = ["include/xnnpack.h"],
3972 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003973 "-Isrc",
3974 "-Iinclude",
3975 ] + select({
3976 ":debug_build": [],
3977 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003978 }) + select({
3979 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3980 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003981 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003982 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003983 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003984 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003985 visibility = xnnpack_visibility(),
3986 deps = [
3987 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003988 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003989 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003990 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003991 ":operator_run",
3992 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003993 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003994 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003995 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003996 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003997 ] + select({
3998 ":emscripten": [],
3999 "//conditions:default": ["@cpuinfo"],
4000 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001)
4002
Marat Dukhan10a38082020-04-17 03:58:35 -07004003xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004004 name = "XNNPACK_test_mode",
4005 srcs = [
4006 "src/init.c",
4007 "src/runtime.c",
4008 "src/subgraph.c",
4009 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004010 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004011 hdrs = ["include/xnnpack.h"],
4012 copts = LOGGING_COPTS + [
4013 "-Isrc",
4014 "-Iinclude",
4015 "-UNDEBUG",
4016 "-DXNN_TEST_MODE=1",
4017 ] + select({
4018 ":debug_build": [],
4019 "//conditions:default": xnnpack_min_size_copts(),
4020 }) + select({
4021 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4022 "//conditions:default": [],
4023 }),
4024 gcc_copts = xnnpack_gcc_std_copts(),
4025 includes = ["include"],
4026 msvc_copts = xnnpack_msvc_std_copts(),
4027 visibility = xnnpack_visibility(),
4028 deps = [
4029 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004030 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004031 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004032 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004033 ":operator_run_test_mode",
4034 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004035 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004036 "@clog",
4037 "@FP16",
4038 "@pthreadpool",
4039 ] + select({
4040 ":emscripten": [],
4041 "//conditions:default": ["@cpuinfo"],
4042 }),
4043)
4044
4045xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004046 name = "xnnpack_f32",
4047 srcs = [
4048 "src/init.c",
4049 "src/runtime.c",
4050 "src/subgraph.c",
4051 "src/tensor.c",
4052 ] + SUBGRAPH_SRCS,
4053 hdrs = ["include/xnnpack.h"],
4054 copts = LOGGING_COPTS + [
4055 "-Isrc",
4056 "-Iinclude",
4057 ] + select({
4058 ":debug_build": [],
4059 "//conditions:default": xnnpack_min_size_copts(),
4060 }) + select({
4061 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4062 "//conditions:default": [],
4063 }),
4064 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004065 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004066 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004067 "XNN_NO_U8_OPERATORS",
4068 "XNN_NO_X8_OPERATORS",
4069 "XNN_NO_F16_OPERATORS",
4070 "XNN_NO_X16_OPERATORS",
4071 ],
4072 gcc_copts = xnnpack_gcc_std_copts(),
4073 includes = ["include"],
4074 msvc_copts = xnnpack_msvc_std_copts(),
4075 visibility = xnnpack_visibility(),
4076 deps = [
4077 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004078 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004079 ":logging_utils",
4080 ":memory_planner",
4081 ":operator_run",
4082 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004083 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004084 "@clog",
4085 "@FP16",
4086 "@pthreadpool",
4087 ] + select({
4088 ":emscripten": [],
4089 "//conditions:default": ["@cpuinfo"],
4090 }),
4091)
4092
4093xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004094 name = "xnnpack_operators_nhwc_f32",
4095 srcs = [
4096 "src/init.c",
4097 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004098 hdrs = ["include/xnnpack.h"],
4099 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004100 "-Isrc",
4101 "-Iinclude",
4102 ] + select({
4103 ":debug_build": [],
4104 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004105 }) + select({
4106 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4107 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004108 }),
4109 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004110 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004111 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004112 "XNN_NO_U8_OPERATORS",
4113 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004114 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004115 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004116 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004117 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004118 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119 visibility = xnnpack_visibility(),
4120 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004121 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004122 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123 ":operator_run",
4124 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004125 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004126 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004128 ] + select({
4129 ":emscripten": [],
4130 "//conditions:default": ["@cpuinfo"],
4131 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004132)
4133
Marat Dukhancf056b22019-10-07 10:26:29 -07004134xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004135 name = "bench_utils",
4136 srcs = ["bench/utils.cc"],
4137 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004138 deps = [
4139 "@com_google_benchmark//:benchmark",
4140 "@cpuinfo",
4141 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004142)
4143
Frank Barchard7e955972019-10-11 10:34:25 -07004144######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004145
4146xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004147 name = "qs8_gemm_bench",
4148 srcs = [
4149 "bench/gemm.h",
4150 "bench/qs8-gemm.cc",
4151 "src/xnnpack/AlignedAllocator.h",
4152 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4153 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4154)
4155
4156xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004157 name = "qs8_requantization_bench",
4158 srcs = [
4159 "bench/qs8-requantization.cc",
4160 "src/xnnpack/requantization-stubs.h",
4161 "src/xnnpack/AlignedAllocator.h",
4162 ] + MICROKERNEL_BENCHMARK_HDRS,
4163 deps = MICROKERNEL_BENCHMARK_DEPS,
4164)
4165
4166xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004167 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004168 srcs = [
4169 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004170 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004171 "src/xnnpack/AlignedAllocator.h",
4172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004173 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004174 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004175)
4176
4177xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004178 name = "qu8_requantization_bench",
4179 srcs = [
4180 "bench/qu8-requantization.cc",
4181 "src/xnnpack/requantization-stubs.h",
4182 "src/xnnpack/AlignedAllocator.h",
4183 ] + MICROKERNEL_BENCHMARK_HDRS,
4184 deps = MICROKERNEL_BENCHMARK_DEPS,
4185)
4186
4187xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004188 name = "f16_igemm_bench",
4189 srcs = [
4190 "bench/f16-igemm.cc",
4191 "bench/conv.h",
4192 "bench/google/conv.h",
4193 "src/xnnpack/AlignedAllocator.h",
4194 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004195 deps = MICROKERNEL_BENCHMARK_DEPS + [
4196 ":indirection",
4197 ":packing",
4198 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004199)
4200
4201xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004202 name = "f16_gemm_bench",
4203 srcs = [
4204 "bench/f16-gemm.cc",
4205 "bench/gemm.h",
4206 "src/xnnpack/AlignedAllocator.h",
4207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004208 deps = MICROKERNEL_BENCHMARK_DEPS + [
4209 ":packing",
4210 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211)
4212
4213xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004214 name = "f16_spmm_bench",
4215 srcs = [
4216 "bench/f16-spmm.cc",
4217 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004218 "src/xnnpack/AlignedAllocator.h",
4219 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004220 deps = MICROKERNEL_BENCHMARK_DEPS,
4221)
4222
4223xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224 name = "f32_igemm_bench",
4225 srcs = [
4226 "bench/f32-igemm.cc",
4227 "bench/conv.h",
4228 "src/xnnpack/AlignedAllocator.h",
4229 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004230 deps = MICROKERNEL_BENCHMARK_DEPS + [
4231 ":indirection",
4232 ":packing",
4233 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004234)
4235
4236xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004237 name = "f16_relu_bench",
4238 srcs = [
4239 "bench/f16-relu.cc",
4240 "src/xnnpack/AlignedAllocator.h",
4241 ] + MICROKERNEL_BENCHMARK_HDRS,
4242 deps = MICROKERNEL_BENCHMARK_DEPS,
4243)
4244
4245xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004246 name = "f32_conv_hwc_bench",
4247 srcs = [
4248 "bench/f32-conv-hwc.cc",
4249 "bench/dconv.h",
4250 "src/xnnpack/AlignedAllocator.h",
4251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004252 deps = MICROKERNEL_BENCHMARK_DEPS + [
4253 ":packing",
4254 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255)
4256
4257xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004258 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004259 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004260 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004261 "bench/dconv.h",
4262 "src/xnnpack/AlignedAllocator.h",
4263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_BENCHMARK_DEPS + [
4265 ":packing",
4266 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004267)
4268
4269xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004270 name = "f16_dwconv_bench",
4271 srcs = [
4272 "bench/f16-dwconv.cc",
4273 "bench/dwconv.h",
4274 "bench/google/dwconv.h",
4275 "src/xnnpack/AlignedAllocator.h",
4276 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004277 deps = MICROKERNEL_BENCHMARK_DEPS + [
4278 ":indirection",
4279 ":packing",
4280 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004281)
4282
4283xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004284 name = "f32_dwconv_bench",
4285 srcs = [
4286 "bench/f32-dwconv.cc",
4287 "bench/dwconv.h",
4288 "src/xnnpack/AlignedAllocator.h",
4289 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004290 deps = MICROKERNEL_BENCHMARK_DEPS + [
4291 ":indirection",
4292 ":packing",
4293 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004294)
4295
4296xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004297 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004299 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004300 "bench/dwconv.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004303 deps = MICROKERNEL_BENCHMARK_DEPS + [
4304 ":indirection",
4305 ":packing",
4306 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307)
4308
4309xnnpack_benchmark(
4310 name = "f32_gemm_bench",
4311 srcs = [
4312 "bench/f32-gemm.cc",
4313 "bench/gemm.h",
4314 "src/xnnpack/AlignedAllocator.h",
4315 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004316 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004317 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318)
4319
4320xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004321 name = "f32_hswish_bench",
4322 srcs = [
4323 "bench/f32-hswish.cc",
4324 "src/xnnpack/AlignedAllocator.h",
4325 ] + MICROKERNEL_BENCHMARK_HDRS,
4326 deps = MICROKERNEL_BENCHMARK_DEPS,
4327)
4328
4329xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004330 name = "f32_raddexpminusmax_bench",
4331 srcs = [
4332 "bench/f32-raddexpminusmax.cc",
4333 "src/xnnpack/AlignedAllocator.h",
4334 ] + MICROKERNEL_BENCHMARK_HDRS,
4335 deps = MICROKERNEL_BENCHMARK_DEPS,
4336)
4337
4338xnnpack_benchmark(
4339 name = "f32_raddextexp_bench",
4340 srcs = [
4341 "bench/f32-raddextexp.cc",
4342 "src/xnnpack/AlignedAllocator.h",
4343 ] + MICROKERNEL_BENCHMARK_HDRS,
4344 deps = MICROKERNEL_BENCHMARK_DEPS,
4345)
4346
4347xnnpack_benchmark(
4348 name = "f32_raddstoreexpminusmax_bench",
4349 srcs = [
4350 "bench/f32-raddstoreexpminusmax.cc",
4351 "src/xnnpack/AlignedAllocator.h",
4352 ] + MICROKERNEL_BENCHMARK_HDRS,
4353 deps = MICROKERNEL_BENCHMARK_DEPS,
4354)
4355
4356xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004357 name = "f32_relu_bench",
4358 srcs = [
4359 "bench/f32-relu.cc",
4360 "src/xnnpack/AlignedAllocator.h",
4361 ] + MICROKERNEL_BENCHMARK_HDRS,
4362 deps = MICROKERNEL_BENCHMARK_DEPS,
4363)
4364
4365xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 name = "f32_rmax_bench",
4367 srcs = [
4368 "bench/f32-rmax.cc",
4369 "src/xnnpack/AlignedAllocator.h",
4370 ] + MICROKERNEL_BENCHMARK_HDRS,
4371 deps = MICROKERNEL_BENCHMARK_DEPS,
4372)
4373
4374xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004375 name = "f32_sigmoid_bench",
4376 srcs = [
4377 "bench/f32-sigmoid.cc",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004380 deps = MICROKERNEL_BENCHMARK_DEPS,
4381)
4382
4383xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_spmm_bench",
4385 srcs = [
4386 "bench/f32-spmm.cc",
4387 "bench/gemm.h",
4388 "src/xnnpack/AlignedAllocator.h",
4389 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390 deps = MICROKERNEL_BENCHMARK_DEPS,
4391)
4392
4393xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004394 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004395 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004396 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004397 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004398 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004399 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004400)
4401
4402xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004403 name = "f32_vscaleexpminusmax_bench",
4404 srcs = [
4405 "bench/f32-vscaleexpminusmax.cc",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + MICROKERNEL_BENCHMARK_HDRS,
4408 deps = MICROKERNEL_BENCHMARK_DEPS,
4409)
4410
4411xnnpack_benchmark(
4412 name = "f32_vscaleextexp_bench",
4413 srcs = [
4414 "bench/f32-vscaleextexp.cc",
4415 "src/xnnpack/AlignedAllocator.h",
4416 ] + MICROKERNEL_BENCHMARK_HDRS,
4417 deps = MICROKERNEL_BENCHMARK_DEPS,
4418)
4419
4420xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004421 name = "f32_vsqrt_bench",
4422 srcs = [
4423 "bench/f32-vsqrt.cc",
4424 "src/xnnpack/AlignedAllocator.h",
4425 ] + MICROKERNEL_BENCHMARK_HDRS,
4426 deps = MICROKERNEL_BENCHMARK_DEPS,
4427)
4428
4429xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430 name = "f32_im2col_gemm_bench",
4431 srcs = [
4432 "bench/f32-im2col-gemm.cc",
4433 "bench/conv.h",
4434 "src/xnnpack/AlignedAllocator.h",
4435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004436 deps = MICROKERNEL_BENCHMARK_DEPS + [
4437 ":im2col",
4438 ":packing",
4439 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004440)
4441
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004442xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004443 name = "rounding_bench",
4444 srcs = [
4445 "bench/rounding.cc",
4446 "src/xnnpack/math-stubs.h",
4447 "src/xnnpack/AlignedAllocator.h",
4448 ] + MICROKERNEL_BENCHMARK_HDRS,
4449 deps = MICROKERNEL_BENCHMARK_DEPS,
4450)
4451
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452########################### Benchmarks for operators ###########################
4453
4454xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004455 name = "average_pooling_bench",
4456 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004457 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004458 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004459 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460)
4461
4462xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004463 name = "bankers_rounding_bench",
4464 srcs = ["bench/bankers-rounding.cc"],
4465 copts = xnnpack_optional_tflite_copts(),
4466 tags = ["nowin32"],
4467 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4468)
4469
4470xnnpack_benchmark(
4471 name = "ceiling_bench",
4472 srcs = ["bench/ceiling.cc"],
4473 copts = xnnpack_optional_tflite_copts(),
4474 tags = ["nowin32"],
4475 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4476)
4477
4478xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479 name = "channel_shuffle_bench",
4480 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004481 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482)
4483
4484xnnpack_benchmark(
4485 name = "convolution_bench",
4486 srcs = ["bench/convolution.cc"],
4487 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_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() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490)
4491
4492xnnpack_benchmark(
4493 name = "deconvolution_bench",
4494 srcs = ["bench/deconvolution.cc"],
4495 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004496 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004497 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004498)
4499
4500xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004501 name = "floor_bench",
4502 srcs = ["bench/floor.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 = "global_average_pooling_bench",
4510 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004511 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512)
4513
4514xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004515 name = "hardswish_bench",
4516 srcs = ["bench/hardswish.cc"],
4517 copts = xnnpack_optional_tflite_copts(),
4518 tags = ["nowin32"],
4519 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4520)
4521
4522xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004523 name = "max_pooling_bench",
4524 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004525 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526)
4527
4528xnnpack_benchmark(
4529 name = "sigmoid_bench",
4530 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004531 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004532 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004533 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534)
4535
4536xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004537 name = "prelu_bench",
4538 srcs = ["bench/prelu.cc"],
4539 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004540 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004541 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004542)
4543
4544xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004545 name = "softmax_bench",
4546 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004547 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004548 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004549 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550)
4551
Marat Dukhan87727142020-06-24 15:24:10 -07004552xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004553 name = "square_root_bench",
4554 srcs = ["bench/square-root.cc"],
4555 copts = xnnpack_optional_tflite_copts(),
4556 tags = ["nowin32"],
4557 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4558)
4559
4560xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004561 name = "truncation_bench",
4562 srcs = ["bench/truncation.cc"],
4563 deps = OPERATOR_BENCHMARK_DEPS,
4564)
4565
Marat Dukhanc068bb62019-10-04 13:24:39 -07004566############################# End-to-end benchmarks ############################
4567
4568cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004569 name = "fp32_mobilenet_v1",
4570 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004571 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004572 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004573 linkstatic = True,
4574 deps = [
4575 ":XNNPACK",
4576 "@pthreadpool",
4577 ],
4578)
4579
4580cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004581 name = "fp16_mobilenet_v1",
4582 srcs = ["models/fp16-mobilenet-v1.cc"],
4583 hdrs = ["models/models.h"],
4584 copts = xnnpack_std_cxxopts(),
4585 linkstatic = True,
4586 deps = [
4587 ":XNNPACK",
4588 "@FP16",
4589 "@pthreadpool",
4590 ],
4591)
4592
4593cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004594 name = "qs8_mobilenet_v1",
4595 srcs = ["models/qs8-mobilenet-v1.cc"],
4596 hdrs = ["models/models.h"],
4597 copts = xnnpack_std_cxxopts(),
4598 linkstatic = True,
4599 deps = [
4600 ":XNNPACK",
4601 "@pthreadpool",
4602 ],
4603)
4604
4605cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004606 name = "qs8_mobilenet_v2",
4607 srcs = ["models/qs8-mobilenet-v2.cc"],
4608 hdrs = ["models/models.h"],
4609 copts = xnnpack_std_cxxopts(),
4610 linkstatic = True,
4611 deps = [
4612 ":XNNPACK",
4613 "@pthreadpool",
4614 ],
4615)
4616
4617cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004618 name = "fp32_mobilenet_v2",
4619 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004620 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004621 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004622 linkstatic = True,
4623 deps = [
4624 ":XNNPACK",
4625 "@pthreadpool",
4626 ],
4627)
4628
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004629cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004630 name = "fp16_mobilenet_v2",
4631 srcs = ["models/fp16-mobilenet-v2.cc"],
4632 hdrs = ["models/models.h"],
4633 copts = xnnpack_std_cxxopts(),
4634 linkstatic = True,
4635 deps = [
4636 ":XNNPACK",
4637 "@FP16",
4638 "@pthreadpool",
4639 ],
4640)
4641
4642cc_library(
4643 name = "fp32_mobilenet_v3_large",
4644 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004645 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004646 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004647 linkstatic = True,
4648 deps = [
4649 ":XNNPACK",
4650 "@pthreadpool",
4651 ],
4652)
4653
4654cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004655 name = "fp16_mobilenet_v3_large",
4656 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4657 hdrs = ["models/models.h"],
4658 copts = xnnpack_std_cxxopts(),
4659 linkstatic = True,
4660 deps = [
4661 ":XNNPACK",
4662 "@FP16",
4663 "@pthreadpool",
4664 ],
4665)
4666
4667cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004668 name = "fp32_mobilenet_v3_small",
4669 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004670 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004671 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004672 linkstatic = True,
4673 deps = [
4674 ":XNNPACK",
4675 "@pthreadpool",
4676 ],
4677)
4678
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004679cc_library(
4680 name = "fp16_mobilenet_v3_small",
4681 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4682 hdrs = ["models/models.h"],
4683 copts = xnnpack_std_cxxopts(),
4684 linkstatic = True,
4685 deps = [
4686 ":XNNPACK",
4687 "@FP16",
4688 "@pthreadpool",
4689 ],
4690)
4691
Marat Dukhanc068bb62019-10-04 13:24:39 -07004692xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004693 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004694 srcs = [
4695 "bench/f32-dwconv-e2e.cc",
4696 "bench/end2end.h",
4697 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004698 deps = MICROKERNEL_BENCHMARK_DEPS + [
4699 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004700 ":fp32_mobilenet_v1",
4701 ":fp32_mobilenet_v2",
4702 ":fp32_mobilenet_v3_large",
4703 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004704 ],
4705)
4706
4707xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004708 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004709 srcs = [
4710 "bench/f32-gemm-e2e.cc",
4711 "bench/end2end.h",
4712 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004713 deps = MICROKERNEL_BENCHMARK_DEPS + [
4714 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004715 ":fp32_mobilenet_v1",
4716 ":fp32_mobilenet_v2",
4717 ":fp32_mobilenet_v3_large",
4718 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004719 ],
4720)
4721
4722xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004723 name = "end2end_bench",
4724 srcs = ["bench/end2end.cc"],
4725 deps = [
4726 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004727 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004728 ":fp16_mobilenet_v1",
4729 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004730 ":fp16_mobilenet_v3_large",
4731 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004732 ":fp32_mobilenet_v1",
4733 ":fp32_mobilenet_v2",
4734 ":fp32_mobilenet_v3_large",
4735 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004736 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004737 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004738 "@pthreadpool",
4739 ],
4740)
4741
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004742#################### Accuracy evaluation for math functions ####################
4743
4744xnnpack_benchmark(
4745 name = "f32_exp_eval",
4746 srcs = [
4747 "eval/f32-exp.cc",
4748 "src/xnnpack/AlignedAllocator.h",
4749 ] + ACCURACY_EVAL_HDRS,
4750 deps = ACCURACY_EVAL_DEPS,
4751)
4752
Marat Dukhan515c9772019-10-17 18:07:57 -07004753xnnpack_benchmark(
4754 name = "f32_expminus_eval",
4755 srcs = [
4756 "eval/f32-expminus.cc",
4757 "src/xnnpack/AlignedAllocator.h",
4758 ] + ACCURACY_EVAL_HDRS,
4759 deps = ACCURACY_EVAL_DEPS,
4760)
4761
Marat Dukhan98ba4412019-10-23 02:14:28 -07004762xnnpack_benchmark(
4763 name = "f32_extexp_eval",
4764 srcs = [
4765 "eval/f32-extexp.cc",
4766 "src/xnnpack/AlignedAllocator.h",
4767 ] + ACCURACY_EVAL_HDRS,
4768 deps = ACCURACY_EVAL_DEPS,
4769)
4770
Marat Dukhan8853b822020-05-07 12:19:01 -07004771xnnpack_unit_test(
4772 name = "f32_roundne_eval",
4773 srcs = [
4774 "eval/f32-roundne.cc",
4775 "src/xnnpack/AlignedAllocator.h",
4776 "src/xnnpack/math-stubs.h",
4777 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004778 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004779 deps = MICROKERNEL_TEST_DEPS,
4780)
4781
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004782xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004783 name = "f32_roundd_eval",
4784 srcs = [
4785 "eval/f32-roundd.cc",
4786 "src/xnnpack/AlignedAllocator.h",
4787 "src/xnnpack/math-stubs.h",
4788 ] + MICROKERNEL_TEST_HDRS,
4789 automatic = False,
4790 deps = MICROKERNEL_TEST_DEPS,
4791)
4792
4793xnnpack_unit_test(
4794 name = "f32_roundu_eval",
4795 srcs = [
4796 "eval/f32-roundu.cc",
4797 "src/xnnpack/AlignedAllocator.h",
4798 "src/xnnpack/math-stubs.h",
4799 ] + MICROKERNEL_TEST_HDRS,
4800 automatic = False,
4801 deps = MICROKERNEL_TEST_DEPS,
4802)
4803
4804xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004805 name = "f32_roundz_eval",
4806 srcs = [
4807 "eval/f32-roundz.cc",
4808 "src/xnnpack/AlignedAllocator.h",
4809 "src/xnnpack/math-stubs.h",
4810 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004811 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004812 deps = MICROKERNEL_TEST_DEPS,
4813)
4814
Marat Dukhan346a9e52019-11-15 09:06:30 -08004815xnnpack_benchmark(
4816 name = "f32_sigmoid_eval",
4817 srcs = [
4818 "eval/f32-sigmoid.cc",
4819 "src/xnnpack/AlignedAllocator.h",
4820 ] + ACCURACY_EVAL_HDRS,
4821 deps = ACCURACY_EVAL_DEPS,
4822)
4823
Marat Dukhan84000762020-06-29 18:38:43 -07004824xnnpack_benchmark(
4825 name = "f32_sqrt_eval",
4826 srcs = [
4827 "eval/f32-sqrt.cc",
4828 "src/xnnpack/AlignedAllocator.h",
4829 ] + ACCURACY_EVAL_HDRS,
4830 deps = ACCURACY_EVAL_DEPS + [
4831 ":bench_utils",
4832 ],
4833)
4834
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835######################### Unit tests for micro-kernels #########################
4836
4837xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004838 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004840 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004841 "test/gemm-microkernel-tester.h",
4842 "src/xnnpack/AlignedAllocator.h",
4843 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004844 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845)
4846
4847xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004848 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004849 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004850 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004851 "test/spmm-microkernel-tester.h",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + MICROKERNEL_TEST_HDRS,
4854 deps = MICROKERNEL_TEST_DEPS,
4855)
4856
4857xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858 name = "f32_argmaxpool_test",
4859 srcs = [
4860 "test/f32-argmaxpool.cc",
4861 "test/argmaxpool-microkernel-tester.h",
4862 "src/xnnpack/AlignedAllocator.h",
4863 ] + MICROKERNEL_TEST_HDRS,
4864 deps = MICROKERNEL_TEST_DEPS,
4865)
4866
4867xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004868 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004870 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871 "test/avgpool-microkernel-tester.h",
4872 "src/xnnpack/AlignedAllocator.h",
4873 ] + MICROKERNEL_TEST_HDRS,
4874 deps = MICROKERNEL_TEST_DEPS,
4875)
4876
4877xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004878 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004879 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004880 "test/f32-ibilinear.cc",
4881 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004882 "src/xnnpack/AlignedAllocator.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 deps = MICROKERNEL_TEST_DEPS,
4885)
4886
4887xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004888 name = "f16_clamp_test",
4889 srcs = [
4890 "test/f16-clamp.cc",
4891 "test/clamp-microkernel-tester.h",
4892 ] + MICROKERNEL_TEST_HDRS,
4893 deps = MICROKERNEL_TEST_DEPS,
4894)
4895
4896xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004897 name = "f32_clamp_test",
4898 srcs = [
4899 "test/f32-clamp.cc",
4900 "test/clamp-microkernel-tester.h",
4901 ] + MICROKERNEL_TEST_HDRS,
4902 deps = MICROKERNEL_TEST_DEPS,
4903)
4904
4905xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004906 name = "f32_igemm_test",
4907 srcs = [
4908 "test/f32-igemm.cc",
4909 "test/gemm-microkernel-tester.h",
4910 "src/xnnpack/AlignedAllocator.h",
4911 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004912 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004913)
4914
4915xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004916 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004917 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004918 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004919 "test/gemm-microkernel-tester.h",
4920 "src/xnnpack/AlignedAllocator.h",
4921 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004922 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004923)
4924
4925xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004926 name = "f32_igemm_minmax_test",
4927 srcs = [
4928 "test/f32-igemm-minmax.cc",
4929 "test/gemm-microkernel-tester.h",
4930 "src/xnnpack/AlignedAllocator.h",
4931 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004932 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004933)
4934
4935xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004936 name = "f16_igemm_minmax_test",
4937 srcs = [
4938 "test/f16-igemm-minmax.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"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004943)
4944
4945xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946 name = "f32_conv_hwc_test",
4947 srcs = [
4948 "test/f32-conv-hwc.cc",
4949 "test/conv-hwc-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 Dukhan1f29b802020-05-15 23:46:39 -07004956 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004957 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004958 "test/f32-conv-hwc2chw.cc",
4959 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 "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 Dukhan08c4a432019-10-03 09:29:21 -07004963)
4964
4965xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004966 name = "f32_dwconv_test",
4967 srcs = [
4968 "test/f32-dwconv.cc",
4969 "test/dwconv-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"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004973)
4974
4975xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004976 name = "f16_dwconv_minmax_test",
4977 srcs = [
4978 "test/f16-dwconv-minmax.cc",
4979 "test/dwconv-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"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004983)
4984
4985xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004986 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004987 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004988 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 "test/dwconv-microkernel-tester.h",
4990 "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 Dukhan1f29b802020-05-15 23:46:39 -07004996 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004998 "test/f32-dwconv-chw.cc",
4999 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000 "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 Dukhan08c4a432019-10-03 09:29:21 -07005003)
5004
5005xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005006 name = "f16_gavgpool_minmax_test",
5007 srcs = [
5008 "test/f16-gavgpool-minmax.cc",
5009 "test/gavgpool-microkernel-tester.h",
5010 "src/xnnpack/AlignedAllocator.h",
5011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005016 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005018 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019 "test/gavgpool-microkernel-tester.h",
5020 "src/xnnpack/AlignedAllocator.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005026 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005028 "test/f32-gavgpool-cw.cc",
5029 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005030 "src/xnnpack/AlignedAllocator.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005036 name = "f32_gemm_test",
5037 srcs = [
5038 "test/f32-gemm.cc",
5039 "test/gemm-microkernel-tester.h",
5040 "src/xnnpack/AlignedAllocator.h",
5041 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005042 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005043)
5044
5045xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005046 name = "f32_gemm_relu_test",
5047 srcs = [
5048 "test/f32-gemm-relu.cc",
5049 "test/gemm-microkernel-tester.h",
5050 "src/xnnpack/AlignedAllocator.h",
5051 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005052 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005053)
5054
5055xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005056 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005058 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059 "test/gemm-microkernel-tester.h",
5060 "src/xnnpack/AlignedAllocator.h",
5061 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005062 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063)
5064
5065xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005066 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005068 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005069 "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 Dukhan08c4a432019-10-03 09:29:21 -07005073)
5074
5075xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005076 name = "f16_hswish_test",
5077 srcs = [
5078 "test/f16-hswish.cc",
5079 "test/hswish-microkernel-tester.h",
5080 ] + MICROKERNEL_TEST_HDRS,
5081 deps = MICROKERNEL_TEST_DEPS,
5082)
5083
5084xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 name = "f32_hswish_test",
5086 srcs = [
5087 "test/f32-hswish.cc",
5088 "test/hswish-microkernel-tester.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 deps = MICROKERNEL_TEST_DEPS,
5091)
5092
5093xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005094 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005096 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005097 "test/maxpool-microkernel-tester.h",
5098 ] + MICROKERNEL_TEST_HDRS,
5099 deps = MICROKERNEL_TEST_DEPS,
5100)
5101
5102xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005103 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005104 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005105 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005106 "test/avgpool-microkernel-tester.h",
5107 "src/xnnpack/AlignedAllocator.h",
5108 ] + MICROKERNEL_TEST_HDRS,
5109 deps = MICROKERNEL_TEST_DEPS,
5110)
5111
5112xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005113 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005115 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005116 "test/gemm-microkernel-tester.h",
5117 "src/xnnpack/AlignedAllocator.h",
5118 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005119 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120)
5121
5122xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005123 name = "f16_prelu_test",
5124 srcs = [
5125 "test/f16-prelu.cc",
5126 "test/prelu-microkernel-tester.h",
5127 "src/xnnpack/AlignedAllocator.h",
5128 ] + MICROKERNEL_TEST_HDRS,
5129 deps = MICROKERNEL_TEST_DEPS,
5130)
5131
5132xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005133 name = "f32_prelu_test",
5134 srcs = [
5135 "test/f32-prelu.cc",
5136 "test/prelu-microkernel-tester.h",
5137 "src/xnnpack/AlignedAllocator.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005143 name = "f32_raddexpminusmax_test",
5144 srcs = [
5145 "test/f32-raddexpminusmax.cc",
5146 "test/raddexpminusmax-microkernel-tester.h",
5147 ] + MICROKERNEL_TEST_HDRS,
5148 deps = MICROKERNEL_TEST_DEPS,
5149)
5150
5151xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005152 name = "f32_raddextexp_test",
5153 srcs = [
5154 "test/f32-raddextexp.cc",
5155 "test/raddextexp-microkernel-tester.h",
5156 ] + MICROKERNEL_TEST_HDRS,
5157 deps = MICROKERNEL_TEST_DEPS,
5158)
5159
5160xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005161 name = "f32_raddstoreexpminusmax_test",
5162 srcs = [
5163 "test/f32-raddstoreexpminusmax.cc",
5164 "test/raddstoreexpminusmax-microkernel-tester.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005170 name = "f32_relu_test",
5171 srcs = [
5172 "test/f32-relu.cc",
5173 "test/vunary-microkernel-tester.h",
5174 ] + MICROKERNEL_TEST_HDRS,
5175 deps = MICROKERNEL_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179 name = "f32_rmax_test",
5180 srcs = [
5181 "test/f32-rmax.cc",
5182 "test/rmax-microkernel-tester.h",
5183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005188 name = "f32_sigmoid_test",
5189 srcs = [
5190 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005191 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005197 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005198 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005199 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200 "test/spmm-microkernel-tester.h",
5201 "src/xnnpack/AlignedAllocator.h",
5202 ] + MICROKERNEL_TEST_HDRS,
5203 deps = MICROKERNEL_TEST_DEPS,
5204)
5205
5206xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005207 name = "f16_vadd_minmax_test",
5208 srcs = [
5209 "test/f16-vadd-minmax.cc",
5210 "test/vbinary-microkernel-tester.h",
5211 ] + MICROKERNEL_TEST_HDRS,
5212 deps = MICROKERNEL_TEST_DEPS,
5213)
5214
5215xnnpack_unit_test(
5216 name = "f16_vaddc_minmax_test",
5217 srcs = [
5218 "test/f16-vaddc-minmax.cc",
5219 "test/vbinaryc-microkernel-tester.h",
5220 ] + MICROKERNEL_TEST_HDRS,
5221 deps = MICROKERNEL_TEST_DEPS,
5222)
5223
5224xnnpack_unit_test(
5225 name = "f16_vdiv_minmax_test",
5226 srcs = [
5227 "test/f16-vdiv-minmax.cc",
5228 "test/vbinary-microkernel-tester.h",
5229 ] + MICROKERNEL_TEST_HDRS,
5230 deps = MICROKERNEL_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
5234 name = "f16_vdivc_minmax_test",
5235 srcs = [
5236 "test/f16-vdivc-minmax.cc",
5237 "test/vbinaryc-microkernel-tester.h",
5238 ] + MICROKERNEL_TEST_HDRS,
5239 deps = MICROKERNEL_TEST_DEPS,
5240)
5241
5242xnnpack_unit_test(
5243 name = "f16_vrdivc_minmax_test",
5244 srcs = [
5245 "test/f16-vrdivc-minmax.cc",
5246 "test/vbinaryc-microkernel-tester.h",
5247 ] + MICROKERNEL_TEST_HDRS,
5248 deps = MICROKERNEL_TEST_DEPS,
5249)
5250
5251xnnpack_unit_test(
5252 name = "f16_vmax_test",
5253 srcs = [
5254 "test/f16-vmax.cc",
5255 "test/vbinary-microkernel-tester.h",
5256 ] + MICROKERNEL_TEST_HDRS,
5257 deps = MICROKERNEL_TEST_DEPS,
5258)
5259
5260xnnpack_unit_test(
5261 name = "f16_vmaxc_test",
5262 srcs = [
5263 "test/f16-vmaxc.cc",
5264 "test/vbinaryc-microkernel-tester.h",
5265 ] + MICROKERNEL_TEST_HDRS,
5266 deps = MICROKERNEL_TEST_DEPS,
5267)
5268
5269xnnpack_unit_test(
5270 name = "f16_vmin_test",
5271 srcs = [
5272 "test/f16-vmin.cc",
5273 "test/vbinary-microkernel-tester.h",
5274 ] + MICROKERNEL_TEST_HDRS,
5275 deps = MICROKERNEL_TEST_DEPS,
5276)
5277
5278xnnpack_unit_test(
5279 name = "f16_vminc_test",
5280 srcs = [
5281 "test/f16-vminc.cc",
5282 "test/vbinaryc-microkernel-tester.h",
5283 ] + MICROKERNEL_TEST_HDRS,
5284 deps = MICROKERNEL_TEST_DEPS,
5285)
5286
5287xnnpack_unit_test(
5288 name = "f16_vmul_minmax_test",
5289 srcs = [
5290 "test/f16-vmul-minmax.cc",
5291 "test/vbinary-microkernel-tester.h",
5292 ] + MICROKERNEL_TEST_HDRS,
5293 deps = MICROKERNEL_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
5297 name = "f16_vmulc_minmax_test",
5298 srcs = [
5299 "test/f16-vmulc-minmax.cc",
5300 "test/vbinaryc-microkernel-tester.h",
5301 ] + MICROKERNEL_TEST_HDRS,
5302 deps = MICROKERNEL_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
5306 name = "f16_vsub_minmax_test",
5307 srcs = [
5308 "test/f16-vsub-minmax.cc",
5309 "test/vbinary-microkernel-tester.h",
5310 ] + MICROKERNEL_TEST_HDRS,
5311 deps = MICROKERNEL_TEST_DEPS,
5312)
5313
5314xnnpack_unit_test(
5315 name = "f16_vsubc_minmax_test",
5316 srcs = [
5317 "test/f16-vsubc-minmax.cc",
5318 "test/vbinaryc-microkernel-tester.h",
5319 ] + MICROKERNEL_TEST_HDRS,
5320 deps = MICROKERNEL_TEST_DEPS,
5321)
5322
5323xnnpack_unit_test(
5324 name = "f16_vrsubc_minmax_test",
5325 srcs = [
5326 "test/f16-vrsubc-minmax.cc",
5327 "test/vbinaryc-microkernel-tester.h",
5328 ] + MICROKERNEL_TEST_HDRS,
5329 deps = MICROKERNEL_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005333 name = "f32_vabs_test",
5334 srcs = [
5335 "test/f32-vabs.cc",
5336 "test/vunary-microkernel-tester.h",
5337 ] + MICROKERNEL_TEST_HDRS,
5338 deps = MICROKERNEL_TEST_DEPS,
5339)
5340
5341xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005342 name = "f32_vadd_test",
5343 srcs = [
5344 "test/f32-vadd.cc",
5345 "test/vbinary-microkernel-tester.h",
5346 ] + MICROKERNEL_TEST_HDRS,
5347 deps = MICROKERNEL_TEST_DEPS,
5348)
5349
5350xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005351 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005353 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005354 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005355 ] + MICROKERNEL_TEST_HDRS,
5356 deps = MICROKERNEL_TEST_DEPS,
5357)
5358
5359xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005360 name = "f32_vadd_relu_test",
5361 srcs = [
5362 "test/f32-vadd-relu.cc",
5363 "test/vbinary-microkernel-tester.h",
5364 ] + MICROKERNEL_TEST_HDRS,
5365 deps = MICROKERNEL_TEST_DEPS,
5366)
5367
5368xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005369 name = "f32_vaddc_test",
5370 srcs = [
5371 "test/f32-vaddc.cc",
5372 "test/vbinaryc-microkernel-tester.h",
5373 ] + MICROKERNEL_TEST_HDRS,
5374 deps = MICROKERNEL_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005378 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005379 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005380 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005381 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005382 ] + MICROKERNEL_TEST_HDRS,
5383 deps = MICROKERNEL_TEST_DEPS,
5384)
5385
5386xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005387 name = "f32_vaddc_relu_test",
5388 srcs = [
5389 "test/f32-vaddc-relu.cc",
5390 "test/vbinaryc-microkernel-tester.h",
5391 ] + MICROKERNEL_TEST_HDRS,
5392 deps = MICROKERNEL_TEST_DEPS,
5393)
5394
5395xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005396 name = "f32_vdiv_test",
5397 srcs = [
5398 "test/f32-vdiv.cc",
5399 "test/vbinary-microkernel-tester.h",
5400 ] + MICROKERNEL_TEST_HDRS,
5401 deps = MICROKERNEL_TEST_DEPS,
5402)
5403
5404xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005405 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005406 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005407 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005408 "test/vbinary-microkernel-tester.h",
5409 ] + MICROKERNEL_TEST_HDRS,
5410 deps = MICROKERNEL_TEST_DEPS,
5411)
5412
5413xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005414 name = "f32_vdiv_relu_test",
5415 srcs = [
5416 "test/f32-vdiv-relu.cc",
5417 "test/vbinary-microkernel-tester.h",
5418 ] + MICROKERNEL_TEST_HDRS,
5419 deps = MICROKERNEL_TEST_DEPS,
5420)
5421
5422xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005423 name = "f32_vdivc_test",
5424 srcs = [
5425 "test/f32-vdivc.cc",
5426 "test/vbinaryc-microkernel-tester.h",
5427 ] + MICROKERNEL_TEST_HDRS,
5428 deps = MICROKERNEL_TEST_DEPS,
5429)
5430
5431xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005432 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005433 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005434 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005435 "test/vbinaryc-microkernel-tester.h",
5436 ] + MICROKERNEL_TEST_HDRS,
5437 deps = MICROKERNEL_TEST_DEPS,
5438)
5439
5440xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005441 name = "f32_vdivc_relu_test",
5442 srcs = [
5443 "test/f32-vdivc-relu.cc",
5444 "test/vbinaryc-microkernel-tester.h",
5445 ] + MICROKERNEL_TEST_HDRS,
5446 deps = MICROKERNEL_TEST_DEPS,
5447)
5448
5449xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005450 name = "f32_vrdivc_test",
5451 srcs = [
5452 "test/f32-vrdivc.cc",
5453 "test/vbinaryc-microkernel-tester.h",
5454 ] + MICROKERNEL_TEST_HDRS,
5455 deps = MICROKERNEL_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005459 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005460 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005461 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005462 "test/vbinaryc-microkernel-tester.h",
5463 ] + MICROKERNEL_TEST_HDRS,
5464 deps = MICROKERNEL_TEST_DEPS,
5465)
5466
5467xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005468 name = "f32_vrdivc_relu_test",
5469 srcs = [
5470 "test/f32-vrdivc-relu.cc",
5471 "test/vbinaryc-microkernel-tester.h",
5472 ] + MICROKERNEL_TEST_HDRS,
5473 deps = MICROKERNEL_TEST_DEPS,
5474)
5475
5476xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005477 name = "f32_vmax_test",
5478 srcs = [
5479 "test/f32-vmax.cc",
5480 "test/vbinary-microkernel-tester.h",
5481 ] + MICROKERNEL_TEST_HDRS,
5482 deps = MICROKERNEL_TEST_DEPS,
5483)
5484
5485xnnpack_unit_test(
5486 name = "f32_vmaxc_test",
5487 srcs = [
5488 "test/f32-vmaxc.cc",
5489 "test/vbinaryc-microkernel-tester.h",
5490 ] + MICROKERNEL_TEST_HDRS,
5491 deps = MICROKERNEL_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
5495 name = "f32_vmin_test",
5496 srcs = [
5497 "test/f32-vmin.cc",
5498 "test/vbinary-microkernel-tester.h",
5499 ] + MICROKERNEL_TEST_HDRS,
5500 deps = MICROKERNEL_TEST_DEPS,
5501)
5502
5503xnnpack_unit_test(
5504 name = "f32_vminc_test",
5505 srcs = [
5506 "test/f32-vminc.cc",
5507 "test/vbinaryc-microkernel-tester.h",
5508 ] + MICROKERNEL_TEST_HDRS,
5509 deps = MICROKERNEL_TEST_DEPS,
5510)
5511
5512xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005513 name = "f32_vmul_test",
5514 srcs = [
5515 "test/f32-vmul.cc",
5516 "test/vbinary-microkernel-tester.h",
5517 ] + MICROKERNEL_TEST_HDRS,
5518 deps = MICROKERNEL_TEST_DEPS,
5519)
5520
5521xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005522 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005523 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005524 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005525 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005526 ] + MICROKERNEL_TEST_HDRS,
5527 deps = MICROKERNEL_TEST_DEPS,
5528)
5529
5530xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005531 name = "f32_vmul_relu_test",
5532 srcs = [
5533 "test/f32-vmul-relu.cc",
5534 "test/vbinary-microkernel-tester.h",
5535 ] + MICROKERNEL_TEST_HDRS,
5536 deps = MICROKERNEL_TEST_DEPS,
5537)
5538
5539xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005540 name = "f32_vmulc_test",
5541 srcs = [
5542 "test/f32-vmulc.cc",
5543 "test/vbinaryc-microkernel-tester.h",
5544 ] + MICROKERNEL_TEST_HDRS,
5545 deps = MICROKERNEL_TEST_DEPS,
5546)
5547
5548xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005549 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005550 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005551 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005552 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 ] + MICROKERNEL_TEST_HDRS,
5554 deps = MICROKERNEL_TEST_DEPS,
5555)
5556
5557xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005558 name = "f32_vmulc_relu_test",
5559 srcs = [
5560 "test/f32-vmulc-relu.cc",
5561 "test/vbinaryc-microkernel-tester.h",
5562 ] + MICROKERNEL_TEST_HDRS,
5563 deps = MICROKERNEL_TEST_DEPS,
5564)
5565
5566xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005567 name = "f16_vmulcaddc_minmax_test",
5568 srcs = [
5569 "test/f16-vmulcaddc-minmax.cc",
5570 "test/vmulcaddc-microkernel-tester.h",
5571 "src/xnnpack/AlignedAllocator.h",
5572 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005573 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005574)
5575
5576xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005577 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005579 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005580 "test/vmulcaddc-microkernel-tester.h",
5581 "src/xnnpack/AlignedAllocator.h",
5582 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005583 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005584)
5585
5586xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005587 name = "f32_vlrelu_test",
5588 srcs = [
5589 "test/f32-vlrelu.cc",
5590 "test/vunary-microkernel-tester.h",
5591 ] + MICROKERNEL_TEST_HDRS,
5592 deps = MICROKERNEL_TEST_DEPS,
5593)
5594
5595xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005596 name = "f32_vneg_test",
5597 srcs = [
5598 "test/f32-vneg.cc",
5599 "test/vunary-microkernel-tester.h",
5600 ] + MICROKERNEL_TEST_HDRS,
5601 deps = MICROKERNEL_TEST_DEPS,
5602)
5603
5604xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005605 name = "f32_vrndne_test",
5606 srcs = [
5607 "test/f32-vrndne.cc",
5608 "test/vunary-microkernel-tester.h",
5609 ] + MICROKERNEL_TEST_HDRS,
5610 deps = MICROKERNEL_TEST_DEPS,
5611)
5612
5613xnnpack_unit_test(
5614 name = "f32_vrndz_test",
5615 srcs = [
5616 "test/f32-vrndz.cc",
5617 "test/vunary-microkernel-tester.h",
5618 ] + MICROKERNEL_TEST_HDRS,
5619 deps = MICROKERNEL_TEST_DEPS,
5620)
5621
5622xnnpack_unit_test(
5623 name = "f32_vrndu_test",
5624 srcs = [
5625 "test/f32-vrndu.cc",
5626 "test/vunary-microkernel-tester.h",
5627 ] + MICROKERNEL_TEST_HDRS,
5628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
5631xnnpack_unit_test(
5632 name = "f32_vrndd_test",
5633 srcs = [
5634 "test/f32-vrndd.cc",
5635 "test/vunary-microkernel-tester.h",
5636 ] + MICROKERNEL_TEST_HDRS,
5637 deps = MICROKERNEL_TEST_DEPS,
5638)
5639
5640xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005641 name = "f32_vscale_test",
5642 srcs = [
5643 "test/f32-vscale.cc",
5644 "test/vscale-microkernel-tester.h",
5645 ] + MICROKERNEL_TEST_HDRS,
5646 deps = MICROKERNEL_TEST_DEPS,
5647)
5648
5649xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005650 name = "f32_vscaleexpminusmax_test",
5651 srcs = [
5652 "test/f32-vscaleexpminusmax.cc",
5653 "test/vscaleexpminusmax-microkernel-tester.h",
5654 ] + MICROKERNEL_TEST_HDRS,
5655 deps = MICROKERNEL_TEST_DEPS,
5656)
5657
5658xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005659 name = "f32_vscaleextexp_test",
5660 srcs = [
5661 "test/f32-vscaleextexp.cc",
5662 "test/vscaleextexp-microkernel-tester.h",
5663 ] + MICROKERNEL_TEST_HDRS,
5664 deps = MICROKERNEL_TEST_DEPS,
5665)
5666
5667xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005668 name = "f32_vsqr_test",
5669 srcs = [
5670 "test/f32-vsqr.cc",
5671 "test/vunary-microkernel-tester.h",
5672 ] + MICROKERNEL_TEST_HDRS,
5673 deps = MICROKERNEL_TEST_DEPS,
5674)
5675
5676xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005677 name = "f32_vsqrdiff_test",
5678 srcs = [
5679 "test/f32-vsqrdiff.cc",
5680 "test/vbinary-microkernel-tester.h",
5681 ] + MICROKERNEL_TEST_HDRS,
5682 deps = MICROKERNEL_TEST_DEPS,
5683)
5684
5685xnnpack_unit_test(
5686 name = "f32_vsqrdiffc_test",
5687 srcs = [
5688 "test/f32-vsqrdiffc.cc",
5689 "test/vbinaryc-microkernel-tester.h",
5690 ] + MICROKERNEL_TEST_HDRS,
5691 deps = MICROKERNEL_TEST_DEPS,
5692)
5693
5694xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005695 name = "f32_vsqrt_test",
5696 srcs = [
5697 "test/f32-vsqrt.cc",
5698 "test/vunary-microkernel-tester.h",
5699 ] + MICROKERNEL_TEST_HDRS,
5700 deps = MICROKERNEL_TEST_DEPS,
5701)
5702
5703xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005704 name = "f32_vsub_test",
5705 srcs = [
5706 "test/f32-vsub.cc",
5707 "test/vbinary-microkernel-tester.h",
5708 ] + MICROKERNEL_TEST_HDRS,
5709 deps = MICROKERNEL_TEST_DEPS,
5710)
5711
5712xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005713 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005714 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005715 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005716 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005717 ] + MICROKERNEL_TEST_HDRS,
5718 deps = MICROKERNEL_TEST_DEPS,
5719)
5720
5721xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005722 name = "f32_vsub_relu_test",
5723 srcs = [
5724 "test/f32-vsub-relu.cc",
5725 "test/vbinary-microkernel-tester.h",
5726 ] + MICROKERNEL_TEST_HDRS,
5727 deps = MICROKERNEL_TEST_DEPS,
5728)
5729
5730xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005731 name = "f32_vsubc_test",
5732 srcs = [
5733 "test/f32-vsubc.cc",
5734 "test/vbinaryc-microkernel-tester.h",
5735 ] + MICROKERNEL_TEST_HDRS,
5736 deps = MICROKERNEL_TEST_DEPS,
5737)
5738
5739xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005740 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005741 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005742 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005743 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005744 ] + MICROKERNEL_TEST_HDRS,
5745 deps = MICROKERNEL_TEST_DEPS,
5746)
5747
5748xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005749 name = "f32_vsubc_relu_test",
5750 srcs = [
5751 "test/f32-vsubc-relu.cc",
5752 "test/vbinaryc-microkernel-tester.h",
5753 ] + MICROKERNEL_TEST_HDRS,
5754 deps = MICROKERNEL_TEST_DEPS,
5755)
5756
5757xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005758 name = "f32_vrsubc_test",
5759 srcs = [
5760 "test/f32-vrsubc.cc",
5761 "test/vbinaryc-microkernel-tester.h",
5762 ] + MICROKERNEL_TEST_HDRS,
5763 deps = MICROKERNEL_TEST_DEPS,
5764)
5765
5766xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005767 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005768 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005769 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005770 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005771 ] + MICROKERNEL_TEST_HDRS,
5772 deps = MICROKERNEL_TEST_DEPS,
5773)
5774
5775xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005776 name = "f32_vrsubc_relu_test",
5777 srcs = [
5778 "test/f32-vrsubc-relu.cc",
5779 "test/vbinaryc-microkernel-tester.h",
5780 ] + MICROKERNEL_TEST_HDRS,
5781 deps = MICROKERNEL_TEST_DEPS,
5782)
5783
5784xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005785 name = "qs8_dwconv_minmax_test",
5786 srcs = [
5787 "test/qs8-dwconv-minmax.cc",
5788 "test/dwconv-microkernel-tester.h",
5789 "src/xnnpack/AlignedAllocator.h",
5790 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5791 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5792)
5793
5794xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005795 name = "qs8_gavgpool_minmax_test",
5796 srcs = [
5797 "test/qs8-gavgpool-minmax.cc",
5798 "test/gavgpool-microkernel-tester.h",
5799 "src/xnnpack/AlignedAllocator.h",
5800 ] + MICROKERNEL_TEST_HDRS,
5801 deps = MICROKERNEL_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005805 name = "qs8_gemm_minmax_test",
5806 srcs = [
5807 "test/qs8-gemm-minmax.cc",
5808 "test/gemm-microkernel-tester.h",
5809 "src/xnnpack/AlignedAllocator.h",
5810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5812)
5813
5814xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005815 name = "qs8_igemm_minmax_test",
5816 srcs = [
5817 "test/qs8-igemm-minmax.cc",
5818 "test/gemm-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(
5825 name = "qs8_requantization_test",
5826 srcs = [
5827 "src/xnnpack/requantization-stubs.h",
5828 "test/qs8-requantization.cc",
5829 "test/requantization-tester.h",
5830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005835 name = "qs8_vadd_minmax_test",
5836 srcs = [
5837 "test/qs8-vadd-minmax.cc",
5838 "test/vadd-microkernel-tester.h",
5839 ] + MICROKERNEL_TEST_HDRS,
5840 deps = MICROKERNEL_TEST_DEPS,
5841)
5842
5843xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005844 name = "qs8_vaddc_minmax_test",
5845 srcs = [
5846 "test/qs8-vaddc-minmax.cc",
5847 "test/vaddc-microkernel-tester.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 deps = MICROKERNEL_TEST_DEPS,
5850)
5851
5852xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005853 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005854 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005855 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856 "test/avgpool-microkernel-tester.h",
5857 "src/xnnpack/AlignedAllocator.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005863 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005864 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005865 "test/qu8-dwconv-minmax.cc",
5866 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005867 "src/xnnpack/AlignedAllocator.h",
5868 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005869 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005870)
5871
5872xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005873 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005874 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005875 "test/qu8-igemm-minmax.cc",
5876 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005877 "src/xnnpack/AlignedAllocator.h",
5878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005879 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005880)
5881
5882xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005883 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005885 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005886 "test/gavgpool-microkernel-tester.h",
5887 "src/xnnpack/AlignedAllocator.h",
5888 ] + MICROKERNEL_TEST_HDRS,
5889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
5892xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005893 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005894 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005895 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 "test/gemm-microkernel-tester.h",
5897 "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 Dukhan5b69f8b2020-07-24 15:26:48 -07005903 name = "qu8_requantization_test",
5904 srcs = [
5905 "src/xnnpack/requantization-stubs.h",
5906 "test/qu8-requantization.cc",
5907 "test/requantization-tester.h",
5908 ] + MICROKERNEL_TEST_HDRS,
5909 deps = MICROKERNEL_TEST_DEPS,
5910)
5911
5912xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005913 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005914 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005915 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005916 "test/vadd-microkernel-tester.h",
5917 ] + MICROKERNEL_TEST_HDRS,
5918 deps = MICROKERNEL_TEST_DEPS,
5919)
5920
5921xnnpack_unit_test(
5922 name = "u8_clamp_test",
5923 srcs = [
5924 "test/u8-clamp.cc",
5925 "test/clamp-microkernel-tester.h",
5926 ] + MICROKERNEL_TEST_HDRS,
5927 deps = MICROKERNEL_TEST_DEPS,
5928)
5929
5930xnnpack_unit_test(
5931 name = "u8_lut32norm_test",
5932 srcs = [
5933 "test/u8-lut32norm.cc",
5934 "test/lut-norm-microkernel-tester.h",
5935 ] + MICROKERNEL_TEST_HDRS,
5936 deps = MICROKERNEL_TEST_DEPS,
5937)
5938
5939xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005940 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005941 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005942 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005943 "test/maxpool-microkernel-tester.h",
5944 ] + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS,
5946)
5947
5948xnnpack_unit_test(
5949 name = "u8_rmax_test",
5950 srcs = [
5951 "test/u8-rmax.cc",
5952 "test/rmax-microkernel-tester.h",
5953 ] + MICROKERNEL_TEST_HDRS,
5954 deps = MICROKERNEL_TEST_DEPS,
5955)
5956
5957xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005958 name = "x32_fill_test",
5959 srcs = [
5960 "test/x32-fill.cc",
5961 "test/fill-microkernel-tester.h",
5962 ] + MICROKERNEL_TEST_HDRS,
5963 deps = MICROKERNEL_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005967 name = "x32_packx_test",
5968 srcs = [
5969 "test/x32-packx.cc",
5970 "test/pack-microkernel-tester.h",
5971 "src/xnnpack/AlignedAllocator.h",
5972 ] + MICROKERNEL_TEST_HDRS,
5973 deps = MICROKERNEL_TEST_DEPS,
5974)
5975
5976xnnpack_unit_test(
5977 name = "x32_pad_test",
5978 srcs = [
5979 "test/x32-pad.cc",
5980 "test/pad-microkernel-tester.h",
5981 ] + MICROKERNEL_TEST_HDRS,
5982 deps = MICROKERNEL_TEST_DEPS,
5983)
5984
5985xnnpack_unit_test(
5986 name = "x32_unpool_test",
5987 srcs = [
5988 "test/x32-unpool.cc",
5989 "test/unpool-microkernel-tester.h",
5990 ] + MICROKERNEL_TEST_HDRS,
5991 deps = MICROKERNEL_TEST_DEPS,
5992)
5993
5994xnnpack_unit_test(
5995 name = "x32_zip_test",
5996 srcs = [
5997 "test/x32-zip.cc",
5998 "test/zip-microkernel-tester.h",
5999 ] + MICROKERNEL_TEST_HDRS,
6000 deps = MICROKERNEL_TEST_DEPS,
6001)
6002
6003xnnpack_unit_test(
6004 name = "x8_lut_test",
6005 srcs = [
6006 "test/x8-lut.cc",
6007 "test/lut-microkernel-tester.h",
6008 ] + MICROKERNEL_TEST_HDRS,
6009 deps = MICROKERNEL_TEST_DEPS,
6010)
6011
6012xnnpack_unit_test(
6013 name = "x8_zip_test",
6014 srcs = [
6015 "test/x8-zip.cc",
6016 "test/zip-microkernel-tester.h",
6017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
Marat Dukhan20c3b922020-03-10 03:45:06 -07006021########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006022
6023xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006024 name = "operator_size_test",
6025 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006026 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006027)
6028
Marat Dukhan20c3b922020-03-10 03:45:06 -07006029xnnpack_binary(
6030 name = "subgraph_size_test",
6031 srcs = ["test/subgraph-size.c"],
6032 deps = [":XNNPACK"],
6033)
6034
6035########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006036
6037xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006038 name = "abs_nc_test",
6039 srcs = [
6040 "test/abs-nc.cc",
6041 "test/abs-operator-tester.h",
6042 ],
6043 deps = OPERATOR_TEST_DEPS,
6044)
6045
6046xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006047 name = "add_nd_test",
6048 srcs = [
6049 "test/add-nd.cc",
6050 "test/binary-elementwise-operator-tester.h",
6051 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006052 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006053)
6054
6055xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006056 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006057 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006058 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006059 "test/argmax-pooling-operator-tester.h",
6060 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006061 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006062)
6063
6064xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006065 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006067 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068 "test/average-pooling-operator-tester.h",
6069 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006070 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071)
6072
6073xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006074 name = "bankers_rounding_nc_test",
6075 srcs = [
6076 "test/bankers-rounding-nc.cc",
6077 "test/bankers-rounding-operator-tester.h",
6078 ],
6079 deps = OPERATOR_TEST_DEPS,
6080)
6081
6082xnnpack_unit_test(
6083 name = "ceiling_nc_test",
6084 srcs = [
6085 "test/ceiling-nc.cc",
6086 "test/ceiling-operator-tester.h",
6087 ],
6088 deps = OPERATOR_TEST_DEPS,
6089)
6090
6091xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006092 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006094 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095 "test/channel-shuffle-operator-tester.h",
6096 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006097 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006098)
6099
6100xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006101 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006102 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006103 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006104 "test/clamp-operator-tester.h",
6105 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006106 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006107)
6108
6109xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006110 name = "constant_pad_nd_test",
6111 srcs = [
6112 "test/constant-pad-nd.cc",
6113 "test/constant-pad-operator-tester.h",
6114 ],
6115 deps = OPERATOR_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006119 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006120 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006121 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006122 "test/convolution-operator-tester.h",
6123 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006124 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006125)
6126
6127xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006128 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006129 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006130 "test/convolution-nchw.cc",
6131 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006133 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006134)
6135
6136xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006137 name = "copy_nc_test",
6138 srcs = [
6139 "test/copy-nc.cc",
6140 "test/copy-operator-tester.h",
6141 ],
6142 deps = OPERATOR_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006146 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006147 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006148 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149 "test/deconvolution-operator-tester.h",
6150 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006151 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152)
6153
6154xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006155 name = "divide_nd_test",
6156 srcs = [
6157 "test/binary-elementwise-operator-tester.h",
6158 "test/divide-nd.cc",
6159 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006160 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006161)
6162
6163xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006164 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006165 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006166 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006167 "test/fully-connected-operator-tester.h",
6168 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006169 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006170)
6171
6172xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006173 name = "floor_nc_test",
6174 srcs = [
6175 "test/floor-nc.cc",
6176 "test/floor-operator-tester.h",
6177 ],
6178 deps = OPERATOR_TEST_DEPS,
6179)
6180
6181xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006182 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006183 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006184 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006185 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006186 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006187 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006188)
6189
6190xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006191 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006192 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006193 "test/global-average-pooling-ncw.cc",
6194 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006195 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006196 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197)
6198
6199xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006200 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006201 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006202 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006203 "test/hardswish-operator-tester.h",
6204 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006205 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006206)
6207
6208xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006209 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006210 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006211 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006212 "test/leaky-relu-operator-tester.h",
6213 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006214 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215)
6216
6217xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006218 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006219 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006220 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006221 "test/max-pooling-operator-tester.h",
6222 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006223 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224)
6225
6226xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006227 name = "maximum_nd_test",
6228 srcs = [
6229 "test/binary-elementwise-operator-tester.h",
6230 "test/maximum-nd.cc",
6231 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006232 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006233)
6234
6235xnnpack_unit_test(
6236 name = "minimum_nd_test",
6237 srcs = [
6238 "test/binary-elementwise-operator-tester.h",
6239 "test/minimum-nd.cc",
6240 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006241 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006242)
6243
6244xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006245 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006246 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006247 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006248 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006249 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006250 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006251)
6252
6253xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006254 name = "negate_nc_test",
6255 srcs = [
6256 "test/negate-nc.cc",
6257 "test/negate-operator-tester.h",
6258 ],
6259 deps = OPERATOR_TEST_DEPS,
6260)
6261
6262xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006263 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006264 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006265 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266 "test/prelu-operator-tester.h",
6267 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006268 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006269)
6270
6271xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006272 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006273 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006274 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006275 "test/resize-bilinear-operator-tester.h",
6276 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006277 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006278)
6279
6280xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006281 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006282 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006283 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284 "test/sigmoid-operator-tester.h",
6285 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006286 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287)
6288
6289xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006290 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006292 "test/softmax-nc.cc",
6293 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006295 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296)
6297
6298xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006299 name = "square_nc_test",
6300 srcs = [
6301 "test/square-nc.cc",
6302 "test/square-operator-tester.h",
6303 ],
6304 deps = OPERATOR_TEST_DEPS,
6305)
6306
6307xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006308 name = "square_root_nc_test",
6309 srcs = [
6310 "test/square-root-nc.cc",
6311 "test/square-root-operator-tester.h",
6312 ],
6313 deps = OPERATOR_TEST_DEPS,
6314)
6315
6316xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006317 name = "squared_difference_nd_test",
6318 srcs = [
6319 "test/binary-elementwise-operator-tester.h",
6320 "test/squared-difference-nd.cc",
6321 ],
6322 deps = OPERATOR_TEST_DEPS,
6323)
6324
6325xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006326 name = "subtract_nd_test",
6327 srcs = [
6328 "test/binary-elementwise-operator-tester.h",
6329 "test/subtract-nd.cc",
6330 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006331 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006332)
6333
6334xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006335 name = "truncation_nc_test",
6336 srcs = [
6337 "test/truncation-nc.cc",
6338 "test/truncation-operator-tester.h",
6339 ],
6340 deps = OPERATOR_TEST_DEPS,
6341)
6342
6343xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006344 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006345 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006346 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006347 "test/unpooling-operator-tester.h",
6348 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006349 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006350)
6351
Chao Mei6ddfc602020-05-13 22:29:36 -07006352############################### Misc unit tests ###############################
6353
6354xnnpack_unit_test(
6355 name = "memory_planner_test",
6356 srcs = [
6357 "test/memory-planner-test.cc",
6358 ],
6359 deps = [
6360 ":XNNPACK",
6361 ":memory_planner",
6362 ],
6363)
6364
Marat Dukhan08c4a432019-10-03 09:29:21 -07006365############################# Build configurations #############################
6366
Marat Dukhanb8642352019-10-30 15:43:02 -07006367# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006368config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006369 name = "xnn_enable_assembly_explicit_true",
6370 define_values = {"xnn_enable_assembly": "true"},
6371)
6372
6373# Disables usage of assembly kernels.
6374config_setting(
6375 name = "xnn_enable_assembly_explicit_false",
6376 define_values = {"xnn_enable_assembly": "false"},
6377)
6378
Marat Dukhan9de90e02020-06-18 16:04:12 -07006379# Enables usage of sparse inference.
6380config_setting(
6381 name = "xnn_enable_sparse_explicit_true",
6382 define_values = {"xnn_enable_sparse": "true"},
6383)
6384
6385# Disables usage of sparse inference.
6386config_setting(
6387 name = "xnn_enable_sparse_explicit_false",
6388 define_values = {"xnn_enable_sparse": "false"},
6389)
6390
Marat Dukhan05702cf2020-03-26 15:41:33 -07006391# Disables usage of HMP-aware optimizations.
6392config_setting(
6393 name = "xnn_enable_hmp_explicit_false",
6394 define_values = {"xnn_enable_hmp": "false"},
6395)
6396
Chao Mei6ddfc602020-05-13 22:29:36 -07006397# Enable usage of optimized memory allocation
6398config_setting(
6399 name = "xnn_enable_memopt_explicit_true",
6400 define_values = {"xnnpack_enable_memopt": "true"},
6401)
6402
6403# Disable usage of optimized memory allocation
6404config_setting(
6405 name = "xnn_enable_memopt_explicit_false",
6406 define_values = {"xnnpack_enable_memopt": "false"},
6407)
6408
Marat Dukhanb8642352019-10-30 15:43:02 -07006409# Builds with -c dbg
6410config_setting(
6411 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006413 "compilation_mode": "dbg",
6414 },
6415)
6416
6417# Builds with -c opt
6418config_setting(
6419 name = "optimized_build",
6420 values = {
6421 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006422 },
6423)
6424
6425config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006426 name = "linux_k8",
6427 values = {"cpu": "k8"},
6428)
6429
6430config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006431 name = "linux_arm",
6432 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006433)
6434
6435config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006436 name = "linux_armeabi",
6437 values = {"cpu": "armeabi"},
6438)
6439
6440config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006441 name = "linux_armhf",
6442 values = {"cpu": "armhf"},
6443)
6444
6445config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006446 name = "linux_armv7a",
6447 values = {"cpu": "armv7a"},
6448)
6449
6450config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006451 name = "linux_aarch64",
6452 values = {"cpu": "aarch64"},
6453)
6454
6455config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006456 name = "android",
6457 values = {"crosstool_top": "//external:android/crosstool"},
6458)
6459
6460config_setting(
6461 name = "android_armv7",
6462 values = {
6463 "crosstool_top": "//external:android/crosstool",
6464 "cpu": "armeabi-v7a",
6465 },
6466)
6467
6468config_setting(
6469 name = "android_arm64",
6470 values = {
6471 "crosstool_top": "//external:android/crosstool",
6472 "cpu": "arm64-v8a",
6473 },
6474)
6475
6476config_setting(
6477 name = "android_x86",
6478 values = {
6479 "crosstool_top": "//external:android/crosstool",
6480 "cpu": "x86",
6481 },
6482)
6483
6484config_setting(
6485 name = "android_x86_64",
6486 values = {
6487 "crosstool_top": "//external:android/crosstool",
6488 "cpu": "x86_64",
6489 },
6490)
6491
6492config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006493 name = "windows_x86_64",
6494 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006495)
6496
6497config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006498 name = "windows_x86_64_clang",
6499 values = {
6500 "compiler": "clang-cl",
6501 "cpu": "x64_windows",
6502 },
6503)
6504
6505config_setting(
6506 name = "windows_x86_64_mingw",
6507 values = {
6508 "compiler": "mingw-gcc",
6509 "cpu": "x64_windows",
6510 },
6511)
6512
6513config_setting(
6514 name = "windows_x86_64_msys",
6515 values = {
6516 "compiler": "msys-gcc",
6517 "cpu": "x64_windows",
6518 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006519)
6520
6521config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006522 name = "macos_x86_64",
6523 values = {
6524 "apple_platform_type": "macos",
6525 "cpu": "darwin",
6526 },
6527)
6528
6529config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006530 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006531 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532)
6533
6534config_setting(
6535 name = "emscripten_wasm",
6536 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006537 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006538 "cpu": "wasm",
6539 },
6540)
6541
6542config_setting(
6543 name = "emscripten_wasmsimd",
6544 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006545 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006546 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006547 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006548 },
6549)
6550
6551config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006552 name = "ios_armv7",
6553 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006554 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006555 "cpu": "ios_armv7",
6556 },
6557)
6558
6559config_setting(
6560 name = "ios_arm64",
6561 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006562 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006563 "cpu": "ios_arm64",
6564 },
6565)
6566
6567config_setting(
6568 name = "ios_arm64e",
6569 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006570 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006571 "cpu": "ios_arm64e",
6572 },
6573)
6574
6575config_setting(
6576 name = "ios_x86",
6577 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006578 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006579 "cpu": "ios_i386",
6580 },
6581)
6582
6583config_setting(
6584 name = "ios_x86_64",
6585 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006586 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006587 "cpu": "ios_x86_64",
6588 },
6589)
6590
6591config_setting(
6592 name = "watchos_armv7k",
6593 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006594 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006595 "cpu": "watchos_armv7k",
6596 },
6597)
6598
6599config_setting(
6600 name = "watchos_arm64_32",
6601 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006602 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006603 "cpu": "watchos_arm64_32",
6604 },
6605)
6606
6607config_setting(
6608 name = "watchos_x86",
6609 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006610 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006611 "cpu": "watchos_i386",
6612 },
6613)
6614
6615config_setting(
6616 name = "watchos_x86_64",
6617 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006618 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006619 "cpu": "watchos_x86_64",
6620 },
6621)
6622
6623config_setting(
6624 name = "tvos_arm64",
6625 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006626 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006627 "cpu": "tvos_arm64",
6628 },
6629)
6630
6631config_setting(
6632 name = "tvos_x86_64",
6633 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006634 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006635 "cpu": "tvos_x86_64",
6636 },
6637)