blob: 4ff076ce5d41a3ee8f8a294ec6addebfb3dfce66 [file] [log] [blame]
Miao Wange9993472020-02-10 15:00:10 -08001// Copyright (C) 2020 The Android Open Source Project
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
Bob Badourc0116292021-02-12 20:34:13 -080015package {
16 default_applicable_licenses: ["external_XNNPACK_license"],
17}
18
19// Added automatically by a large-scale-change
20// See: http://go/android-license-faq
21license {
22 name: "external_XNNPACK_license",
23 visibility: [":__subpackages__"],
24 license_kinds: [
25 "SPDX-license-identifier-BSD",
26 ],
27 license_text: [
28 "LICENSE",
29 ],
30}
31
Miao Wange9993472020-02-10 15:00:10 -080032OPERATOR_SRCS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -070033 "src/operators/argmax-pooling-nhwc.c",
34 "src/operators/average-pooling-nhwc.c",
35 "src/operators/binary-elementwise-nd.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070036 "src/operators/channel-shuffle-nc.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070037 "src/operators/constant-pad-nd.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070038 "src/operators/convolution-nchw.c",
39 "src/operators/convolution-nhwc.c",
40 "src/operators/deconvolution-nhwc.c",
Miao Wang5eea8312020-12-07 09:12:40 -080041 "src/operators/depth-to-space-nchw2nhwc.c",
42 "src/operators/depth-to-space-nhwc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070043 "src/operators/fully-connected-nc.c",
44 "src/operators/global-average-pooling-ncw.c",
45 "src/operators/global-average-pooling-nwc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070046 "src/operators/leaky-relu-nc.c",
47 "src/operators/max-pooling-nhwc.c",
48 "src/operators/prelu-nc.c",
Miao Wang5eea8312020-12-07 09:12:40 -080049 "src/operators/resize-bilinear-nchw.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070050 "src/operators/resize-bilinear-nhwc.c",
51 "src/operators/sigmoid-nc.c",
52 "src/operators/softmax-nc.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070053 "src/operators/unary-elementwise-nc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070054 "src/operators/unpooling-nhwc.c",
Miao Wange9993472020-02-10 15:00:10 -080055]
56
Miao Wang86f5fbe2020-07-24 11:16:10 -070057LOGGING_SRCS = [
58 "src/operator-strings.c",
59 "src/subgraph-strings.c",
60]
61
62SUBGRAPH_SRCS = [
63 "src/subgraph/abs.c",
64 "src/subgraph/add2.c",
65 "src/subgraph/argmax-pooling-2d.c",
66 "src/subgraph/average-pooling-2d.c",
67 "src/subgraph/bankers-rounding.c",
68 "src/subgraph/ceiling.c",
69 "src/subgraph/clamp.c",
70 "src/subgraph/convolution-2d.c",
71 "src/subgraph/deconvolution-2d.c",
Miao Wang5eea8312020-12-07 09:12:40 -080072 "src/subgraph/depth-to-space.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070073 "src/subgraph/depthwise-convolution-2d.c",
74 "src/subgraph/divide.c",
Miao Wang55abe392021-02-03 14:54:41 -080075 "src/subgraph/elu.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070076 "src/subgraph/floor.c",
Miao Wang5eea8312020-12-07 09:12:40 -080077 "src/subgraph/fully-connected.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070078 "src/subgraph/global-average-pooling-2d.c",
79 "src/subgraph/hardswish.c",
80 "src/subgraph/leaky-relu.c",
81 "src/subgraph/max-pooling-2d.c",
82 "src/subgraph/maximum2.c",
83 "src/subgraph/minimum2.c",
84 "src/subgraph/multiply2.c",
85 "src/subgraph/negate.c",
86 "src/subgraph/prelu.c",
87 "src/subgraph/sigmoid.c",
88 "src/subgraph/softmax.c",
Miao Wang5eea8312020-12-07 09:12:40 -080089 "src/subgraph/square-root.c",
90 "src/subgraph/square.c",
91 "src/subgraph/squared-difference.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070092 "src/subgraph/static-constant-pad.c",
93 "src/subgraph/static-reshape.c",
94 "src/subgraph/static-resize-bilinear-2d.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070095 "src/subgraph/subtract.c",
96 "src/subgraph/unpooling-2d.c",
97]
98
Miao Wange9993472020-02-10 15:00:10 -080099TABLE_SRCS = [
100 "src/tables/exp2-k-over-64.c",
101 "src/tables/exp2-k-over-2048.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800102 "src/tables/exp2minus-k-over-4.c",
103 "src/tables/exp2minus-k-over-8.c",
104 "src/tables/exp2minus-k-over-16.c",
105 "src/tables/exp2minus-k-over-64.c",
106 "src/tables/exp2minus-k-over-2048.c",
Miao Wange9993472020-02-10 15:00:10 -0800107]
108
109SCALAR_UKERNELS = [
110 "src/f32-argmaxpool/4x-scalar-c1.c",
111 "src/f32-argmaxpool/9p8x-scalar-c1.c",
112 "src/f32-argmaxpool/9x-scalar-c1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700113 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
114 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700115 "src/f32-clamp/gen/scalar-x1.c",
116 "src/f32-clamp/gen/scalar-x2.c",
117 "src/f32-clamp/gen/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800118 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
119 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700120 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700121 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
122 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800123 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
124 "src/f32-dwconv/gen/up1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700125 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
126 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800127 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
128 "src/f32-dwconv/gen/up1x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700129 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
130 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800131 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x25-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700133 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
134 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800135 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
136 "src/f32-dwconv/gen/up2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700137 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700141 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800143 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
144 "src/f32-dwconv/gen/up2x25-scalar.c",
145 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
146 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
147 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
148 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
149 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
150 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
151 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
152 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
153 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
154 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
155 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
156 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
157 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
158 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
159 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
160 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
161 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
162 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
163 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
164 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
165 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
166 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
167 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
168 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
169 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
170 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
171 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
173 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
174 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
175 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
176 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
177 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
178 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
179 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
180 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
181 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
182 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700183 "src/f32-gavgpool-cw/scalar-x1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700184 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
185 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
186 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
187 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
188 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700189 "src/f32-gemm/gen/1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800190 "src/f32-gemm/gen/1x4-relu-scalar.c",
191 "src/f32-gemm/gen/1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700192 "src/f32-gemm/gen/2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800193 "src/f32-gemm/gen/2x4-relu-scalar.c",
194 "src/f32-gemm/gen/2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700195 "src/f32-gemm/gen/4x2-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800196 "src/f32-gemm/gen/4x2-relu-scalar.c",
197 "src/f32-gemm/gen/4x2-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700198 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800199 "src/f32-gemm/gen/4x4-relu-scalar.c",
200 "src/f32-gemm/gen/4x4-scalar.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700201 "src/f32-hswish/gen/hswish-scalar-x1.c",
202 "src/f32-hswish/gen/hswish-scalar-x2.c",
203 "src/f32-hswish/gen/hswish-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800204 "src/f32-ibilinear-chw/gen/scalar-p1.c",
205 "src/f32-ibilinear-chw/gen/scalar-p2.c",
206 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700207 "src/f32-ibilinear/gen/scalar-c1.c",
208 "src/f32-ibilinear/gen/scalar-c2.c",
209 "src/f32-ibilinear/gen/scalar-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700210 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800211 "src/f32-igemm/gen/1x4-relu-scalar.c",
212 "src/f32-igemm/gen/1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700213 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800214 "src/f32-igemm/gen/2x4-relu-scalar.c",
215 "src/f32-igemm/gen/2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700216 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800217 "src/f32-igemm/gen/4x2-relu-scalar.c",
218 "src/f32-igemm/gen/4x2-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700219 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800220 "src/f32-igemm/gen/4x4-relu-scalar.c",
221 "src/f32-igemm/gen/4x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700222 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
223 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
224 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
225 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
226 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
227 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
228 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800229 "src/f32-prelu/gen/scalar-2x1.c",
230 "src/f32-prelu/gen/scalar-2x4.c",
231 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Miao Wange9993472020-02-10 15:00:10 -0800232 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800233 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Miao Wange9993472020-02-10 15:00:10 -0800234 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
235 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800236 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800237 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Miao Wange9993472020-02-10 15:00:10 -0800238 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800239 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Miao Wange9993472020-02-10 15:00:10 -0800240 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
241 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800242 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700243 "src/f32-relu/gen/scalar-x1.c",
244 "src/f32-relu/gen/scalar-x2.c",
245 "src/f32-relu/gen/scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800246 "src/f32-relu/gen/scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800247 "src/f32-rmax/scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800248 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
249 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
250 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800251 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
252 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
253 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800254 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
255 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
256 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700257 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
258 "src/f32-spmm/gen/1x1-minmax-scalar.c",
259 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
260 "src/f32-spmm/gen/2x1-minmax-scalar.c",
261 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
262 "src/f32-spmm/gen/4x1-minmax-scalar.c",
263 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
264 "src/f32-spmm/gen/8x1-minmax-scalar.c",
265 "src/f32-spmm/gen/8x2-minmax-scalar.c",
266 "src/f32-spmm/gen/8x4-minmax-scalar.c",
267 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
268 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
269 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800270 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
271 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
272 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
273 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
274 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700275 "src/f32-vbinary/gen/vadd-scalar-x1.c",
276 "src/f32-vbinary/gen/vadd-scalar-x2.c",
277 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800278 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700279 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
280 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
281 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800282 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
283 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
284 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
285 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
286 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700287 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
288 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
289 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800290 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700291 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
292 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
293 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800294 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
295 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
296 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
297 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
298 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700299 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
300 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
301 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800302 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700303 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800306 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
307 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
310 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700311 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
312 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
313 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800314 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800315 "src/f32-vbinary/gen/vmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800318 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800319 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
320 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
321 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800322 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800323 "src/f32-vbinary/gen/vmin-scalar-x1.c",
324 "src/f32-vbinary/gen/vmin-scalar-x2.c",
325 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800326 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800327 "src/f32-vbinary/gen/vminc-scalar-x1.c",
328 "src/f32-vbinary/gen/vminc-scalar-x2.c",
329 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800330 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700331 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800334 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
335 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
336 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
337 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
338 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700339 "src/f32-vbinary/gen/vmul-scalar-x1.c",
340 "src/f32-vbinary/gen/vmul-scalar-x2.c",
341 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800342 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700343 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
344 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
345 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800346 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
347 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
348 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
349 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
350 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700351 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
352 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
353 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800354 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700355 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
356 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
357 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800358 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
359 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
360 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
361 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
362 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700363 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
364 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
365 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800366 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700367 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
368 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
369 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800370 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
371 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
372 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
373 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
374 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700375 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
376 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
377 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800378 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700379 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
380 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
381 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800382 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700383 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
384 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
385 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800386 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700387 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
388 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
389 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800390 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
391 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
392 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
393 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
394 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700395 "src/f32-vbinary/gen/vsub-scalar-x1.c",
396 "src/f32-vbinary/gen/vsub-scalar-x2.c",
397 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800398 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700399 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
400 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
401 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800402 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
403 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
404 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
405 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
406 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700407 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
408 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
409 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800410 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Miao Wang55abe392021-02-03 14:54:41 -0800411 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
412 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
413 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
414 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
415 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
416 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
417 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
418 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
419 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
420 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
421 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
422 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700423 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
424 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
425 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700426 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
427 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
428 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700429 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
430 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
431 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800432 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
435 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
438 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
440 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700441 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
442 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
443 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
444 "src/f32-vunary/gen/vabs-scalar-x1.c",
445 "src/f32-vunary/gen/vabs-scalar-x2.c",
446 "src/f32-vunary/gen/vabs-scalar-x4.c",
447 "src/f32-vunary/gen/vneg-scalar-x1.c",
448 "src/f32-vunary/gen/vneg-scalar-x2.c",
449 "src/f32-vunary/gen/vneg-scalar-x4.c",
450 "src/f32-vunary/gen/vsqr-scalar-x1.c",
451 "src/f32-vunary/gen/vsqr-scalar-x2.c",
452 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800453 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
454 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
455 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
456 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
457 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
458 "src/math/expm1minus-scalar-rr2-p5.c",
459 "src/math/expm1minus-scalar-rr2-p6.c",
460 "src/math/expminus-scalar-rr2-lut64-p2.c",
461 "src/math/expminus-scalar-rr2-lut2048-p1.c",
462 "src/math/expminus-scalar-rr2-p5.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700463 "src/math/roundd-scalar-addsub.c",
464 "src/math/roundd-scalar-cvt.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700465 "src/math/roundd-scalar-floor.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800466 "src/math/roundne-scalar-addsub.c",
467 "src/math/roundne-scalar-nearbyint.c",
468 "src/math/roundne-scalar-rint.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700469 "src/math/roundu-scalar-addsub.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700470 "src/math/roundu-scalar-ceil.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700471 "src/math/roundu-scalar-cvt.c",
472 "src/math/roundz-scalar-addsub.c",
473 "src/math/roundz-scalar-cvt.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700474 "src/math/roundz-scalar-trunc.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800475 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
476 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
477 "src/math/sigmoid-scalar-rr2-p5-div.c",
478 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
479 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
480 "src/qs8-requantization/fp32-scalar-lrintf.c",
481 "src/qs8-requantization/fp32-scalar-magic.c",
482 "src/qs8-requantization/precise-scalar-signed64.c",
483 "src/qs8-requantization/precise-scalar-unsigned32.c",
484 "src/qs8-requantization/precise-scalar-unsigned64.c",
485 "src/qs8-requantization/q31-scalar.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700486 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
487 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
488 "src/qu8-dwconv/up1x9-minmax-scalar.c",
489 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
490 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
491 "src/qu8-gemm/2x2-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800492 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
493 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700494 "src/qu8-igemm/2x2-minmax-scalar.c",
495 "src/qu8-requantization/fp32-scalar-lrintf.c",
496 "src/qu8-requantization/fp32-scalar-magic.c",
497 "src/qu8-requantization/precise-scalar-signed64.c",
498 "src/qu8-requantization/precise-scalar-unsigned32.c",
499 "src/qu8-requantization/precise-scalar-unsigned64.c",
500 "src/qu8-requantization/q31-scalar.c",
501 "src/qu8-vadd/minmax-scalar.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700502 "src/u8-clamp/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800503 "src/u8-lut32norm/scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700504 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Miao Wange9993472020-02-10 15:00:10 -0800505 "src/u8-rmax/scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800506 "src/x8-lut/scalar.c",
507 "src/x8-zip/x2-scalar.c",
508 "src/x8-zip/x3-scalar.c",
509 "src/x8-zip/x4-scalar.c",
510 "src/x8-zip/xm-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800511 "src/x32-depthtospace2d-chw2hwc/scalar.c",
512 "src/x32-fill/scalar-float.c",
513 "src/x32-fill/scalar-int.c",
514 "src/x32-packx/x2-scalar.c",
515 "src/x32-packx/x3-scalar.c",
516 "src/x32-packx/x4-scalar.c",
517 "src/x32-pad/scalar-float.c",
518 "src/x32-pad/scalar-int.c",
519 "src/x32-unpool/scalar.c",
520 "src/x32-zip/x2-scalar.c",
521 "src/x32-zip/x3-scalar.c",
522 "src/x32-zip/x4-scalar.c",
523 "src/x32-zip/xm-scalar.c",
524 "src/xx-copy/memcpy.c",
Miao Wange9993472020-02-10 15:00:10 -0800525]
526
Miao Wange9993472020-02-10 15:00:10 -0800527// ISA-specific micro-kernels
528NEON_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -0700529 "src/f32-argmaxpool/4x-neon-c4.c",
530 "src/f32-argmaxpool/9p8x-neon-c4.c",
531 "src/f32-argmaxpool/9x-neon-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700532 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
533 "src/f32-avgpool/9x-minmax-neon-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700534 "src/f32-clamp/gen/neon-x4.c",
535 "src/f32-clamp/gen/neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700536 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700537 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800538 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700539 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
540 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700541 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800542 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700543 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Miao Wang55abe392021-02-03 14:54:41 -0800544 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700545 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800546 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700547 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800548 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700549 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800550 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
551 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
552 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
553 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
554 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700555 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800556 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
557 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
558 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
559 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
560 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
561 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
562 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
563 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
564 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
565 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
566 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
567 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
568 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
569 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
570 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
571 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
572 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
573 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
574 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
575 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
576 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
577 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
578 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
579 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
580 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
581 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
582 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
583 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
584 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
585 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
586 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
587 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
588 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
589 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
590 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
591 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
592 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
593 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
594 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
595 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
596 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
597 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700598 "src/f32-gavgpool-cw/neon-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700599 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
600 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700601 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800602 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
603 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700604 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800605 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
606 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
607 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
608 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
609 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700610 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
611 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800612 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
613 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -0700614 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
615 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800616 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
617 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
618 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
619 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
620 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
621 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
622 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
623 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
624 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
625 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
626 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
627 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
628 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
629 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
630 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
631 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700632 "src/f32-hswish/gen/hswish-neon-x4.c",
633 "src/f32-hswish/gen/hswish-neon-x8.c",
634 "src/f32-hswish/gen/hswish-neon-x16.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700635 "src/f32-ibilinear/gen/neon-c4.c",
636 "src/f32-ibilinear/gen/neon-c8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800637 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700638 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800639 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700640 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
641 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700642 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800643 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
644 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
645 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
646 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700647 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
648 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800649 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
650 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -0700651 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
652 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
653 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
654 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
655 "src/f32-pavgpool/9x-minmax-neon-c4.c",
656 "src/f32-ppmm/gen/4x8-minmax-neon.c",
657 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700658 "src/f32-prelu/gen/neon-1x4.c",
659 "src/f32-prelu/gen/neon-1x8.c",
660 "src/f32-prelu/gen/neon-1x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800661 "src/f32-prelu/gen/neon-2x4.c",
662 "src/f32-prelu/gen/neon-2x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700663 "src/f32-prelu/gen/neon-2x16.c",
664 "src/f32-prelu/gen/neon-4x4.c",
665 "src/f32-prelu/gen/neon-4x8.c",
666 "src/f32-prelu/gen/neon-4x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800667 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800668 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800669 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800670 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
671 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800672 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Miao Wange9993472020-02-10 15:00:10 -0800673 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
674 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800675 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800676 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
677 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800678 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
679 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
680 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
681 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
682 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
683 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
684 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
685 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
686 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
687 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
688 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
689 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
690 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700691 "src/f32-relu/gen/neon-x4.c",
692 "src/f32-relu/gen/neon-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800693 "src/f32-rmax/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800694 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
695 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
696 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
697 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
698 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
699 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
700 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
701 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
702 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
703 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
704 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
705 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800706 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
707 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
708 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
709 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
710 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
711 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -0800712 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
713 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
714 "src/f32-spmm/gen/4x1-minmax-neon.c",
715 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
716 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
717 "src/f32-spmm/gen/8x1-minmax-neon.c",
718 "src/f32-spmm/gen/12x1-minmax-neon.c",
719 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
720 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
721 "src/f32-spmm/gen/16x1-minmax-neon.c",
722 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
723 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
724 "src/f32-spmm/gen/32x1-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700725 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
726 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
727 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
728 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800729 "src/f32-vbinary/gen/vmax-neon-x4.c",
730 "src/f32-vbinary/gen/vmax-neon-x8.c",
731 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
732 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
733 "src/f32-vbinary/gen/vmin-neon-x4.c",
734 "src/f32-vbinary/gen/vmin-neon-x8.c",
735 "src/f32-vbinary/gen/vminc-neon-x4.c",
736 "src/f32-vbinary/gen/vminc-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700737 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
738 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
739 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
740 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
741 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
742 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700743 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
744 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
745 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
746 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700747 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
748 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
749 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
750 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Miao Wang55abe392021-02-03 14:54:41 -0800751 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
752 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
753 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
754 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
755 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
756 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
757 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
758 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
759 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
760 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
761 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
762 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700763 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
764 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700765 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
766 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700767 "src/f32-vrnd/gen/vrndd-neon-x4.c",
768 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800769 "src/f32-vrnd/gen/vrndne-neon-x4.c",
770 "src/f32-vrnd/gen/vrndne-neon-x8.c",
771 "src/f32-vrnd/gen/vrndu-neon-x4.c",
772 "src/f32-vrnd/gen/vrndu-neon-x8.c",
773 "src/f32-vrnd/gen/vrndz-neon-x4.c",
774 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700775 "src/f32-vunary/gen/vabs-neon-x4.c",
776 "src/f32-vunary/gen/vabs-neon-x8.c",
777 "src/f32-vunary/gen/vneg-neon-x4.c",
778 "src/f32-vunary/gen/vneg-neon-x8.c",
779 "src/f32-vunary/gen/vsqr-neon-x4.c",
780 "src/f32-vunary/gen/vsqr-neon-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800781 "src/math/expm1minus-neon-rr2-lut16-p3.c",
782 "src/math/expm1minus-neon-rr2-p6.c",
783 "src/math/roundd-neon-addsub.c",
784 "src/math/roundd-neon-cvt.c",
785 "src/math/roundne-neon-addsub.c",
786 "src/math/roundu-neon-addsub.c",
787 "src/math/roundu-neon-cvt.c",
788 "src/math/roundz-neon-addsub.c",
789 "src/math/roundz-neon-cvt.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800790 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
791 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
792 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
793 "src/math/sqrt-neon-nr1rsqrts.c",
794 "src/math/sqrt-neon-nr2rsqrts.c",
795 "src/math/sqrt-neon-nr3rsqrts.c",
796 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
797 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
798 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
799 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
800 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
801 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
802 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
803 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
804 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
805 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
806 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
807 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
808 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
809 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
810 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
811 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
812 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
813 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
814 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
815 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
816 "src/qs8-requantization/fp32-neon.c",
817 "src/qs8-requantization/precise-neon.c",
818 "src/qs8-requantization/q31-neon.c",
819 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
820 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
821 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
822 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
823 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
824 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
825 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
826 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700827 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
828 "src/qu8-avgpool/9x-minmax-neon-c8.c",
829 "src/qu8-dwconv/up8x9-minmax-neon.c",
830 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
831 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
832 "src/qu8-gemm/4x8-minmax-neon.c",
833 "src/qu8-gemm/8x8-minmax-neon.c",
834 "src/qu8-igemm/4x8-minmax-neon.c",
835 "src/qu8-igemm/8x8-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700836 "src/qu8-requantization/fp32-neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800837 "src/qu8-requantization/precise-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700838 "src/qu8-requantization/q31-neon.c",
839 "src/qu8-vadd/minmax-neon.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700840 "src/u8-clamp/neon-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700841 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Miao Wange9993472020-02-10 15:00:10 -0800842 "src/u8-rmax/neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800843 "src/x8-zip/x2-neon.c",
844 "src/x8-zip/x3-neon.c",
845 "src/x8-zip/x4-neon.c",
846 "src/x8-zip/xm-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700847 "src/x32-fill/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800848 "src/x32-packx/x4-neon-st4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700849 "src/x32-pad/neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700850 "src/x32-unpool/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800851 "src/x32-zip/x2-neon.c",
852 "src/x32-zip/x3-neon.c",
853 "src/x32-zip/x4-neon.c",
854 "src/x32-zip/xm-neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800855]
856
857NEONFMA_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -0800858 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
859 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
860 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
861 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
862 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
863 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
864 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
865 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
866 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
867 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
868 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
869 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
870 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
871 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
872 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
873 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
874 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
875 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
876 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
877 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
878 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
879 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
880 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
881 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
882 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
883 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
884 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
885 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
886 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
887 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700888 "src/f32-ibilinear/gen/neonfma-c4.c",
889 "src/f32-ibilinear/gen/neonfma-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700890 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800891 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -0700892 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800893 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
894 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -0700895 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
896 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -0700897 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
898 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -0700899 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
900 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Miao Wange9993472020-02-10 15:00:10 -0800901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Miao Wange9993472020-02-10 15:00:10 -0800907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
908 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800909 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800910 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
911 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800912 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
920 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
921 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
922 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
923 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
924 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Miao Wange9993472020-02-10 15:00:10 -0800925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Miao Wange9993472020-02-10 15:00:10 -0800937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
938 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
939 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
940 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
941 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
942 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Miao Wange9993472020-02-10 15:00:10 -0800943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Miao Wange9993472020-02-10 15:00:10 -0800955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
956 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
957 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
958 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
959 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
960 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
974 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
975 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
976 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
977 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
978 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -0800979 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
980 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
981 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
982 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
983 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
984 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
985 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
986 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
987 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
988 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
989 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
990 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
991 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
993 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
994 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
995 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
996 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
997 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
999 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1000 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1001 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1002 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1003 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -07001004 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1005 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1011 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1012 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1013 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1014 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1015 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1021 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1022 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1023 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1024 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1025 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001026 "src/math/exp-neonfma-rr2-lut64-p2.c",
1027 "src/math/exp-neonfma-rr2-p5.c",
Miao Wang55abe392021-02-03 14:54:41 -08001028 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1029 "src/math/expm1minus-neonfma-rr1-p6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001030 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1031 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1032 "src/math/expminus-neonfma-rr2-p5.c",
Miao Wange9993472020-02-10 15:00:10 -08001033 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1034 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1035 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001036 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1037 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1038 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001039 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1040 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1041 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001042 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1043 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1044 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001045 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1046 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1047 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001048 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1049 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1050 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001051 "src/math/sqrt-neonfma-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001052 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001053 "src/math/sqrt-neonfma-nr2fma.c",
1054 "src/math/sqrt-neonfma-nr2fma1adj.c",
1055 "src/math/sqrt-neonfma-nr3fma.c",
Miao Wange9993472020-02-10 15:00:10 -08001056]
1057
1058AARCH64_NEONFMA_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -08001059 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1060 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1061 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
1062 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
1063 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1064 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1065 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
1066 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
1067 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
1071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
1072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
1073 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1074 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1075 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1076 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1077 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
1078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
1081 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
1082 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
1083 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1084 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1085 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
1086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
1090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
1091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
1093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
1094 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
1095 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
1096 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
1097 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1098 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
1099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1104 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1105 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1106 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
1107 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
1108 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Miao Wang400e4042020-04-17 10:15:59 -07001109 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001110 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001111 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001112 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1113 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1114 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001115 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1116 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1117 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1118 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1119 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1120 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1121 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001122 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1123 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1124 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001125 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001126 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001127 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1128 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Miao Wange9993472020-02-10 15:00:10 -08001129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1130 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1131 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1132 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1133 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1134 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1136 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1137 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1138 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1139 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1140 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1142 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1143 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1144 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1145 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1146 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -07001147 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1148 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -07001149 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1150 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001151 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1152 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001153 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1154 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001155 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1156 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
1157 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1158 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1159 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1160 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1161 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1162 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001163 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1164 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001165 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001166 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001167 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001168 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001169 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001170 "src/math/sigmoid-neonfma-rr2-p5-div.c",
1171]
1172
Miao Wangc0aa11a2020-06-10 13:41:26 -07001173NEONV8_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -07001174 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1175 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001176 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1177 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1178 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1179 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1180 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1181 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001182 "src/math/roundd-neonv8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001183 "src/math/roundne-neonv8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001184 "src/math/roundu-neonv8.c",
1185 "src/math/roundz-neonv8.c",
1186]
1187
Miao Wange9993472020-02-10 15:00:10 -08001188AARCH64_NEONFP16ARITH_UKERNELS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -07001189 "src/f16-clamp/gen/neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001190 "src/f16-clamp/gen/neonfp16arith-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001191 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1192 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1193 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1194 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001195 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1196 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1197 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1198 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1199 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1200 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1201 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1202 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001203 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1204 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001205 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1206 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1207 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1208 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1209 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1210 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1211 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1212 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1213 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1214 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1215 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1216 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1217 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1218 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1219 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1220 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001221 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001222 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1223 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1224 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1225 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1226 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1227 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1228 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1229 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1230 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001231 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001232 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001233 "src/f16-relu/gen/neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001234 "src/f16-relu/gen/neonfp16arith-x16.c",
1235 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001236 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001237 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001238 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001239 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001240 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001241 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001242 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001243 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1244 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1245 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1246 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1247 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1248 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1249 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1250 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1251 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1252 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1253 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1254 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1255 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1256 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1257 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1258 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1259 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1260 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1261 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1262 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1263 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1264 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1265 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1266 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1267 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1268 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1269 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1270 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1271 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1272 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1273]
1274
1275NEONDOT_UKERNELS = [
1276 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
1277 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
1278 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
1279 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
1280 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
1281 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
1282 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1283 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
1284 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1285 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
1286 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
1287 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
1288 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
1289 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
1290 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
1291 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1292 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
1293 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Miao Wange9993472020-02-10 15:00:10 -08001294]
1295
1296SSE_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001297 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1298 "src/f32-avgpool/9x-minmax-sse-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001299 "src/f32-clamp/gen/sse-x4.c",
1300 "src/f32-clamp/gen/sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001301 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1302 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001303 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1304 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1305 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1306 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001307 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1308 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001309 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1310 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1311 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1312 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001313 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1314 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1318 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1319 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
1320 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1321 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1322 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1323 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1324 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1326 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1327 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
1328 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
1329 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
1330 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1331 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1332 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
1333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
1337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
1338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
1339 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
1340 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
1341 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
1342 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
1343 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
1344 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
1345 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
1346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
1347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
1348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
1349 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
1350 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
1351 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
1352 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
1353 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
1354 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
1355 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001356 "src/f32-gavgpool-cw/sse-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -07001357 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1358 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001359 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1360 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1361 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001362 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
1363 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
1364 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001365 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1366 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1367 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001368 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
1369 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
1370 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001371 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1372 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1373 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001374 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
1375 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
1376 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001377 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1378 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1379 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1380 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001381 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
1382 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
1383 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001384 "src/f32-hswish/gen/hswish-sse-x4.c",
1385 "src/f32-hswish/gen/hswish-sse-x8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001386 "src/f32-ibilinear/gen/sse-c4.c",
1387 "src/f32-ibilinear/gen/sse-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001388 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1389 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1390 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001391 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
1392 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
1393 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001394 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1395 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1396 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1397 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001398 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
1399 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
1400 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001401 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1402 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1403 "src/f32-pavgpool/9x-minmax-sse-c4.c",
1404 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001405 "src/f32-prelu/gen/sse-2x4.c",
1406 "src/f32-prelu/gen/sse-2x8.c",
1407 "src/f32-relu/gen/sse-x4.c",
1408 "src/f32-relu/gen/sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001409 "src/f32-rmax/sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001410 "src/f32-spmm/gen/4x1-minmax-sse.c",
1411 "src/f32-spmm/gen/8x1-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001412 "src/f32-spmm/gen/16x1-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001413 "src/f32-spmm/gen/32x1-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001414 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1415 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1416 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1417 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1418 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1419 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1420 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1421 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001422 "src/f32-vbinary/gen/vmax-sse-x4.c",
1423 "src/f32-vbinary/gen/vmax-sse-x8.c",
1424 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1425 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1426 "src/f32-vbinary/gen/vmin-sse-x4.c",
1427 "src/f32-vbinary/gen/vmin-sse-x8.c",
1428 "src/f32-vbinary/gen/vminc-sse-x4.c",
1429 "src/f32-vbinary/gen/vminc-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001430 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1431 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1432 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1433 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1434 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1435 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1436 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1437 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001438 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1439 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1440 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1441 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001442 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1443 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1444 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1445 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001446 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1447 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001448 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1449 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001450 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1451 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
1452 "src/f32-vunary/gen/vabs-sse-x4.c",
1453 "src/f32-vunary/gen/vabs-sse-x8.c",
1454 "src/f32-vunary/gen/vneg-sse-x4.c",
1455 "src/f32-vunary/gen/vneg-sse-x8.c",
1456 "src/f32-vunary/gen/vsqr-sse-x4.c",
1457 "src/f32-vunary/gen/vsqr-sse-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001458 "src/math/roundd-sse-addsub.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001459 "src/math/roundne-sse-addsub.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001460 "src/math/roundu-sse-addsub.c",
1461 "src/math/roundz-sse-addsub.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001462 "src/math/sqrt-sse-hh1mac.c",
1463 "src/math/sqrt-sse-nr1mac.c",
1464 "src/math/sqrt-sse-nr2mac.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001465 "src/x32-fill/sse.c",
1466 "src/x32-packx/x4-sse.c",
1467 "src/x32-pad/sse.c",
Miao Wange9993472020-02-10 15:00:10 -08001468]
1469
1470SSE2_UKERNELS = [
Miao Wange9993472020-02-10 15:00:10 -08001471 "src/f32-argmaxpool/4x-sse2-c4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001472 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Miao Wange9993472020-02-10 15:00:10 -08001473 "src/f32-argmaxpool/9x-sse2-c4.c",
Miao Wang55abe392021-02-03 14:54:41 -08001474 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
1475 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
1476 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
1477 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
1478 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
1479 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
1480 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
1481 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
1482 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
1483 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
1484 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
1485 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Miao Wange9993472020-02-10 15:00:10 -08001486 "src/f32-prelu/gen/sse2-2x4.c",
1487 "src/f32-prelu/gen/sse2-2x8.c",
1488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Miao Wange9993472020-02-10 15:00:10 -08001489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001493 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Miao Wange9993472020-02-10 15:00:10 -08001494 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1495 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001496 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001497 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1498 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001499 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1500 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1501 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1502 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1503 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1504 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1505 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001506 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
1507 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
1508 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
1509 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
1510 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1511 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -08001512 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
1513 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
1514 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
1515 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
1516 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
1517 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
1518 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
1519 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
1520 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
1521 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
1522 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
1523 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001524 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1525 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001526 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1527 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001528 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1529 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1530 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1531 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1532 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1533 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1534 "src/math/exp-sse2-rr2-lut64-p2.c",
1535 "src/math/exp-sse2-rr2-p5.c",
1536 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
1537 "src/math/expm1minus-sse2-rr2-p6.c",
1538 "src/math/expminus-sse2-rr2-p5.c",
1539 "src/math/roundd-sse2-cvt.c",
1540 "src/math/roundne-sse2-cvt.c",
1541 "src/math/roundu-sse2-cvt.c",
1542 "src/math/roundz-sse2-cvt.c",
1543 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1544 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1545 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1546 "src/math/sigmoid-sse2-rr2-p5-div.c",
1547 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1548 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
1549 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1550 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1551 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
1552 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1553 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1554 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
1555 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1556 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1557 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
1558 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1559 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1560 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1561 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1562 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1563 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1564 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
1565 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
1566 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1567 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
1568 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
1569 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
1570 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
1571 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
1572 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
1573 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1574 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1575 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1576 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1577 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
1578 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
1579 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
1580 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
1581 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
1582 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
1583 "src/qs8-requantization/fp32-sse2.c",
1584 "src/qs8-requantization/precise-sse2.c",
1585 "src/qs8-requantization/q31-sse2.c",
1586 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1587 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1588 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1589 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
1590 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1591 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1592 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1593 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001594 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1595 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001596 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1597 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1598 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1599 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1600 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001601 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001602 "src/qu8-requantization/fp32-sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001603 "src/qu8-requantization/precise-sse2.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001604 "src/qu8-requantization/q31-sse2.c",
1605 "src/qu8-vadd/minmax-sse2.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001606 "src/u8-clamp/sse2-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001607 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Miao Wange9993472020-02-10 15:00:10 -08001608 "src/u8-rmax/sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001609 "src/x8-zip/x2-sse2.c",
1610 "src/x8-zip/x3-sse2.c",
1611 "src/x8-zip/x4-sse2.c",
1612 "src/x8-zip/xm-sse2.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001613 "src/x32-unpool/sse2.c",
Miao Wange9993472020-02-10 15:00:10 -08001614 "src/x32-zip/x2-sse2.c",
1615 "src/x32-zip/x3-sse2.c",
1616 "src/x32-zip/x4-sse2.c",
1617 "src/x32-zip/xm-sse2.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001618]
1619
1620SSSE3_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -08001621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
1622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
1623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
1624 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
1625 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
1626 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
1627 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
1628 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
1629 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
1630 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
1631 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1632 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1633 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
1634 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1635 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1636 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
1637 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1638 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1639 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
1640 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1641 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1642 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1643 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1644 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1645 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1646 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
1647 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
1648 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1649 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
1650 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
1651 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
1652 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
1653 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
1654 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
1655 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1656 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1657 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1658 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1659 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
1660 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
1661 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
1662 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
1663 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
1664 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
1665 "src/qs8-requantization/precise-ssse3.c",
1666 "src/qs8-requantization/q31-ssse3.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001667 "src/qu8-requantization/precise-ssse3.c",
1668 "src/qu8-requantization/q31-ssse3.c",
Miao Wange9993472020-02-10 15:00:10 -08001669]
1670
1671SSE41_UKERNELS = [
1672 "src/f32-prelu/gen/sse41-2x4.c",
1673 "src/f32-prelu/gen/sse41-2x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001674 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
1675 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
1676 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
1677 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
1678 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
1679 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001680 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
1681 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
1682 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
1683 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
1684 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1685 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -08001686 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
1687 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
1688 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
1689 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
1690 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
1691 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
1692 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
1693 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
1694 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
1695 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
1696 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
1697 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001698 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1699 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001700 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1701 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001702 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1703 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1704 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1705 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1706 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1707 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001708 "src/math/roundd-sse41.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001709 "src/math/roundne-sse41.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001710 "src/math/roundu-sse41.c",
1711 "src/math/roundz-sse41.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001712 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1713 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1714 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
1715 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1716 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1717 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
1718 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1719 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1720 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
1721 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1722 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1723 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1724 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1725 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1726 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1727 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
1728 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
1729 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1730 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
1731 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
1732 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
1733 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
1734 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
1735 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
1736 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1737 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1738 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1739 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1740 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
1741 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
1742 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
1743 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
1744 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
1745 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
1746 "src/qs8-requantization/fp32-sse4.c",
1747 "src/qs8-requantization/precise-sse4.c",
1748 "src/qs8-requantization/q31-sse4.c",
1749 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1750 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1751 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1752 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
1753 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
1754 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
1755 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
1756 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
1757 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
1758 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
1759 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
1760 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
1761 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
1762 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
1763 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
1764 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
1765 "src/qu8-requantization/precise-sse4.c",
1766 "src/qu8-requantization/q31-sse4.c",
Miao Wange9993472020-02-10 15:00:10 -08001767]
1768
1769AVX_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -07001770 "src/f32-clamp/gen/avx-x8.c",
1771 "src/f32-clamp/gen/avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001772 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1773 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001774 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1775 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001776 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1777 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001778 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1779 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1780 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1781 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1782 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1783 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001784 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001785 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1786 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001787 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001788 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001789 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001790 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001791 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1792 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1793 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1794 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1795 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1796 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1797 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1798 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1799 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1800 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1801 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001802 "src/f32-hswish/gen/hswish-avx-x8.c",
1803 "src/f32-hswish/gen/hswish-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001804 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001805 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1806 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001807 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001808 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001809 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001810 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001811 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1812 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001813 "src/f32-prelu/gen/avx-2x8.c",
1814 "src/f32-prelu/gen/avx-2x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001815 "src/f32-relu/gen/avx-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001816 "src/f32-relu/gen/avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001817 "src/f32-rmax/avx.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001818 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
1819 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
1820 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
1821 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
1822 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
1823 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
1824 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
1825 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
1826 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
1827 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
1828 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
1829 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
1830 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
1831 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
1832 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
1833 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
1834 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
1835 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
1836 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
1837 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Miao Wang400e4042020-04-17 10:15:59 -07001838 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1839 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1840 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1841 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1842 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1843 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1844 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1845 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001846 "src/f32-vbinary/gen/vmax-avx-x8.c",
1847 "src/f32-vbinary/gen/vmax-avx-x16.c",
1848 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1849 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1850 "src/f32-vbinary/gen/vmin-avx-x8.c",
1851 "src/f32-vbinary/gen/vmin-avx-x16.c",
1852 "src/f32-vbinary/gen/vminc-avx-x8.c",
1853 "src/f32-vbinary/gen/vminc-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001854 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1855 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1856 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1857 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1858 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1859 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1860 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1861 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001862 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1863 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1864 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1865 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001866 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1867 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1868 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1869 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Miao Wang55abe392021-02-03 14:54:41 -08001870 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
1871 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
1872 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
1873 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
1874 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
1875 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
1876 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
1877 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
1878 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
1879 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
1880 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
1881 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
1882 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
1883 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
1884 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
1885 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
1886 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
1887 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001888 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1889 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001890 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1891 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001892 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1893 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1894 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1895 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1896 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1897 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1898 "src/f32-vscale/avx-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001899 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1900 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
1901 "src/f32-vunary/gen/vabs-avx-x8.c",
1902 "src/f32-vunary/gen/vabs-avx-x16.c",
1903 "src/f32-vunary/gen/vneg-avx-x8.c",
1904 "src/f32-vunary/gen/vneg-avx-x16.c",
1905 "src/f32-vunary/gen/vsqr-avx-x8.c",
1906 "src/f32-vunary/gen/vsqr-avx-x16.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001907 "src/math/exp-avx-rr2-p5.c",
1908 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
1909 "src/math/expm1minus-avx-rr2-lut16-p3.c",
1910 "src/math/expm1minus-avx-rr2-p6.c",
1911 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
1912 "src/math/sigmoid-avx-rr2-p5-div.c",
1913 "src/math/sigmoid-avx-rr2-p5-nr1.c",
1914 "src/math/sigmoid-avx-rr2-p5-nr2.c",
1915]
1916
1917XOP_UKERNELS = [
1918 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1919 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1920 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1921 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1922 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1923 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1924 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
1925 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
1926 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1927 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
1928 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
1929 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1931 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
1932 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
1933 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1934 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1935 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1936 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1937 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
1938 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
1939 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
1940 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
1941 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1942 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1943 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
1944 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
1945 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
1946 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
1947 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
1948 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
1949 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
1950 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08001951]
1952
1953FMA3_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001954 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1955 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07001956 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1957 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07001958 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1959 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001960 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1961 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1962 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1963 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1964 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1965 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07001966 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001967 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1968 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1969 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1970 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001971 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001972 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1973 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001974 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001975 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1976 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001977 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1978 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1979 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001980 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1981 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1982 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1983 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1984 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1985 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1986 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1987 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1988 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1989 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1990 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1991 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1992 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1993 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001994 "src/f32-hswish/gen/hswish-fma3-x8.c",
1995 "src/f32-hswish/gen/hswish-fma3-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001996 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001997 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1998 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1999 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2000 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002001 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002002 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2003 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002004 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002005 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2006 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002007 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2008 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2009 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002010 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2011 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2012 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2013 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2014 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2015 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2016 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2017 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
2018 "src/math/sqrt-fma3-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002019 "src/math/sqrt-fma3-nr1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002020 "src/math/sqrt-fma3-nr2fma.c",
Miao Wange9993472020-02-10 15:00:10 -08002021]
2022
2023AVX2_UKERNELS = [
Miao Wange9993472020-02-10 15:00:10 -08002024 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2025 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002026 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002027 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002028 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002029 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2030 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002031 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002032 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2033 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2034 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002035 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002036 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2037 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002038 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002039 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002040 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002041 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2042 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002043 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002044 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2045 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2046 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002047 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002048 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2049 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002050 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002051 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002052 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002053 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2054 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002055 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002056 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2057 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002059 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002060 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2061 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2062 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2063 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2064 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2065 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2066 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2067 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2068 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2069 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2070 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2071 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2072 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2073 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2074 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2083 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2084 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2085 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2086 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2087 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2088 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2089 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Miao Wang55abe392021-02-03 14:54:41 -08002090 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2091 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2092 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2093 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2094 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2095 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2096 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2097 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2098 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2099 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2100 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2101 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2102 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2103 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2104 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2105 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2106 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2107 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2108 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2109 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2110 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2111 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2112 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2113 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2114 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2115 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2116 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2117 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2118 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2119 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2120 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2121 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2122 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2123 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2124 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2125 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2126 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2127 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2128 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2129 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002130 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2131 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2132 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2133 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2134 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2135 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2136 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2137 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2138 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2139 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2140 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2141 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2142 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2143 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2144 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2145 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2146 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2147 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2148 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2149 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2150 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2151 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2152 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2153 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002154 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2155 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2156 "src/math/exp-avx2-rr2-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002157 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2158 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2159 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2160 "src/math/expm1minus-avx2-rr1-p6.c",
Miao Wang55abe392021-02-03 14:54:41 -08002161 "src/math/expminus-avx2-rr2-p5.c",
Miao Wange9993472020-02-10 15:00:10 -08002162 "src/math/extexp-avx2-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002163 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2164 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2165 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2166 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
Miao Wange9993472020-02-10 15:00:10 -08002167 "src/math/sigmoid-avx2-rr1-p5-div.c",
Miao Wange9993472020-02-10 15:00:10 -08002168 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002169 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2170 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2171 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2172 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2173 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2174 "src/math/sigmoid-avx2-rr2-p5-div.c",
2175 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2176 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2177 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2178 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2179 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2180 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2181 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
2182 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
2183 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2184 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2185 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2186 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2187 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
2188 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
2189 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2190 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2191 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
2192 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2193 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2194 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2195 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2196 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2197 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2198 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2199 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002200]
2201
2202AVX512F_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -07002203 "src/f32-clamp/gen/avx512f-x16.c",
2204 "src/f32-clamp/gen/avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002205 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2206 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002207 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2208 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002209 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2210 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002211 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2212 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2213 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2214 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2215 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2216 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002217 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2218 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2219 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2220 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2221 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2222 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002223 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2224 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2225 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2226 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2227 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2228 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002229 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2230 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002231 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2232 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2233 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2234 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2235 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2236 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07002237 "src/f32-prelu/gen/avx512f-2x16.c",
2238 "src/f32-prelu/gen/avx512f-2x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002239 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2240 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002241 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002242 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002243 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002244 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2245 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002246 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002247 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2248 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2249 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002250 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Miao Wange9993472020-02-10 15:00:10 -08002251 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2252 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002253 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002254 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002255 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002256 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2257 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002258 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002259 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2260 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2261 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002262 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Miao Wange9993472020-02-10 15:00:10 -08002263 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2264 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002265 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002266 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002267 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002268 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2269 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002270 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002271 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2272 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2273 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002274 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002275 "src/f32-relu/gen/avx512f-x16.c",
2276 "src/f32-relu/gen/avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002277 "src/f32-rmax/avx512f.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002278 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2279 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2280 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2281 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2282 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2283 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2284 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2285 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2286 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2287 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2288 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2289 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2290 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2291 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2292 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2293 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2294 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2295 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2296 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2297 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2298 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2299 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2300 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2301 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2302 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2303 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2304 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2305 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2306 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2307 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2308 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2309 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2310 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2311 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2312 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2313 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2314 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2315 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2316 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2317 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2318 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2319 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2320 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2321 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2322 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2323 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2324 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2325 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Miao Wang400e4042020-04-17 10:15:59 -07002326 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2327 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2328 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2329 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2330 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2331 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2332 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2333 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002334 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2335 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2336 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2337 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2338 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2339 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2340 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2341 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002342 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2343 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2344 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2345 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2346 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2347 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2348 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2349 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002350 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2351 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2352 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2353 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002354 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2355 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2356 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2357 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Miao Wang55abe392021-02-03 14:54:41 -08002358 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
2359 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
2360 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
2361 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
2362 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
2363 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
2364 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
2365 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
2366 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
2367 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
2368 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
2369 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
2370 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
2371 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
2372 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
2373 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002374 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2375 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002376 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2377 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2378 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2379 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2380 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2381 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2382 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2383 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2384 "src/f32-vscale/avx512f-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002385 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2386 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2387 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2388 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2389 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2390 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2391 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2392 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2393 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2394 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2395 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2396 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2397 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2398 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2399 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2400 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2401 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2402 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2403 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2404 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2405 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2406 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2407 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2408 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002409 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2410 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2411 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2412 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2413 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2414 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2415 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2416 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
2417 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2418 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2419 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2420 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2421 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2422 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002423 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2424 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2425 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2426 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2427 "src/math/exp-avx512f-rr2-p5-scalef.c",
2428 "src/math/exp-avx512f-rr2-p5.c",
2429 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
2430 "src/math/expm1minus-avx512f-rr1-p6.c",
Miao Wange9993472020-02-10 15:00:10 -08002431 "src/math/extexp-avx512f-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002432 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2433 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2434 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2435 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2436 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2437 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2438 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
2439 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
2440 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2441 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2442 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2443 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2444 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2445 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2446 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2447 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2448 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2449 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2450 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
2451 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
2452 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2453 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2454 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2455 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002456 "src/math/sqrt-avx512f-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002457 "src/math/sqrt-avx512f-nr1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002458 "src/math/sqrt-avx512f-nr2fma.c",
2459]
2460
2461AVX512SKX_UKERNELS = [
2462 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2463 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
2464 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2465 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2466 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2467 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2468 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2469 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2470 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2471 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
Miao Wange9993472020-02-10 15:00:10 -08002472]
2473
2474AARCH32_ASM_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -07002475 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
2476 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002477 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2478 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002479 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Miao Wang400e4042020-04-17 10:15:59 -07002480 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002481 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002482 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002483 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2484 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002485 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2486 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2487 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2488 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Miao Wange9993472020-02-10 15:00:10 -08002489]
2490
2491AARCH64_ASM_UKERNELS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -07002492 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002493 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002494 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002495 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002496 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002497 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002498 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002499 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2500 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2501 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2502 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2503 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2504 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2505 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002506 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2507 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Miao Wang400e4042020-04-17 10:15:59 -07002508 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2510 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002511 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2512 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002513 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002517 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002518 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2519 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002520 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2521 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Miao Wang400e4042020-04-17 10:15:59 -07002524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Miao Wang400e4042020-04-17 10:15:59 -07002526 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002527 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002528 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2529 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2530 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2531 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2532 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2533 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2534 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2535 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2536 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2537 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2538 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2539 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2540 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2541 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2542 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2543 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2544 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2545 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2546 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2547 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2548 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2549 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Miao Wang400e4042020-04-17 10:15:59 -07002550 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002551 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002552 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2553 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002554 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2555 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2557 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2558 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2559 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002560 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2561 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2562 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2563 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002564 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2565 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002566 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
2567 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
2568 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Miao Wange9993472020-02-10 15:00:10 -08002569]
2570
2571cc_defaults {
2572 name: "xnnpack_internal_default",
2573 vendor_available: true,
2574 sdk_version: "current",
2575 local_include_dirs: [
2576 "include",
2577 "src",
2578 ],
2579 cflags: [
2580 "-std=c99",
2581 "-DXNN_LOG_LEVEL=2",
2582 "-Wno-unused-parameter",
2583 "-Wno-missing-field-initializers",
2584 "-Wno-pointer-arith",
Miao Wangef75bc92020-07-25 09:15:12 -07002585 "-Wno-implicit-function-declaration",
Miao Wange9993472020-02-10 15:00:10 -08002586 ],
2587 stl: "libc++_static",
2588}
2589
2590cc_library_static {
2591 name: "xnnpack_tables",
2592 defaults: ["xnnpack_internal_default"],
2593 srcs: TABLE_SRCS,
2594}
2595
2596cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07002597 name: "xnnpack_logging_utils",
2598 defaults: ["xnnpack_internal_default"],
2599 srcs: LOGGING_SRCS,
2600 header_libs: [
2601 "fp16_headers",
2602 ],
2603 static_libs: [
2604 "libclog",
2605 "libpthreadpool",
2606 ],
2607}
2608
2609cc_library_static {
2610 name: "xnnpack_memory_planner",
2611 defaults: ["xnnpack_internal_default"],
2612 srcs: [
2613 "src/memory-planner.c",
2614 ],
2615 cflags: [
2616 "-DXNN_ENABLE_MEMOPT=1",
2617 ],
2618 header_libs: [
2619 "fp16_headers",
2620 ],
2621 static_libs: [
2622 "libpthreadpool",
2623 "xnnpack_logging_utils",
2624 ],
2625}
2626
2627cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002628 name: "xnnpack_im2col",
2629 defaults: ["xnnpack_internal_default"],
2630 srcs: [
2631 "src/im2col.c",
2632 ],
2633}
2634
2635cc_library_static {
2636 name: "xnnpack_indirection",
2637 defaults: ["xnnpack_internal_default"],
2638 srcs: [
2639 "src/indirection.c",
2640 ],
2641 header_libs: [
2642 "fp16_headers",
2643 "fxdiv_headers",
2644 ],
2645 static_libs: [
2646 "libpthreadpool",
2647 ],
2648}
2649
2650cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07002651 name: "xnnpack_packing",
2652 defaults: ["xnnpack_internal_default"],
2653 srcs: [
2654 "src/packing.c",
2655 ],
2656 header_libs: [
2657 "fp16_headers",
2658 "fxdiv_headers",
2659 ],
2660 static_libs: [
2661 "libpthreadpool",
2662 ],
2663}
2664
2665cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002666 name: "xnnpack_operator_run",
2667 defaults: ["xnnpack_internal_default"],
2668 srcs: [
2669 "src/operator-run.c",
2670 ],
2671 cflags: [
2672 "-Wno-vla",
2673 ],
2674 header_libs: [
2675 "fp16_headers",
2676 "fxdiv_headers",
2677 ],
2678 static_libs: [
2679 "libclog",
2680 "libpthreadpool",
2681 ],
2682}
2683
2684cc_library_static {
2685 name: "xnnpack_operators",
2686 defaults: ["xnnpack_internal_default"],
2687 srcs: OPERATOR_SRCS + [
2688 "src/memory.c",
2689 "src/operator-delete.c",
2690 ],
2691 header_libs: [
2692 "fp16_headers",
2693 "fxdiv_headers",
2694 ],
2695 static_libs: [
2696 "libclog",
2697 "libpthreadpool",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002698 "xnnpack_logging_utils",
2699 "xnnpack_packing",
Miao Wange9993472020-02-10 15:00:10 -08002700 ],
2701 whole_static_libs: [
2702 "xnnpack_indirection",
2703 ],
2704}
2705
2706cc_library_static {
2707 name: "xnnpack_scalar_ukernels",
2708 defaults: ["xnnpack_internal_default"],
2709 srcs: SCALAR_UKERNELS,
2710 header_libs: [
2711 "fp16_headers",
2712 "fxdiv_headers",
2713 ],
2714 static_libs: [
2715 "libpthreadpool",
2716 "xnnpack_tables",
2717 ],
2718}
2719
2720cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002721 name: "xnnpack_neon_ukernels",
2722 defaults: ["xnnpack_internal_default"],
2723 arch: {
2724 arm: {
2725 srcs: NEON_UKERNELS,
2726 cflags: [
2727 "-marm",
2728 "-mfpu=neon",
2729 ],
2730 },
2731 arm64: {
2732 srcs: NEON_UKERNELS,
2733 },
2734 x86: { enabled: false, },
2735 x86_64: { enabled: false, },
2736 },
2737 header_libs: [
2738 "fp16_headers",
2739 ],
2740 static_libs: [
2741 "libpthreadpool",
2742 "xnnpack_tables",
2743 ],
2744}
2745
2746cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08002747 name: "xnnpack_neondot_ukernels",
2748 defaults: ["xnnpack_internal_default"],
2749 arch: {
2750 arm: {
2751 srcs: NEONDOT_UKERNELS,
2752 cflags: [
2753 "-marm",
2754 "-march=armv8.2-a+dotprod",
2755 "-mfpu=neon-fp-armv8",
2756 ],
2757 },
2758 arm64: {
2759 srcs: NEONDOT_UKERNELS,
2760 cflags: [
2761 "-march=armv8.2-a+dotprod",
2762 ],
2763 },
2764 x86: { enabled: false, },
2765 x86_64: { enabled: false, },
2766 },
2767 header_libs: [
2768 "fp16_headers",
2769 ],
2770 static_libs: [
2771 "libpthreadpool",
2772 "xnnpack_tables",
2773 ],
2774}
2775
2776cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002777 name: "xnnpack_neonfma_ukernels",
2778 defaults: ["xnnpack_internal_default"],
2779 arch: {
2780 arm: {
2781 srcs: NEONFMA_UKERNELS,
2782 cflags: [
2783 "-marm",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002784 "-march=armv7-a",
Miao Wange9993472020-02-10 15:00:10 -08002785 "-mfpu=neon-vfpv4",
2786 ],
2787 },
2788 arm64: {
2789 srcs: NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
2790 },
2791 x86: { enabled: false, },
2792 x86_64: { enabled: false, },
2793 },
2794 header_libs: [
2795 "fp16_headers",
2796 ],
2797 static_libs: [
2798 "libpthreadpool",
2799 "xnnpack_tables",
2800 ],
2801}
2802
2803cc_library_static {
Miao Wangc0aa11a2020-06-10 13:41:26 -07002804 name: "xnnpack_neonv8_ukernels",
2805 defaults: ["xnnpack_internal_default"],
2806 arch: {
2807 arm: {
2808 srcs: NEONV8_UKERNELS,
2809 cflags: [
2810 "-marm",
2811 "-march=armv8-a",
2812 "-mfpu=neon-fp-armv8",
2813 ],
2814 },
2815 arm64: {
2816 srcs: NEONV8_UKERNELS,
2817 },
2818 x86: { enabled: false, },
2819 x86_64: { enabled: false, },
2820 },
2821 header_libs: [
2822 "fp16_headers",
2823 ],
2824 static_libs: [
2825 "libpthreadpool",
2826 "xnnpack_tables",
2827 ],
2828}
2829
2830cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002831 name: "xnnpack_neonfp16arith_ukernels",
2832 defaults: ["xnnpack_internal_default"],
2833 arch: {
2834 arm: { enabled: false, },
2835 arm64: {
2836 srcs: AARCH64_NEONFP16ARITH_UKERNELS,
2837 cflags: [
2838 "-march=armv8.2-a+fp16",
2839 ],
2840 },
2841 x86: { enabled: false, },
2842 x86_64: { enabled: false, },
2843 },
2844 header_libs: [
2845 "fp16_headers",
2846 ],
2847 static_libs: [
2848 "libpthreadpool",
2849 "xnnpack_tables",
2850 ],
2851}
2852
2853cc_library_static {
2854 name: "xnnpack_asm_ukernels",
2855 defaults: ["xnnpack_internal_default"],
2856 arch: {
2857 arm: {
2858 srcs: AARCH32_ASM_UKERNELS,
2859 },
2860 arm64: {
2861 srcs: AARCH64_ASM_UKERNELS,
Miao Wang400e4042020-04-17 10:15:59 -07002862 clang_asflags: [
Miao Wang5eea8312020-12-07 09:12:40 -08002863 "-march=armv8.2-a+fp16+dotprod",
Miao Wang400e4042020-04-17 10:15:59 -07002864 ],
Miao Wange9993472020-02-10 15:00:10 -08002865 },
2866 x86: { enabled: false, },
2867 x86_64: { enabled: false, },
2868 },
2869}
2870
2871cc_library_static {
2872 name: "xnnpack_sse2_ukernels",
2873 defaults: ["xnnpack_internal_default"],
2874 arch: {
2875 arm: { enabled: false, },
2876 arm64: { enabled: false, },
2877 x86: {
2878 srcs: SSE_UKERNELS + SSE2_UKERNELS,
2879 cflags: [
2880 "-msse2",
2881 ],
2882 },
2883 x86_64: {
2884 srcs: SSE_UKERNELS + SSE2_UKERNELS,
2885 cflags: [
2886 "-msse2",
2887 ],
2888 },
2889 },
2890 header_libs: [
2891 "fp16_headers",
2892 ],
2893 static_libs: [
2894 "libpthreadpool",
2895 "xnnpack_tables",
2896 ],
2897}
2898
2899cc_library_static {
Miao Wang2534c2f2020-03-16 11:58:04 -07002900 name: "xnnpack_ssse3_ukernels",
2901 defaults: ["xnnpack_internal_default"],
2902 arch: {
2903 arm: { enabled: false, },
2904 arm64: { enabled: false, },
2905 x86: {
2906 srcs: SSSE3_UKERNELS,
2907 cflags: [
2908 "-mssse3",
2909 ],
2910 },
2911 x86_64: {
2912 srcs: SSSE3_UKERNELS,
2913 cflags: [
2914 "-mssse3",
2915 ],
2916 },
2917 },
2918 header_libs: [
2919 "fp16_headers",
2920 ],
2921 static_libs: [
2922 "libpthreadpool",
2923 "xnnpack_tables",
2924 ],
2925}
2926
2927cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002928 name: "xnnpack_sse41_ukernels",
2929 defaults: ["xnnpack_internal_default"],
2930 arch: {
2931 arm: { enabled: false, },
2932 arm64: { enabled: false, },
2933 x86: {
2934 srcs: SSE41_UKERNELS,
2935 cflags: [
2936 "-msse4.1",
2937 ],
2938 },
2939 x86_64: {
2940 srcs: SSE41_UKERNELS,
2941 cflags: [
2942 "-msse4.1",
2943 ],
2944 },
2945 },
2946 header_libs: [
2947 "fp16_headers",
2948 ],
2949 static_libs: [
2950 "libpthreadpool",
2951 "xnnpack_tables",
2952 ],
2953}
2954
2955cc_library_static {
2956 name: "xnnpack_avx_ukernels",
2957 defaults: ["xnnpack_internal_default"],
2958 arch: {
2959 arm: { enabled: false, },
2960 arm64: { enabled: false, },
2961 x86: {
2962 srcs: AVX_UKERNELS,
2963 cflags: [
2964 "-mavx",
2965 ],
2966 },
2967 x86_64: {
2968 srcs: AVX_UKERNELS,
2969 cflags: [
2970 "-mavx",
2971 ],
2972 },
2973 },
2974 header_libs: [
2975 "fp16_headers",
2976 ],
2977 static_libs: [
2978 "libpthreadpool",
2979 "xnnpack_tables",
2980 ],
2981}
2982
2983cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08002984 name: "xnnpack_xop_ukernels",
2985 defaults: ["xnnpack_internal_default"],
2986 arch: {
2987 arm: { enabled: false, },
2988 arm64: { enabled: false, },
2989 x86: {
2990 srcs: XOP_UKERNELS,
2991 cflags: [
2992 "-mxop",
2993 ],
2994 },
2995 x86_64: {
2996 srcs: XOP_UKERNELS,
2997 cflags: [
2998 "-mxop",
2999 ],
3000 },
3001 },
3002 header_libs: [
3003 "fp16_headers",
3004 ],
3005 static_libs: [
3006 "libpthreadpool",
3007 "xnnpack_tables",
3008 ],
3009}
3010
3011cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08003012 name: "xnnpack_fma3_ukernels",
3013 defaults: ["xnnpack_internal_default"],
3014 arch: {
3015 arm: { enabled: false, },
3016 arm64: { enabled: false, },
3017 x86: {
3018 srcs: FMA3_UKERNELS,
3019 cflags: [
3020 "-mfma",
3021 ],
3022 },
3023 x86_64: {
3024 srcs: FMA3_UKERNELS,
3025 cflags: [
3026 "-mfma",
3027 ],
3028 },
3029 },
3030 header_libs: [
3031 "fp16_headers",
3032 ],
3033 static_libs: [
3034 "libpthreadpool",
3035 "xnnpack_tables",
3036 ],
3037}
3038
3039cc_library_static {
3040 name: "xnnpack_avx2_ukernels",
3041 defaults: ["xnnpack_internal_default"],
3042 arch: {
3043 arm: { enabled: false, },
3044 arm64: { enabled: false, },
3045 x86: {
3046 srcs: AVX2_UKERNELS,
3047 cflags: [
3048 "-mfma",
3049 "-mavx2",
3050 ],
3051 },
3052 x86_64: {
3053 srcs: AVX2_UKERNELS,
3054 cflags: [
3055 "-mfma",
3056 "-mavx2",
3057 ],
3058 },
3059 },
3060 header_libs: [
3061 "fp16_headers",
3062 ],
3063 static_libs: [
3064 "libpthreadpool",
3065 "xnnpack_tables",
3066 ],
3067}
3068
3069cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003070 name: "xnnpack_avx512skx_ukernels",
3071 defaults: ["xnnpack_internal_default"],
3072 arch: {
3073 arm: { enabled: false, },
3074 arm64: { enabled: false, },
3075 x86: {
3076 srcs: AVX512SKX_UKERNELS,
3077 cflags: [
3078 "-mavx512f",
3079 "-mavx512cd",
3080 "-mavx512bw",
3081 "-mavx512dq",
3082 "-mavx512vl",
3083 ],
3084 },
3085 x86_64: {
3086 srcs: AVX512SKX_UKERNELS,
3087 cflags: [
3088 "-mavx512f",
3089 "-mavx512cd",
3090 "-mavx512bw",
3091 "-mavx512dq",
3092 "-mavx512vl",
3093 ],
3094 },
3095 },
3096 header_libs: [
3097 "fp16_headers",
3098 ],
3099 static_libs: [
3100 "libpthreadpool",
3101 "xnnpack_tables",
3102 ],
3103}
3104
3105cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08003106 name: "xnnpack_avx512f_ukernels",
3107 defaults: ["xnnpack_internal_default"],
3108 arch: {
3109 arm: { enabled: false, },
3110 arm64: { enabled: false, },
3111 x86: {
3112 srcs: AVX512F_UKERNELS,
3113 cflags: [
3114 "-mavx512f",
3115 ],
3116 },
3117 x86_64: {
3118 srcs: AVX512F_UKERNELS,
3119 cflags: [
3120 "-mavx512f",
3121 ],
3122 },
3123 },
3124 header_libs: [
3125 "fp16_headers",
3126 ],
3127 static_libs: [
3128 "libpthreadpool",
3129 "xnnpack_tables",
3130 ],
3131}
3132
3133cc_library_static {
3134 name: "xnnpack_ukernels",
3135 defaults: ["xnnpack_internal_default"],
3136 arch: {
3137 arm: {
3138 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003139 "xnnpack_neon_ukernels",
3140 "xnnpack_neonfma_ukernels",
Miao Wangc0aa11a2020-06-10 13:41:26 -07003141 "xnnpack_neonv8_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003142 "xnnpack_neondot_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003143 "xnnpack_asm_ukernels",
3144 ],
3145 },
3146 arm64: {
3147 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003148 "xnnpack_neon_ukernels",
3149 "xnnpack_neonfma_ukernels",
Miao Wangc0aa11a2020-06-10 13:41:26 -07003150 "xnnpack_neonv8_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003151 "xnnpack_neonfp16arith_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003152 "xnnpack_neondot_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003153 "xnnpack_asm_ukernels",
3154 ],
3155 },
3156 x86: {
3157 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003158 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07003159 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003160 "xnnpack_sse41_ukernels",
3161 "xnnpack_avx_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003162 "xnnpack_xop_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003163 "xnnpack_fma3_ukernels",
3164 "xnnpack_avx2_ukernels",
3165 "xnnpack_avx512f_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003166 "xnnpack_avx512skx_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003167 ],
3168 },
3169 x86_64: {
3170 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003171 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07003172 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003173 "xnnpack_sse41_ukernels",
3174 "xnnpack_avx_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003175 "xnnpack_xop_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003176 "xnnpack_fma3_ukernels",
3177 "xnnpack_avx2_ukernels",
3178 "xnnpack_avx512f_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003179 "xnnpack_avx512skx_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003180 ],
3181 },
3182 },
3183 whole_static_libs: [
3184 "xnnpack_scalar_ukernels",
3185 "xnnpack_tables",
3186 ],
3187}
3188
3189cc_library_static {
3190 name: "libXNNPACK",
3191 defaults: ["xnnpack_internal_default"],
3192 export_include_dirs: ["include"],
3193 srcs: [
3194 "src/init.c",
3195 "src/runtime.c",
3196 "src/subgraph.c",
3197 "src/tensor.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003198 ] + SUBGRAPH_SRCS,
3199 header_libs: [
3200 "fp16_headers",
Miao Wange9993472020-02-10 15:00:10 -08003201 ],
3202 whole_static_libs: [
3203 "libclog",
3204 "libcpuinfo",
3205 "libpthreadpool",
3206 "xnnpack_ukernels",
3207 "xnnpack_operator_run",
3208 "xnnpack_operators",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003209 "xnnpack_logging_utils",
3210 "xnnpack_memory_planner",
3211 "xnnpack_packing",
Miao Wange9993472020-02-10 15:00:10 -08003212 ],
3213}
3214
3215// Tests and benchmarks
3216cc_defaults {
3217 name: "xnnpack_tests_default",
3218 vendor_available: true,
3219 stl: "libc++_static",
3220 local_include_dirs: [
3221 "bench",
3222 "models",
3223 "test",
3224 "src",
3225 ],
3226 cflags: [
Miao Wang400e4042020-04-17 10:15:59 -07003227 "-Wno-unused-function",
3228 "-Wno-unused-parameter",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003229 "-Wno-unused-private-field",
Miao Wange9993472020-02-10 15:00:10 -08003230 ],
3231 header_libs: [
3232 "fp16_headers",
3233 ],
3234 static_libs: [
3235 "libXNNPACK",
3236 "libpthreadpool",
3237 "libgmock",
3238 ],
3239 shared_libs: [
3240 "liblog",
3241 ],
3242}
3243
3244cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003245 name: "xnnpack_mobilenet_v1_fp32",
Miao Wange9993472020-02-10 15:00:10 -08003246 defaults: ["xnnpack_tests_default"],
3247 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003248 "models/fp32-mobilenet-v1.cc",
Miao Wange9993472020-02-10 15:00:10 -08003249 ],
3250}
3251
3252cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003253 name: "xnnpack_qs8_mobilenet_v1",
3254 defaults: ["xnnpack_tests_default"],
3255 srcs: [
3256 "models/qs8-mobilenet-v1.cc",
3257 ],
3258}
3259
3260cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003261 name: "xnnpack_mobilenet_v1_fp16",
Miao Wange9993472020-02-10 15:00:10 -08003262 defaults: ["xnnpack_tests_default"],
3263 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003264 "models/fp16-mobilenet-v1.cc",
Miao Wange9993472020-02-10 15:00:10 -08003265 ],
3266}
3267
3268cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003269 name: "xnnpack_qs8_mobilenet_v2",
3270 defaults: ["xnnpack_tests_default"],
3271 srcs: [
3272 "models/qs8-mobilenet-v2.cc",
3273 ],
3274}
3275
3276cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003277 name: "xnnpack_mobilenet_v2_fp32",
Miao Wange9993472020-02-10 15:00:10 -08003278 defaults: ["xnnpack_tests_default"],
3279 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003280 "models/fp32-mobilenet-v2.cc",
Miao Wange9993472020-02-10 15:00:10 -08003281 ],
3282}
3283
3284cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003285 name: "xnnpack_mobilenet_v2_fp16",
Miao Wange9993472020-02-10 15:00:10 -08003286 defaults: ["xnnpack_tests_default"],
3287 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003288 "models/fp16-mobilenet-v2.cc",
3289 ],
3290}
3291
3292
3293cc_library_static {
3294 name: "xnnpack_mobilenet_v3_large_fp32",
3295 defaults: ["xnnpack_tests_default"],
3296 srcs: [
3297 "models/fp32-mobilenet-v3-large.cc",
3298 ],
3299}
3300
3301cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003302 name: "xnnpack_mobilenet_v3_large_fp16",
3303 defaults: ["xnnpack_tests_default"],
3304 srcs: [
3305 "models/fp16-mobilenet-v3-large.cc",
3306 ],
3307}
3308
3309cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003310 name: "xnnpack_mobilenet_v3_small_fp32",
3311 defaults: ["xnnpack_tests_default"],
3312 srcs: [
3313 "models/fp32-mobilenet-v3-small.cc",
Miao Wange9993472020-02-10 15:00:10 -08003314 ],
3315}
3316
Miao Wang5eea8312020-12-07 09:12:40 -08003317cc_library_static {
3318 name: "xnnpack_mobilenet_v3_small_fp16",
3319 defaults: ["xnnpack_tests_default"],
3320 srcs: [
3321 "models/fp16-mobilenet-v3-small.cc",
3322 ],
3323}
3324
Miao Wange9993472020-02-10 15:00:10 -08003325cc_benchmark {
3326 name: "xnnpack_end2end_bench",
3327 defaults: ["xnnpack_tests_default"],
3328 srcs: [
3329 "bench/end2end.cc",
3330 "bench/utils.cc",
3331 ],
3332 cflags: [
3333 "-Wno-unused-result"
3334 ],
3335 static_libs: [
3336 "libcpuinfo",
3337 "libgoogle-benchmark",
Miao Wang5eea8312020-12-07 09:12:40 -08003338 "xnnpack_qs8_mobilenet_v1",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003339 "xnnpack_mobilenet_v1_fp32",
3340 "xnnpack_mobilenet_v1_fp16",
Miao Wang5eea8312020-12-07 09:12:40 -08003341 "xnnpack_qs8_mobilenet_v2",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003342 "xnnpack_mobilenet_v2_fp32",
3343 "xnnpack_mobilenet_v2_fp16",
3344 "xnnpack_mobilenet_v3_large_fp32",
Miao Wang5eea8312020-12-07 09:12:40 -08003345 "xnnpack_mobilenet_v3_large_fp16",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003346 "xnnpack_mobilenet_v3_small_fp32",
Miao Wang5eea8312020-12-07 09:12:40 -08003347 "xnnpack_mobilenet_v3_small_fp16",
Miao Wange9993472020-02-10 15:00:10 -08003348 ],
3349}
3350
3351cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003352 name: "xnnpack_abs_nc_test",
Miao Wange9993472020-02-10 15:00:10 -08003353 defaults: ["xnnpack_tests_default"],
3354 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003355 "test/abs-nc.cc",
Miao Wange9993472020-02-10 15:00:10 -08003356 ],
3357 test_suites: [
3358 "general-tests",
3359 ],
3360}
3361
3362cc_test {
3363 name: "xnnpack_add_nd_test",
3364 defaults: ["xnnpack_tests_default"],
3365 srcs: [
3366 "test/add-nd.cc",
3367 ],
3368 test_suites: [
3369 "general-tests",
3370 ],
3371}
3372
3373cc_test {
3374 name: "xnnpack_argmax_pooling_nhwc_test",
3375 defaults: ["xnnpack_tests_default"],
3376 srcs: [
3377 "test/argmax-pooling-nhwc.cc",
3378 ],
3379 test_suites: [
3380 "general-tests",
3381 ],
3382}
3383
3384cc_test {
3385 name: "xnnpack_average_pooling_nhwc_test",
3386 defaults: ["xnnpack_tests_default"],
3387 srcs: [
3388 "test/average-pooling-nhwc.cc",
3389 ],
3390 test_suites: [
3391 "general-tests",
3392 ],
3393}
3394
3395cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003396 name: "xnnpack_bankers_rounding_nc_test",
Miao Wange9993472020-02-10 15:00:10 -08003397 defaults: ["xnnpack_tests_default"],
3398 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003399 "test/bankers-rounding-nc.cc",
3400 ],
3401 test_suites: [
3402 "general-tests",
3403 ],
3404}
3405
3406cc_test {
3407 name: "xnnpack_ceiling_nc_test",
3408 defaults: ["xnnpack_tests_default"],
3409 srcs: [
3410 "test/ceiling-nc.cc",
Miao Wange9993472020-02-10 15:00:10 -08003411 ],
3412 test_suites: [
3413 "general-tests",
3414 ],
3415}
3416
3417cc_test {
3418 name: "xnnpack_channel_shuffle_nc_test",
3419 defaults: ["xnnpack_tests_default"],
3420 srcs: [
3421 "test/channel-shuffle-nc.cc",
3422 ],
3423 test_suites: [
3424 "general-tests",
3425 ],
3426}
3427
3428cc_test {
3429 name: "xnnpack_clamp_nc_test",
3430 defaults: ["xnnpack_tests_default"],
3431 srcs: [
3432 "test/clamp-nc.cc",
3433 ],
3434 test_suites: [
3435 "general-tests",
3436 ],
3437}
3438
3439cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003440 name: "xnnpack_constant_pad_nd_test",
3441 defaults: ["xnnpack_tests_default"],
3442 srcs: [
3443 "test/constant-pad-nd.cc",
3444 ],
3445 test_suites: [
3446 "general-tests",
3447 ],
3448}
3449
3450cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003451 name: "xnnpack_convolution_nhwc_test",
3452 defaults: ["xnnpack_tests_default"],
3453 srcs: [
3454 "test/convolution-nhwc.cc",
3455 ],
3456 test_suites: [
3457 "general-tests",
3458 ],
3459}
3460
3461cc_test {
3462 name: "xnnpack_convolution_nchw_test",
3463 defaults: ["xnnpack_tests_default"],
3464 srcs: [
3465 "test/convolution-nchw.cc",
3466 ],
3467 test_suites: [
3468 "general-tests",
3469 ],
3470}
3471
3472cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003473 name: "xnnpack_copy_nc_test",
3474 defaults: ["xnnpack_tests_default"],
3475 srcs: [
3476 "test/copy-nc.cc",
3477 ],
3478 test_suites: [
3479 "general-tests",
3480 ],
3481}
3482
3483cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003484 name: "xnnpack_deconvolution_nhwc_test",
3485 defaults: ["xnnpack_tests_default"],
3486 srcs: [
3487 "test/deconvolution-nhwc.cc",
3488 ],
3489 test_suites: [
3490 "general-tests",
3491 ],
3492}
3493
3494cc_test {
3495 name: "xnnpack_divide_nd_test",
3496 defaults: ["xnnpack_tests_default"],
3497 srcs: [
3498 "test/divide-nd.cc",
3499 ],
3500 test_suites: [
3501 "general-tests",
3502 ],
3503}
3504
3505cc_test {
3506 name: "xnnpack_fully_connected_nc_test",
3507 defaults: ["xnnpack_tests_default"],
3508 srcs: [
3509 "test/fully-connected-nc.cc",
3510 ],
3511 test_suites: [
3512 "general-tests",
3513 ],
3514}
3515
3516cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003517 name: "xnnpack_floor_nc_test",
3518 defaults: ["xnnpack_tests_default"],
3519 srcs: [
3520 "test/floor-nc.cc",
3521 ],
3522 test_suites: [
3523 "general-tests",
3524 ],
3525}
3526
3527cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003528 name: "xnnpack_global_average_pooling_nwc_test",
3529 defaults: ["xnnpack_tests_default"],
3530 srcs: [
3531 "test/global-average-pooling-nwc.cc",
3532 ],
3533 test_suites: [
3534 "general-tests",
3535 ],
3536}
3537
3538cc_test {
3539 name: "xnnpack_global_average_pooling_ncw_test",
3540 defaults: ["xnnpack_tests_default"],
3541 srcs: [
3542 "test/global-average-pooling-ncw.cc",
3543 ],
3544 test_suites: [
3545 "general-tests",
3546 ],
3547}
3548
3549cc_test {
3550 name: "xnnpack_hardswish_nc_test",
3551 defaults: ["xnnpack_tests_default"],
3552 srcs: [
3553 "test/hardswish-nc.cc",
3554 ],
3555 test_suites: [
3556 "general-tests",
3557 ],
3558}
3559
3560cc_test {
3561 name: "xnnpack_leaky_relu_nc_test",
3562 defaults: ["xnnpack_tests_default"],
3563 srcs: [
3564 "test/leaky-relu-nc.cc",
3565 ],
3566 test_suites: [
3567 "general-tests",
3568 ],
3569}
3570
3571cc_test {
3572 name: "xnnpack_max_pooling_nhwc_test",
3573 defaults: ["xnnpack_tests_default"],
3574 srcs: [
3575 "test/max-pooling-nhwc.cc",
3576 ],
3577 test_suites: [
3578 "general-tests",
3579 ],
3580}
3581
3582cc_test {
3583 name: "xnnpack_maximum_nd_test",
3584 defaults: ["xnnpack_tests_default"],
3585 srcs: [
3586 "test/maximum-nd.cc",
3587 ],
3588 test_suites: [
3589 "general-tests",
3590 ],
3591}
3592
3593cc_test {
3594 name: "xnnpack_minimum_nd_test",
3595 defaults: ["xnnpack_tests_default"],
3596 srcs: [
3597 "test/minimum-nd.cc",
3598 ],
3599 test_suites: [
3600 "general-tests",
3601 ],
3602}
3603
3604cc_test {
3605 name: "xnnpack_multiply_nd_test",
3606 defaults: ["xnnpack_tests_default"],
3607 srcs: [
3608 "test/multiply-nd.cc",
3609 ],
3610 test_suites: [
3611 "general-tests",
3612 ],
3613}
3614
3615cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003616 name: "xnnpack_negate_nc_test",
3617 defaults: ["xnnpack_tests_default"],
3618 srcs: [
3619 "test/negate-nc.cc",
3620 ],
3621 test_suites: [
3622 "general-tests",
3623 ],
3624}
3625
3626cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003627 name: "xnnpack_prelu_nc_test",
3628 defaults: ["xnnpack_tests_default"],
3629 srcs: [
3630 "test/prelu-nc.cc",
3631 ],
3632 test_suites: [
3633 "general-tests",
3634 ],
3635}
3636
3637cc_test {
3638 name: "xnnpack_resize_bilinear_nhwc_test",
3639 defaults: ["xnnpack_tests_default"],
3640 srcs: [
3641 "test/resize-bilinear-nhwc.cc",
3642 ],
3643 test_suites: [
3644 "general-tests",
3645 ],
3646}
3647
3648cc_test {
3649 name: "xnnpack_sigmoid_nc_test",
3650 defaults: ["xnnpack_tests_default"],
3651 srcs: [
3652 "test/sigmoid-nc.cc",
3653 ],
3654 test_suites: [
3655 "general-tests",
3656 ],
3657}
3658
3659cc_test {
3660 name: "xnnpack_softmax_nc_test",
3661 defaults: ["xnnpack_tests_default"],
3662 srcs: [
3663 "test/softmax-nc.cc",
3664 ],
3665 test_suites: [
3666 "general-tests",
3667 ],
3668}
3669
3670cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003671 name: "xnnpack_square_nc_test",
3672 defaults: ["xnnpack_tests_default"],
3673 srcs: [
3674 "test/square-nc.cc",
3675 ],
3676 test_suites: [
3677 "general-tests",
3678 ],
3679}
3680
3681cc_test {
3682 name: "xnnpack_square_root_nc_test",
3683 defaults: ["xnnpack_tests_default"],
3684 srcs: [
3685 "test/square-root-nc.cc",
3686 ],
3687 test_suites: [
3688 "general-tests",
3689 ],
3690}
3691
3692cc_test {
3693 name: "xnnpack_square_difference_nd_test",
3694 defaults: ["xnnpack_tests_default"],
3695 srcs: [
3696 "test/squared-difference-nd.cc",
3697 ],
3698 test_suites: [
3699 "general-tests",
3700 ],
3701}
3702
3703cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003704 name: "xnnpack_subtract_nd_test",
3705 defaults: ["xnnpack_tests_default"],
3706 srcs: [
3707 "test/subtract-nd.cc",
3708 ],
3709 test_suites: [
3710 "general-tests",
3711 ],
3712}
3713
3714cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003715 name: "xnnpack_truncation_nc_test",
3716 defaults: ["xnnpack_tests_default"],
3717 srcs: [
3718 "test/truncation-nc.cc",
3719 ],
3720 test_suites: [
3721 "general-tests",
3722 ],
3723}
3724
3725cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003726 name: "xnnpack_unpooling_nhwc_test",
3727 defaults: ["xnnpack_tests_default"],
3728 srcs: [
3729 "test/unpooling-nhwc.cc",
3730 ],
3731 test_suites: [
3732 "general-tests",
3733 ],
3734}