nihuini
cd4be6d0fa
call vulkan create_pipeline on the vkdev condition, drop opt_cpu hacks
2019-10-23 10:50:28 +08:00
nihuini
81a028547a
fix bus error on armv7
2019-10-22 19:03:13 +08:00
daichuanliang
6176ada9f0
update ncnn2int8.cpp ( #1315 )
...
Fix compile issue with ncnn2int8
2019-10-22 18:47:30 +08:00
nihuini
19d75955d6
arm neon assembly optimization for conv3x3s1 winograd pack4to1
2019-10-22 17:53:03 +08:00
bindog
04b4b02324
[WIP] add reduce op support for onnx ( #1308 )
...
* [WIP] add reduce op support for onnx
* extend reduction to support 1,2-dim reduction and keepdims
* fix compile error
* split type to 3 flags && split keepdims to another function
2019-10-22 12:53:12 +08:00
nihuini
22a2be4e6c
fix crop pack4 with reference blob
2019-10-22 10:55:54 +08:00
nihuini
6a8e5c58da
fix build on armv7
2019-10-22 10:42:15 +08:00
nihuini
e63e2449fd
arm neon assembly optimization for conv7x7s2 pack1to4
2019-10-21 12:49:33 +08:00
nihui
56fd26a2da
arm neon assembly optimization for conv1x1s1 pack4to1
2019-10-20 17:07:51 +08:00
nihui
7ad514917b
fix potential out of write on unroll 12 remainder
2019-10-20 16:56:29 +08:00
nihuini
15e86dc8e9
reduce pack4 weight memory usage for specialized kernel, reduce runtime memory usage in conv3x3s1 winograd
2019-10-18 10:53:33 +08:00
nihuini
581a06d471
since innerproduct pack4 always consumes flattened blob, which layout is same as pack1 branch, so reuse pack1 implementation to reduce memory usage
2019-10-18 10:29:55 +08:00
nihuini
c5f1dc3fe4
arm neon assembly optimization for conv3x3s1 pack4to1
2019-10-17 17:29:46 +08:00
nihui
2f8b31c3b4
unroll outch 2 for conv3x3s1 pack1to4
2019-10-16 21:50:44 +08:00
ShuangLiu1992
396b057248
fix opencv compatibility issue with major verion > 2 for the quantisation tool ( #1302 )
...
* Update CMakeLists.txt
enable quant
* fix opencv compatibility issue with major higher than 2
2019-10-15 06:41:55 +08:00
nihui
e0f6e3f669
pre-interleave 8-channel weight data on aarch64, conv1x1s1 version
2019-10-14 23:00:18 +08:00
nihuini
d11bf14d44
pre-interleave 8-channel weight data on aarch64
2019-10-14 19:29:04 +08:00
nihui
7173b6e38e
arm neon assembly optimization for conv3x3s2 pack4
2019-10-14 08:34:43 +08:00
nihuini
cf0c49dd71
arm neon assembly optimization for conv5x5s1 pack4 and conv5x5s2 pack4
2019-10-12 19:21:58 +08:00
bindog
53d54115a8
add split op support for onnx2ncnn ( #1299 )
2019-10-12 10:15:41 +08:00
nihui
9e529354fb
arm neon optimization for conv1x1s2 pack4
2019-10-11 23:41:42 +08:00
tpoisonooo
25979c1bdb
move ncnnoptimize doc to docs directory ( #1297 )
2019-10-11 16:25:38 +08:00
nihuini
f8f3b0b5aa
shufflechannel pack4
2019-10-11 15:51:43 +08:00
nihuini
50d5896ce7
reshape pack4
2019-10-10 17:15:05 +08:00
nihuini
624291e2b2
use subop optimization for group convolution deconvolution pack4 family
2019-10-10 12:37:40 +08:00
nihui
48e3e7d49c
move neon activation into a wrapper function
2019-10-09 23:21:15 +08:00
nihui
8c1b87b1a2
fallback to cpu if no vulkan device found
2019-10-08 22:55:31 +08:00
nihui
b37ecab630
auto flatten before innerproduct pack4
2019-10-08 22:48:19 +08:00
nihui
afd1f08194
arm neon assembly optimization for pooling2x2s2 max pack4
2019-10-06 17:44:26 +08:00
nihui
e19b7097df
arm neon assembly optimization for conv3x3s1 pack1to4
2019-10-06 17:19:37 +08:00
nihui
3ac6335ba3
hardsigmoid and hardswish pack4
2019-10-05 16:11:15 +08:00
nihui
21e74487b4
arm neon optimization for convdw5x5 pack4
2019-10-05 16:03:40 +08:00
volvet
ecd64fb36b
Fixed lots of compile warnings ( #1286 )
...
* Fixed lots of compile warnings
* refine the unused warning change
2019-10-04 21:15:33 +08:00
nihui
3e1bad4880
arm neon assembly optimization for pooling3x3s2 max pack4
2019-10-04 14:27:01 +08:00
nihui
08a97c169f
arm neon assembly optimization for relu pack4
2019-10-04 12:41:48 +08:00
nihui
a1bd88fb4a
arm neon assembly optimization for padding constant pack4
2019-10-04 10:47:37 +08:00
nihui
17f343e7e4
convdw3x3 pack4 arm neon assembly optimization
2019-10-03 09:39:43 +08:00
nihui
6703286831
the very long ld1, one less load
2019-09-30 21:30:04 +08:00
nihui
22a3ade6ce
unroll size 12 for conv1x1s1 and conv3x3s1 winograd pack4 on aarch64
2019-09-29 22:50:00 +08:00
tpoisonooo
e5eb3e427c
Update CMakeLists.txt ( #1279 )
2019-09-29 15:14:31 +08:00
nihuini
c3eea4cc0b
fuse onnx unsqueeze-batchnorm-squeeze as batchnorm
2019-09-29 14:51:47 +08:00
nihui
3a452f734a
arm neon assembly optimization for conv3x3s2 pack1to4
2019-09-28 18:27:32 +08:00
nihui
6edd42f566
arm neon assembly for conv1x1s1 and conv3x3s1 winograd pack4
2019-09-28 09:11:38 +08:00
nihuini
e147debe07
add retinaface example
2019-09-27 16:20:54 +08:00
nihuini
87ae1e9a2a
%e in ncnnoptimize
2019-09-27 14:49:24 +08:00
nihuini
7d3ebe2d69
%g may output number without dot, stick to %e
2019-09-26 19:38:03 +08:00
nihuini
a5b6826a16
workaround duplicated op name, fusion function, fix wrong layer count when the last layer is Reshape
2019-09-26 17:32:13 +08:00
nihuini
583f0a7c2f
eliminate noop
2019-09-26 14:43:22 +08:00
Howave
50f69f1755
mxnet hardsigmoid hardswish op fusion ( #1266 )
2019-09-19 23:05:20 +08:00
nihuini
f7808b2dfc
write pad_value
2019-09-19 11:12:28 +08:00
nihuini