* dnn: LSTM optimisation
This uses the AVX-optimised fastGEMM1T for matrix multiplications where available, instead of the standard cv::gemm.
fastGEMM1T is already used by the fully-connected layer. This commit involves two minor modifications:
- Use unaligned access. I don't believe this involves any performance hit in on modern CPUs (Nehalem and Bulldozer onwards) in the case where the address is actually aligned.
- Allow for weight matrices where the number of columns is not a multiple of 8.
I have not enabled AVX-512 as I don't have an AVX-512 CPU to test on.
* Fix warning about initialisation order
* Remove C++11 syntax
* Fix build when AVX(2) is not available
In this case the CV_TRY_X macros are defined to 0, rather than being undefined.
* Minor changes as requested:
- Don't check hardware support for AVX(2) when dispatch is disabled for these
- Add braces
* Fix out-of-bounds access in fully connected layer
The old tail handling in fastGEMM1T implicitly rounded vecsize up to the next multiple of 8, and the fully connected layer implements padding up to the next multiple of 8 to cope with this. The new tail handling does not round the vecsize upwards like this but it does require that the vecsize is at least 8. To adapt to the new tail handling, the fully connected layer now rounds vecsize itself at the same time as adding the padding(which makes more sense anyway).
This also means that the fully connected layer always passes a vecsize of at least 8 to fastGEMM1T, which fixes the out-of-bounds access problems.
* Improve tail mask handling
- Use static array for generating tail masks (as requested)
- Apply tail mask to the weights as well as the input vectors to prevent spurious propagation of NaNs/Infs
* Revert whitespace change
* Improve readability of conditions for using AVX
* dnn(lstm): minor coding style changes, replaced left aligned load
* dnn(ocl4dnn): fix LRN layer accuracy problems
- FP16 intermediate computation is not accurate and may provide NaN values
* dnn(test): update tolerance for FP16
fix bug: wrong output dimension when "keep_dims" is false in pooling layer.
* fix bug in max layer
* code align
* delete permute layer and add test case
* add name assert
* check other cases
* remove c++11 features
* style:add "const" remove assert
* style:sanitize file names
* dnn: fix unaligned memory access crash on armv7
The getTensorContent function would return a Mat pointing to some
member of a Protobuf-encoded message. Protobuf does not make any
alignment guarantees, which results in a crash on armv7 when loading
models while bit 2 is set in /proc/cpu/alignment (or the relevant
kernel feature for alignment compatibility is disabled). Any read
attempt from the previously unaligned data member would send SIGBUS.
As workaround, this commit makes an aligned copy via existing clone
functionality in getTensorContent. The unsafe copy=false option is
removed. Unfortunately, a rather crude hack in PReLUSubgraph in fact
writes(!) to the Protobuf message. We limit ourselves to fixing the
alignment issues in this commit, and add getTensorContentRefUnaligned
to cover the write case with a safe memcpy. A FIXME marks the issue.
* dnn: reduce amount of .clone() calls
* dnn: update FIXME comment
Co-authored-by: Alexander Alekhin <alexander.a.alekhin@gmail.com>
Add support for YOLOv4x-mish
* backport to 3.4 for supporting yolov4x-mish
* add YOLOv4x-mish test
* address review comments
Co-authored-by: Guo Xu <guoxu@1school.com.cn>
