Merge pull request #17885 from alalek:dnn_ocl_slice_update
authorAlexander Alekhin <alexander.a.alekhin@gmail.com>
Mon, 3 Aug 2020 14:13:34 +0000 (17:13 +0300)
committerGitHub <noreply@github.com>
Mon, 3 Aug 2020 14:13:34 +0000 (14:13 +0000)
commit1c8ee3f957f1d8bcb9ac051f0068b9d3aca00856
treedd0fc10164db56b29f2b18038faf2646f31f5b5e
parent922108060d48742ff4f2397af9bf8d89c87acab9
Merge pull request #17885 from alalek:dnn_ocl_slice_update

DNN: OpenCL/slice update

* dnn(ocl/slice): make slice kernel VTune friendly

- more unique names
- inline code of copy functions

* dnn(ocl/slice): prefer to spawn more work groups

- even in case with 1D copy
- perf improvement up to 2x of kernel time (due to changed configuration 128x1x1 => 128x32x1)

* dnn(ocl/slice): cache kernel exec info
modules/dnn/src/layers/slice_layer.cpp
modules/dnn/src/opencl/slice.cl