Closed Ken1256 closed 5 years ago
Did you use the latest version? (Commit bbdac46)
51043ef, Commit bbdac46 solve the problem. Other question if KNLMeasCL_cupy temporal part done, nvidia graphic cards KNLMeasCL_cupy will faster than KNLMeasCL OpenCL?(e.g d=2) Thanks
Maybe the CuPy version could run faster than the OpenCL version if I further optimize the code, but it seems that my current implementation runs a bit slower than the OpenCL version, maybe because I didn't take advantage of CUDA and/or the OpenCL version has been fully optimized and it's hard to optimize further.
The CuPy version currently makes it easier for me to experiment with various ideas and practice programming.
Maybe add Dither_bilateral16 subspl and thr can skip some process to speed up KNLMeasCL? Dither_bilateral16, warpsharp need avsproxy pass to VapourSynth it slow but simple and effective.
Here is a native implementation of awarpsharp2 in VapourSynth.
subspl
will certainly be useful in reducing the number of memory operations, and numerous other acceleration techniques exist (e.g. a fast approximation in OpenCV). I will check them later.
In my test warpsharp/Dither_bilateral16 must better than VapourSynth awarpsharp2/VapourSynth-BilateralGPU, I think it is completely different filters.