CUDA binaries for multiview reconstruction Windows 10

fiji
imagej
plugin

#1

Hi All,
I have successfully used Multiview Reconstruction with SeparableConvolutionCUDALib and FourierConvolutionCUDALib on a Windows 7 x64 machine using the released .dll files.

I have now invested in a much more powerful analysis PC and this is running Windows 10 x64. I tried using the same .dll files to run the Multiview Reconstruction pipeline that I had working on the Windows 7 x64 machine but get the error that it fails to load JNA library.

I have had no luck in recompiling these for Windows 10 x64 and wonder if anyone out there as successfully recompiled SeparableConvolutionCUDA.dll and Convolution3D_fftCUDAlib.dll for Windows 10 x64 and would be happy to share them?

Thanks,
Ben


#2

Hello @mnqslbs2 ,

What GPU are you using? I can give you the ones I compiled here for Win 10 x64 with a GTX1080 Ti

FourierConvolutionCUDALib.zip (520.7 KB)

I haven’t used SeparableConvolutionCUDALib though, I was just interested in the fourier convolution one…

I documented my struggles with Win10 and Visual Studio here
https://c4science.ch/w/bioimaging_and_optics_platform_biop/image-processing/deconvolution/cuda-deconvolution/


#3

Hello @oburri,
Thank you for sharing.
I am using a GTX 1080 (actually have a dual card setup with an SLI bridge, although not sure if I can get Fiji to use both cards at once), so your compiled FourierConvolutionCUDALib should work.
I’ll give this a try and let you know.
Thanks again.
Best,
Ben


#4

Hello @oburri,
I have managed to compile a SeparableConvolutionCUDA.dll on Windows 10 x64 using CUDA Toolkit 10 and Visual Studio 2017, or those that might be intersted here are the files as a zip:
SeperableConvolutionCUDALib Win 10 x64

I sucessfully compiled FourierConvolutionCUDALib on my system with the aid of your instructions.
However the when I run a deconvolution fusion, using the Multiview Reconstruction plugin from @StephanPreibisch, I get a very odd result when using my freshly complied FourierConvolutionCUDALib and the GPU, see representative single slice from the resulting GPU processed image stack (124_GPU_8bit):


If I run the same deconvolution fusion on the CPU I get the expected result, see same representative single slice from the resulting CPU processed image stack (124_GPU_8bit):

Do you, or anyone else out there, have any ideas what has happened here? Could it be due to the use of CUDA Toolkit 10? I am unsure of where to start with troubleshooting.

Thanks,
Ben