How to enable VkFFTBackend when building itk 5.3.0 on windows

[zhusihan-python]

hello, im trying to build itk 5.3.0 with ITK_USE_CUFFTW ON on windows, but get some problems with that.
is the VkFFTBackend a substitude for cufft, if yes, how to enable VkFFTBackend in itk? will the itk montage benefit from VkFFTBackend set ON

my environment: windows 10 cmake 3.26.1 cuda v12.1

[dzenanz]

Montage might benefit from doing FFTs on the GPU, but the GPU needs to be more powerful than the CPU and tiles need to be big enough to justify the time it takes to transfer them to the GPU. The only sure way to to know is to try it.

@Leengit or @tbirdso might be able to help with building, but you need to better describe the problem, e.g. provide the error message.

[zhusihan-python]

thanks for the reply, here is my cmake config

cmake configure output
cmake_configure.txt
build log
itk_build_log.txt
one of the error message:

683>ITKOptimizersv4-5.4.lib(ITKOptimizersv4-5.4.dll) : error LNK2005: "public: class itk::LBFGSOptimizerBaseHelperv4<class vnl_lbfgs> * __cdecl itk::LBFGSOptimizerBasev4<class vnl_lbfgs>::GetOptimizer(void)" (?GetOptimizer@?$LBFGSOptimizerBasev4@Vvnl_lbfgs@@@itk@@QEAAPEAV?$LBFGSOptimizerBaseHelperv4@Vvnl_lbfgs@@@2@XZ) 已经在 itkLBFGSOptimizerBasev4Python.obj 中定义
683>  正在创建库 E:/BUILD/master-cuda/itk/lib/Release/ITKOptimizersv4Python.lib 和对象 E:/BUILD/master-cuda/itk/lib/Release/ITKOptimizersv4Python.exp
683>E:\BUILD\master-cuda\itk\Wrapping\Generators\Python\itk\_ITKOptimizersv4Python.pyd : fatal error LNK1169: 找到一个或多个多重定义的符号

sor for the Chinese in the message, fyi:
已经在 itkLBFGSOptimizerBasev4Python.obj 中定义：already defined in itkLBFGSOptimizerBasev4Python.obj
找到一个或多个多重定义的符号：one or more multiply defined symbols found

[tbirdso]

Hi @zhusihan-python, a few thoughts:

im trying to build itk 5.3.0 with ITK_USE_CUFFTW ON on windows, but get some problems with that

Does ITK 5.3.0 compile successfully for you when ITK_USE_CUFFTW is turned off? Have you tested to verify that this is the flag that causes the compilation failure you noted above?

is the VkFFTBackend a substitude for cufft

ITKVkFFTBackend is a wrapper to use the VkFFT cross-platform library in ITK image processing pipelines. VkFFT itself sits on top of FFT implementations such as cuFFT, meaning ITKVkFFTBackend is not a substitute for cuFFT, but can provide an alternate path for compiling ITK with cuFFT support.

how to enable VkFFTBackend in itk?

To compile ITKVkFFTBackend as a remote module alongside ITK, set the Module_VkFFTBackend parameter to ON in your ITK-build cmake configuration.

[zhusihan-python]

hi tom @tbirdso , i already build ITK 5.3.0 and ITK 5.3.0 with ITK_USE_MKL ON successful on windows 10. but failed compile it with ITK_USE_CUFFTW ON or FFTWD FFTWF ITK_USE_FFTWF_DEFAULT ON or ITK_USE_CUFFTW FFTWD FFTWF ITK_USE_FFTWF_DEFAULT ON.

as i want to compare the performance of itk Montage CompleteMontage of the default fft、mkl fft、cuda fft backends.
the first two is ready, compiling ITK with cuFFT support is exactly what i want now. so i want to try ITKVkFFTBackend

by default the Module_VkFFTBackend didn't show in the configure, then i tried move this part from Modules/Remote/CmakeLists to the root CmakeLists

file(GLOB Modules/Remote/remotes "*.remote.cmake")
foreach(remote_module ${remotes})
  include(${remote_module})
endforeach()

or add an entry by hand

it seems both not working, so my problem now is how to set Module_VkFFTBackend ON in cmake gui

[dzenanz]

Try setting ITK_MINIMUM_COMPLIANCE_LEVEL to 1. That should expose more module ON/OFF settings in CMake GUI.

[zhusihan-python]

i found that cmake advanced mode will get the remote Module entries, then i enabled Module_VkFFTBackend and Module_Montage, removed ITK_WRAP_PYTHOH,
still get compile error:

272>PhaseCorrelationImageRegistration.obj : error LNK2001: unresolved external symbol "void __cdecl itk::FFTWFFTImageFilterInitFactoryRegister__Private(void)" (?FFTWFFTImageFilterInitFactoryRegister__Private@itk@@YAXXZ)
272>E:\BUILD\master-cuda\itk\Wrapping\Generators\Python\itk\PhaseCorrelationImageRegistration.exe : fatal error LNK1120: 1 unresolved externals
271>RefineMontage.obj : error LNK2001: unresolved external symbol "void __cdecl itk::FFTWFFTImageFilterInitFactoryRegister__Private(void)" (?FFTWFFTImageFilterInitFactoryRegister__Private@itk@@YAXXZ)
271>E:\BUILD\master-cuda\itk\Wrapping\Generators\Python\itk\RefineMontage.exe : fatal error LNK1120: 1 unresolved externals

fulllog: build_vkfft_nopy.txt

[dzenanz]

Is this InsightSoftwareConsortium/ITKMontage#214 reappearing?

Does the error persist if you do a clean build, in a new directory?

[zhusihan-python]

Is this InsightSoftwareConsortium/ITKMontage#214 reappearing?

Does the error persist if you do a clean build, in a new directory?

yes sir. I compiled several times in seperate dst dir from seperate source code. the difference is last time in the issue you referred I didn't set the cufft ON.

[zhusihan-python]

the error message seems similar FFT image filter but not exactly the same. maybe it's caused by the same reason

[tbirdso]

Hi @zhusihan-python , for your use case would it be reasonable to install and use prebuilt ITKVkFFTBackend OpenCL Python packages instead of building them yourself, similar to the workaround in InsightSoftwareConsortium/ITKMontage#214?

$ python -m pip install itk==5.3.0 itk-montage itk-vkfft

itk-vkfft packages available on PyPI will look for an OpenCL implementation on your machine by default. Your CUDA installation likely provides this OpenCL DLL. The ITK Python factory loading mechanism takes care of setting up the accelerated FFT filters at runtime, so you will be able to use accelerated FFTs with ITKMontage filters without any additional lines of Python code.

[zhusihan-python]

hi @tbirdso
the python version itk is kind of slow in stitching. 150 3072*2048 images reading and stitching take more than 55s whie c++CompleteMontage cost less than 12s.

i can use 5.2.1 cuda version now. just curious about the speed of 5.3.0 with vkfft backends. hope i can try it in following releases.

[dzenanz]

ITKPython has a big startup cost. Just loading all the relevant DLLs takes ~10 seconds. But most of the computation is done in C++ libraries. I wonder whether you have a relatively slow graphics card, which would make the Python+vkfft slower.

@tbirdso is there a way to disable vkFFT via environment variable or something similar? That way @zhusihan-python could compare C++ CPU vs Python CPU, and test my theory of a slow graphics card.

[zhusihan-python]

the python version itk what i test is 5.3.0 installed from pypi, i think the vkfft backend is not enabled by default, right.

[dzenanz]

Correct, it is not enabled by default. Could it be that Python version is single-threaded? If you look at CPU usage when you run, does Python version use all CPU cores?

[zhusihan-python]

i think it does use multiple CPU cores, 132 3000*4096 images reading、compute transfrom and Resampling cost 53.8s, writing output not included