DTolm
How to reduce the initialization time?
The initialization time of plan is too long. How to reduce this time? For example, can the compiled kernel be saved? I saw the saveApplicationToString option, which can save the entire plan. Is there any other way?
Hello,
No, there is no other way. saveApplicationToString only saves the binaries, not the plan and this is precisely what you are looking for.
Best regards, Dmitrii
Is it possible to save the internally compiled kernel function in a directory so that it can be used directly next time without compiling it again? Is such a solution feasible in vkfft? Is it difficult?
Hello,
saveApplicationToString saves the binaries, which you can later load with loadApplicationFromString configuration option. See pages 64-65 of the documentation.
Best regards, Dmitrii
save-application cannot adapt to all situations. I want to save the internal kernel, not the application. I compile all the kernels in advance, and any subsequent application that requires the same kernel can be used directly without compiling again.
I am sorry, I don't understand. Which situations it can't adapt to? Please provide an example configuration. From what I read, this functionality is exactly what you want to do.
I want to implement a function such that any size and any stride can be quickly initialized. My approach is to prepare all the kernels in advance and put them in a directory. When you use it later, you don't need to compile it again. It is impossible to save all apps, there are too many apps. But the internal kernel is universal and limited, and can be saved in advance.
The internal kernel is not universal and can't be saved in advance. The thing that you call kernel is a sequence of CPU calls that create the code for a particular FFT and compile it later.
I extracted the generic kernel and compiled it in advance. Subsequent can call directly without compilation at runtime. Can such a feature be achieved by modifying the code? If possible, I'll try to try it.
No, it is not possible to create an uberkernel that will work for all system configurations - it will require a big redesign of the library and won't work with all the algorithms.
Just for HIP. And it’s ok if it can cover most of the algorithms. Can such a feature be implemented?
This feature is not on the radar of my development, as it will require too much time to implement for no particular benefits. If you want to experiment with it - you are free to do so.
As long as it can be achieved and it takes no more than one month, that I will give it a try.