Clarification on implemented user case for convolution mode

[Vandermode]

Hi Dmitrii,

Could you clarify the current state of the convolution mode?
It seems not all use cases are supported, the information of the current state is relatively sparse, and I gather a few from the past issues, e.g.,
#79
#159
#193
#66

I focus on 2D C2C convolution in my interest.

I have run several experiments, and my observations are

1. Radix transform is well supported; however, when it comes to multi-upload, one must be careful to either rely on vkFFT disabling fourStepReorder to compute the transformed convolutional kernel or manually transpose the kernel to match the data layout.

2. Other non-radix transform seems not to be supported at all (I am not sure if this is a bug or not)
   For example, if I perform a 1234 x 1234 convolution, I always get a error code 4031 corresponding to VKFFT_ERROR_FAILED_TO_COMPILE_PROGRAM

3. For my NVIDIA A100 GPU, the transition to the multi-load algorithm seems to occur in 8192.
   The following is the result I got:

VkFFTApp[CUDA]: (8192,8192) C2C/s/i/⨂ [rr] [21] buf= 0
app.nb_axis_upload: [1, 2]
app.use_bluestein_fft: [False, False]
app.tmp_buffer_nbytes: 0
app.axis_split [[8192 0 0 0]
[ 128 64 0 0]]

As you can see, I inspected the axisSplit variable, but it's unclear to me why only the second dimension entails a 2-axis upload, though the first dimension size is the same as the second.

Thank you!

[vincefn]

Hi @Vandermode - you may also want to look at vincefn/pyvkfft#33 - I added convolution support to pyvkfft (in a separate branch), with the corresponding systematic tests.

As you can see, not all radix sizes work yet. See @DTolm's answer in that thread too.

[Vandermode]

Hi @vincefn , thanks for the reply. You can see I checked the discussion you pointed out vincefn/pyvkfft#33 (above your last reply there)

I am adding functionalities into the pyvkfft, by exposing more control parameters, such as performZeropadding into the python end.

To this end, I implement the exact python ctypes.Structure counterpart of VkFFTConfiguration Struct (and VkFFTPlan, VkFFTApplication)
See https://github.com/Vandermode/pyvkfft/blob/e27a4d438c68bc85f90d5c98244b04dacdf27057/pyvkfft/cuda.py#L45

This will enable generic support to access arbitrary fields in the vkfft struct in python without the need to modify vkfft_cuda.cu

I am also working on implementing jax/pytorch backend support of pyvkfft.