What's the k in AoSoA?

luli · 8 February 2023 06:45

Hello, I have some trouble in understanding the format of AoSoA. From the introduction of pyfr data structure, I know that the memory arrangement of AoSoA is as shown in the figure.I guess the k in the picture corresponds to soasz on cuda platform, and its value is 32. Is that right?

fdw · 8 February 2023 14:54

Yes; if you look at:

github.com

PyFR/PyFR/blob/83c972d7af2c8263f9dfe305a7889cfe3f8263c7/pyfr/backends/base/types.py#L36


      
          if ndim == 2:
              nrow, ncol = shape
          
          
    # Alignment requirement for the leading dimension
              ldmod = csubsz if 'align' in self.tags else 1
              leaddim = csubsz if backend.blocks else ncol - (ncol % -ldmod)
          
          
    nblocks = (ncol - (ncol % -leaddim)) // leaddim
              datashape = [nblocks, nrow, leaddim]
          else:
              nvar, narr, k = shape[-2], shape[-1], soasz
              nparr = narr - narr % -csubsz
          
          
    nrow = shape[0] if ndim == 3 else shape[0]*shape[1]
              ncol = nvar*nparr
              leaddim = nvar*csubsz if backend.blocks else ncol
          
          
    nblocks = (ncol - (ncol % -leaddim)) // leaddim
              datashape = [nblocks, *shape[:-2], nparr // (nblocks*k), nvar, k]
          
          
# Assign

where:

github.com

PyFR/PyFR/blob/83c972d7af2c8263f9dfe305a7889cfe3f8263c7/pyfr/backends/base/types.py#L24


      
          self.backend = backend
          self.tags = self._base_tags | tags
          
          
self.dtype = dtype
          self.itemsize = np.dtype(dtype).itemsize
          
          
# Our shape and dimensionality
          shape, ndim = list(ioshape), len(ioshape)
          
          
# SoA and block column size
          soasz, csubsz = backend.soasz, backend.csubsz
          
          
if ndim == 2:
              nrow, ncol = shape
          
          
    # Alignment requirement for the leading dimension
              ldmod = csubsz if 'align' in self.tags else 1
              leaddim = csubsz if backend.blocks else ncol - (ncol % -ldmod)
          
          
    nblocks = (ncol - (ncol % -leaddim)) // leaddim
              datashape = [nblocks, nrow, leaddim]

Regards, Freddie.

luli · 7 March 2023 01:32

It seems this value equal to wrap size on cuda platform, so in hip platform this value should be 64?

fdw · 7 March 2023 01:42

Yes, although some (consumer grade) HIP devices have a size of 32, rather than 64.

Regards, Freddie.

luli · 7 March 2023 04:31

Thank you very much!

Topic		Replies	Views
Matrix dimensions and cublas kernels Development	2	232	13 December 2022
Problem with CUDA backend General	2	445	26 October 2020
Find input of kernels General	2	485	17 June 2021
How does pyfr organize its data Development	1	138	29 August 2023
Uncoalescing data format has a bad impact on the GPU kernel performance General	7	392	8 November 2021

What's the k in AoSoA?

Related topics