Matrix dimensions and cublas kernels

Development
1

Dear PyFR team,

I am trying to use cublas in PyFR and I am using the Euler vortex case for testing. However, I am confused because when I print the matrix that is input to cublas, I found that it is three-dimensional (64, 4, 400). I am using the print function print(A.get()) to print the matrix. Shouldn’t the input matrix to cublas be two-dimensional? And I also noticed that A.ncol is 1664. How is this number calculated? I am having a hard time understanding this.

I would greatly appreciate it if you could provide some clarification on this issue.

Best regards,

2

I know the reason now, in file backend.base.type , pyfr reshape the 3D matrix. And I found a parameter soasz =32 , is this number related to the hardware arch? How to get this number?

3

We use an AoSoA data layout with variable width. On GPUs it makes sense to set this to 32 as there are typically 32 threads in a warp, so you can get efficient memory access with this width.