To test this feature, visit your live site.

tridiag_solve unused in celeris webgpu?

Hi, first of all thanks for sharing celeris webgpu.

I was checking that most if not all examples have a tridiag_solve = 2 option, but when inspecting the code I could not find any mention to this name, not in constants_load_calc.js nor Run_Tridiag_Solver.js.

Am I missing something maybe?

Thanks

2 comments

Comments (2)

José Galaz

3日前

Thanks for the explanation!

Patrick Lynett

Dec 31, 2024

Hi Jose-

You are right! This is a relic flag that I used when I was creating the GPU-based Bous model. The options that "tridiag_solve" used to choose between - different ways to solve the tridaigonal system - have all been removed. You can ignore "tridiag_solve", and I will remove it from the scripts in future releases. For more background, getting the parallel cyclic reduction (PCR) solver to work properly was very difficult. So, as I was trying to get this working, I also implemented different solvers, including a CPU-based Thomas algorithm and a GPU-based G-S iterative solver. I used these to check other parts of the code... However both of these options are much slower than the PCR on GPU, and there was no reason to keep them around.

-Pat