Repository navigation
Decide between hand-written CUDA and code generation before the guiding-centre kernels #690
Description
The guiding-centre pushers and evaluation kernels are the largest kernels to port. Before porting them, compare:
-
writing CUDA by hand (the current approach, 1:1 with pyccel)
-
generating kernels from the Python source (
cupyx.jit, numba-cuda, or a pyccel CUDA backend) -
port one guiding-centre kernel both ways
-
compare code size, speed, and how the parity tests and the CPU emulation fit
-
record the decision in
CUDA_STRATEGY.md
Part of #650 (see CUDA_STRATEGY.md).
Activity
Merge mhd-equilibria-cupy (#714) into cuda-codegen-spike
Remove the numba code-generation prototype (#690 decided: hand-writte…
Port bstar_parallel_3form to CUDA by hand (decision for #690: no code…
max-models commentedon Oct 14, 2026
Decided in #718: no code generation, every CUDA kernel is written by hand.
bstar_parallel_3formwas ported both ways for the comparison, and the reasons are inCUDA_STRATEGY.md, section Hand-written CUDA (decision for #690).Port bstar_parallel_3form to CUDA by hand (decision for #690: no code…