6D array views in cunumpy for the matrix accumulations

Matrix accumulations write 6D stencil-matrix data, but cunumpy's array views stop at 4D (Array1D–Array4D, CArray1D–CArray4D). This blocks:

vlasov_maxwell (VlasovAmpereCoupling), which finishes Vlasov–Ampère/Maxwell on the GPU
linear_vlasov_ampere (EfieldWeightsCoupling), the δf variant

Tasks:

cunumpy: Array5D/Array6D (and the C-contiguous versions), or a documented raw-pointer convention for higher-dimensional data
struphy: CUDA versions of vlasov_maxwell and linear_vlasov_ampere, with parity cases
feectools: remove the unshifted 2p+1-diagonal assumption in the 3D stencil kernels (stencil_dot_3d, stencil_transpose_3d), which today raise NotImplementedError for matrices that don't fit

Part of #650 (see CUDA_STRATEGY.md).

Done:

cunumpy 0.6.1 has Array6D (and the C-contiguous views).
CUDA version of the linear_vlasov_ampere accumulation #705 adds the CUDA version of linear_vlasov_ampere, CUDA version of the vlasov_maxwell accumulation #713 the one of vlasov_maxwell, both with parity cases.
Stencil 3D device kernels: 6D matrix views instead of the 2p+1 assumption feectools#97 removes the unshifted 2p+1-diagonal assumption in stencil_dot_3d/stencil_transpose_3d (6D matrix views, diagonals taken from the data).

Navigation Menu