towards release 2.5 to-do list #670
Description
- consider whether the big array fwbatch should be malloced at plan or at execute. #480 allocating fwbatch workspace on the fly each exec call. Always do it (saves RAM in multiple plans).
- adjoints Add the capability to do adjoint transforms #633 all done except C++ tester for the adjoint property.
- finally replace GPL-licensed Gauss quadrature code with Apache-2-compatible code. #689
- cuda type-1 output-driven (meth=3) cuFINUFFT ouput driven #661
- Unify CPU/GPU quadrature code #696
- 2.4.1 fails to build:
finufft_core.h: error: field 'mstu' has incomplete type 'std::array<long long int, 3>'#711 . restore fallback spreadinterp code when xsimd unsupported. - constness Add more const-ness and access control #647
- kernel math on CPU: switch onedim_ft to use kernel
- on-the-fly kernel coeff generation - use Robert's inverse Vandermonde evaluator to avoid LSQ solve lin alg dependency?
- Adjoint on GPU since saving memory there is crucial there