CUDA

[fgerick]

Following #5 this implements the CUDA transforms. Here is a preliminary example, using a cuda compiled SHTns library:

import Pkg; Pkg.activate(@__DIR__)
using Revise

ENV["SHTNS_PATH"] = "/home/eh/gerickf/programs/shtns/libshtns_cuda.so"
using CUDA
using SHTns

function spat_func_test(theta, phi)
    return 1.0 + 0.01*cos(theta)  + 0.1*(3cos(theta)*cos(theta) - 1.0) +	#// Y00, Y10, Y20
    (cos(phi) + 0.3*sin(phi)) * sin(theta)	+ #// Y11
    (cos(2phi) + 0.1*sin(2phi)) * sin(theta)*sin(theta) * (7.0* cos(theta)*cos(theta) - 1.0) * 3/8 	#// Y42
end

nlat, nphi, lmax = 512,1024,7

cfg_gpu = SHTnsCfg(lmax, 3, 1, nlat, nphi; shtype=SHTns.QuickInit(; gpu=true, contiguous_lat = true, contiguous_phi=false, padding=false, float32=false) )
cfg_cpu = SHTnsCfg(lmax, 3, 1, nlat, nphi; shtype=SHTns.QuickInit(; gpu=false, contiguous_lat = true, contiguous_phi=false, padding=false) )

θ,ϕ = SHTns.grid(cfg_cpu)

d = spat_func_test.(θ,ϕ')
q = SHTns.analys(cfg_cpu, d)

dg = CuArray(zero(d))
qg = CuArray(q)

SHTns.synth!(cfg_gpu, qg, dg)

Array(dg) ≈ d #true

SHTns.analys!(cfg_gpu, dg, qg)

Array(qg) ≈ q #true

SHTns.synth!(cfg_cpu, qg, dg) #throws assertion error, need to define SHTnsCfg with shtype with gpu=true.