CUDA eigenvalues of a sparse matrix

Specific Domains GPU
6

The full thing takes 800ms here (also about 500ms using Arpack), so still slower but not that big of a different (while avoiding expensive memory copies).

For the sparse case, isn’t csreigvsi the API you need?

julia> @benchmark CUSOLVER.csreigvsi(dA, rand(T), CUDA.rand(T, 3000), 1e-6, Cint(1000), 'O')
BenchmarkTools.Trial: 
  memory estimate:  1.47 KiB
  allocs estimate:  64
  --------------
  minimum time:     1.384 ms (0.00% GC)
  median time:      2.278 ms (0.00% GC)
  mean time:        4.155 ms (0.00% GC)
  maximum time:     328.102 ms (0.00% GC)
  --------------
  samples:          1203
  evals/sample:     1

(Note that these wrappers are a little rough, and would benefit from a clean-up / higher-level functions.)

5 Likes