Sparse Arrays

AMDGPU.jl wraps AMD's rocSPARSE and stores sparse matrices on the device with the AMDGPU.rocSPARSE array types:

• ROCSparseMatrixCSR — compressed sparse row (the usual default for GPU work).

• ROCSparseMatrixCSC — compressed sparse column.

• ROCSparseMatrixCOO — coordinate format.

• ROCSparseVector — sparse vector.

Construct them by converting a host SparseMatrixCSC, and convert back the same way:

using AMDGPU, SparseArrays
using AMDGPU.rocSPARSE

S  = sprand(Float32, 1000, 1000, 0.01)
dS = ROCSparseMatrixCSR(S)        # upload to the GPU

SparseMatrixCSC(dS)               # download back to the host

Operations

Sparse matrix–vector (SpMV) and matrix–matrix (SpMM) products use the standard * operator:

x = AMDGPU.rand(Float32, 1000)
y = dS * x                        # SpMV

B = AMDGPU.rand(Float32, 1000, 8)
Y = dS * B                        # SpMM

rocSPARSE has no symmetric or Hermitian products, so a product with a Symmetric or Hermitian wrapper first materializes the full matrix on the device. That copy is made on every product; code that multiplies by the same wrapped matrix repeatedly should build it once with sparse(Symmetric(dS)) and reuse the result.

Incomplete-factorization preconditioners (ic0, ilu0) and further sparse routines (conversions between formats, triangular solves, gather/scatter) are available in the AMDGPU.rocSPARSE submodule for building iterative solvers on the GPU.