Skip to main content

Module spmv

Module spmv 

Source
Expand description

Sparse matrix-vector multiplication (SpMV).

Computes y = alpha * A * x + beta * y where A is a sparse CSR matrix and x, y are dense vectors stored as raw device pointers.

Three kernel strategies are available:

  • Scalar: one thread per row (best for very sparse rows, < 4 nnz/row)
  • Vector: one warp per row with shuffle reduction (best for moderate sparsity)
  • Adaptive: auto-selects based on average nnz per row

Enums§

SpMVAlgo
Algorithm selection for SpMV.

Functions§

spmv
Sparse matrix-vector multiplication: y = alpha * A * x + beta * y.