mirror of
https://github.com/sfilippone/psblas3.git
synced 2026-10-16 19:02:09 +00:00
Stack-1
cd0fa59647
Generated from template-psblas, where the unrolled csmv_inner kernel was
added to the type-generic X_csr_impl.P90. The d version produced by the
template is byte-identical to the hand-written one in f7db723f, so this
commit only carries s/c/z.
Same dispatch as d: mean row length >= 16 nonzeros/row uses four
accumulators to break the FMA latency chain, shorter rows keep the
original loop. Applies to the beta==0/alpha==1 and beta==1/alpha==1
branches of psb_?_csr_csmv_inner.
Checked with mpif90 -O3 -fopenmp on all three types; the unrolled and
reference loops agree to rounding on rows of every length mod 4,
including empty rows.