mirror of
https://github.com/sfilippone/psblas3.git
synced 2026-10-19 12:22:10 +00:00
Stack-1
f5f894222e
dpdegenmv converts the matrix ncnv times before the timed loop. Those conversions are neither timed nor printed, but on large matrices they dominate the wall time: at idim=480 an HLL run spent 21 minutes, of which 62 s generating the matrix and 21 s on the timed products. SPMV_NCNV overrides the default, which stays at 3, so the behaviour without the variable is unchanged. Also add sweep_fixed.sh, the serial sweep used to measure CSR, HLL and HDIA after the DIA/HDIA fixes: it resumes where it stopped, takes a lock so two instances cannot share the pinned core, and skips a size when the machine does not have the memory for it.