Script.setmaxnreg

Script.setmaxnreg

Script.setmaxnreg(nregs, action)

Adjust the maximum number of registers per thread of the current warp group.

Lets warp-specialized kernels rebalance the register file: a producer warp group that only issues TMA can release registers with action='dec', and consumer warp groups that hold large accumulators can claim them with action='inc'.

Parameters:

  • nregs (int) – The new per-thread register limit. Must be a multiple of 8 in [24, 256].

  • action (str) – 'inc' to raise the limit (blocks until enough registers are free) or 'dec' to lower it and release registers. Candidates: 'inc', 'dec'.

Return type:

None

Notes

  • Thread group: Must be executed by whole warp groups (a multiple of 128 threads).

  • Hardware: Requires compute capability 9.0a+ (sm_90a).

  • PTX: setmaxnreg.{inc,dec}.sync.aligned.u32

Contents