For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

warpgroup_reg_dealloc

def warpgroup_reg_dealloc[count: Int]()

Deallocates additional registers for the executing warp group.

Hints to the system to decrease per-thread registers owned by the executing warp. Releases extra registers to reduce the absolute per-thread maximum register count from its current value to the specified count.

Note:

• Only supported on NVIDIA SM90+ GPUs.
• Performance optimization hint that may be ignored by the hardware.
• Pair with warpgroup_reg_alloc() when extra registers are needed.

Parameters:

• ​count (Int): The desired number of registers per thread. Must be:
  • A multiple of 8.
  • Between 24 and 256 (inclusive).