helion.language.register_block_size
- helion.language.register_block_size(min_or_max, max_or_none=None, /)[source]
Explicitly register a block size that should be autotuned and can be used for allocations and inside hl.tile(…, block_size=…).
This is useful if you have two loops where you want them to share a block size, or if you need to allocate a kernel tensor before the hl.tile() loop.
- The signature can one of:
hl.register_block_size(max) hl.register_block_size(min, max)
Where min and max are integers that control the range of block_sizes searched by the autotuner. Max may be a symbolic shape, but min must be a constant integer.