github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-04-05 03:01:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

ROCM IFU: Extend input to 32-bit when necessary

Browse Source 
Note: we'll need to check this later if we can use i8 for some
reduction operations
This commit is contained in:
Lixun Zhang
2023-09-29 11:30:00 -05:00
committed by Jason Furmanek
parent 92edee723b
commit a41f13adcd
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/triton/language/core.py
View File

@@ -1377,6 +1377,11 @@ def reduce(input, axis, combine_fn, _builder=None, _generator=None):
@builtin
def _promote_reduction_input(t, _builder=None):
scalar_ty = t.type.scalar
# input is extended to 32-bits if necessary
# this increases numerical accuracy and can be done pretty much for free
# on GPUs
if scalar_ty.is_int() and scalar_ty.int_bitwidth < 32:
return t.to(int32, _builder=_builder)
# hardware doesn't support FMAX, FMIN, CMP for bfloat16
if scalar_ty is bfloat16: