Lowering arith::TruncFOp
from fp32 to bf16 is harder than it should be.
#1111
Labels
Milestone
arith::TruncFOp
from fp32 to bf16 is harder than it should be.
#1111
arith::TruncFOp
is currently lowered to a number of bitwise operations since thellvm-spirv
translation ofis
with half defined as
This is partially solved with PR #1074, but the Intel SPIR-V intrinsics only supports round-to-nearest-even so there are still some bit operations for truncation with round-to-zero.
One of the big barriers for the SPIR-V Translation Tools (
llvm-spirv
) is that there isn't even abf16
type in base SPIR-V, we just have the intel extensions to truncate to abf16
format but the type isi16
.Currently we don't see patching
llvm-spirv
as a solution for internal reasons.The text was updated successfully, but these errors were encountered: