[PATCH] D126953: Promote bf16 to f32 when the target doesn't support it
Benjamin Kramer via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Fri Jun 3 02:24:30 PDT 2022
bkramer created this revision.
bkramer added a reviewer: t.p.northover.
Herald added subscribers: jsji, Enna1, pengfei, hiraditya, mgorny.
Herald added a project: All.
bkramer requested review of this revision.
Herald added projects: Sanitizers, LLVM.
Herald added subscribers: llvm-commits, Sanitizers.
This is modeled after the half-precision fp support. Two new nodes are
introduced for casting from and to bf16. Since casting from bf16 is a
simple operation I opted to always directly lower it to integer
arithmetic. The other way round is more complicated if you want to
preserve IEEE semantics, so it's handled by a new __truncsfbf2
compiler-rt builtin.
This is of course very bare bones, but sufficient to get a semi-softened
fadd on x86.
Possible future improvements:
- Targets with bf16 conversion instructions can now make fp_to_bf16 legal
- The software conversion to bf16 can be replaced by a trivial implementation under fast math.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D126953
Files:
compiler-rt/lib/builtins/CMakeLists.txt
compiler-rt/lib/builtins/fp_trunc.h
compiler-rt/lib/builtins/truncsfbf2.c
llvm/include/llvm/CodeGen/ISDOpcodes.h
llvm/include/llvm/IR/RuntimeLibcalls.def
llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
llvm/lib/CodeGen/TargetLoweringBase.cpp
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/bfloat.ll
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D126953.433987.patch
Type: text/x-patch
Size: 11359 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220603/10062c47/attachment.bin>
More information about the llvm-commits
mailing list