Updated design details
# Details on legalization
Since most of the HW has no native support for computation on bf16, we added a
pass `BF16Legalization` to use fp32 computing bf16 data. It has 3 sub-passes:
`Promotion`, `Elimilination` and `Lowering`.
## BF16Promotion
It adds `cast_to_fp32()
ok, working on it to implement a new legalization pass
---
[Visit Topic](https://discuss.tvm.ai/t/rfc-add-bfloat16-data-type/6778/4) to
respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.ai/email/unsub
@tqchen As you have proposed to eliminate "bfloat16" dtype after the
legalization pass, I have two concerns.
* The legalization pass will be more compilicated. It has to check every TIR
node to replace the bf16 dtype
* As the casting between bf16 and fp32 will be lowered as function calls as
Hi all,
We have completed a workable draft of bfloat16 (bf16) in TVM and the bf16
related codegen in LLVM.
We add bfloat16 as a new type named "bf16" in the frontend. Completed LLVM
backend for generating bf16.
* Use int16 as the storage type in LLVM
* Add legalization to enable computation
