Support for block floating point (bfp16)
newling opened this issue · 0 comments
newling commented
There are 2 kinds of bfp16 number formats which AIE2p supports (as announced at computex 2024..). They are exponent:mantissa
ratios of 1:8
and 1:16
.
We need to support this at multiple levels.
In IREE:
- CPU emulation (will we be able to run with llvm-cpu backend?)
- storage formats
- etc.
In iree-amd-aie:
- exponent and mantissa now separate tensors to be tiled, and moved through DMAs separately.
In peano:
- general support
In numpy and C++:
- You wish!
Realistically we might not get all of this for quite a long time, so ideas on prioritization are welcome here. Please feel free to modify this task