ara
ara copied to clipboard
Add support for vector floating-point division and square-root instructions
Currently, Ara supports the most important FP operations to run a matmul
kernel.
All the other vector floating-point instructions still need to be added.
Note: Some of them, namely:
-
vfclass
-
vfdiv
-
vfsqrt
are not implemented due to issues with the FPU (fpnew
) when operating in vectorial (SIMD) mode. Once these problems are solved, the corresponding instructions will be added as well.
Is there a corresponding issue in the fpnew
repo? As far as I know, this is not an issue with Ara, but rather on how the fpnew
handles the vector execution of those instructions.
This is a reference to the synch issue that affects vfdiv
in fpnew
: https://github.com/pulp-platform/fpnew/issues/45
We will add the new issue for vfclass
soon; I have opened an issue here as well just as a reminder, as the implementation of these instructions in Ara is stalled by the issues in fpnew
Currently, Ara supports the most important FP operations to run a
matmul
kernel. All the other vector floating-point instructions still need to be added.Note: Some of them, namely:
vfclass
vfdiv
vfsqrt
are not implemented due to issues with the FPU (
fpnew
) when operating in vectorial (SIMD) mode. Once these problems are solved, the corresponding instructions will be added as well.
Due to a delay in updating the FPU, these instructions were delayed as well. We are merging them now with PR https://github.com/pulp-platform/ara/pull/100.