A few improvements to the onxx 2 hlo converter. - Githubissues

LeelaChessZero / lc0

The rewritten engine, originally for tensorflow. Now all other backends have been ported here.

GNU General Public License v3.0

2.38k stars 525 forks source link

A few improvements to the onxx 2 hlo converter. #2047

Closed mooskagh closed 2 weeks ago

mooskagh commented 1 month ago

In compile time, only cast from/to f32/f64/s32/s64.
Constant evaluation of OpMul.
Added support of Max op, also with constant evaluation.
Const ReduceProd also does constant folding for the final reshape.
OpReshape also handles constant evaluation now.