Closed akroviakov closed 2 weeks ago
This PR adds support for loading 1x16xf16 tiles by specifying the number of destination vector-registers (64B each) to be at least 1 (even for the cases when the actual payload is <64B).
1x16xf16
This PR adds support for loading
1x16xf16
tiles by specifying the number of destination vector-registers (64B each) to be at least 1 (even for the cases when the actual payload is <64B).