Support fp16 for more cublas/cusolver?

Hello, is there a future plan on adding geam and qr in half precision? If not, is there a good source where I can find how to optimize my kernels so that I can to be as close as possible to the cublas version in terms of wall clock times?

This is part of my masters thesis if it plays any part

Thank you!

I’m not able to comment on future plans. You will find recent discussions here on the library board that discuss optimizing kernels to come close to geam behavior. You can also express your desire for new features to the development team using the bug reporting facility.