Lunderberg commented on issue #8294: URL: https://github.com/apache/tvm/issues/8294#issuecomment-866243787
From what I can tell, the float16 values are packed into uint32 when not in use, and are cast to the float16 when used. I think there will need to be some special handling to pad out the calls to `make_uintN` (e.g. `make_uint2` for a length-3 array of float16s packed into 64-bits with the last 16 bits empty), or the packing will need to change. Either way, looks like there isn't an immediate fix, and I don't know the cuda codegen as well as the vulkan side to say whether other issues would come up as well. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
