search

oneapi-src / oneDNN

oneAPI Deep Neural Network Library (oneDNN)
https://uxlfoundation.org
Apache License 2.0
3.64k stars 1.01k forks source link

supported matmul data types #2196

Closed jinz2014 closed 1 week ago

jinz2014 commented 4 weeks ago

https://oneapi-src.github.io/oneDNN/dev_guide_matmul.html

Is FP64 supported ?

Does the destination include f32 when the type of input and weight is fp16 ?

f16 | f16, u8, s8, u4, s4 | f16, u8, s8

Thanks for updating the table with all supported types.

raistefintel commented 4 weeks ago

Hi @jinz2014. Yes, these data types are supported. Please note that f64 is only supported on Intel(R) Data Center GPU Max Series. We will update the table accordingly. Thank you for your feedback.

jinz2014 commented 4 weeks ago

oneDNN is portable. So it will be helpful to know the CUDA/HIP support.

raistefintel commented 3 weeks ago

Hi @jinz2014, you can find general information on Nvidia/AMD backend support, including supported data types, in the NVIDIA backend readme and AMD backend readme.

vpirogov commented 1 week ago

Fixed by #2217.