-
-
Notifications
You must be signed in to change notification settings - Fork 205
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Gemv with half precision cannot get correct result. #561
Comments
Thanks for sharing the complete test results here. I looked at your code briefly but I don't see anything obviously wrong. A few things to try:
But perhaps a better thing is to try to run the CLBlast tests themselves. Run CMake with
For example on my test device the output would look like this (for the HGEMV part):
So no failures here. If you run that, at least we know if the problem is in your test code or in the combination of CLBlast with your device. BTW, you can also test speed with the CLBlast 'clients':
|
Through this Issue I selected the expected high-performance operator, but found a problem. If the data type is float, the calculation result is correct, but if the data type is half, the calculation result is incorrect.
And I got the output log with
VERBOSE=ON
compile option, the fp16 output is obviously incorrect. I suspect data overflow, but 18000 actually does not exceed the range of FP16, so do you have any ideas?The text was updated successfully, but these errors were encountered: