-
Notifications
You must be signed in to change notification settings - Fork 3.6k
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
The recognition results with Vulkan are so bad #2400
Comments
And the Vulkan version spend lots of time to process it |
Just adding here I've also found that Vulkan transcriptions can be bad on some GPUs and good on others. E.g. on an AMD RX 7900 XT it gives a very poor result (as far as I can see nothing correct, and way too short). Whereas on the integrated AMD GPU on the same machine it gave a result that was correct, I think identical to the CPU result. And I've also found the Vulkan implementation to be slow, e.g. something that takes around 1 minute on the CPU, takes about 40 seconds with Vulkan. Whereas CUDA (same GPU as when running Vulkan, Nvidia RTX 3080 Ti) and Metal (M1) both take around 17 seconds. |
I've found that Vulkan killed performance compared to OpenCL. A transcription result that took only 13 seconds with OpenCL now takes 40+ with Vulkan, but it's still faster than CPU, which is roughly 60 seconds. |
Yep. I am using v1.7.1 and can confirm that performance with Vulkan is terrible(only 1.5x real time). GPU is RX 7600. I agree removing OpenCL support was a mistake. Could you tell me the version that you using that had support for OpenCL? |
cbacb76 is the last one with support for OpenCL |
Hello guys, I tried to compile it with the latest version with Vulkan but failed: Failed to compile it with Vulkan
Thus I reverted the codes to 32f88af
So I compiled it with Vulkan version and run main.exe to process the wav file, but the poor quality of recognition is unacceptable.
Here are the results of Vulkan version:
Here are the results with CPU:
The text was updated successfully, but these errors were encountered: