[ISSUE]: DirectML AMD GPU is not used for Index file + overflow issue? #222
Description
Voice Changer Version
(B2332) voice-changer-windows-amd64-dml.zip
Operational System
Windows 11 24H2
GPU
AMD Radeon RX 6700XT
Read carefully and check the options
- I've tried to Clear Settings
- Sample/Default Models are working
- I've tried to change the Chunk Size
- GUI was successfully launched
- I've read the tutorial
- I've tried to extract to another folder (or re-extract) the .zip file
Model Type
RVC
Issue Description
This likely applies to all AMD GPU users.
Basically the issue is, if you're including and using the Index file, all of it is handed over to the CPU. This uses a way too much processing power, especially on smaller chunksizes. It also increases the expected chunksize by a large sum if the index file size is over 1000MB.
To put it in a bit of a perspective:
100MiB index file needs +50ms chunk delay
400MiB index file needs about + 100ms chunk delay
1000MiB index files need about +800ms chunk delay
Surprisingly, even though an increased chunksize is demanded, the CPU isn't actually that busy with it all after a delay of above 300ms (which is between 22%~12%), yet at around 150, it uses at least 60% CPU bandwidth.
The issue is actually worse on ONNX RVC models. With ONNX models, everything requires a ton more delay even without the index file. I think this may be a related thing. Converting to ONNX is a bad idea as such.
I have only tested this with regular pht (safetensors) and onnx files.
Application Screenshot
(Note: All of these involve regular pht (safetensors) RVC model files.)
Testing without Index file.
Testing with index file turned on. Index filesize 112MiB.
Testing with index file turned on. Index filesize 400MiB
Testing with index file turned on. Index filesize 1.2GiB (notice the severe amount of chunk size / delay I need to add)
