Quote:
Originally Posted by LeXXuz
Probably stupid question, but are there any differences in quality between these indexing methods? Or is it first of all a matter of speed? To be honest, I don't understand the detail of this indexing process at all. Logic tells me that it could not affect quality until indexing is not meant in the matter of just finding desired information in the source file. Multiple ways lead to Rome, right?
|
Not quite right. DG tools are designed from the ground up for frame accuracy. DSS2() and the others are not and run into issues in several common scenarios. For example, they might serve the correct frame's data, but it shows artifacts because the source filter has not injected the needed SPS/PPS. Or they can simply deliver the wrong frame's data because the timestamps are not perfectly monotonic. Or they get screwed up by trim() in your script. If you need assured accurate random frame access, then DG tools are the way to go. DISCLAIMER: I have an interest in the DG tools. But you can find similar assessments by others on the web.
Quote:
|
I never thought about GPU indexing, but if it comes with a noticeable speed increase, why not. Is a 8400 fast enough for this task or would a more expensive card speed up things further more?
|
The indexing is not done on the GPU. Decoding the video is done on the GPU. So the Avisynth source filter's workload is offloaded to the GPU. Indexing simply enables frame accuracy in the source filter.
For purely decoding, a fast CPU can beat the GPU (although Fermi is starting to challenge that wisdom). But in the vast majority of cases where Avisynth is used, the user is *transcoding* and the decoding rate is not the bottleneck, so offloading the decode to the GPU can give an overall performance gain to the transcoding operation by leaving more CPU bandwidth for the encoder.
The Fermi cards will outperform the 8400GS but are MUCH more expensive. The 8400GS has a VP3 engine and is very fast; it's the current sweet spot for price/performance. Refer to the DGDecNV Benchmarking thread for more details.