Software-based decoding is a different beast altogether. So long as the instructions sets are the same, we are dealing with identical images, regardless of who manufactures the hardware.
When you use a hardware-based decoder, the video data is processed through a specific path that has DXVA API calls for hardware-accelerated decoding. To a certain degree, the flow of data can still be handled differently on dissimilar pieces of hardware using the same DXVA accelerated decoder. For the consumer, you have no way of knowing how much of the DXVA pipeline has been implemented (and to be fair, you probably don't care). That is why hardware decoding on WMP12 and PowerDVD can still produce a different image, even though they both use EVR and enjoy hardware-accelerated decoding.
For software-based decoding, we are using FrameShots to capture specific frames. Since it uses software-based decoders, there was no way we could do the first part of our analysis with it. Furthermore, we aren't able to compare it head-to-head against the hardware-accelerated decoding shots generated from WMP12. Why? This program doesn't use EVR. In fact, it uses Video Mixing Renderer 9 (VMR9). For that reason, we are only able to compare two software codecs against one other.
FrameShots uses a custom DirectShow filter that sits on the DS filter tree between the video decoder and the renderer. This means we are actually eliminating the video renderer as a variable to a certain extent, something we could not do with WMP12. The difference now is that we're using the DS filter to capture a specific subset of video data.
It is hard to pick out the differences here, but if you look at the edges of objects like the Humvees and the nose of the white plane to the right, aliasing seems a bit heavier with ffdshow. Even though we captured the same frame in both cases, aliasing is occurring in different places. This makes it all the harder to call out a clear winner here. After all, a jagged line is a jagged line.
These two images look basically identical when you overlay them. There are only two noticeable differences. With ffdshow, you get a bit more detail on the light reflecting off the top of the car. Yet, strangely, the decoder (or the renderer) seems to drop the top part of the colon in the top left time stamp.
Skipping directly to GP, we see a bit of a difference. MainConcept shows less detail and appears a bit smoother if you are looking at it pixel by pixel. In ffdshow, Gwyneth's hair appears a bit sharper, but overall, the picture looks grainier, too. Strangely, MainConcept is the one that drops half a colon in the time stamp.
What would image quality comparison be without explosions? This seems to be another one of those cases where anything in high motion shows little variation. Honestly, we see more differences creep up in the slower scenes. I wish I could have shown the explosion screenshots for hardware decoding, but the native screen capture exceeds the limits of our image server. We have posted them on ZumoDrive, if you want to examine them yourself.
On a final note, when you use FrameShots without having installed ffdshow tryouts, it will default to the next available decoder. On our system, this happens to be the H.264 decoder from our installation of MainConcept Reference v18.104.22.1685. The decoder is an unknown version, so we are simply listing the program version it was included with, and we disabled hardware-accelerated decoding in our quality comparisons of the MainConcept H.264 decoder. We are aware that ffdshow tryouts recently added a limited degree of DXVA support, but it isn't part of the latest stable build.
There are a plethora of software decoders available on the market. We are only selecting ffdshow tryouts (build 3154) and MainConcept to make a point; all software decoders are not created equal.
On that note, we want to present an interesting chart. It seems that, even in the same software-based decoder, we can get different performance results. When we play back our unprotected H.264 source in PowerDVD (build 10.0.2325.21), the results for hardware-accelerated decoding fall to 5% CPU utilization and under for all three graphics configurations.
But something strange happens when you disable hardware-accelerated decoding. In theory, everything should be running on the Core i5-2500K. Yet, in software-only mode, the numbers indicate otherwise. Somehow, simply dropping in a GeForce GTX 580 results in the lowest CPU use (remember, this is all running on the host processor). With Intel's integrated graphics enabled, the utilization is only marginally higher. Perhaps that's a result of the HD Graphics engine using resources that'd otherwise be freed up for the processing cores. More alarming, though, is that adding a Radeon HD 6970 spikes CPU utilization more noticeably. Indeed, if you look at the graph, you can see the CPU spikes occur in the same places as the software-based decoder is dealing with processing-hungry scenes at the same times.
We raised this issue to AMD and CyberLink, but we still have no satisfactory answer as to why it's occurring. We'll update this space should we get a clear answer.
[Update 2/4/2011]: It turns out this was a bug, likely on Cyberlink's side. We contacted Corel for a copy of WinDVD. The results speak for themselves. No matter what GPU you drop in, CPU decoding is should be the same provided you use the same software decoder.
- Image Quality: Examined
- Intel, AMD, And Nvidia: Decode And Encode Support
- Transcoding Quality Revisited: CUDA Problems?
- Test Setup
- Hardware Decoder Quality: Examined
- Software Decoding: All CPU, All the Time
- Full Blu-ray Transcoding Speed: APP Versus CUDA Versus Quick Sync
- Small Clip Transcoding Speed: APP Versus CUDA Versus Quick Sync
- Transcoding Quality: APP Versus CUDA Versus Quick Sync
- Transcoding Quality: Rated By Software Title
- Playing Devil's Advocate: "There is No Spoon"
- Inside The Black Box: GPGPU Encoding
- Final Words