Intel QuickSync Decoder - HW accelerated FFDShow decoder with video processing - Page 3

Honestly, thats what images have a stride for, so that even non-mod16 images can be accessed with aligned memory functions. Forcing a border thats not there in the source material is just wrong.
What i do in those cases is setup a temporary buffer with the appropriate stride to copy the image to from the "slow" GPU source, and afterwards if its required to remove the added stride pixels, do a plain memcpy to get rid of it again. Its alot faster then trying to get it right in the first step (because the "slow" source only works properly with aligned access) - and if running against a renderer, it will usually request a stride which is at least mod 16, most of the time even a power of 2.

You don't want to know about the filters, yet you do ugly stuff to make it work with broken filters? Your logic escapes me.
If you expect your decoder to be used as-is, it should not do anything except deliver the untouched image.

In theory, you can actually propose a stride yourself, make biWidth mod 16, and specify the actual width in rcTarget (like a video decoder would request it), but in reality thats something ffdshow should already do for you - i suppose its too bad for that, though.
There is also the question if the typical post-processing filters properly handle that, anyhow. I haven't tested it, but maybe i should. The MSDN at least defines the behaviour that way.

I am having a strange problem that I will try to explain the best I can.

I am using the Quicksync decoder on a dual GPU setup. I have my screen connected to my AMD 6570 and have my Core i3 disconnected. Using your small guide I have my system setup to display on my AMD 6570 but use the Sandy Bridge GPU to decode the video. For a splitter I am using LAV 0.37.

The problem I have is with WMC. I have certain MKVs that if I start them in WMC while WMC is full screen the video will never start playing, it just sits on the blue WMC background indefinitely. If I start the video in WMC while WMC is windowed the video plays back normally, I can even set WMC to full screen after the video starts and everything works fine.

I am going to look for a way to cut up one of the MKV's in questions because as it sits the file is 24 gigs. If there is anything else you need from me to help sort this out, just let me know.

I am using WMC that comes with Windows 7 and it is the x64 version. I also uploaded a 100 MB clip for you here: http://www.multiupload.com/DFEXJLCCW0

Practical possibly, but it does violate the original image, and is the worst you can do for quality. If a post-processing filter needs a special stride, it should insist on getting it - and not just go tits up if it doesn't. Thats my attitude to the whole issue.

In any case, it does look like i'll need to re-implement the MSDK decoder then instead of using your DLL (can still use it as a template, i suppose), i'm not willing to accept such compromise.

I use Win7DSFilterTweaker to use FFDShow over the built in windows decoder. You can grab it at http://www.codecguide.com/windows7_p...er_tweaker.htm. You can also use it to disable the built in windows codecs.

You are correct. I double checked this and in fullscreen ffdshow isn't loaded if I have QuickSync selected. If I use dxva or software rendering in ffdshow everything works fine.

I'd like to report some ffdshow bugs that exist in your builds only.
1. some audio, such as TrueHD causes a crash in ffdshow.ax - this does not happen with sourcforge builds.
2. The ffdshow dxva decoder is broken and If I replace the ffmpeg.dll with one from a sourceforge build, it fixes it.

http://www.demo-world.eu/trailers/hi...n-trailers.php

Search for truehd and you'll find a few trailers.

My issues(s) may have been specific to your .17 release.

I have since moved on from those binaries. If you have a testbuild that you can hook me up with that is not using the .17 binaries I'd be interested in giving that a shot for testing purposes (both x64 and x64). If you're close to releasing .18, I'll test them and report back to this thread.

Here is the TrueHD sample that crashed on ffdshow.ax
When I installed a sourceforge build, It no longer crashed.
The DXVA problem MAY exists in the trunk - I need to test further when I have the time.

nice work ericgur. my ffdshow issues with your builds are resolved.
I have only tested 32bit so far, but it does look good.

Celeron/Pentium SNB are too weak to use with AMD GPU under madVR because of the lack of a hardware video decoder equivalent to CUDA Video Decoder (there exists an API, however, called OpenVideo Decode (OVD)). Now ericgur's hybrid HW accelerated system instantly fixes the problem. This method can be also used with a GeForce card, in particular, that does not have enough CUDA cores (e.g. 48) to do deinterlacing and post-processing simultaneously under CUDA Video Decoder.

	Decode	Deinterlacing	Post-Processing
libavcodec+yadif+madVR	CPU	CPU	Shader Engine
Intel QuickSync+yadif+madVR	Intel MFX	CPU	Shader Engine
CUVID+madVR (NVIDIA only)	Video Processor	CUDA cores	CUDA cores

(Shader engine = CUDA cores [NVIDIA's term])

System used

- Processor: Celeron G530
- Motherboard: H67 chipset mb (GIGABYTE GA-H67MA-UD2H-B3)
- Memory: DDR3-1066 2 x 2GB (actually DDR3-1600, but it runs only at 1066 under G530+H67)
- Graphics: Radeon HD 5570 DDR3-1800 1GB

- Player: MPC HomeCinema
- Source/Splitter: LAV Source Splitter
- Video decoder: ffdshow video decoder (libavcodec or Intel QuickSync)
- Deinterlacer: ffdshow video decoder (yadif)
- Video Renderer: madVR with HQ (SoftCubic 100/Lanczos 4 taps/Lanczos 4 taps)
- Test clips: the same as this post

libavcodec+yadif+madVR

	SD film	SD video	HD film	HD i video	HD p video
Dropped frames	0	0	0	2774	3616
CPU usage (average)	22%	27%	56%	95%	99%
GPU usage (average)	27%	53%	25%	5%	6%
Rendering time (average)	7.47 ms	7.30 ms	4.11 ms	6.15 ms	7.77 ms
Power consumption (average)	66W	73W	73W	81W	81W

CPU is completely taxed for HD i video and HD p video, resulting in lots of dropped frames:

- HD i video: 1080i60 decode + deinterlacing
- HD p video: 1080p60 decode

Intel QuickSync+yadif+madVR

	SD film	SD video	HD film	HD i video	HD p video
Dropped frames	0	0	0	0	0
CPU usage (average)	23%	28%	33%	77%	57%
GPU usage (average)	27%	53%	25%	46%	48%
Rendering time (average)	7.53 ms	7.44 ms	4.57 ms	4.33 ms	4.16 ms
Power consumption (average)	66W	73W	68W	81W	79W

Decode is now done by Intel MFX (Multi-Format Codec) fixed function in Intel HD Graphics, that offloads CPU considerably.

Remarks

1. This is basically the discrete GPU version of this post, where Intel HD Graphics + madVR MQ is used instead of dGPU + madVR HQ.

2. CyberLink Video Decoder (HAM) + yadif won't work for interlaced contents because CyberLink Video Decoder (HAM) outputs video as progressive.

Yes, Intel QS + EVR works and all AMD's post-processors will work (well, AMD [or NVIDIA or Intel]'s post-processors will work under *any* video decoder [NV12 output] + EVR ). But if you use EVR, why not use AMD UVD + EVR under DXVA API? This is even processor-independent (i.e. works with any processor, AMD or Intel). Is there any significant advantage of Intel MFX over AMD UVD?

Personally I see Intel QS or CUVID is useful only under a non-EVR video decoder, in particular madVR (or if I want to insert a filter between decoder and EVR, such as subtitle filter).

Whatever-based EVR, whichever GPU's post-processors will work with whatever video decoder outputting NV12 + EVR (I have been testing throughly).


Nowadays DXVA works pretty well, in particular with TMT, PowerDVD and any DirectShow media player with Microsoft DTV-DVD Video Decode/ArcSoft Video Decoder/CyberLink Video Decoder, even if you limit yourself to Intel HD Graphics (let alone AMD and NVIDIA). When was your last experience with DXVA?

In my experience, Microsoft DTV-DVD Video Decoder (from the latest build of Windows 7) works pretty well with any GPU (AMD, NVIDIA or Intel) under DXVA+EVR, that's my favorite DXVA video decoder (for AVC and MPEG-2). What problem do you see in Microsoft DTV-DVD Video Decoder?

Hmm, if you don't mind, will you post a couple of your lots of test clips that can't be decoded by Intel HD Graphics + Microsoft DTV-DVD Video Decoder somewhere? (You say "a lot". How many such clips do you find?) I would like to see them myself (out of curiosity).