PDA

View Full Version : Accelerating video card for de-bayering



laco
05-21-2012, 04:04 PM
I'm sure that most people will shoot Prores422HQ - including myself.
But when it comes to RAW post processing, good quality de-bayering needs a lot of processing power.

for the $25.000 RED ONE, RED came out with the $4750 Red Rocket which gave the users real-time 4K de-bayering.
Of course the priority is the get the Magic Cam ready and shipping by July, but will BMD ever come out with a video card that accelerates the de-bayering of the 2.5K CinemaDNG files?
Or is it needed?

Just a sidenote: I tested the 1920x1080 CinemaDNG RAW files from the Ikonoskop. In Davinci 8 I was able to get 20(!) nodes with real-time playback....

Barry Green
05-21-2012, 04:24 PM
Well, remember the Red is dealing with almost 4x as much data, and that computers and graphics cards are getting faster all the time. The BMC is using 1/4 as much data as the Red Epic in 5K. I'm sure that it won't be too very long before we can count on realtime support with modern graphics cards. But we should all be asking the various NLE manufacturers what their plans are for supporting graphics card acceleration of raw files and specifically CinemaDNG files.

laco
05-21-2012, 05:14 PM
Well, remember the Red is dealing with almost 4x as much data, and that computers and graphics cards are getting faster all the time. The BMC is using 1/4 as much data as the Red Epic in 5K. I'm sure that it won't be too very long before we can count on realtime support with modern graphics cards. But we should all be asking the various NLE manufacturers what their plans are for supporting graphics card acceleration of raw files and specifically CinemaDNG files.

I know it's not related.. but they say, with 3:1 compression Epic5K is 140MB/s. So the data rate is similar, but the number of pixels are much-much more, as you said. Also when converting R3D-s there are two tasks for the CPU: de-bayering, and uncompressing frames.
BMC files would only need the de-bayering part:)

Tzedekh
05-21-2012, 05:39 PM
Well, remember the Red is dealing with almost 4x as much data . . .
But is it? The Scarlet tops out at 55 MB/sec, whereas the BMC recording CinemaDNG tops out at 150 MB/sec (5 MB/frame x 30 fps). At its highest data rate, the Epic processes 50% more data than the BMC, 225 vs. 150 MB/sec. But at 24 fps, at the more common 5:1 compression ratio, the Epic churns through about 6 GB/minute while the BMC does 7.2 GB/minute -- 20% more.

taz291819
05-21-2012, 06:52 PM
But is it? The Scarlet tops out at 55 MB/sec, whereas the BMC recording CinemaDNG tops out at 150 MB/sec (5 MB/frame x 30 fps). At its highest data rate, the Epic processes 50% more data than the BMC, 225 vs. 150 MB/sec. But at 24 fps, at the more common 5:1 compression ratio, the Epic churns through about 6 GB/minute while the BMC does 7.2 GB/minute -- 20% more.

I think it comes down to how sophisticated the algorithm is. For instance, decoding 40Mbps MPEG-2 actually takes less CPU resources than 20Mbps h.264.

And since Adobe created CinemaDNG, one would think it would be a given for them to add GPU-acceleration for it in CS6.

Kevin Marshall
05-21-2012, 07:06 PM
Just as a side note: someone at Red (I think it was Stuart), clarified that the Rocket doesn't accelerate the debayering, but rather the decompression of the wavelet codec (RedCode). I'll try and find the quote and post it here.

In the BMC's case, the RAW is uncompressed, so wouldn't need an accelerator card like the Rocket.

EDIT: Found it. It was from Rob, though, not Stuart (my bad):


I want to correct a piece of wrong information. The reason we have a Rocket has nothing to do with the debayer/demosiac, it has everything to do with the decompression. Bruce: GPU can't do our decompression faster, trust me, I've tried. All those other cameras don't do compression + 4K/5K. ProRes 1080p ain't the same game.

Link (http://reduser.net/forum/showthread.php?71391-WHY!!!!-are-people-still-using-color-baked-DPX-s-F-amp-*-*-K!&p=930202&viewfull=1#post930202)

Tzedekh
05-21-2012, 07:09 PM
I think it comes down to how sophisticated the algorithm is. For instance, decoding 40Mbps MPEG-2 actually takes less CPU resources than 20Mbps h.264.

And since Adobe created CinemaDNG, one would think it would be a given for them to add GPU-acceleration for it in CS6.
Adobe doesn't really support CinemaDNG very well in CS6 -- it's still limited to 8 bits.

Note Suwanchote
05-21-2012, 07:29 PM
Adobe doesn't really support CinemaDNG very well in CS6 -- it's still limited to 8 bits.

what nle has better support? maybe CC in resolve then export?
it seems as if thats lame sinc adobe created dng

does the MPE support dng on premiere?

laco
05-22-2012, 07:32 AM
Just as a side note: someone at Red (I think it was Stuart), clarified that the Rocket doesn't accelerate the debayering, but rather the decompression of the wavelet codec (RedCode). I'll try and find the quote and post it here.

In the BMC's case, the RAW is uncompressed, so wouldn't need an accelerator card like the Rocket.

EDIT: Found it. It was from Rob, though, not Stuart (my bad):



Link (http://reduser.net/forum/showthread.php?71391-WHY!!!!-are-people-still-using-color-baked-DPX-s-F-amp-*-*-K!&p=930202&viewfull=1#post930202)

Ah I see.. then we are good to go:)

Brad Ferrell
05-22-2012, 09:01 AM
I'm sure Adobe will support the BMCC as it comes out or shortly thereafter. No worries. It's their MO.

Andrew_HD
05-22-2012, 11:15 AM
There is no need for any special card for debayering 2.5K DNG files- good CPU is enough. Data rate (of course when we have fast enough RAID) has nothing to do with debayering- it's resolution. It's easier debayer uncompressed RAW data than compressed eg. like RED one. With RED problem is resolution, with BMC 2.5K it's not a real problem anymore, so it can be uncompressed. Well optimized piece of software can debayer 2.5K files on modern CPU even few x faster than RT. There is simply NO good solution for DNG, as there was never need for it.

laco
05-22-2012, 01:50 PM
There is no need for any special card for debayering 2.5K DNG files- good CPU is enough. Data rate (of course when we have fast enough RAID) has nothing to do with debayering- it's resolution. It's easier debayer uncompressed RAW data than compressed eg. like RED one. With RED problem is resolution, with BMC 2.5K it's not a real problem anymore, so it can be uncompressed. Well optimized piece of software can debayer 2.5K files on modern CPU even few x faster than RT. There is simply NO good solution for DNG, as there was never need for it.

That was my thought when I played the Ikonoskop footage in Davinci.
Hope to get the same performance in NLEs too...

Peter Chamberlain
05-22-2012, 08:47 PM
DaVinci Resolve uses CUDA based GPU's for realtime debayer of CinemaDNG Raw, F65 4K Raw, Phantom Cineform and Arri Alex Raw. The decompression, if required, is handled by CPUs. RED's r3d is the exception where there are two choices, decompress and debayer in CPU or both in the RED Rocket card. There are no opportunities for debayer of the r3d in GPU.
Peter

laco
05-23-2012, 03:04 AM
DaVinci Resolve uses CUDA based GPU's for realtime debayer of CinemaDNG Raw, F65 4K Raw, Phantom Cineform and Arri Alex Raw. The decompression, if required, is handled by CPUs. RED's r3d is the exception where there are two choices, decompress and debayer in CPU or both in the RED Rocket card. There are no opportunities for debayer of the r3d in GPU.
Peter

Oh I see. I thought CUDA only takes care of the grading.
But if this is the case, then any application that uses CUDA can be adjusted to accelerate CinemaDNG RAW.

Peter Chamberlain
05-23-2012, 08:50 PM
They could, and a single GPU could do this, however, I'm not aware of other grading applications that can parallel process across multiple CUDA GPU's and this is a significant Resolve feature. The debayer algorithms are complex as is writing with CUDA at the low level we do. DaVinci Resolve has been image processing with multiple GPUs since 2008 so we have a lot of experience in the single application, extremely high quality image processing using CUDA with multiple parallel processes across multiple GPU cards.

Peter

laco
05-24-2012, 01:55 AM
They could, and a single GPU could do this, however, I'm not aware of other grading applications that can parallel process across multiple CUDA GPU's and this is a significant Resolve feature. The debayer algorithms are complex as is writing with CUDA at the low level we do. DaVinci Resolve has been image processing with multiple GPUs since 2008 so we have a lot of experience in the single application, extremely high quality image processing using CUDA with multiple parallel processes across multiple GPU cards.

Peter

I meant NLEs under "other applications". I have a Davinci system, so I don't want to change:)
Also clients like it a lot...
I'm interested what will be the new things in Davinci9, I hardly can wait!