Need for Speed

[Steve Sherrick]

So what is the biggest hurdle that software developers will be facing when writing programs for transcoding R3D files (when SDK is released of course)?

Is it GPU, CPU, RAM, combination of all three? With the amazing amount of processing speed out there, and oodles of RAM that can be installed, I'm a bit surprised by how long it takes to process RED footage. No slight against Red or Assimilate meant here, as I'm just trying to understand what issues stand in the way of a high quality, fast, and affordable software transcoder. I know the old saying about cheap, fast, and good pick two, but in this day and age I'd hope we could get all three to process r3d files. If someone could shed some light on what it takes to process these files, it would help me put things into perspective and maybe help me set my expectations for what we may see in the future.

Please, no attacks on current solutions. I just want to know what is possible from people who are in the software development realm.

Thanks!

 

[Jay A. Kelley]

I can tell you that it's a choice. Not a requirement. Some will use just the CPU and others may use the GPU as well.

In the case of Cineform, the initial transcode from r3d to a cineform RAW file with which you can edit is BLAZING fast. I think I've heard up to 11 frames per second or better.

This is a file which would allow you to edit, color correct and then complete post on your project.. HOWEVER at the end of that chain, you will need to RENDER to your final chosen output format, and that will be damn slow, but better to do it at the end, rather than the beginning.

Jay

 

[Adam Glick]

...HOWEVER at the end of that chain, you will need to RENDER to your final chosen output format, and that will be damn slow, but better to do it at the end, rather than the beginning. ..

Not a trying to be "smart" here Jay, but in what circumstances would this workflow (a short transcode up front and a long render at the end) be any less painful (or faster) than any other of the emerging workflows for r3d post?

There are ALWAYS tradeoffs.

 

[Martin Weiss]

Well, in the end you would have much less material to transcode. Especially in documentary work there would be huge benefits.

But then, patience has the tendency of running out towards the end of a project.

 

[Dan Hudgins]

Harddrive speed... RED CODE to TIFF speed?

Harddrive speed... RED CODE to TIFF speed?

I was surprised to find that most of the speed issue in my uncompressed 4K workflow was with the OS, something like 60% to 80% of the time for doing CC of 4K frames was in the loading and saving of the image files, not in the processing of the pixels.

You can check the harddrive speed by copying a uncompressed set of maybe 24 4K TIFF frames from one harddrive to another. And from one folder to another (having the result file on another drive may speed things up). Doing your I/O from a large RAM disk or Solid State HD array may also help. If you cannot copy 4K TIFF frame files faster than one every 5 seconds, then any processing will be on top of that I/O speed.

I have cut the time down quite a bit so that on my older machine I can CC-resize-letter_box 4K uncompressed frames in about 29 seconds +/-. With RED ONE (tm) data being compressed to 10:1 you would expect that about 3 seconds per 4K frame copy with CC would be possable, but doing the full de-Bayer might add up to 15 seconds if it was going in just one GPU or CPU, with parallel processing you might be able to go from RED CODE to TIFF in 3 to 6 seconds per frame if the harddrive array was fast enough.

Question: What processing speed throughput are you now getting going from RED CODE camera 4K footage to TIFF 4K frame set color corrected in frames per minute?

 

[REDHKSC]

As a decent post business chef,

Why not RED has an Uncompressed recording in DPX file format ? 2K 1st then 3K and 4K in future.

The day I trusted RED at IBC 2006 was the RAW Uncompressed port option from RED one, and may I ask Jim that where have the Uncompressed ports gone ?

It seems RED has treated me, and I am a serious imaging post gardener in my life. Was I joking ? NO.

Dalsa / ARRI / VR / SI are all have uncompressed RAW output in file
fomats and many DFRs ( Digital Field Recorder ) to support them.

Stew.

 

[Steve Sherrick]

My times have varied depending on what settings I'm using but on a Mac Pro Quadcore 2.8ghz with 16gb RAM with full debayering I seem to get about 2 frames per second.

 

[Joel Kaye]

Not a trying to be "smart" here Jay, but in what circumstances would this workflow (a short transcode up front and a long render at the end) be any less painful (or faster) than any other of the emerging workflows for r3d post?

There are ALWAYS tradeoffs.

The end transcode is the smallest transcode since it's only the final selected material.

The upfront transcode to an editable codec is the killer because it's ALL the raw material usually. A pretty economical shoot has 10 hours of footage for 1 hour finished. Right now a full RED debayer to ProRes would take 200 hours to get that 10 hours converted. So any workflow that presents a fast upfront transcode is vastly superior.

The best solution would be turn REDCODE into a true Codec that could be accessed by any app with no up front transcode.

 

[Adam Glick]

It's not the prettiest workflow -and it may not be suitable for ALL projects, but there are facilities that are playing r3d out of the redBOXX HD-SDI into a KONA-equipped MacPro and creating full-quality HD /2K ProRes or DPX in realtime.

There is some "manual labor" required for naming the files & directories created during capture, and much of the metadata is lost, but it's WAAAAY faster than any software transcoding alternatives.

It would be very cool if REDCODE RAW could be implemented a true CODEC, but if that were possible or feasible, I'm sure RED would have delivered (or at least discussed) it as a future option.

 

[M Most]

In the case of Cineform, the initial transcode from r3d to a cineform RAW file with which you can edit is BLAZING fast. I think I've heard up to 11 frames per second or better.

Well, yes. Because it's not doing any kind of debayering or image reconstruction. It's decompressing, then recompressing on raw data - so among other things, it's working with only 1/3 the amount of information as it would have to if it were creating RGB images.

 

[M Most]

The upfront transcode to an editable codec is the killer because it's ALL the raw material usually. A pretty economical shoot has 10 hours of footage for 1 hour finished. Right now a full RED debayer to ProRes would take 200 hours to get that 10 hours converted.

Not if you do your offline in standard def - like, oh, about 98% of network television programs and about 75% of studio motion pictures. That conversion would, in all likelihood, take less than the 10 hours of material to complete.

 

[Joel Kaye]

redBOXX HD-SDI into a KONA-equipped MacPro and creating full-quality HD /2K ProRes or DPX in realtime.

I thought the debayer was CPU based. Are your sure it's a realtime 2K full debayer? Or is it just pulling the 2k stream and converting it? If that's the case I think RedCine going out a Matrox could do something similar. Or you can just drop your proxies in Compressor and make a 2k or 1K ProRes proxy fairly quickly. (edit: Yes, Mike you're right you can do this, thanks for the lesson.)

Mmost brings up a good point that Cineform is apparently not doing a debayer either.

RED has mentioned REDCODE being supported in many apps but hasn't clarified what that really means.

 

[DanWood]

Technology is moving fast. soon we will have 4K video cameras in our mobile phones.
I am reading "Motorola, Kodak Launches 5 Megapixel Camera Phone ‘ZN5’ "
That is video capable? What would be next?

This is unrelated to the thread. Sorry.

 

[Dan Hudgins]

Why no raw...

Why no raw...

Why not RED has an Uncompressed recording in DPX file format?

The answer seems to be that most or many users will not settle (it is supposed) for 24fps since they are into slow-mo or non cinema film out.

The bandwidth issues are easy to do for 24fps, but not for 120fps.

For myself I care not for much over 24fps, but it was not my decision. There are easy ways to do raw recording. Since one CF card can record 10:1 compressed, 12 or 24 CF cards in parallel could record 12 bits uncompressed.

The RED engineers seem to be oblivious to the obvious simple solutions to uncompressed recording at 24fps or for marketing reasons have decided not to offer a low cost simple RAW recording option. If it is a technical issue I could help them with how to do it, if it is a marketing reason that they want to do a over blown disk array that will cost so much no one will be able to buy one, nothing I can do about that...

==

About my through put question, how long does it take to process the RED CODE 4k footage into 4k TIFF files color corrected, like the example before if you have 10 hours of 4K RED CODE, how many hours does it take to make 10 hours of 4k uncompressed 16bit (48bit RGB) TIFF files color corrected? (and on what computer setup)

I am interested in the through put question to set a benchmark for my goals, since they are starting with 10:1 compression you would think that the processing time would be half what I get with 48 bit uncompressed 4K (input 10:1 output 1:3), but the de-bayer takes some time, so the figure should be closer to my uncompressed-to-uncompresed maximum maybe?

 

[Dan Hudgins]

4K to 2K w/o de-Bayer

4K to 2K w/o de-Bayer

Are your sure it's a realtime 2K full debayer?

If the camera has 2K or HD res output that can be done without doing a de-Bayer, the signal comes from four pixel groups, that can result in more Aliasing, which may not be appparent with the signals getting low pass filtered in hardware.

A high quality de-Bayer like DCRAW (tm) can take about 30 seconds per 4K frame. Just doing a color output 1/4 size with hardware can be live.

The fastest way to get an edit might be to make SD proxies live from the video output, and later conform on the timecode for only the 4K frames that are needed?

This link from Arri talks about doing 4:2:2 YUV live and 4:4:4 RGB slower in software,

http://www.arri.de/news/newsletter/articles/09211103/d20.htm

 

[Steve Sherrick]

Although the issue of uncompressed output is a valid question to raise, I prefer we stick on subject which is, is it possible to write a software transcoder that can achieve high quality, fast exports (much faster than Redcine), and be fairly affordable (that is subjective, so let's just assume that it would be in line with the Final Cut Pro/Apple approach). I'm wondering if the processing power is there to do this in a desktop system. If not, and if many of the programmers on here think it's just not possible yet without the implementation of several machines linked together in a render farm system, then that helps me as I look for affordable post solutions. At the moment, Scratch Cine is just outside the budget or I would have already brought that in.

 

[REDHKSC]

So why not take a look at SPEEDGRADE XR ?
http://www.speedgrade.com/xr/

Do all we know SDK is gonna to be released to others in AUGUST 2008.

As a leader as RED one camera owners , and we should have our options to play and fight........

I want to BREAK FREE....

STEW.

Although the issue of uncompressed output is a valid question to raise, I prefer we stick on subject which is, is it possible to write a software transcoder that can achieve high quality, fast exports (much faster than Redcine), and be fairly affordable (that is subjective, so let's just assume that it would be in line with the Final Cut Pro/Apple approach). I'm wondering if the processing power is there to do this in a desktop system. If not, and if many of the programmers on here think it's just not possible yet without the implementation of several machines linked together in a render farm system, then that helps me as I look for affordable post solutions. At the moment, Scratch Cine is just outside the budget or I would have already brought that in.

 

[Dan Hudgins]

Speed? High Quality? At the same time?

Speed? High Quality? At the same time?

Although the issue of uncompressed output is a valid question to raise, I prefer we stick on subject which is, is it possible to write a software transcoder that can achieve high quality, fast exports (much faster than Redcine), and be fairly affordable (that is subjective, so let's just assume that it would be in line with the Final Cut Pro/Apple approach). I'm wondering if the processing power is there to do this in a desktop system. If not, and if many of the programmers on here think it's just not possible yet without the implementation of several machines linked together in a render farm system, then that helps me as I look for affordable post solutions. At the moment, Scratch Cine is just outside the budget or I would have already brought that in.

If you want an idea of speed, try loading a JPG or JPG2000 frame into an photo edit program, time it with a stop watch, did it take more that 1/24th of a second? It takes twice as long to load and save a frame.

"High quality"? If you load a compressed image alter the pixels and save it compressed you get "double compression losses". Re-compression of a compressed image can look nasty, try loading a JPG, alter it and save it back as a JPG, then compair the compression artifacts of the compression artifacts. This gets multiplied when the program is released compressed, then gets recompressed at a lower rate for broadcast, and then maybe the cable company re-compresses it again to save more bandwidth.

A real time 4K RGB 12x3 bits image has this bandwidth:

4096x2304x36x24=8,153,726,976 bits/sec

If processing that on a cpu was about average one bit per clock you would maybe need a 8GHz+ processor (you figure it takes at least 8 clock cycles per pixel to do something useful). The computer memory bandwidth cannot work with data rates like that. HD playback is done using GPU in parallel processing, the compressed data is moved into the parallel GPU so the amount of data is 100:1 less or greater.

There are very real reasons to go from RED CODE to uncompressed TIFF for all post operations right up to the film recorder to try to get a high quality result, and that would be non real time post.

Someone here shoud know how long it takes to convert 10 hours of RED CODE to 10 hours of TIFF 4K uncompressed frames, that would set the long limit, the short limit is live to record video out as HD/SD, and everything else would be between.