Let's petition Canon to make a CUDA version of DPP

Currenly with an Intel Quad Q6600 processing 1000 images from raw files to large jpeg using Canon DPP takes about 30 minutes. Using something as cheap as an nVidia 8800GT with CUDA support this would only take 5 minutes!

Contact:
http://www.usa.canon.c​om/consumer/controller​?act=ContactCanonAct

Would be of absolutely no benefit to me.
1. I've never shot 1000 images that needed identical processing in my life. Two or three at most.
2. Even seemingly similar shots can end up being processed totally differently.
3. I only use DPP as a RAW converter.

1) CUDA requires massively parallel computing, RAW encoding is generally only splittable to about 16 threads.
2) CUDA requires an nVidia graphics card, matrox, ati, etc can't support it
3) CUDA is limited by the buffer of the graphics card. each raw is 10mb (mkiii is 14 average, 5dmkii about 24, so giving you a generous lead there), so assuming that the total overhead is just 12mb (likely ten times that), your average 512mb 8800 card can do at most 50 files at a time. Assuming 16 threads though, we're down to just 8 files at a time. Your processors can probably do 4 files at the same speed, so in reality you get 2x, not 6x.
4) Costs a lot of money to do that. I would rather have canon spend that money to get better compression into dslrs for video, so instead of 1080p at 45mbps we have it at 15mbps, and that gives us the option of 90fps

is DPP even multi-threaded yet?

Wow,
Guess I have to educate you fellows...

• I edit each and every file in DPP and they are all different. When I am finished I convert them to .jpeg. I then edit in another editor if needed.
  
• Using DPP as a raw converter takes time, which I want to speed up. Good grief!
  
• CUDA enables massive parallel computing which a RAW converter will take advantage of.
  
• CUDA does require an nVidia card, true. However code written in CUDA can be converted to OpenCL in short order.
  
• At the most 50 files? This limitation bothers you? That would be very fast! But that limitation is not correct. The onboard video RAM while only having 512 can pull from system memory, just a touch slower. Good grief!
  
• Yes DPP does use multi-core hence my testing confirming what it does with 500 files on my Quad. It is very nice. I would rather see 50 or more files worked on at the same time instead of how it is currenly programmed.

pcunite wrote in post #8007767

• CUDA enables massive parallel computing which a RAW converter will take advantage of.

CUDA does enable, but is also requires it to be of any practical use. Unless you are talking about batch processing, the gains will not be nearly as high as you expect. And CUDA only works with video memory unless in vista, in which case all memory is pooled anyway and CUDA can work as it was originally intended. Which brings up another issue, what of the mac users? They won't see anything, especially since opencl support is very limited and video display model in macs is outdated.

Hey man, I was just asking... I also use a program called BreezeBrowser which is not multi threaded, it uses the Canon engine for conversion so I thought maybe that = DPP also not being multithreaded, but I forget DPP is different from Canon's other converter utility.. ie: better.

I'm all for more speed,. I was one of the early 64 bit adopters,. "but CDS, why woyukd you ever need more than 3GB of system ram?"
Early multi -proceesor system adopter (my first was a Dual Pentium II 333MHZ LX board.. Tyan Thunder II) " But CDS, what apps will ever be multithreaded"?
I use a ramdrive and Solid state Drive for swap files to work around non 64 bit apps ram limitations "But CDS, why would you need a scratch disk that runs at system memory speeds"?
And an Early Multi Core (still running my Dually Dual Core Opteron Board now.. ) .. ie: 4 cores before you could get a four core CPU.

I was just reading about these Tesla cards NVidia makes for unreal added CPU power,.
http://www.nvidia.com/​object/IO_43499.html

.. and would love a new Dual socket I7 board with the PCIExpress 2 16 lane slots for a Tesla... if i could get my RAW and PSCS to use it all!

So do we need "cuda" to use those Teslas on our RAW converter?

pcunite wrote in post #8007767
Wow,
Guess I have to educate you fellows...

• I edit each and every file in DPP and they are all different. When I am finished I convert them to .jpeg. I then edit in another editor if needed.
  
• Using DPP as a raw converter takes time, which I want to speed up. Good grief!
  
• CUDA enables massive parallel computing which a RAW converter will take advantage of.
  
• CUDA does require an nVidia card, true. However code written in CUDA can be converted to OpenCL in short order.
  
• At the most 50 files? This limitation bothers you? That would be very fast! But that limitation is not correct. The onboard video RAM while only having 512 can pull from system memory, just a touch slower. Good grief!
  
• Yes DPP does use multi-core hence my testing confirming what it does with 500 files on my Quad. It is very nice. I would rather see 50 or more files worked on at the same time instead of how it is currenly programmed.

Hey you fellows just keep cluthing to your old ways... write a book on why things don't need to improve! I for one would love to see GPU computing make it way into photography applications.

CyberDyneSystems wrote in post #8007985
I was just reading about these Tesla cards NVidia makes for unreal added CPU power,.
http://www.nvidia.com/​object/IO_43499.html

.. and would love a new Dual socket I7 board with the PCIExpress 2 16 lane slots for a Tesla... if i could get my RAW and PSCS to use it all!

So do we need "cuda" to use those Teslas on our RAW converter?

Currently... until OpenCL matures yes you will need CUDA drivers installed which only work with nVidia 88xx series, Tesla cards. Quadro FX 5600 or later as well. The catch is that Adobe CS4 only works with Quadro or Tesla cards. There is no reason for this other than nVidia makes them I guess.

I want RAW converters to batch to jpeg using CUDA for all supported nVidia cards cutting our time down to nothing.

pcunite wrote in post #8009013
Currently... until OpenCL matures yes you will need CUDA drivers installed which only work with nVidia 88xx series, Tesla cards. Quadro FX 5600 or later as well. The catch is that Adobe CS4 only works with Quadro or Tesla cards. There is no reason for this other than nVidia makes them I guess.

I want RAW converters to batch to jpeg using CUDA for all supported nVidia cards cutting our time down to nothing.

Not true. quadro+tesla have extra acceleration, but the biggest speed enhancement is opengl support in photoshop, which any graphics card (even my 10 year old tnt2) can have (well, tnt2 doesn't have support for ogl 2 or 3, and you need ogl 3 to get the most of accelerations)

CyberDyneSystems wrote in post #8007985
.. and would love a new Dual socket I7 board with the PCIExpress 2 16 lane slots for a Tesla... if i could get my RAW and PSCS to use it all!

So do we need "cuda" to use those Teslas on our RAW converter?

Why dual i7? Just get dual socket xeon mobo (supporting the new 6 and 8 core i7 based xeons), 2x 16+1x8 pci-e, FBDIMM, sata3 raid 6 of half a dozen ssd, three quadro 5600 based teslas in sli. It'll put out about 500gflop from the processors alone, and another 5 tflop from the tesla boards. I'm sure you could edit all your photos while hosting potn, and entire cal-i tournament of counterstrike source, crysis, act as a server for 100 of your halo 3 playing buddies, and have enough power left over to save the world with protien folding Yea it'll cost you more than a p60+, but it'll be worth it to be hanging off the other side of the bleeding edge.

CyberDyneSystems wrote in post #8007985
Hey man, I was just asking... I also use a program called BreezeBrowser which is not multi threaded, it uses the Canon engine for conversion so I thought maybe that = DPP also not being multithreaded,

Haven't tried lately, but ISTR that older versions weren't, but you could get it to use more then one processor by not batching all files at once, but rather batch in 2 parts.

Edit: Google still works
Link

basroil wrote in post #8009258
Not true. quadro+tesla have extra acceleration, but the biggest speed enhancement is opengl support in photoshop, which any graphics card (even my 10 year old tnt2) can have (well, tnt2 doesn't have support for ogl 2 or 3, and you need ogl 3 to get the most of accelerations)

Dude... really... I am not talking or even implying about graphic accelerations. I am talking about number crunching which any GeForce 88xx series can do because they are all GPGPU now. Adobe has hinted that they will use the professional line of nVidia cards to do encoding (number crunching) in the future if not already on version CS4 (which I don't have to test).

Please, either support this thread by saying "yeah that would be great" or optionally "nah I am happy with the current trend of using serial CPU processors because I like to goof off while waiting for it to finish".

I am using my GeForce right now in a distributed computing effort and it is much faster than the latest Quad CPU chips. RAW converting can be made parallel.

Again, for the textually challenged... I could write the software myself... but would rather Canon did it... to use a GPU instead of a CPU.

basroil wrote in post #8009258
Why dual i7? Just get dual socket xeon mobo (supporting the new 6 and 8 core i7 based xeons)...

This is what I meant by "dual i7",. the only "Dual socket" i7 boards are Xeon I assume.. so I thought this was self evident in my post.

So yes, if I were upgrading and had the dosh, it would be a dual socket Xeon board with two i7 based multi core Xeons..

pcunite wrote in post #8010756
[SIZE=2]...

I am using my GeForce right now in a distributed computing effort and it is much faster than the latest Quad CPU chips. ....

Man that is very cool! I could have used that kind of horsepower a few years back!
Are you on a team?
What are you crunching?

(...used to work with Genome, Folding and SETI on a team some years ago,... )

Dang, I do parallel RAW conversions now, with Lightroom. What's the fuss?

As for how long it takes, here's an old-timers view: have the computer do it when you're doing something else (like sleeping, or watching TV). Then it doesn't matter how long it takes the computer to do it's job. There's no way that I'd wait around for 1000 images to process.

Granted, I'd never process 1000 images. I've taken 52,000 pictures in the last 5 years, and there's only a few hundred that I'd bother converting to TIFF, let alone JPG.

Oops, what I mean to say was:

"nah I am happy with the current trend of using serial CPU processors because I won't wait for it to finish"