PTGui and GPU processing or how to build a new computer

[Thomas Krueger]

Shopping for a new workstation I'm not sure about the graphic card, Nividia or ATI. But it seems that PTGui doesn't support GPU processing like Autopano. There is only the option to use the CUDA enabled PTStitcherNG together with a NVIDIA graphic card. So building up a fast computer for stitching with PTGui a NVIDIA graphic card and a SSD scratch disc is recommended looking at the results of Bernhard Vogels speed test at http://hdview.at/speedtest/results.html ??

[michael crane]

the CUDA thing looked like the way to go but I don't have the resources at
present to do testing.

mick

[Joergen Geerds]

my question for you is: PTGui with it's current speed isn't fast
enough for your project? does it matter if something renders in 4
instead of 3 minutes?
also keep in mind that many tiles from a gigapixel project across the
horizon need very little transformation, and hence render much faster
than tiles from a fisheye lens where pixels have a much heavier
transformation. also, the choice of interpolators affects the
rendering speed big time, much more than disk I/O.

a nice pair of SSD in a RAID0 setup takes care of most of the I/O
bottlenecks for PTGui swap, and feeding the source tiles from a
"traditional" SATA disk RAID0 helps also a lot, together with enough
RAM.

CUDA/GPU rendering would be great for the detail viewer, but I have
not looked at reliable numbers in terms of CPU vs GPU rendering for a
variety of panoramas.

joergen
newyorkpanorama.com

[Joergen Geerds]

just as a reference, since my results are not included in the
speedtest results yet:
PTGui 9.01 needs about
5min 40sec with 26GB RAM assigned
8min 27sec with 1.6GB RAM assigned

blending is the process that needs the most ram, warping isn't so
critical with RAM.
interestingly, PTGui does some form of pre-blending while warping. if
you pick enblend for blending, warping happens in 1/3 of the time, but
enblend itself takes forever (I aborted it after 18 minutes of blend
time)
the interpolator choice doesn't matter so much for this speedtest,
since there is almost nothing to warp anyway.

joergen
newyorkpanorama.com

[Jim Watters]

On 2011-02-15 1:44 PM, Joergen Geerds wrote:

also keep in mind that many tiles from a gigapixel project across the
horizon need very little transformation, and hence render much faster
than tiles from a fisheye lens where pixels have a much heavier
transformation. 

Whether it is a small or large transformation, just the fact that a transformation is being done, will add a small amount of distortion or degradation to the image.  The amount of transformation required is not going to effect speed.  Not correcting lens distortion at all might speed it a bit but going to be very small.  Images that have viewpoint correction will take longer but that is because PTGui processes more pixels (I think the entire output size).

the interpolator choice doesn't matter so much for this speedtest,
since there is almost nothing to warp anyway.

The interpolator will effect the time needed to process an image.  The bottleneck  may be on disk io but has nothing to do with the amount of warping that is needed.

joergen
newyorkpanorama.com

-- 
Jim Watters
http://photocreations.ca

[Joergen Geerds]

On Feb 15, 1:30 pm, Jim Watters <jwatt...@photocreations.ca> wrote:
> Whether it is a small or large transformation, just the fact that a
> transformation is being done, will add a small amount of distortion or
> degradation to the image.  The amount of transformation required is not going to
> effect speed.  

I think it's faster to transform a 21mpx tile to a 21 mpx tile (i.e.
tiles of a gigapan near the horizon, just fixing the lens distortion)
vs. transforming a 21mpx fisheye to a 60+mpx tile in a 360x180.
I did a quick test with the gigapan speedtest file, doubling the
output size quadrupled the time necessary for warping, as one would
expect.

> Images that have viewpoint correction will take longer
> but that is because PTGui processes more pixels (I think the entire output size).

that is correct

> The interpolator will effect the time needed to process an image.  

mostly true, the basic interpolators have all the same rendering
speed, while the more precise ones take much longer:
interpolators warping 337 6mpx tiles:
bicubic normal: 3min 40sec
ptgui default: 3min 43sec
lanczos2/sinc16: 3min 40sec
lanczos4/sinc64: 4min 35sec
lanczos16/sinc1024: 22min 47sec

> The bottleneck may be on disk io but has nothing to do with the amount of warping that is needed.

there was no significant disk activity in my tests during warping,
everything happened in RAM.

joergen
newyorkpanorama.com

[Hans]

On Feb 15, 7:30 pm, Jim Watters <jwatt...@photocreations.ca> wrote:
> On 2011-02-15 1:44 PM, Joergen Geerds wrote:> also keep in mind that many tiles from a gigapixel project across the
> > horizon need very little transformation, and hence render much faster
> > than tiles from a fisheye lens where pixels have a much heavier
> > transformation.
>
> Whether it is a small or large transformation, just the fact that a
> transformation is being done, will add a small amount of distortion or
> degradation to the image.  The amount of transformation required is not going to
> effect speed.

Sorry Jim but that is not true at all.
The warping of a fullframe 8 bit 21 mp fisheye takes around 20 sec on
my iMac.
Same as a rectilinear from a gigapixel pano with very little warping
takes less than 5 sec each.

The gigapixel test warps in 6 min. They are 337 6 mp images. Thats 1
sec each.
I just made atest with 8 fullframe fisheyes 6mp. Time 22 sec. Thats
2.75 sec each.

Hans

[Thomas Krueger]

On Anandtech was a review between Intel's Quick Sync and CUDA processing related to video decoding, but I don't how if this means also something for stitching and Photoshop work:
http://www.anandtech.com/show/4083/the-sandy-bridge-review-intel-core-i7-2600k-i5-2500k-core-i3-2100-tested/9

The upcoming Intel Z68 chipset supports also SSD caching, so instead of waiting for a 1155 motherboard I will wait a couple of weeks more for a mainboard with the Z68 chipset.
http://news.softpedia.com/news/Intel-May-Launch-the-Z68-Chipset-In-Early-April-183618.shtml

@Joergen: waiting one minute more is no problem, I just want to be on the safe side choosing the components for the next computer (which will stay the next couple of years in the office).

[PTGui Support]

Intel's QuickSync is a fixed pipeline for video de/encoding only.
Photoshop uses OpenGL which is available on all graphics cards.
PTStitcherNG uses CUDA which is NVidia only.
But I expect the future standard will be OpenCL, which is kind of a
cross platform implementation of CUDA.

Joost

> --
> You received this message because you are subscribed to the Google
> Groups "PTGui" group.
> To post to this group, send email to pt...@googlegroups.com
> To unsubscribe from this group, send email to
> ptgui+un...@googlegroups.com
> Please do not add attachments to your posts; instead you may upload files at
> http://groups.google.com/group/ptgui/files
> For more options, visit this group at http://groups.google.com/group/ptgui

[Thomas Krueger]

BTW, the Autopano rendering engine inside 2.5 is OpenCL based:
http://www.autopano.net/wiki-en/action/view/Computer_system_guide

2011/2/16 PTGui Support <sup...@ptgui.com>:

[Jim Watters]

On 2011-02-15 8:49 PM, Hans wrote:
>
> On Feb 15, 7:30 pm, Jim Watters<jwatt...@photocreations.ca> wrote:
>> On 2011-02-15 1:44 PM, Joergen Geerds wrote:>
>>> also keep in mind that many tiles from a gigapixel project across the
>>> horizon need very little transformation, and hence render much faster
>>> than tiles from a fisheye lens where pixels have a much heavier
>>> transformation.
>> Whether it is a small or large transformation, just the fact that a
>> transformation is being done, will add a small amount of distortion or
>> degradation to the image. The amount of transformation required is not going to
>> effect speed.
> Sorry Jim but that is not true at all.
> The warping of a fullframe 8 bit 21 mp fisheye takes around 20 sec on
> my iMac.
> Same as a rectilinear from a gigapixel pano with very little warping
> takes less than 5 sec each.
>
> The gigapixel test warps in 6 min. They are 337 6 mp images. Thats 1
> sec each.
> I just made atest with 8 fullframe fisheyes 6mp. Time 22 sec. Thats
> 2.75 sec each.
>
> Hans

I am very surprised. Hans, Sorry for not running my own test.
Looking at my own 8 fisheye image pano. Only three of them do not cross the 360
boundary. And all of the tiles created are larger than the input image because
of the way they are projected onto the equirectangular image. They take up
nearly half the output width. The rectilinear image projected onto a gigapano
image will be roughly the same number of pixels.
When I was referring to large transformation I was talking about big numbers for
rpyabcde.
I bet your fisheye time would drop if you could somehow stop from wrapping the
image around the 360 boundary. I also bet that the gigapixels times would
increase to about double it the images were all rotated an additional 30 deg so
the tiles generated were twice as big.

One of the advantages of PTStitherNG will be that all the pixels that are
already accounted for by already warped images will not be processed. If going
to a blended pano and not tiles that is.

Sorry again for not running the test, but I am interested in the outcome if you
do test it.