dmmartindale@watcgl.waterloo.edu (Dave Martindale) (04/27/89)
I will soon be doing some work with very high resolution colour images
(4000x3000 pixels and more) and need to buy hardware capable of
processing and displaying at least portions of such images.
The only suitable hardware I know of at the moment is the Pixar Image
Computer, the Pixar II, and the Transcept/Sun TAAC-1 board. If anyone
has used one or more of these and has any comments about them, I'd
really appreciate hearing from you. I've already read the manual set
for both Pixar and Taac, so I know what the hardware looks like - I'm
more interested in the sort of knowledge that comes from real
experience with using something.
Pointers to other hardware meeting the requirements are also very
welcome. I'll summarize the replies.
My requirements:
Since I'll be working with such large images, I want a
horizontal resolution of at least 1024, to minimize the number
of pieces I have to cut the full images into. 512 or 640 just
isn't acceptable; more than 1024 would be nice.
Some images will ultimately end up on film. In order to avoid
quantization problems and the Mach bands created by them, the
system needs the ability to work with pixels that have 11-12
significant bits per colour component at any point where
arithmetic is being done on components. 8 bits just isn't good
enough for computations.
Because of the non-linear response of a CRT, it is not
necessary to have 12 bits per component going to the video DACs
in the frame buffer - 8 seems sufficient for practical
purposes. Less than 8 is not acceptable.
The hardware should also have considerable computational
horsepower so it can do pixel-oriented operations on large
images in reasonable time (Imagine doing a 5x5 convolution on a
4000x3000 image).
What I know so far about the Pixar Image Computer (hereafter called
Pixar I), Pixar II, and TAAC:
The Pixar I is the only machine that will allow panning across
a 4000x3000 image in memory, even in theory. (Can you get more
than 48 Mb of memory on a Pixar I now?) The Pixar II frame
buffer controller is integrated with a memory board, so the
largest displayable image is 2048x1024 without copying data
from off-screen memory. The TAAC is limited to 1024x2048, with
no possibility of adding extra memory.
Both Pixars store 11 significant bits per component, so you can
directly view intermediate results when applying a series of
image-processing steps to an image - this is great for anything
interactive. The TAAC is 8-bit oriented - you can do many
operations on 16-bit components, but you have to transform your
12- or 16-bit intermediate results into an 8-bit second image
to display it. When working with components wider than 8 bits,
the TAAC effectively has half the memory of the Pixar II and
1/4 that of the Pixar I. This will make interactive
manipulation of images on the TAAC a pain.
The Pixars seem harder to program at a low level - they are
SIMD machines, and pipeline delays are visible to Chap
programs. Image memory can't be addressed directly; pixels
must be brought into the scratchpad for use. There is no
floating point. Most arithmetic is 16-bit.
The TAAC is SISD (for each processor), with no pipelining.
Image memory can be randomly accessed by the processors, and
used for simple 32-bit word storage as well as pixels.
Arithmetic is 32-bit, and a floating point unit is available.
The Pixar has a multi-layer set of software libraries, and you
can do a lot without ever writing code that runs on the Chap.
However, it will take quite a while to become familiar enough
with the hundreds of routines to select the "best" way of doing
something. The TAAC software structure is much simpler.
The TAAC fits in 3 slots on a Sun, requiring no separate power,
cooling, or space. The Pixar II is a separate standalone box.
The Pixar I mounts in a rack.
Anything that you can add to the information above would be appreciated.
Subjective impressions ("It took me 3 months to figure out the video
registers") are particularly interesting.
Dave Martindale
watmath!watcgl!dmmartindale
dmmartindale@math.waterloo.edupxl@lanl.gov (Peter Lomdahl) (04/29/89)
In article <9400@watcgl.waterloo.edu>, dmmartindale@watcgl.waterloo.edu (Dave Martindale) writes: > > The only suitable hardware I know of at the moment is the Pixar Image > Computer, the Pixar II, and the Transcept/Sun TAAC-1 board. If anyone > has used one or more of these and has any comments about them, I'd > really appreciate hearing from you. I just took delivery of a Pixar II image computer and I'm trying to get up to speed using it. My requirements are a little different than yours, - I want to do movie-loops of regular 2D and volume rendered data. Anyway, here are some of my impressions. (1) The installation of the Pixar II is relatively easy. It goes basically like the instructions say. I had one problem with the configuration of the driver for the high-speed-interface in the kernel. I called Pixar customer support and my problem was fixed. It turned out that the documentation was wrong. The person that I talked to seemed quite knowledgeable about the product. (2) I'm in the process of getting some of my own data up on the Pixar, and I was indeed faced with the problem you mention. It isn't entirely obvious which is the best way of doing things. I guess it takes about a 3-4 days careful perusal of the documentation to get familiar with things and decide what will do the job for you. I hope I'm doing the right thing now :-) - I'm using picio(3H). I'm so far pleased with the tight UNIX-like 'feel' of the software. (3) It is my impression that it would not be too much of a problem to use the off-screen-memory (OSM) for you application. I'm certainly planning on doing just that for movie-loops, and there is a program called 'tileblast' that does that. Also, there is a demo program that does an image-roam on a 2Kx1K picture. I suspect it wouldn't be too hard to modify that to take advantage of OSM. Another possibility you might consider - if you budget allows it - is to get the parallel transfer disks. I've seen them demoed on a 16Kx16K image-roam - that was quite impressive. They run about $50K list for 1 Gbyte. Btw, the Pixar I takes 192Mb now. (4) Some people here have TAAC boards, and it's my impression that the user interface (the integration with SunView) makes them easy to prototype on. On the other hand these people are also complaining about the limited 8M of memory. I'll let you know more when I get a little more experience with the Pixar. -- Peter Lomdahl -- Los Alamos National Laboratory -- pxl@lanl.gov