I've got a dual opteron 242 setup, and gimp works beautifully. Note, however, that i'm running gentoo, so Gimp has been compiled from scratch with a 64-bit gcc.

As for improvements over 32-bit, I can't really say, my best comparison is comparing a 2x1.6gHz 64bit machine to a 1x0.9gHz 32 bit machine, so no real basis for comparison.

Also, I don't know if any of the filters in GIMP have been optimized with hand-written assembly, or if they are straight c/c++/whatever. If they're assembly, they may actually run slower as-is on amd64 than x86 (this has been the case with several media (de|en)coders), but that will change when someone writes an amd64 assembly version.

Hi,

Ketil Froyn writes:

Thanks for the responses. I guess one major limiting factor could be the types used internally by the GIMP. For instance, if it does operations on uint32 type variables internally, I'd think the compiler needs to be pretty clever to make use of 64 bit operations. But now I'm getting into deep waters here here...:) I was hoping someone in the know could confirm whether there's any point going to 64 bit to make the GIMP faster, rather than just getting a faster 32 bit CPU.

The only real advantage you will get from a 64bit CPU is that you can address more than 4GB of RAM and will be able to handle very large images in GIMP provided that you actually stuff more than 4GB of RAM into your hardware.

Sven

On Friday 08 April 2005 17:41, Sven Neumann wrote:

Hi,

Ketil Froyn writes:

Thanks for the responses. I guess one major limiting factor could be the types used internally by the GIMP. For instance, if it does operations on uint32 type variables internally, I'd think the compiler needs to be pretty clever to make use of 64 bit operations. But now I'm getting into deep waters here here...:) I was hoping someone in the know could confirm whether there's any point going to 64 bit to make the GIMP faster, rather than just getting a faster 32 bit CPU.

The only real advantage you will get from a 64bit CPU is that you can address more than 4GB of RAM and will be able to handle very large images in GIMP provided that you actually stuff more than 4GB of RAM into your hardware.

Actually, these 64 bit CPUs have more registers as well. I'd expect the compiler to make a better usage of them. Does this happen?

Ketil, as for the 64 bit integers - there could not possibly be a program benefiting from this directly. They still do one operation a time. Multiple operations with a single instruction (SIMD) is what would speed up an app CPU usage are the correlate processing units like 3Dnow, MMX and SSE. I'd guess the GIMP takes advantadge of MMX, SSE, SSE2, and these had not improved on the new 64 bit chips (btw, the registers relevant to SIMD were already 64 or 128 bit wide way before the move to 64 bit on the general purpose registers).

Certainly the SSE2 code might get some improvements..there is room for implementing 3Dnow code as well - it may not even get faster than SSE2 code on Athlons, but it would quite a task for one to perform. Do you have some spare time to learn coding 64/128 bit SIMD? :-)

Regards,

JS ->

Sven

On Thu, Apr 07, 2005 at 11:12:13PM +0200, Ketil Froyn wrote:

Thanks for the responses. I guess one major limiting factor could be the types used internally by the GIMP. For instance, if it does operations on uint32 type variables internally, I'd think the compiler needs to be pretty clever to make use of 64 bit operations. But now I'm getting into deep waters here here...:) I was hoping someone in the know could confirm whether there's any point going to 64 bit to make the GIMP faster, rather than just getting a faster 32 bit CPU.

As mentioned in other parts of this thread, amd64 processors have twice the amount of general purpose registers, which means gcc can generate more efficient code. It also uses sse and sse2 instructions for floating point code, which is a big performance win over x87. But neither of these are anything 64-bit specific, and *all* applications will benefit without doing anything special.

The only 64-bit specific thing that GIMP will benefit from is if you have > 4 GB of memory and you're dealing with very large images. Since it can address that much memory directly, the image can be loaded entirely into RAM without any disk swapping.

You have a misconception that using 64-bits for everything internally will magically make things faster. In reality, this would make things slower, since data units will be 8 bytes vs. 4 bytes, which means you're using up more memory bandwidth and precious cache space. If your app only needs 32 bits of precision, then it's a waste of space and a waste of time to shuffle the extra bits back and forth.

Probably for this reason, amd64 has a default operand size of 32 bits, even in 64-bit mode. As a result, code size isn't siginificantly larger than x86 (you take a hit on the wider pointers, but that's offset by the doubling of the registers; gcc can do more with less code).

So there isn't much in GIMP that needs 64 bit integers, especially not in performance critical sections. There's the mmx, sse, sse2 acceleration for some functions, but those use 64-bit and 128-bit stuff on both x86 and amd64.

To summarize, an amd64 platform will give you better performance over an x86 platform, but mainly to due other aspects of its design, and not much to do with 64-bit wide integer operations.

-Yosh