[gmx-users] 64-bit vs. 32-bit

[Erik Lindahl]

Hi,

On Sep 3, 2004, at 12:00 PM, Berk Hess wrote:

>> Is it really impossible to have the same gain when going to 64 bit as 
>> you
>> see for non-assembly-optimized code? I am just curious on what is 
>> causing
>> this "barrier".
>
> 64-bit computation is always slower than 32-bit.
> An Athlon 64 can be (much) faster than an Athlon XP/MP,
> because of a better architecture.
> But with the Athlon 64 you can have direct access to double
> the amount of registers in 64-bit mode, not in 32-bit mode.
> So programs could run faster in 64-bit mode, not due to the 64 bits,
> but due to the amount of registers.

The main reason why I haven't bothered to release the 64-bit SSE loops 
yet is that I've been busy benchmarking and trying to use all 16 SSE 
registers.

Unfortunately I really don't gain anything from using 16 instead of 8 
registers; this is partly because the loops are pretty streamlined 
already (not primarily register constrained), but also because of the 
hardware register renaming - internally the CPU has over 100 registers 
that are allocated dynamically - the 64-bit Athlons just have a 
slightly bigger virtual "window" towards the hardware registers 
compared to 32-bit code.

Anyway, this is almost done, and then I'll add the AMD-64 loops to the 
distribution - just don't expect any huge magical performance boost 
compared to the 32-bit code running on AMD64!

Cheesr,

Erik
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2363 bytes
Desc: not available
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20040903/d9b20bdd/attachment.p7s>