Re: LuaJIT and Floating Point Determinism
- From: malkia <malkia@xxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Mon, 13 Jan 2014 21:03:07 -0800
It's possible that the code generated by luajit would work the same
way on different cpus.
But it's a real problem (outside) luajit. For example read this:
http://www-archive.xenproject.org/files/xensummit_germany09/AMD.pdf#page=15
(This is a problem not only for games, but for virtual machines
(moving from one host machine to another)).
"SSE reciprocal and reciprocal-square root approximation
instructions: AMD and Intel provide different numbers of significant
bits – RCPPS, RCPSS, RSQRTPS, RSQRTSS.
x87 instructions with infinite series implementations can return
answers whose results differ in the units in the last place (ULP). –
Trigonometric functions – ranscendental functions – Reciprocals"
(I think luajit still uses some x87 - at least that's what vm_x86.dasc
shows but I could be wrong).
If you happen to use external library called by luajit, or something
that changes rounding modes without your knowledge (directx used to be
such offender), then you would have differences.
Here is a fun cpuid detection code
found here:
http://randomascii.wordpress.com/2013/07/16/floating-point-determinism/
// CPUID is for wimps:
__m128 input = { -997.0f };
input = _mm_rcp_ps(input);
int platform = (input.m128_u32[0] >> 16) & 0xf;
switch (platform)
{
case 0×0: printf(“Intel.\n”); break;
case 0×7: printf(“AMD Bulldozer.\n”); break;
case 0×8: printf(“AMD K8, Bobcat, Jaguar.\n”); break;
default: printf(“Dunno\n”); break;
}
On Mon, Jan 13, 2014 at 11:32 AM, Alex <initrd.gz@xxxxxxxxx> wrote:
>> It is not so hopeless
>
> I never said it was hopeless; I just wanted to be sure it wasn't. Also,
> thanks for the library, I'll take a peek at it.
>
> --
> Sincerely,
> Alex Parrill
--
Dimiter "malkia" Stanev,
ICQ: 21875894
malkia@xxxxxxx
malkia@xxxxxxxxx
Other related posts:
