On Mon, Apr 16, 2012 at 5:50 AM, Joanna Rutkowska
<address@hidden> wrote:
On 04/16/12 14:20, Joanna Rutkowska wrote:
/.../
>
> So, it seems like it is not a Xen issue, but instead that the kernel
> I'm using in the VM (essentially a vanilla 3.0.4) is not enabling AVX
> in XCR0. It would be interesting if anybody could try this on a
> non-Xen system with a similarly old kernel as I have (and on the
> AVX-capable processor, of course).
>
Interestingly, I tried another kernel in the VM (essentially a vanilla
3.2.7), and again the AVX seemed to be disabled in the VM (volk used
sse4 again).
So, I quickly looked through the kernel sources and I think that the
kernel enables AVX in this code:
static void __init xstate_enable_boot_cpu(void)
{
/.../
cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
pcntxt_mask = eax + ((u64)edx << 32);
if ((pcntxt_mask & XSTATE_FPSSE) != XSTATE_FPSSE) {
printk(KERN_ERR "FP/SSE not shown under xsave features 0x%llx\n",
pcntxt_mask);
BUG();
}
/*
* Support only the state known to OS.
*/
pcntxt_mask = pcntxt_mask & XCNTXT_MASK;
xstate_enable();
The XSTATE_FPSSE and XCNTXT_MASK are defined as follows:
#define XSTATE_FP 0x1
#define XSTATE_SSE 0x2
#define XSTATE_YMM 0x4
#define XSTATE_FPSSE (XSTATE_FP | XSTATE_SSE)
#define XCNTXT_MASK (XSTATE_FP | XSTATE_SSE | XSTATE_YMM)
The YMM corresponds to the AVX enable flag (not sure why they call it
differently?).
Anyway, this shows that, assuming CPUID returns correct values (that
indicate AVX to be enabled), then the (guest) kernel should enable AVX
(and Xen should emulate this and allow for this, as indicated in the
previous message). And if CPUID was not returning correct values (i.e.
omitting the AVX flag), then we would not have this whole discussion.
So, what am I missing?
Can you point out which kernels do you use that you have AVX working fine?
Linux smidgen 3.0.0-17-generic #30-Ubuntu SMP Thu Mar 8 20:45:39 UTC 2012 x86_64 x86_64 x86_64 GNU/Linux
I really do think it's a Xen issue; do you have a kernel you can try to run outside the hypervisor?
In any case, it's mostly a moot point, as there are very few Volk kernels which make use of AVX, and those which do don't appear to be significantly faster than their SSE3 counterparts. I think that's due to either memory bandwidth or instruction promotion of SSE3 intrinsics, but that's a different story.
--n
joanna.
_______________________________________________
Discuss-gnuradio mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/discuss-gnuradio