Research Question

[64NOMIS]

Lets say you had an application that takes advantage of AVX, SSE 4.1, and SSE 4.2.

What is the greatest improvement in performance you might achieve?

1. Define application/s
2. Define method of difference test
3. Screenshots

This is a research question, no actual benching necessary.

[charged3800z24]

Not sure if I can come up with something or not Much smarter people will be in soon I suspect..
I know that with out AVX support a 256bit FP unit is pointless, isn't Bulldozer going to have a Flex-FP that functions with out AVX support? I guess I'm going OT...

[stangracin3]

Not sure if I can come up with something or not Much smarter people will be in soon I suspect..
I know that with out AVX support a 256bit FP unit is pointless, isn't Bulldozer going to have a Flex-FP that functions with out AVX support? I guess I'm going OT...

idk but i know everytime he posts, it seems to be a teaser right before he puts things out...

[FlanK3r]

hm, I think, some benefits could be in some video encoding aplications. I heard, x264 support AVX a bit. Others, i dont know . Maybe some searching sequence in videos? It is still new instruction set and think, we will later, how will be with better support. And we must waiting in this thread to some smarter answer

[NEOAethyr]

PCSX2 now uses a tiny bit of avx, at 1st they said it was completely useless but I guess they found some use for it , I've seen a few code updates with avx support added, minor ones but still, they are using it.

"SSSE3", not "SSE3" had been in use for along time by the same program.
Same goes for SSE4.
Now, finally, they have added a few lines of code that can use SSE3, the stuff the amd's have.

Exact comment.
Posted by gabest:

GSdx: finally, some use for hsubps (SSE3).

[Grinder]

Sounds like a can of floating point whoop ass Wider, three operand support, allows optimization to one common standard if you have the latest intel or amd cpus.

So scientific apps/graphics rendering should benefit. I see SETI@home has an experimental AVX optimized version, but I'm not familiar with using it as a benchmark. Could we have it do the same work units on the AVX version versus others?

Better yet, maybe Cinebench has AVX support? It doesn't look like 11.5 does. Also, I think I read that the latest SiSoft Sandra did...

The trick will be making apples to apples comparisons "app" wise Unless we can turn AVX on/off, we'd be limited to using different versions of benches with/without avx support, and who knows what other differences between them.

I'll say run Sandra with 4.1, then 4.2, then avx enabled, and compare results. Downloading sandra lite to play with.

Here we go, just to get the ball rolling. You can enable/disable AVX for the modules that support it, and sometimes 4.1/4.2:





So, whether it's Sandra or some other bench, gather your data for each supported test, with avx/whatever else you can enable/disable, then plot results and %improvement. We could throw cpu frequency into the mix to see how it scales. Now we need avx capable hardware

Nice article here : http://blogs.amd.com/developer/2009/...ing-a-balance/

[Dirk Diggler]

Only app I can think of that has been mentioned is x264

[charged3800z24]

I run WCG almost 24/7, only off when I am benching or decide to play a game. I know there should be advantages for this type of application, but how to test I am not sure.

Downloading Sandra lite though...

[DedEmbryonicCe1]

AIDA64 has benchmarks with AVX acceleration but I believe he is asking more for real world usage scenarios. I don't think AMD is adding AVX just for our benchmarking pleasure.

[Grinder]

I did notice AIDA64 I'm sure it would be fine as well.

As long as it's measureable, and creates excitement about the capability, and the new hardware, it's win/win. Especially if it kicks intel ass. I did see some numbers on the Sandra site comparing SB vs. previous non avx generation.

[FlanK3r]

AVX= x264 benchmark, as I said...and maybe mediaExpress. Others, I dont now, its about support marketing and near future software.

[NEOAethyr]

I wouldn't worry so much about benching it.
Unless you are coding, in which case you could find out for your self which pathway is best .

The intel sandy's support avx.
The new am3+ cpu's supposedly support it.

If a benchmark or other program has support for it, I would blindly trust in good faith it's for the best and think no more of it .

In theory it could double performance.
In all likely hood the performance is probably 5% better if used correctly and nothing more.

Avx is 256bit while sse is 128bit.
I'm not sure how many cycles it will take to process one cpu to mem opcode compared to say 4x 64bit generic reg to mem opcodes.
Let alone the methods of turning it on and off, etc, on some of that fpu like stuff you need to have everything aligned and make sure things are correct once you stop using it, wasted time.
You have to use it in such a way to make it worth while.

[64NOMIS]

Questions. Questions that need answering.

[Dirk Diggler]

list of supported functionality from x264

#define X264_CPU_CACHELINE_32   0x000001  /* avoid memory loads that span the border between two cachelines */
#define X264_CPU_CACHELINE_64   0x000002  /* 32/64 is the size of a cacheline in bytes */
#define X264_CPU_ALTIVEC        0x000004
#define X264_CPU_MMX            0x000008
#define X264_CPU_MMXEXT         0x000010  /* MMX2 aka MMXEXT aka ISSE */
#define X264_CPU_SSE            0x000020
#define X264_CPU_SSE2           0x000040
#define X264_CPU_SSE2_IS_SLOW   0x000080  /* avoid most SSE2 functions on Athlon64 */
#define X264_CPU_SSE2_IS_FAST   0x000100  /* a few functions are only faster on Core2 and Phenom */
#define X264_CPU_SSE3           0x000200
#define X264_CPU_SSSE3          0x000400
#define X264_CPU_SHUFFLE_IS_FAST 0x000800 /* Penryn, Nehalem, and Phenom have fast shuffle units */
#define X264_CPU_STACK_MOD4     0x001000  /* if stack is only mod4 and not mod16 */
#define X264_CPU_SSE4           0x002000  /* SSE4.1 */
#define X264_CPU_SSE42          0x004000  /* SSE4.2 */
#define X264_CPU_SSE_MISALIGN   0x008000  /* Phenom support for misaligned SSE instruction arguments */
#define X264_CPU_LZCNT          0x010000  /* Phenom support for "leading zero count" instruction. */
#define X264_CPU_ARMV6          0x020000
#define X264_CPU_NEON           0x040000  /* ARM NEON */
#define X264_CPU_FAST_NEON_MRC  0x080000  /* Transfer from NEON to ARM register is fast (Cortex-A9) */
#define X264_CPU_SLOW_CTZ       0x100000  /* BSR/BSF x86 instructions are really slow on some CPUs */
#define X264_CPU_SLOW_ATOM      0x200000  /* The Atom just sucks */
#define X264_CPU_AVX            0x400000  /* AVX support: requires OS support even if YMM registers
                                           * aren't used. */

[Tom128]

I can't tell if this is a riddle, a challenge, or a "can you guess the numbers I already know" lol. All three of those instruction sets can see very big performance gains working with HD Codecs as well as encryption (especially SSE 4.2).

[NBF]

1. Linpack
2. Test: i5 2500k running latest Linpack library under Windows 7. Run test without SP1 for AVX disabled result and with SP1 applied for AVX enabled result.
3. Micro Screenshots

- No AVX support ->

- AVX support enabled ->

[charged3800z24]

I am looking for a way to run the BOINC benchmark with options to see if we can see the difference. I am having trouble finding anything online.

[wez]

1. Linpack
2. Test: i5 2500k running latest Linpack library under Windows 7. Run test without SP1 for AVX disabled result and with SP1 applied for AVX enabled result.
3. Micro Screenshots

- No AVX support ->

- AVX support enabled ->

Almost double GFlops, now thats a pretty neat speed-up

Cant think of anything else than crunching that I personally could benefit from with AVX.
But at the same time I'm sure we all use some application that could take advantage of it, we just dont know about, and would require the sw developers to put some focus on it, which might not happen until enough people starts putting out requests for it..

[64NOMIS]

NBF is on track. Thanks Grinder.

[beta]

1. C-ray. According to phoronix c-ray benefits from avx:



2. Compile c-ray under linux with avx/sse4.1/sse4.2, then compare completion time with results of other c-ray binaries built with different fp math options.

3. Well I dont have sandy bridge, so the last benchmark failed without avx:

[Grinder]

NBF is on track. Thanks Grinder.

My pleasure