amd phenom II x4 955 BE @ 3.20 GHz:
x87: 3745 ms
sse1: 5008 ms
sse2: 4774 ms

So this basically compared how my CPU handles instruction sets huh? So if programs have multiple versions, I shouldn't be choosing ssix versions and instead go for the FPU/x87 version? If so, this topic has helped me in 2 ways now

Core i5-480M @ 2.66 GHz

Core i5 M520 @ 2.4 GHz (laptop)


---------- testing started ----------
testing x87

Starting matrix multiplication loop...
Elapsed time: 3697 ms
testing sse1

Starting matrix multiplication loop...
Elapsed time: 3635 ms
testing sse2

Starting matrix multiplication loop...
Elapsed time: 3510 ms
---------- testing finished ----------

Wee!

Intel i7 980X @ 4.2GHz

C:\temp>echo off---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 1888 mstesting sse1Starting matrix multiplication loop...Elapsed time: 1919 mstesting sse2Starting matrix multiplication loop...Elapsed time: 1763 ms---------- testing finished ----------

I find it kind of funny how everyone jumped all over them for using x87 code, but that wasn't that bad of a decision to begin with apparently. It appears to be best for AMD machines, and only slightly worse than SSE2 for Intel machines.

So, wait, I'm confused. If x87 is the most optimized for my AMD CPU and that is vanilla, then why am I getting a boost from tesval/skyboost?

SSE2 on my Core2Duo is 27% better the x87 (2215 vs. 3042 ms for the test). That's a lot more than "slightly worse". The situation is different on the i5 and i7 processors, but SSE2 all the way for me.

Because there is a bunch of other stuff the mod is doing that doesn't involve math.

Inline functions.

Don't know if you still need more but here are my results

Intel Core i5-750 @2.67GHz

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 3884 mstesting sse1Starting matrix multiplication loop...Elapsed time: 3962 mstesting sse2Starting matrix multiplication loop...Elapsed time: 3932 ms---------- testing finished ----------

It's also worth noting that if Bethesda compiled with optimisations enabled -- which include a whole variety of optimisation techniques, not just inlining -- we'd be getting better performance than TESVAL and Skyboost provide for both AMD and Intel platforms.

AMD Phenom II X4 3GHZ

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 3963 mstesting sse1Starting matrix multiplication loop...Elapsed time: 5429 mstesting sse2Starting matrix multiplication loop...Elapsed time: 5366 ms

I7 740Q @ 1.73GHz.
x87 - 4774
SSE1 - 4353
SSE2 - 4196

You can't say you didn't get fast feedback.....

i5 2500K - OC (4.7Ghz)

---------- testing started ----------
testing x87

Starting matrix multiplication loop...
Elapsed time: 1498 ms
testing sse1

Starting matrix multiplication loop...
Elapsed time: 1560 ms
testing sse2

Starting matrix multiplication loop...
Elapsed time: 1467 ms
---------- testing finished ----------

C2D E7400 @ 3 GHz

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 3588 mstesting sse1Starting matrix multiplication loop...Elapsed time: 3089 mstesting sse2Starting matrix multiplication loop...Elapsed time: 2683 ms---------- testing finished ----------

Intel Core2 Quad Q8200 3.1 ghz

---------- testing started ----------
testing x87

Starting matrix multiplication loop...
Elapsed time: 3401 ms
testing sse1

Starting matrix multiplication loop...
Elapsed time: 2948 ms
testing sse2

Starting matrix multiplication loop...
Elapsed time: 2574 ms
---------- testing finished ----------

Oh youre Right, what the hell was i looking at then? haha...

Thanks, this took me lots of tweaking and lots of testing!

Anyway, here's a load of tests. Taken from computers around the house.

CPU: AMD Athlon 64 X2 4800+ @ 2.5GHz

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 5141 mstesting sse1Starting matrix multiplication loop...Elapsed time: 6203 mstesting sse2Starting matrix multiplication loop...Elapsed time: 5938 ms---------- testing finished ----------

CPU: Intel Pentium Dual Core E5200 @ 2.5GHz

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 4218 mstesting sse1Starting matrix multiplication loop...Elapsed time: 3594 mstesting sse2Starting matrix multiplication loop...Elapsed time: 3172 ms---------- testing finished ----------

CPU: Intel Core 2 Duo E8400 @ 3.0GHz

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 3682 mstesting sse1Starting matrix multiplication loop...Elapsed time: 3074 mstesting sse2Starting matrix multiplication loop...Elapsed time: 2652 ms---------- testing finished ----------

That's it, should help.

*shakes tiny fist*

i2600K@4.8Ghz (I won't normally run this, my heatsink can't deal with 8 cores transcoding at this in a comfy temperature )

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 1466 mstesting sse1Starting matrix multiplication loop...Elapsed time: 1513 mstesting sse2Starting matrix multiplication loop...Elapsed time: 1436 ms---------- testing finished ----------Press any key to continue . . .

I am starting to think that this CPU test is completely useless - systems that should be VERY NEAR - like i5 750 and i7 920 @3.8-4Ghz with HT disabled are getting completely different results:

for example:

On i7 920 @3.8Ghz and HT disabled I am getting:

---------- testing started ----------testing x87Starting matrix multiplication loop...Elapsed time: 2091 mstesting sse1Starting matrix multiplication loop...Elapsed time: 2137 mstesting sse2Starting matrix multiplication loop...Elapsed time: 1981 ms---------- testing finished ----------

The only real difference between these CPU's (with HT disabled) is that i920 has triple channel DDR3 memory controller and i5 - dual channel. I can't run it under Vtune, but most probably it is hitting some memory hierarchy / prefetch bottlenecks and not testing performance.

Of course user could have got his results wrong as well

Something to keep in mind: This is more of a want to see the differences between those three stats per system, not really trying to compare systems themselves to one another. The other is that the more things open or going on in the background the higher the numbers will be. Even turning off firefox with the skyrimnexus page open dropped my numbers about 100ms. Again though: This is for comparising the different methods, not the different systems. It is not an [censored] benchmark.

ED: wait, e-youknowhat is censored? hahahaha.

Yeah I don't think we need any more, the test shows:

AMD -> x87

Intel -> SSE2

The flawed micro benchmarks lead to flawed conclusions about what versions to build for what CPUs. It doesn't take assembly god to google around to see things like this:

http://www.geeks3d.com/20100711/test-simple-x87-vs-sse2-performance-test-with-matrix-multiplication/

Where users are getting double performance increase in same micro benchmark ( multiplying matrixes) going from x87 to SSE2 on Intel machines. I am not seeing such gains here and that leads to conclusion that something is wrong.

What you're not taking into account in that case is that this mod does more then just change those methods, it also inlines a tonne of extraneous function calls that spend more time getting to and from the function the recovering the memory location the function itself was written to return. That's what the core of tesval did and what skyboost extends upon with further stuff like what he's asking us for feedback upon now.

ED: Er sorry wait scratch that, you're not talking about the mod but how the different methods vary.