WinUAE 440 slower than 281?

[Xtro]

Hi everyone, looking at some Lightwave 3.5 benchmarks tried to replicate on my system with WinUAE 440 64bit and saw my render times were about 2 time slower than on comparable systems. Then I tried WinUAE 281 32bit and render times were reduced to half. The WinUAE config was the same for both WinUAE versions, 68040 + FPU + JIT.
Don't know if this affects other aspects of CPU emulation performance, or that 440 version is more compatible.

Kind regards

[Toni Wilen]

JIT should be more or less same. Does 32-bit version return different results? Make also sure FPU mode is "Host (64-bit)" in CPU panel. 80-bit will be slower (and usually not needed)

[Xtro]

Originally Posted by Toni Wilen

JIT should be more or less same. Does 32-bit version return different results? Make also sure FPU mode is "Host (64-bit)" in CPU panel. 80-bit will be slower (and usually not needed)

Also tried 440 32-bit with the same exact result than 440 64-bit. FPU was always "Host (64-bit)" in all my tests.

Tried other WinUAE versions, and the last one that gave "fast" results was 361 64-bit. 400 64-bit is even slower thant 440. There should be something between 361 and 400 that made JIT slower.

[Toni Wilen]

Could you try without JIT FPU enabled? Do you get similar (slow) results? (=problem is JIT FPU)

[Xtro]

Tried again same test with 440 & 361 64-bit, now 68040 JIT FPU disabled and again 440 takes about double time to render scene, 440 (39 minutes) & 361 (20 minutes).

[Rotareneg]

Just did a little benchmarking and found a similar (although not quite so large) performance regression in 4.5.0 vs 3.6.1 (both 64 bit.)

Rendering a test scene in Sculpt Animate 4D:

3.6.1 with '030 68882 JIT+FPU, Integer version renders in 6.8 seconds, Floating point in 8.7 seconds. With JIT FPU off FP version jumps to 9.6 seconds.

4.5.0 beta 2 with '030 68882 JIT+FPU, Integer version takes 7.5 seconds and FP 10 seconds. With JIT FPU off it's 11 seconds for FP render.

[Xtro]

Tried Real 3D Pro 1.4.2 with candlescene rendering all by default settings, some weird results:

361 INT 55.3s ------- 440 INT 137s
361 FPU 23.5s ------- 440 FPU 65.1s
361 INT JIT 10.2s --- 440 INT JIT 38.4s
361 FPU JIT 3.1s ---- 440 FPU JIT 32.4s

So in this test 361 FPU JIT 10x faster than 440 FPU JIT
even 361 JIT disabled is faster than 440 JIT enabled

[Toni Wilen]

Hmm.. It must be config/hardware/something related because everyone would have complained if it was common problem.

Could you narrow down the version that introduced this difference?

[Xtro]

For me version 400 introduced this "slow" performance issue

[Rotareneg]

I just tried with VistaPro 3 (as its FPU version is actually significantly faster than the non-FPU one, unlike with Sculpt4D) and I couldn't find any performance difference between 3.6.1, 4.4.0, and 4.5.0, so I don't know.

The 32 bit versions are always faster though, my test scene rendering in 12 seconds with the x64 version and only 9 with the 32 bit version.

My CPU is an old i7 860, in case that's relevant.

[Toni Wilen]

Could you attach your config file?

It is possible some CPU related config entry was ignored in JIT mode with older versions but are now in use, resulting in slower speed.

[Xtro]

Here is the config I am using
Thanks for checking Toni

[Toni Wilen]

Disable vsync first. Vsync and JIT won't work very nicely together.

Disabling JIT hard flush option might increase speed a bit.

[Xtro]

OK, I think that did it. Only did a quick test with Lightwave, I had enabled Lagless Vsync (Beam-Raced) and that was the cause of slow performance, 3.6.1 did not have that option so it was fast.

[Rotareneg]

I wasn't using vsync, but I'm dubious of the results I was getting as I can't reproduce them now... don't troubleshoot when sleepy! I did find a bug with 80 bit FPU in 4.5.0, I'll post it in the normal beta thread.