AIBB 6.5 - FMath test

[michaljarz]

Hi.
I'm wondering how the "FMath" test is built in AIBB 6.5. While other tests are several dozen times ahead of the A4000 (e.g. in JIT mode - 74x emutest, 164x matrix, 118x beachball), FMath (using a coprocessor - FP Math) gives me a result only 2x (3x in JIT mode) faster than the A4000. Is the construction of this test based on the commands that the MC68881 has and x86 has to emulate them in a rather complicated way? Even more interesting results come out in the "SC Math" mode - almost 8x faster in JIT and almost 2x slower without JIT.

By the way, I don't think it's something that makes life difficult - in my opinion WinUAE is the best I'm just curious what could be the reason for such discrepancies compared to other tests...

Tests were done on Win10 with i5-7400 processor (WinUAE 4.10.0).

(sorry - google translate helped me)

[Toni Wilen]

I wouldn't trust any of those tests because even without JIT most of them execute far too fast.

I get 0.00 second result from FMath (and few others) with JIT on and crazy overflowed result..

If loop is quick enough, in JIT mode it might appear to execute "immediately". JIT (and also possible in fastest possible modes) might execute hundreds of m68k instructions in single set if none of them do any IO accesses and there is no pending interrupts.

[rutra80]

Yesterday I tested all KS and WB booted from original floppies from 1.3 to 3.2 and my FMath AIBB results are similar to michaljarz:

Without JIT:
SC FMath slower than A4000
CP FMath 2x faster than A4000

With JIT:
SC FMath 10x faster than A4000
CP FMath 4x faster than A4000

Other tests are hundreds or thousand times faster.

Fastest possible x64 WinUAE without CE.
Intel Core i7-4790K.