Re: [hatari-devel] Hatari data cache tests

Hi,

One strange thing I noticed w.r.t data cache hits/misses is the apparent lack of impact it has on cycle timings.. Hits seem to cost the same as misses at least in terms of the profiler results... in this case two different instructions with very different hit ratios have almost exactly the same cycle sum.

$000362a0 :             move.w    a3,d0                      0.09% (166384, 665736, 670, 0)
$000362a2 :             move.w    d0,d1                      0.09% (166384, 663084, 22, 0)
$000362a4 :             add.w     d2,d0                      0.09% (166384, 665844, 695, 0)
$000362a6 :             move.b    (a0)+,d1                   0.09% (166384, 1331344, 44, 212) <- hit ratio of 0.1%
$000362a8 :             move.w    (a6,d1.l*2),(a1)+          0.09% (166384, 1999512, 715, 66190)
$000362ac :             move.w    d0,d1                      0.09% (166384, 2768, 660, 0)
$000362ae :             add.w     d2,d0                      0.09% (166384, 663048, 24, 0)
$000362b0 :             move.b    (a0)+,d1                   0.09% (166384, 1331444, 679, 154776) <- hit ratio of 93%
$000362b2 :             move.w    (a6,d1.l*2),(a1)+          0.09% (166384, 1999176, 712, 82447)

On 19 June 2015 at 09:03, Nicolas Pomarède <npomarede@xxxxxxxxxxxx> wrote:

Le 19/06/2015 01:22, Douglas Little a écrit :

Hi,

Here are some results from my Falcon using -f030 -videl config switches.
It was captured on an RGB monitor but it shouldn't matter since videl
isn't fetching.

My Falcon has failed NVRAM battery so it boots into random 50/60hz. It's
not worth trying to capture metrics with Videl enabled without knowing
which mode its in since the results will differ, and my VGA is not handy
to connect right now. So somebody else should do that bit if possible :)
Might be better to wait until the tests are more exhaustive anyway.

The tests themselves are all 'fragment' tests i.e. its set up to test
groups of instructions executed together - but many of them are simple
single instruction fragments (repeated adjacently many times of course).
There are a few multi-instruction fragments (like movem
src,[regpattern], movem regpattern,[dst]), each of which would need
detailing to make sensible comparisons but there's probably enough to
chew on for now with the single moves and addressing modes alone :)

D.

Hi

That's great, thanks !
I don't think I will have time to take this into account for next Hatari 1.9, but it will definitely be useful after that.

Nicolas