Re: [hatari-devel] Hatari data cache tests

[ Thread Index | Date Index | More Archives ]


The tests I did were in CE/PF mode. It seemed like the i-cache was affecting cycles but the d-cache was not (cycle sum unaffected), even if the d-cache hit ratio would change.

However I only looked at that one sequence of code - I'll look closer elsewhere and see if there is some additional context.


On 19 June 2015 at 23:47, Nicolas Pomarède <npomarede@xxxxxxxxxxxx> wrote:
Le 20/06/2015 00:41, Nicolas Pomarède a écrit :
So, the first move takes 8 cycles, it's a miss since no one accessed
$40000 before. Next moves will be some hits and we see they now take 6

At least, we see that cache has an effect on number of cycles (but I
didn't check 68030 doc, are 8 cycles for miss and 6 cycles for hit the
expected values when accessing 16 bit ram ?)

same test for move.b as in your example :

114@198 : 000C4F06 41f9 0005 0000           LEA.L $00050000,A0
118@198 : 000C4F0C 1010                     MOVE.B (A0),D0
124@198 : 000C4F0E 4e71                     NOP
126@198 : 000C4F10 1010                     MOVE.B (A0),D0
130@198 : 000C4F12 4e71                     NOP
132@198 : 000C4F14 1010                     MOVE.B (A0),D0
136@198 : 000C4F16 4e71                     NOP

1st move.b takes 6 cycles ; then the other ones also take 2 cycles less, ie 4 cycles.


Mail converted by MHonArc 2.6.19+