|Re: [hatari-devel] Hatari profiler updates and CPU cycle questions|
[ Thread Index |
| More lists.tuxfamily.org/hatari-devel Archives
- To: hatari-devel@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [hatari-devel] Hatari profiler updates and CPU cycle questions
- From: Douglas Little <doug694@xxxxxxxxxxxxxx>
- Date: Thu, 31 Jan 2013 22:20:44 +0000
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=googlemail.com; s=20120113; h=mime-version:x-received:in-reply-to:references:date:message-id :subject:from:to:content-type; bh=hDmWEdUB20w9dZzAilWnzlePqzLYvLziCyWGbpkZ4C8=; b=ztjDiHUsHRP3kFQJ6B8q+6lKeyMSXTTyqrBBW+jyFO6Tc7MM3VOUxVhYZt4lFPVbxQ QipL3fJS6BpFMuYT4I2kg+UVpamRuXon6RU4BMoB07uuORs/Idvq5/4hoY8y2AO0tj3D 5FzfCZT6egpGZpH0o913KnPZ9EtJwSc6Ick04tCzHx5xKaJZ5UtWSarHA1svE0K1GIDc lZzb6mN05S0wXS/LHqOxEUQrkJU1etVHMqJ9rdagzpa2siXclZEtYJVzZBRmJpASBaYm hcq76d0gxAE+8JgwSY16Vb9Pou9kBUTt6S2hrpgXJ8Q8iBBP2ROfeQoiNjNysvoe4VTC 8O5A==
Could you somehow verify that it's not off-by-one, like the cycles
If it is correct in other respects then yes I think so. If it suffers from similar problems to the cycle timings it could perhaps be a challenge.
The difficulty with the cache miss stuff is that it is naturally erratic - some instructions will repeatedly miss, where subsequent instructions will not (especially if they are small) since longword chunks of memory (or whole lines, if UAE is assuming burst mode) get fetched together - it requires some 'reading between the lines' to decide if it's valid, and it's easy to convince yourself one way or the other, more than any of the other figures.
I will study it though to see if I can figure out how accurate it is from specific cases.
Then there's the data cache thing, but as earlier discussed,
that would probably need separate statistics for reads and writes.
Yes this one is more tricky to implement because unlike i-cache, not all instructions have a mapping to the d-cache. Only instructions which reference memory. So it would be necessary to record BOTH misses AND hits in order to get ratios. Misses would not be enough I think, unless hits could be deduced from known instruction types which 'did not miss' (recording hits seems easier if UAE provides it).
Anyway this is probably best left alone until the cycle timing and i-cache misses are verified and any issues fixed, as it will just complicate everything I expect :-)