Bug 67062

Created attachment 82652 [details]
dmesg info

Created attachment 82653 [details]
Xorg.0.log

I presume this regression will be recovered once mesa is setting the correct mocs.

Did you retest with latest mesa and its mocs?

With latest mesa, the performance improved.

EgyptHD_FixedTime_Offscreen   Performance
-----------------------------------------------------
                     (queued_50b44a4)(queued_0d8ff1) (queued_86281e) latest
(master)00d32cd5              97          86            86
(master)19031294 (latest)     103         97            98

But still a little regression, albeit less so. Afaik MOCS on ivb/hsw is rather limited and can't set the age control stuff. So I guess we still need to correct this in the kernel.

Right, but iirc, the mocs should be setting the cache age to 0 - why I thought it would restore the regression (and posssibly even improve). Maybe still a missing mocs? Or some other secondary effect.

Just to clarify, mocs sets age3. 3 is the youngest age and stays in longest. 0 is the oldest and first to evict.

So my understanding is totally backwards, and the PTE should set 0 with userspace using mocs to mark high priority textures.

Created attachment 83060 [details] [review]
try eLLC default age of 3

Created attachment 83061 [details] [review]
try LLC default age of 3

Ye Tian,

Please try both patches in all 3 configurations:
with one patch
with the second patch
with both patches.

Thanks.

To clarify: Mesa currently doesn't set any MOCS overrides for LLC/eLLC.  It sets them to "Use the PTE values."  (The recent MOCS additions only set L3 cacheability, which is different and separate.)

We're relying on the kernel to set everything to be LLC+eLLC WB-cacheable (or WT cacheable where appropriate).  In the future, we may try some heuristics to play with ages, but it's hard to know what the right settings are, and there are no concrete plans currently.

(In reply to comment #12)
> To clarify: Mesa currently doesn't set any MOCS overrides for LLC/eLLC.  It
> sets them to "Use the PTE values."  (The recent MOCS additions only set L3
> cacheability, which is different and separate.)
> 
> We're relying on the kernel to set everything to be LLC+eLLC WB-cacheable
> (or WT cacheable where appropriate).  In the future, we may try some
> heuristics to play with ages, but it's hard to know what the right settings
> are, and there are no concrete plans currently.

I was wondering about this. I can't find any mention of L3 + MOCS. Where did you guys find this? Docs make be believe 01b is uncached.

(In reply to comment #11)

Performance for the two patches:
First patch:  100 fps
second patch: 105 fps (LLC default )
Both patches: 105 fps 

But I don't know how to try both patches:
eLLC default 
 	if (level != I915_CACHE_NONE)
-		pte |= HSW_WB_ELLC_LLC_AGE0;
+		pte |= HSW_ELLC;
 
LLC default 
  	if (level != I915_CACHE_NONE)
-		pte |= HSW_WB_LLC_AGE0;
+		pte |= HSW_LLC;

Does it mean:
if (level != I915_CACHE_NONE)
-		pte |= HSW_WB_ELLC_LLC_AGE0;
+		pte |= HSW_ELLC;
+               pte |= HSW_LLC;

(In reply to comment #14)
> (In reply to comment #11)
> 
> Performance for the two patches:
> First patch:  100 fps
> second patch: 105 fps (LLC default )
> Both patches: 105 fps 
> 
> But I don't know how to try both patches:

Don't worry about it... I'll get this merged to -next-queued/-nightly ASAP

http://cgit.freedesktop.org/~bwidawsk/drm-intel/commit/?h=drm-intel-next-queued&id=b59f98153cd27b72b25abb9a3d5d50e1cd68b2a4

Should be fixed with:

commit 87a6b688ccc78b2c54bee56879c6d195d2457ebe
Author: Ben Widawsky <ben@bwidawsk.net>
Date:   Sun Aug 4 23:47:29 2013 -0700

    drm/i915/hsw: Change default LLC age to 3

verified it.

(give the credit to Ben)

You mean the patch that you later decided caused a 20% regression in the very same benchmark...