Re: [PATCH] drm/i915: Revert RPS UP_EI value for SandyBridge and IvyBridge

[Ben Widawsky <ben@bwidawsk.net>]

On Sun, 21 Oct 2012 15:44:02 +0100
Chris Wilson <chris@chris-wilson.co.uk> wrote:

> Even though we do not use the EI mode for determining when to change GPU
> frequencies for RPS, changing this value causes no up interrupts to be
> generated whilst an OpenGL client runs.
> 
> Fixes regression from commit 1ee9ae3244c4789f3184c5123f3b2d7e405b3f4c
> Author: Daniel Vetter <daniel.vetter@ffwll.ch>
> Date:   Wed Aug 15 10:41:45 2012 +0200
> 
>     drm/i915: use hsw rps tuning values everywhere on gen6+
> 
> Reported-by: Eric Anholt <eric@anholt.net>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Eric Anholt <eric@anholt.net>
> Cc: Ben Widawsky <ben@bwidawsk.net>
> Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
> Cc: stable@vger.kernel.org
> ---
>  drivers/gpu/drm/i915/intel_pm.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> index 81e88c2..15b585e 100644
> --- a/drivers/gpu/drm/i915/intel_pm.c
> +++ b/drivers/gpu/drm/i915/intel_pm.c
> @@ -2493,7 +2493,7 @@ static void gen6_enable_rps(struct drm_device *dev)
>  
>  	I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
>  	I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
> -	I915_WRITE(GEN6_RP_UP_EI, 66000);
> +	I915_WRITE(GEN6_RP_UP_EI, IS_HASWELL(dev) ? 66000 : 100000);
>  	I915_WRITE(GEN6_RP_DOWN_EI, 350000);
>  
>  	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);




I've not verified this with interrupts, but simply reading back the
current frequency using the sysfs interface.

What I've seen running xonotic on Ivybridge is that we do bump the
frequency initially, and then stops. Nearing the end of the demo, we
again raise the frequency. Note that on my system, both before, and
after this patch, I am able to get to the max GPU frequency with the
xonotic demo.

Specifically, on my IVB which has a range of 350->1100 with an RP1 of
650. I see the following (the demo is roughly 2 minutes)

without patch:
Within a few seconds we cycle up to 750
Nothing for about 30 seconds
very slowly cycle up to 1100 (*just* before the demo ends)
demo ends; throttle down to 350 quickly

with patch:
Within a few seconds we cycle up to 1000
Nothing for about 30 seconds
cycle up to 1100
demo ends; throttle down to 350 slowly

I think if this fixes someones critical issue, it's great, but
unfortunately I do not see the problem the patch claims to fix.
Furthermore, none of us can really make sense of why this has the effect
that it does, but I believe a lot of that is because the workloads we
run (in this case xonotic) are very blackbox.

Personally, on this IVB, I think the behavior before the patch is more
desirable because it stays near RP1 for a longer period of time, and
drops to RP0 quickly (but it's definitely a matter of opinion).

-- 
Ben Widawsky, Intel Open Source Technology Center