On 04/08/2013 06:28 AM, Chris Wilson wrote:
> Enabling context support increases SwapBuffers latency by about 20%
> (measured on an i7-3720qm). We can offset that loss slightly by enabling
> faster caching for the contexts. As they are not backed by any
> particular cache (such as the sampler or render caches) our only option
> is to select the generic mid-level cache. This reduces the latency of
> the swap by about 5%.
>
> Oddly this effect can be observed running smokin-guns on IVB at
> 1280x1024:
> Using BLT copies for swaps: 151.67 fps
> Using Render copies for swaps (unpatched): 141.70 fps
> With contexts disabled: 150.23 fps
> With contexts in L3$: 150.77 fps
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Ben Widawsky <ben at bwidawsk.net>
> Cc: Kenneth Graunke <kenneth at whitecape.org>
> ---
> drivers/gpu/drm/i915/i915_gem_context.c | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
> index 94d873a..a1e8ecb 100644
> --- a/drivers/gpu/drm/i915/i915_gem_context.c
> +++ b/drivers/gpu/drm/i915/i915_gem_context.c
> @@ -152,6 +152,13 @@ create_hw_context(struct drm_device *dev,
> return ERR_PTR(-ENOMEM);
> }
>
> + if (INTEL_INFO(dev)->gen >= 7) {
> + ret = i915_gem_object_set_cache_level(ctx->obj,
> + I915_CACHE_LLC_MLC);
> + if (ret)
> + goto err_out;
> + }
> +
> /* The ring associated with the context object is handled by the normal
> * object tracking code. We give an initial ring value simple to pass an
> * assertion in the context switch code.
Sounds good to me, Chris. Is this also useful on Sandybridge? I don't
see why we couldn't use both L3 & LLC there as well.
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
On a tangent, it would be nice to clean up the I915_CACHE_* enums. The
term "MLC" isn't really used in the modern documentation, and there's
even text that says "Ivybridge doesn't have MLC". But what it really
means here is LLC + L3. It's probably not worth reworking until we can
send out the new Haswell bits, though.
Thanks for doing this!
--Ken
