[Intel-gfx] [PATCH 07/27] drm/i915: Squash repeated awaits on the same fence

[Chris Wilson]

On Wed, Apr 26, 2017 at 01:13:41PM +0100, Tvrtko Ursulin wrote:
> I was thinking of exactly the same thing as this patch does, u64
> context id as key, u32 seqnos (wrapped in a container with
> hlist_node).

#define NSYNC 32
struct intel_timeline_sync { /* kmalloc-256 slab */
	struct hlist_node node;
        u64 prefix;
	u32 bitmap;
	u32 seqno[NSYNC];
};
DECLARE_HASHTABLE(sync, 7);

If I squint, the numbers favour the idr. ;)

Tbh, the precence of the squash is noticeable and well above the noise,
the difference between a hashtable and the idr, far below the noise
floor (in a testcase intended to stress the efficacy of this patch). The
cost of reference counting in execbuffer and the reservation_object hide
all ills. :(

What I am not happy with is the 1<<7 buckets I'm currently using.
Thinking about the idle pruning, there shouldn't be any reason to go
above 1<<3, I hope?

Do we start on GEM_STATS?
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre