[Intel-gfx] [RFC][PATCH] Userptr benchmark

[Chris Wilson]

On Wed, Feb 05, 2014 at 12:41:45PM +0000, Tvrtko Ursulin wrote:
> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> 
> A simple userptr benchmark measuring creation and destruction of userptr
> surfaces and also impact of having a different number of them in the
> process address space.
> 
> Example test output from i7-4550U running Android is below.
> 
> Questions, comments and ideas are welcome.

One silly idea I wanted to test was that speed of read/write access to
ptr was not affected by wrapping it up in a userptr.
 
> IGT-Version: 1.5-NOT-GIT (android-ia) (Linux: 3.10.20-g667dce8-dirty x86_64)
unsync vs sync
- unsync is much faster at creating userptr
  (not having to hook up and search the mmu-notifier)
- unsync is reasonably faster for destroying userptr
- there is no scaling issue with unsync, and minor (log(n)) scaling
  factor for sync

That seems in line with our expectations of i915_gem_userptr.c, which is
reassuring.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre