[Intel-gfx] [PATCH i-g-t 3/3] gem_memory_bandwidth: Add test

Chris Wilson chris at chris-wilson.co.uk
Tue Oct 6 02:25:14 PDT 2015

On Mon, Oct 05, 2015 at 02:42:25PM +0300, Antti Koskipaa wrote:
> This is a benchmark for testing the GPU read and write bandwidth.

I don't see any pass/fail logic so why put in amongst the regression

Please don't advertise this as gem_memory_bandwidth either, it is just a
rendercopy speed test. Note that LOOPS is critical here, setting it to 1
and we measure dispatch overhead (also v. important), setting it to
10,000 we saturate the RCS ring, force RPS boosting etc. Similarly,
measuring performance at min/max clocks is also interesting. There are a
lot of implicit variables here at play that are not taken account of.

I would like to have a set of perf tests as regression tests. What I
have in mind is basically doing a fixed workload inside igt and have the
core automatically compare the runtime of a subtest against a local
database of results and s/PASS/SLOW/ (or outright FAIL) if it exceeds
some threshold. 

Chris Wilson, Intel Open Source Technology Centre

More information about the Intel-gfx mailing list