[Intel-gfx] [PATCH i-g-t 00/29] Update i915 perf tests for Gen8+

Lionel Landwerlin lionel.g.landwerlin at intel.com
Tue Apr 25 22:32:32 UTC 2017


Hi,

Apologies for this unfriendly series, in the end squashing everything
into a single commit didn't look too good.

We went through quite a few iterations to figure all the instability
issues on Gen8+. This has finally reached a state where you can run
the tests on a machine with other applications using the GPU in
parallel. This gives us confidence that we now have a pretty good
understanding of how the filtering of reports should be done as well
as things that might affect periodic report timings.

Most of Rob's patches are already reviewed or co-authored. I'm not
really expecting a someone to thoroughly review all of these changes
as it takes a fair amount of time to get into all of the fiddly
details, but if someone could look at the end result and quickly read
through to check there isn't something terribly wrong, that would be
helpful.

Thanks a lot,

Lionel Landwerlin (8):
  igt/perf: add utility function for checking periodic reports
  igt/perf: make stream_fd a global variable
  igt/perf: update max buffer size for reading reports
  igt/perf: rework oa-exponent test
  igt/perf: make enable-disable more reliable
  igt/perf: make buffer-fill more reliable
  igt/perf: load gt_boost_freq_mhz as max gt frequency
  igt/perf: remove unused frequency functions

Robert Bragg (21):
  igt/perf: generalize lookup for test metric set
  igt/perf: improve robustness of polling/blocking tests
  igt/perf: init timestamp freq and oa format per devid
  igt/perf: update init_sys_info for skl with per-gt configs
  igt/perf: add gen8 formats
  igt/perf: fix a counter indexing
  igt/perf: generalize checks for undefined A counters
  igt/perf: generalize reading gpu ticks from reports
  igt/perf: move timebase + oa exponent utilities up
  igt/perf: wrap emission of MI_REPORT_PERF_COUNT
  igt/perf: handling printing gen8 formats
  igt/perf: avoid assumptions about oa exponent <-> freq mappings
  igt/perf: allow 10% margin matching oa/sysfs freq in test_oa_exponents
  igt/perf: s/test_perf_ctx_mi_rpc/hsw_test_single_ctx_counters/
  igt/perf: don't assume constant of 40 EUs
  igt/perf: consider ctx-switch reports while polling/blocking
  igt/perf: factor out oa report sanity checking
  igt/perf: print [un]slice freq and report reasons in debug
  igt/perf: update print_reports to print context ID
  igt/perf: add per context filtering test for gen8+
  igt/perf: fix rc6 test

 tests/perf.c | 2896 +++++++++++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 2339 insertions(+), 557 deletions(-)

--
2.11.0


More information about the Intel-gfx mailing list