By simply removing volatile qualifier in draw() from uint32_t *ptr, memory
accesses improve a lot in that test. We practically saw 1 order of magnitude in
the performance of VGEM and the speed was equally on using drm_intel specific
ioctls.
The inclusion of volatile has forced the compiler to load and store the
variable from memory on every run, hurting quite badly the performance.
Therefore, performance is dropped due many read operations in the mapped
pointer which is a no-no when objects are write-combining mapped.
This patch also removes sleep, that bothers performance measurements
This CL introduces -p option to vgem_fb_test. For now it only prints the
time to mmap imported vgem buffers, which is particularly useful to
benchmark and compare with other driver implementations.
Signed-off-by: Tiago Vignatti <tiago.vignatti@intel.com>
Measure the overhead of execution when doing nothing, switching between
a pair of contexts, or creating a new context every time.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
It is possible to race between unreference of the underlying BO and
importing it from prime_fd/name. Verify that the behaviour of libdrm
is consistent for prime/flink.
v2: more comments in source file, dropped extra whitespace
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Cc: Thomas Wood <thomas.wood@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
By measuring both the query and the event round trip time, we can make a
reasonable estimate of how long it takes for the query to send the
vblank following an interrupt.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In my haste to merge the two gem_pread/gem_pwrite, I forgot to write up
the command line switch to getopt.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
SKL changed state base address command.
Cc: Thomas Wood <thomas.wood@intel.com>
Signed-off-by: Dominik Zeromski <dominik.zeromski@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
BDW changed structure of surface state and interface descriptors.
Commands like state base address, gpgpu walker were extended.
Cc: Thomas Wood <thomas.wood@intel.com>
Signed-off-by: Dominik Zeromski <dominik.zeromski@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
The gpgpu fill utility functions are used in separate test so it's
logical to keep them in separate file. This is similar to what media
spin test did in the past.
Functionally only gpgpu kernel changed. Send instruction payload size
was reduced. Since offset is incremented by 0x10 bytes there is no point
in using larger writes.
Cc: Thomas Wood <thomas.wood@intel.com>
Signed-off-by: Dominik Zeromski <dominik.zeromski@intel.com>
[Thomas: Fix typo of gpgpu_fill.h in Makefile.sources]
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
Disable the tools / demo code that do not currently build
for android until they can be fixed.
Affected tools / demos
intel_display_crc
intel_sprite_on
v2: intel_display_crc compiled conditionally on ANDROID_HAS_CAIRO
flag.
v3: removed intel_reg from the skip list as Thomas has prepared
a patch to fix it for Android.
Signed-off-by: Derek Morton <derek.j.morton@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
Exit immediately if the test is already exiting and igt_fail is called.
This can happen if an igt_assert fails in an exit handler.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91349
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
Based on an idea from Jani Nikula.
Cc: Jani Nikula <jani.nikula@intel.com>
Cc: Derek Morton <derek.j.morton@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
A rudimentary tool on top of the igt_stats library. Reads a list of
numbers from stdin or from a file and prints the estimate of the central
location, aka average.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We don't always have precise integers with which to store, so allow
degrading to double precision floating point based on available input.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When reading the timestamp register with single 64b read, we are observing
invalid values on x86_64:
[f = valid counter value | X = garbage]
i386: 0x0000000fffffffff
x86_64: 0xffffffffXXXXXXXX
Test checks if the counter is moving and increasing.
Add a check to see if we can use (reg | 1) flag to get a proper 36b timestamp,
shifting the value on x86_64 if we can't.
v2: More iterations of monotonic test, comments, minor fixups (Chris)
v3: Skip tests if reg_read is not supported
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
If we are waiting on a handle, and its request is cancelled due to a
hang, we should report that explicitly through the ABI.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This test should test the interactions between fbcon and the
frontbuffer tracking infrastructure.
Right now the PSR test fails, but as soon as we merge the following
kernel patches, the test wills tart passing:
- drm/i915: PSR: Flush means invalidate + flush
- drm/i915: fbdev restore mode needs to invalidate frontbuffer
- drm/i915: fbdev_set_par reliably invalidating frontbuffer
I didn't want to make this a subtest of kms_frontbuffer_tracking just
because when I wrote it, I really didn't have in mind the fact that
someone might just close the DRM fd in the middle of a subtest.
After this commit we'll have a little bit of duplicated code among
tests. I'll clean this up later.
v2:
- Refactor the code a little bit.
- Add the suspend subtests.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
So we don't need to initialize CRCs we won't use. This makes the
modesetfrombusy and badstride tests a little faster.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
And drop the now unnecessary universal plane definitions. The 2.4.55
version is already quite old, so we should be fine.
Thanks to Thomas Wood for quickly spotting this.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@>
Since on BDW the big modes work fine, I want to test them without
having to recompile the test every time. Let's make std_1024_mode part
of opt.small_modes.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
After the last commit, it's not really needed anymore, so let's get
rid of it: less code lines and less non-standard enums.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Make sure the features are disabled and stay disabled. Otherwise we
could just wrongly think they are disabled while they are just in the
process of being enabled.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
I tried explaining these to people twice and realized how bad the
current naming was. Let's hope the new one is better.
The downside is that the tests were renamed and patches mentioning
them can't be fixed.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
So we can make sure all code paths are properly exercised and lead to
proper frontbuffer tracking.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Only really assert FBC_DISABLED in case the primary plane is actually
disabled. If the sprite plane is covering the whole screen but the
primary plane is still enabled, FBC can remain enabled.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
While debugging the test failures I tried different values for the
offsets, so having a central place to change them is the only sane
way.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
SKL and BXT have some new plane capabilities that previous generations
didn't have; we need to update some of our universal plane tests to
expect success rather than failure when running on these platforms.
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
