[Intel-gfx] [PATCH i-g-t] tests/i915/kms_big_fb: trigger async flip with a dummy flip

[Murthy, Arun R]

> On 5.7.2022 12.49, Karthik B S wrote:
> > On 7/5/2022 3:08 PM, Murthy, Arun R wrote:
> >>> On 6/28/2022 4:34 PM, Arun R Murthy wrote:
> >>>> In oder to trigger the async flip, a dummy flip is required after
> >>>> sync flip so as to update the watermarks for async in KMD which
> >>>> happens as part of this dummy flip. Thereafter async memory update
> >>>> will act as a trigger register.
> >>>> Capturing the CRC is done after the async flip as async flip at
> >>>> some times can consume fairly a vblank period time.
> >>>>
> >>>> Signed-off-by: Arun R Murthy <arun.r.murthy at intel.com>
> >>>> ---
> >>>>    tests/i915/kms_big_fb.c | 29 +++++++++++++++++++----------
> >>>>    1 file changed, 19 insertions(+), 10 deletions(-)
> >>>>
> >>>> diff --git a/tests/i915/kms_big_fb.c b/tests/i915/kms_big_fb.c
> >>>> index
> >>>> d50fde45..6caf3c31 100644
> >>>> --- a/tests/i915/kms_big_fb.c
> >>>> +++ b/tests/i915/kms_big_fb.c
> >>>> @@ -465,7 +465,7 @@ static bool test_pipe(data_t *data)
> >>>>    static bool
> >>>>    max_hw_stride_async_flip_test(data_t *data)
> >>>>    {
> >>>> -    uint32_t ret, startframe;
> >>>> +    uint32_t ret, frame;
> >>>>        const uint32_t w =
> >>>> data->output->config.default_mode.hdisplay,
> >>>>                   h = data->output->config.default_mode.vdisplay;
> >>>>        igt_plane_t *primary;
> >>>> @@ -519,7 +519,19 @@ max_hw_stride_async_flip_test(data_t *data)
> >>>>
> >>> DRM_MODE_ATOMIC_ALLOW_MODESET, NULL);
> >>>>            igt_wait_for_vblank(data->drm_fd, data-
> >>>> display.pipes[primary->pipe->pipe].crtc_offset);
> >>>> -        startframe = kmstest_get_vblank(data->drm_fd, data->pipe,
> >>> 0) + 1;
> >>>> +        /*
> >>>> +         * In older platforms (<= Gen10), async address update bit
> >>>> +is
> >>> double buffered.
> >>>> +         * So flip timestamp can be verified only from the second
> >>>> flip.
> >>>> +         * The first async flip just enables the async address update.
> >>>> +         * In platforms greater than DISPLAY13 the first async
> >>>> +flip
> >>>> will
> >>> be discarded
> >>>> +         * in order to change the watermark levels as per the
> >>> optimization. Hence the
> >>>> +         * subsequent async flips will actually do the
> >>>> +asynchronous
> >>> flips.
> >>>> +         */
> >>>> +        ret = drmModePageFlip(data->drm_fd, data->output-
> >>>> config.crtc->crtc_id,
> >>>> +                              data->big_fb_flip[i].fb_id,
> >>>> +
> >>> DRM_MODE_PAGE_FLIP_ASYNC, NULL);
> >>>> +        igt_wait_for_vblank(data->drm_fd, data-
> >>>> display.pipes[primary->pipe->pipe].crtc_offset);
> >>>> +        frame = kmstest_get_vblank(data->drm_fd, data->pipe, 0) +
> >>> 1;
> >>>
> >>> Hi,
> >>>
> >>> Should this be added inside a gen check similar to kms_async_flips?
> >> Yes sorry missed it!
> >>
> >>>>            for (int j = 0; j < 2; j++) {
> >>>>                do {
> >>>> @@ -528,23 +540,20 @@ max_hw_stride_async_flip_test(data_t
> *data)
> >>>>
> >>> DRM_MODE_PAGE_FLIP_ASYNC, NULL);
> >>>>                } while (ret == -EBUSY);
> >>>>                igt_assert(ret == 0);
> >>>> +            igt_pipe_crc_get_for_frame(data->drm_fd, data-
> >>>> pipe_crc,
> >>>> +                       frame, &compare_crc);
> >>>>
> >>>> +            frame = kmstest_get_vblank(data->drm_fd, data-
> >>>> pipe, 0) + 1;
> >>>>                do {
> >>>>                    ret = drmModePageFlip(data->drm_fd, data-
> >>>> output->config.crtc->crtc_id,
> >>>>                                  data->big_fb.fb_id,
> >>>>
> >>> DRM_MODE_PAGE_FLIP_ASYNC, NULL);
> >>>>                } while (ret == -EBUSY);
> >>>>                igt_assert(ret == 0);
> >>>> +            igt_pipe_crc_get_for_frame(data->drm_fd, data-
> >>>> pipe_crc,
> >>>> +                       frame, &async_crc);
> >>> We should be moving this IMHO. By waiting for vblank after each
> >>> async flip to capture CRC, what is the intended result? Seems to be
> >>> completely changing the existing test logic.
> >> We are getting the vblank count and based on that getting the crc.
> >> Now we know for async flip at certain times it can consume a time
> >> equivalent to a vblank period. So in those scenarios getting crc
> >> based on the vblank goes for a toss. Hence capturing the vblank start
> >> time stamp at the beginning of each flip.
> >
> > Hi,
> >
> > But what is the the reference CRC we're expecting? The original logic
> > is designed to match on one iteration and mismatch on the next using a
> > combination of fb's by using async flips. But with these changes that
> > whole logic goes for a toss?
> >
> >>
> >>> Also, seems like we are overwriting the CRC captured for j = 0, as
> >>> comparison is done outside this loop. Is this done on purpose?
> >> Now with the changing the vblank start frame capture being added
> >> before the async flip, CRC can be captured outside the loop as well,
> >> but makes no sense. To maintain this order moving the CRC Capture
> >> also after each frame.
> >
> > The CRC comparison is still outside the loop, so no point capturing
> > CRC's if we finally discard it anyway?
> >
> 
> I think generally the idea Arun is changing is correct but he's missed how the
> test work.
> 
> I see there's Ville's change on kernel side for display_ver >=13 first async_flip
> is unconditionally (intentionally) missed, this is at
> intel_plane_do_async_flip(..) so this test will need adjustment
> 
> What Arun seem to have missed is on test those nested loops how they
> work, that part probably should've originally been commented in code bit
> better.
> 
> On original code there's after loop for j two time
> igt_pipe_crc_get_for_frame(..), first will capture crc from duration of loop of
> j, second will *wait* and capture current crc while there's framebuffer
> "big_fb" on screen (wait is about 1,5 frames). Inside j loop, against i is async
> flipped with fb from "big_fb_flip[i]" where one one fb look exactly like
> "big_fb" and check with this crc doesn't change when flip with same content.
> Then another "big_fb_flip[i]" is green color and check crc will change.
> 
> Now when inside "j" loop Arun added igt_pipe_crc_get_for_frame(..) this test
> becomes completely different because async flipping no more is tested do
> perform back'n'forth flipping.
> 

Thanks for your comments JP, atleast its clear that this fix is required. But as commented
I am not aware of the concept behind this test, It would be better if someone can take up this
fix and correct the CRC capture sequence as well.

Thanks and Regards,
Arun R Murthy
--------------------