Precise vblank timestamping for VC4 kms.
On 06/23/2016 09:28 AM, Daniel Vetter wrote: > On Thu, Jun 23, 2016 at 8:17 AM, Mario Kleiner > wrote: >> The following patch implements precise vblank timestamping >> for RaspberryPi's VC4, at least for standard progressive >> scan display modes. >> >> It has been tested on the HDMI output with half a dozen different >> video modes using special hardware measurement equipment to compare >> generated time stamps against reality. According to the tests it >> works well in its current form. >> >> Due to hw limitations of the VC4, timestamps can't be scanline >> accurate when taken within vblank, as explained in the patch, >> but at least they will never be off by more than 1 vblank >> duration, and are typically still accurate to ~0.1 msecs >> for the common case when the timestamping is triggered from >> vblank interrupt. >> >> The patch exposed some problems with how the drm core handles >> calculation of vblank timestamping constants for interlaced >> video modes in drm_calc_timestamping_constants(). Seems it cuts >> the expected frame duration framedur_ns into half for interlaced >> modes two times, so it ends up expecting a field duration half >> of what it should be and then miscalculates vblank counter increments >> as soon as vblank timestamping is supported and the core tries to >> derive vblank counts from it. To work around this bug, for the >> moment the vblank timestamping will disable itself for interlaced >> modes and only work for regular progressive scan. >> >> Eric: In the patch i need to calculate fifo_lines as the capacity >> of what seems to be a multi-line fifo line buffer for composited output >> scanlines between the HVS and the PV, or maybe an input fifo between >> the framebuffer(s) and the HVS? The formula i currently use is ad-hoc, >> found by trial and error. It works reasonably well for a range of >> video modes i could test with my measurement equipment, but it would >> be good to replace it by one that is actually accurately derived from >> your hardware docs. >> >> Other than that, this should be good to go. > > Just out of curiosity: Is there no timestamp register that samples a > refclock on each vblank that could be used instead? That seems to be a > somewhat common feature afaict (but I don't know about vc4). > -Daniel > Not according to what Eric told me. I also couldn't find anything in the documented register set from vc4_regs.h, or poking around the neighborhood of those regs. The HVS register i use now seems to be the only thing one can use to approximate the typical scanoutpos method well enough to be useful and precise enough. It would be good to know what other SoCs support such timetamp regs. Using them probably wouldn't be much more simple than the current scanoutpos method, with the need to then synchronize/map the gpu clocks time to the hosts CLOCK_MONOTONIC time, but having the precision timestamps would be valuable, also for modern compositors. -mario
Precise vblank timestamping for VC4 kms.
On Thu, Jun 23, 2016 at 08:17:49AM +0200, Mario Kleiner wrote: > The following patch implements precise vblank timestamping > for RaspberryPi's VC4, at least for standard progressive > scan display modes. > > It has been tested on the HDMI output with half a dozen different > video modes using special hardware measurement equipment to compare > generated time stamps against reality. According to the tests it > works well in its current form. > > Due to hw limitations of the VC4, timestamps can't be scanline > accurate when taken within vblank, as explained in the patch, > but at least they will never be off by more than 1 vblank > duration, and are typically still accurate to ~0.1 msecs > for the common case when the timestamping is triggered from > vblank interrupt. > > The patch exposed some problems with how the drm core handles > calculation of vblank timestamping constants for interlaced > video modes in drm_calc_timestamping_constants(). Seems it cuts > the expected frame duration framedur_ns into half for interlaced > modes two times, so it ends up expecting a field duration half > of what it should be and then miscalculates vblank counter increments > as soon as vblank timestamping is supported and the core tries to > derive vblank counts from it. To work around this bug, for the > moment the vblank timestamping will disable itself for interlaced > modes and only work for regular progressive scan. The code should be correct for i915. The framedur_ns is actually the duration of a single field (since we get a vblank interrupt for each field). But it's going to be wrong for anyone that populates the crtc_ timings with CRTC_INTERLACE_HALVE_V. So to fix it, you'd need the driver to tell drm_calc_timestamping_constants() whether the vertical timings were already halved or not. -- Ville Syrjälä Intel OTC
Precise vblank timestamping for VC4 kms.
On Thu, Jun 23, 2016 at 8:17 AM, Mario Kleiner wrote: > The following patch implements precise vblank timestamping > for RaspberryPi's VC4, at least for standard progressive > scan display modes. > > It has been tested on the HDMI output with half a dozen different > video modes using special hardware measurement equipment to compare > generated time stamps against reality. According to the tests it > works well in its current form. > > Due to hw limitations of the VC4, timestamps can't be scanline > accurate when taken within vblank, as explained in the patch, > but at least they will never be off by more than 1 vblank > duration, and are typically still accurate to ~0.1 msecs > for the common case when the timestamping is triggered from > vblank interrupt. > > The patch exposed some problems with how the drm core handles > calculation of vblank timestamping constants for interlaced > video modes in drm_calc_timestamping_constants(). Seems it cuts > the expected frame duration framedur_ns into half for interlaced > modes two times, so it ends up expecting a field duration half > of what it should be and then miscalculates vblank counter increments > as soon as vblank timestamping is supported and the core tries to > derive vblank counts from it. To work around this bug, for the > moment the vblank timestamping will disable itself for interlaced > modes and only work for regular progressive scan. > > Eric: In the patch i need to calculate fifo_lines as the capacity > of what seems to be a multi-line fifo line buffer for composited output > scanlines between the HVS and the PV, or maybe an input fifo between > the framebuffer(s) and the HVS? The formula i currently use is ad-hoc, > found by trial and error. It works reasonably well for a range of > video modes i could test with my measurement equipment, but it would > be good to replace it by one that is actually accurately derived from > your hardware docs. > > Other than that, this should be good to go. Just out of curiosity: Is there no timestamp register that samples a refclock on each vblank that could be used instead? That seems to be a somewhat common feature afaict (but I don't know about vc4). -Daniel -- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
Precise vblank timestamping for VC4 kms.
The following patch implements precise vblank timestamping for RaspberryPi's VC4, at least for standard progressive scan display modes. It has been tested on the HDMI output with half a dozen different video modes using special hardware measurement equipment to compare generated time stamps against reality. According to the tests it works well in its current form. Due to hw limitations of the VC4, timestamps can't be scanline accurate when taken within vblank, as explained in the patch, but at least they will never be off by more than 1 vblank duration, and are typically still accurate to ~0.1 msecs for the common case when the timestamping is triggered from vblank interrupt. The patch exposed some problems with how the drm core handles calculation of vblank timestamping constants for interlaced video modes in drm_calc_timestamping_constants(). Seems it cuts the expected frame duration framedur_ns into half for interlaced modes two times, so it ends up expecting a field duration half of what it should be and then miscalculates vblank counter increments as soon as vblank timestamping is supported and the core tries to derive vblank counts from it. To work around this bug, for the moment the vblank timestamping will disable itself for interlaced modes and only work for regular progressive scan. Eric: In the patch i need to calculate fifo_lines as the capacity of what seems to be a multi-line fifo line buffer for composited output scanlines between the HVS and the PV, or maybe an input fifo between the framebuffer(s) and the HVS? The formula i currently use is ad-hoc, found by trial and error. It works reasonably well for a range of video modes i could test with my measurement equipment, but it would be good to replace it by one that is actually accurately derived from your hardware docs. Other than that, this should be good to go. thanks, -mario
