[PATCH v3 0/2] MR2226: gdi32: Track ticks since draw start per window_surface.
Track ticks since draw start per window_surface, instead of per DC as is currently the case. This change helps reduce visual glitches caused by badly timed flushes when drawing to the same window_surface from multiple DCs, e.g., child windows (see first commit), or in some special scenarios when the window_surface is updated (see second commit). (Much) more information about this issue and the rationale for the changes can be found in the commit messages. Note that since this is an inherently timing related issue, the visual glitches depend on the application specific draw patterns, drawing speed (and thus processor performance and load) etc. Here is a capture which exhibits the issue with the current implementation: [resizing-regedit-per-dc-ticks.mkv](/uploads/d7d30d009cc4db1337cbc3266df5ae17/resizing-regedit-per-dc-ticks.mkv) And here is a capture of the same scenario with the proposed changes applied, which shows the improvement: [resizing-regedit-per-surface-ticks.mkv](/uploads/6f878568862f5fb5157567341e4e452f/resizing-regedit-per-surface-ticks.mkv) -- v3: win32u: Reset draw_start_ticks for new window_surface. gdi32: Track ticks since draw start per window_surface. https://gitlab.winehq.org/wine/wine/-/merge_requests/2226
From: Alexandros Frantzis <alexandros.frantzis(a)collabora.com> Track ticks since draw start per window_surface, instead of per DC as is currently the case. This change helps reduce visual glitches caused by badly timed flushes when drawing to the same window_surface from multiple DCs (e.g., for child windows). This approach is a better fit for the current heuristic for forcing flushing, which consults the shared window_surface bounds to decide whether this is the start of a draw in order to reset the (currently per DC) draw start time. The problem in the current implementation occurs when a drawing to a DC begins with an already damaged window_surface, e.g., due to draws from other DCs targeting that window_surface. In such a case, the DC draw start time is not reset and refers to the start of some previous draw sequence using this DC, thus increasing the chances that the 50ms time flush limit will be eventually exceeded in the middle of the current draw sequence. In other words, the state of the (shared) window_surface damage is not a reliable indicator of the beginning (or not) of a draw to a DC. An example, assuming DC1 and DC2 target the same window_surface: DC1.start_ticks = 0 DC2.start_ticks = 0 FLUSH_PERIOD = 50 0 flush 1 draw to DC1 -> DC1.start_ticks = 1 ... [no flush] ... 2 draw to DC2 -> DC2.start_ticks remains 0 since surface is damaged ... 50 flush 51 draw to DC1 -> DC1.start_ticks = 51 ... [no flush] ... 52 draw to DC2 -> DC2.start_ticks remains 0 since surface is damaged, current - DC2.start_ticks > FLUSH_PERIOD so we are forced to flush in the middle of the drawing sequence => potential glitch Tracking the draw start per window_surface ameliorates the problem because the beginning of a draw on a DC targeting an undamaged window_surface resets the start time for all DCs targeting that window_surface: ... 50 flush 51 draw to DC1 -> surface.draw_ticks = 51 ... [no flush] ... 52 draw to DC2 -> surface.draw_ticks remains 51 since surface is damaged, but current - surface.draw_ticks < FLUSH_PERIOD, so we do not flush Signed-off-by: Alexandros Frantzis <alexandros.frantzis(a)collabora.com> --- dlls/win32u/dibdrv/dc.c | 6 +++--- include/wine/gdi_driver.h | 1 + 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/dlls/win32u/dibdrv/dc.c b/dlls/win32u/dibdrv/dc.c index 68a009f576e..f96374c1229 100644 --- a/dlls/win32u/dibdrv/dc.c +++ b/dlls/win32u/dibdrv/dc.c @@ -731,7 +731,6 @@ struct windrv_physdev struct gdi_physdev dev; struct dibdrv_physdev *dibdrv; struct window_surface *surface; - DWORD start_ticks; }; static const struct gdi_dc_funcs window_driver; @@ -745,13 +744,14 @@ static inline void lock_surface( struct windrv_physdev *dev ) { /* gdi_lock should not be locked */ dev->surface->funcs->lock( dev->surface ); - if (IsRectEmpty( dev->dibdrv->bounds )) dev->start_ticks = NtGetTickCount(); + if (IsRectEmpty( dev->dibdrv->bounds )) dev->surface->draw_start_ticks = NtGetTickCount(); } static inline void unlock_surface( struct windrv_physdev *dev ) { + BOOL should_flush = NtGetTickCount() - dev->surface->draw_start_ticks > FLUSH_PERIOD; dev->surface->funcs->unlock( dev->surface ); - if (NtGetTickCount() - dev->start_ticks > FLUSH_PERIOD) dev->surface->funcs->flush( dev->surface ); + if (should_flush) dev->surface->funcs->flush( dev->surface ); } static void CDECL unlock_bits_surface( struct gdi_image_bits *bits ) diff --git a/include/wine/gdi_driver.h b/include/wine/gdi_driver.h index 3de7e20af9a..fe0894b2820 100644 --- a/include/wine/gdi_driver.h +++ b/include/wine/gdi_driver.h @@ -217,6 +217,7 @@ struct window_surface struct list entry; /* entry in global list managed by user32 */ LONG ref; /* reference count */ RECT rect; /* constant, no locking needed */ + DWORD draw_start_ticks; /* start ticks of fresh draw */ /* driver-specific fields here */ }; -- GitLab https://gitlab.winehq.org/wine/wine/-/merge_requests/2226
From: Alexandros Frantzis <alexandros.frantzis(a)collabora.com> Normally, a new window_surface has no damage so its draw_start_ticks value is reset at the time of the first lock_surface. However, if contents from the old window_surface are copied to the new one when the window_surface for a window is updated, the new window_surface will be initially damaged. We want to reset the draw_start_ticks value in this case too, to avoid flushes occuring soon after such updates, during the initial redraw of the window_surface. Signed-off-by: Alexandros Frantzis <alexandros.frantzis(a)collabora.com> --- dlls/win32u/dibdrv/dc.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/dlls/win32u/dibdrv/dc.c b/dlls/win32u/dibdrv/dc.c index f96374c1229..23ef7f64390 100644 --- a/dlls/win32u/dibdrv/dc.c +++ b/dlls/win32u/dibdrv/dc.c @@ -744,7 +744,8 @@ static inline void lock_surface( struct windrv_physdev *dev ) { /* gdi_lock should not be locked */ dev->surface->funcs->lock( dev->surface ); - if (IsRectEmpty( dev->dibdrv->bounds )) dev->surface->draw_start_ticks = NtGetTickCount(); + if (IsRectEmpty( dev->dibdrv->bounds ) || dev->surface->draw_start_ticks == 0) + dev->surface->draw_start_ticks = NtGetTickCount(); } static inline void unlock_surface( struct windrv_physdev *dev ) -- GitLab https://gitlab.winehq.org/wine/wine/-/merge_requests/2226
participants (2)
-
Alexandros Frantzis -
Alexandros Frantzis (＠afrantzis)