mirror of
https://github.com/libretro/RetroArch.git
synced 2026-11-12 18:45:33 -10:00
Davey Hughes
56952c9a25
* rgui: upload the menu again to the video driver a reinit makes Closing content with RGUI up left the screen black until a key press. The staged content load presents frames between its stages, so RGUI can upload its framebuffer to the old video driver and clear FB_DIRTY just before the driver stage replaces that driver. The new one has no menu texture, nothing in the menu changes, and RGUI never uploads again. command_event_reinit sets FB_DIRTY after its reinit, but the staged load and video-only resets reach the menu only through context_reset, so RGUI marks its framebuffer dirty there. The menu_icon_ownership harness gets an rgui lane that reinits the video driver and checks the next upload reaches the new one. * video: draw a frame queued before a view count shrink with its views Under threaded video the wrapper ran SET_VIEW_COUNT ahead of the frame it had already claimed, so a closing core's last queued frame drew after its view chains were gone: whole, both screens stacked, until the menu's first frame replaced it. The wrapper still answers at once, but a smaller count reaches the driver only once the frames queued before it have drawn. A grow applies at once as before, with the main thread waiting for the preset parse. Vulkan and glcore therefore build view chains only when the count grows, so a deferred shrink never parses a preset. They also keep the headset's whole-frame chain across a count change: it does not depend on the count, and with chains built only on a grow nothing would rebuild it after a shrink to 0. * android: don't make a second screen's EGL surface once Java took it back Java can take the second screen's surface back between the poll that hands it over and the frame that makes its EGL surface. Skip making one then; the screens go back to the main window as they do when a make fails. Java can still take the surface back during the make itself; the node's own ANativeWindow reference keeps the window alive through that. * android: ask a display move by display id and number it A move named its display by its place in the list, and Java looked that up later on the UI thread against whatever list was newest, so a display coming or going in between could send RetroArch elsewhere. Java now gets the display id and fails the move when that display is gone. Native code now counts the ids when it compares display lists, so a display unplugged and plugged back is a new list and its failures clear. A failure was a bare flag, so one move's failure could end the next move and mark its display unable. Each move now carries a number and a failure counts only for the move it names. The move code also loses two stores that were always overwritten. * android: end the refocus hold after a move only when it is done A move onto a display that is no longer valid now fails at once instead of waiting out its timeout. The hold that keeps RetroArch's window unfocusable after a move ends only on the focus loss Android answers it with, not on a gain that was already queued; an earlier hold ends only once the new move has started; and onStop() ends a running hold rather than leaving it to its timer. * video_windows: tell an external monitor by its display type Android lists the AYN Thor's two panels and a DP monitor with the same unknown connection type, so Auto took the monitor for a second panel: on a one-screen phone Game Screen 2 went to it. Java now reads the hidden Display.getType() by reflection where it can and passes internal or external with each display. A display typed internal is never a monitor; where the type is unknown, a non-default display with an unknown connection and a product name now counts as one. * video: give a Presentation on another display its own UI wait The menu waits up to a second for the Presentation it is laid out for to show, but the wait was only counted from the first one. Once a Presentation had shown, one made on another display had no wait left and the menu fell back to the main window at once. * input: keep the menu's mouse where it was while the menu is in a window With Mouse Support on, the menu moving into the Presentation read as a mouse move to (0,0) for a frame, which marks the pointer active there. The view keeps the main window's last mouse position instead. * input: match a window's touches only from a windowed publication A publication without windows after one with them left the old window ids in place, so a window's touch lookup could still match them. Only touches published in a window match now. No Android path publishes that way today. * vulkan: copy a waiting window's views on a new frame A screen window that was up but had no free image on a new frame did not get its views copied, so a later repeat that drew it, as a pause with the menu there does, left black areas where those views go. They are copied on every new frame now. * video: load the idle logo from the assets directory without widgets The idle logo's path came from the widgets, which only start with On-Screen Notifications on, so with them off the idle window stayed black. Without the widgets' path it is built from the assets directory the way the widgets build it. * android: hold the focus when Android moves RetroArch between displays When the display RetroArch is on, or moving to, goes away, Android moves the Activity to the default display and can send that display's focus gain before the removed display's loss. A task moved to another display and straight back gets no configuration change, and can read the old display's loss last too. With the loss last, RetroArch pauses while Android keeps it focused, and no later focus event wakes it. Count these as moves, as RetroArch's own moves are: a removal, and any change of RetroArch's display it did not ask for. From Android 10, with more than one display, so is losing and regaining the top-resumed place while focused, or regaining it unfocused within a second of losing it after a loss no hold asked for. The refocus hold that covers a move's late loss then covers these too: the window is made unfocusable until Android has taken the focus, then gets the gain again. A gain already on its way skips the hold, and a move marked while RetroArch is focused ends after a second with no focus event. * samples: close test gaps in video_windows and android_focus The is_default check set the flag and read it back, so it could not fail; Auto's fallback to the default display now shows the flag at work. Cover a display listed without an ordinal, and display keys at sizes 0 to 3, two unnamed displays and a 63-character name. Drop a stale note from android_focus_test's header. * video: fix wording left over from the display rows The Screen Display sublabel still named the Separate mode, a setting the display rows removed; it now says "when it has a window of its own". The warnings for a screen window that can't be drawn, or whose surface is gone, said its screens go back to the main window, but a window that shows only the menu reaches them too. They now read like the one for a window that can't be made: what it shows stays in the main window. Comments that no longer matched the code are reworded. pointer_frame's model of the driver before the publication is named was_* rather than today_*, and a wrong commit hash and a plan tag go from the tests. video_driver_free_internal() uses VIDEO_ANDROID_RELEASE_KEPT_WINDOWS() rather than open-coding it, with no change in behaviour.