From e199d7a4214e8965a399640c29a6e6e9ff1fd5e7 Mon Sep 17 00:00:00 2001 From: The412Banner Date: Fri, 25 Sep 2026 07:39:16 -0400 Subject: [PATCH] HUD: follow the game window, not a program's cursor or menus A program's small surfaces (gamescope's 1x1 cursor, Qt's menus and tooltips) announced themselves after the game's window and took the HUD, which then read 0 fps while melonDS ran at 60. A new window now takes the HUD only when it is at least as big, and a window quiet for a second gives way to the next one that draws. --- app/src/main/cpp/waylandcomp/src/compositor.c | 34 +++++++++++++------ 1 file changed, 23 insertions(+), 11 deletions(-) diff --git a/app/src/main/cpp/waylandcomp/src/compositor.c b/app/src/main/cpp/waylandcomp/src/compositor.c index 5b7642bb7d..2199e275d2 100644 --- a/app/src/main/cpp/waylandcomp/src/compositor.c +++ b/app/src/main/cpp/waylandcomp/src/compositor.c @@ -686,9 +686,11 @@ static void take_shm(struct surface *s, struct wl_shm_buffer *shm, struct wl_res if (ci && ci->asked_feedback) ci->shm_frames++; /* since it asked for GPU buffers */ } -/* The window the app's performance HUD follows: the latest one to start presenting GPU frames - * (X11 binds the HUD to the _MESA_DRV window and counts X presents instead). JNI upcalls. */ +/* The window the app's performance HUD follows: the latest one at least as big as the last to + * start presenting GPU frames (take_dmabuf) (X11 binds the HUD to the _MESA_DRV window and counts X presents instead). JNI upcalls. */ static struct surface *g_hud_surface; +/* When the HUD's window last presented a frame (now_ns), to let a replacement take over. */ +static int64_t g_hud_last_ns; extern void banner_on_game_surface(const char *window, const char *gpu); /* window NULL = gone */ extern void banner_on_game_frame(void); /* The program behind that window: its Linux pid (the Wayland client's credentials) and executable name @@ -731,6 +733,7 @@ static void take_dmabuf(struct surface *s, struct dmabuf_buffer *b, struct wl_re s->buf_h = b->height; s->has_content = b->img != NULL; g_stat_dmabuf++; + const int announced_now = !s->announced_vulkan; if (!s->announced_vulkan) { char name[160]; s->announced_vulkan = 1; @@ -751,14 +754,23 @@ static void take_dmabuf(struct surface *s, struct dmabuf_buffer *b, struct wl_re else banner_log("error", "could not import GPU frames from %s (%dx%d, modifier %#llx)", name, b->width, b->height, (unsigned long long)b->modifier); - /* The HUD follows the window whether its frames are copied or go straight to the layer: - * a zero-copy frame the compositor never imported is still a presented game frame. */ - if (b->img || ahb_swapchain_has_ahb(b)) { - g_hud_surface = s; - banner_on_game_surface(name, vkp_gpu_name()); - struct client_info *ci = client_info_of(wl_resource_get_client(s->resource)); - if (ci) banner_on_game_program((int)ci->pid, strncmp(ci->name, "pid ", 4) ? ci->name : ""); - } + } + /* The HUD follows the window whether its frames are copied or go straight to the layer: a + * zero-copy frame the compositor never imported is still a presented game frame. A + * program's small surfaces - gamescope's 1x1 cursor, Qt's menus and tooltips (melonDS) - + * came after the game's and drew too rarely to count, which left the HUD on 0 fps while the + * game ran at 60. So a new window takes the HUD only when it is at least as big (a + * full-screen gamescope over the desktop still does), and a window gone quiet for a second + * gives way to whatever draws next. */ + if (s != g_hud_surface && (b->img || ahb_swapchain_has_ahb(b)) && + (!g_hud_surface || now_ns() - g_hud_last_ns > 1000000000LL || + (announced_now && (int64_t)b->width * b->height >= (int64_t)g_hud_surface->buf_w * g_hud_surface->buf_h))) { + char name[160]; + describe(s, name, sizeof(name)); + g_hud_surface = s; + banner_on_game_surface(name, vkp_gpu_name()); + struct client_info *ci = client_info_of(wl_resource_get_client(s->resource)); + if (ci) banner_on_game_program((int)ci->pid, strncmp(ci->name, "pid ", 4) ? ci->name : ""); } /* HDR session (gate open): a DXVK game switches to HDR by REBUILDING its swapchain on the same * surface, so the one-shot line above never sees the 10-bit buffers - name every format change. */ @@ -780,7 +792,7 @@ static void take_dmabuf(struct surface *s, struct dmabuf_buffer *b, struct wl_re * Asking whether the compositor could import it, as this used to, is a question about the * copy path and not about whether a frame happened: a game whose buffers go straight to * the display layer draws on screen while the counter sat at 0.0 fps and 1000.0 ms. */ - if (s == g_hud_surface) banner_on_game_frame(); + if (s == g_hud_surface) { g_hud_last_ns = now_ns(); banner_on_game_frame(); } } /* ---- hooks for ahb_swapchain.c (zero-copy layers) */