perf: cache Cocoa software MTLBuffer and optimize Wayland output/focus loops

3 files changed, 27 insertions, 16 deletions

diff --git a/src/cocoa/window.m b/src/cocoa/window.m
index 78ab0bd..8856fde 100644
--- a/src/cocoa/window.m
+++ b/src/cocoa/window.m
@@ -53,6 +53,8 @@ struct cocoa_surface {
     YaslView *view;
     CAMetalLayer *layer;
     id<MTLCommandQueue> software_queue;
+    id<MTLBuffer> software_buffer;
+    NSUInteger software_buffer_capacity;
     CADisplayLink *link;
 
     struct cocoa_surface *parent;
@@ -1031,15 +1033,20 @@ bool cocoa_surface_present(struct cocoa_surface *w, const struct yasl_framebuffe
             return false;
         NSUInteger row_bytes = ((NSUInteger)w->size.x * 4 + 255) & ~(NSUInteger)255;
         NSUInteger length = row_bytes * (NSUInteger)w->size.y;
-        id<MTLBuffer> buffer = [w->layer.device newBufferWithLength:length
-                                                options:MTLResourceStorageModeShared];
+        if (!w->software_buffer || w->software_buffer_capacity < length) {
+            [w->software_buffer release];
+            w->software_buffer = [w->layer.device newBufferWithLength:length
+                                                    options:MTLResourceStorageModeShared];
+            w->software_buffer_capacity = w->software_buffer ? length : 0;
+        }
+        id<MTLBuffer> buffer = w->software_buffer;
         if (!buffer) return false;
         for (int y = 0; y < w->size.y; y++)
             memcpy((char *)buffer.contents + (NSUInteger)y * row_bytes,
                    w->pixels + (size_t)y * w->size.x, (size_t)w->size.x * 4);
         id<MTLCommandBuffer> command = [w->software_queue commandBuffer];
         id<MTLBlitCommandEncoder> blit = [command blitCommandEncoder];
-        if (!blit) { [buffer release]; return false; }
+        if (!blit) return false;
         id<CAMetalDrawable> drawable = w->drawable;
         [blit copyFromBuffer:buffer sourceOffset:0 sourceBytesPerRow:row_bytes
            sourceBytesPerImage:length sourceSize:MTLSizeMake(w->size.x, w->size.y, 1)
@@ -1048,7 +1055,6 @@ bool cocoa_surface_present(struct cocoa_surface *w, const struct yasl_framebuffe
         [blit endEncoding];
         [command presentDrawable:drawable];
         [command commit];
-        [buffer release];
     }
     yasl_clock_presented(&w->clock, yasl_loop_now_ns(), false);
     w->did_present = true;
@@ -1394,6 +1400,7 @@ struct cocoa_surface *cocoa_surface_create_toplevel(struct cocoa_display *d,
 static void surface_free(struct cocoa_surface *w)
 {
     [w->link release];
+    [w->software_buffer release];
     [w->software_queue release];
     [w->view release];
     [w->controller release];
diff --git a/src/wayland/buffer.c b/src/wayland/buffer.c
index 0561da3..4cddfa5 100644
--- a/src/wayland/buffer.c
+++ b/src/wayland/buffer.c
@@ -200,24 +200,27 @@ static const struct wp_fractional_scale_v1_listener frac_scale_listener = {
     .preferred_scale = fractional_preferred_scale,
 };
 
-static void surface_follow_refresh(struct wayland_surface *w)
+void surface_follow_outputs(struct wayland_surface *w)
 {
     uint32_t refresh = 0;
-    for (const struct wl_conn_output *o = w->conn->outputs; o; o = o->next)
-        if ((o->mask & w->outputs) && o->refresh_ns &&
-            (!refresh || o->refresh_ns < refresh))
-            refresh = o->refresh_ns;
+    int32_t max_scale = 0;
+    bool need_scale = !w->frac_scale && !surface_gets_preferred_scale(w->surface);
+
+    for (const struct wl_conn_output *o = w->conn->outputs; o; o = o->next) {
+        if (o->mask & w->outputs) {
+            if (o->refresh_ns && (!refresh || o->refresh_ns < refresh))
+                refresh = o->refresh_ns;
+            if (need_scale && o->scale > max_scale)
+                max_scale = o->scale;
+        }
+    }
     if (!refresh && w->conn->outputs)
         refresh = w->conn->outputs->refresh_ns;
     if (refresh)
         w->clock.refresh_ns = refresh;
-}
 
-void surface_follow_outputs(struct wayland_surface *w)
-{
-    surface_follow_refresh(w);
-    if (!w->frac_scale && !surface_gets_preferred_scale(w->surface))
-        surface_request_scale(w, conn_output_scale_120(w->conn, w->outputs));
+    if (need_scale)
+        surface_request_scale(w, (uint32_t)max_scale * YASL_SCALE_UNIT);
 }
 
 static void surface_enter(void *data, struct wl_surface *surface, struct wl_output *output)
diff --git a/src/wayland/keyboard.c b/src/wayland/keyboard.c
index f7e2691..11c7f3f 100644
--- a/src/wayland/keyboard.c
+++ b/src/wayland/keyboard.c
@@ -170,7 +170,8 @@ static void seat_settle_focus(struct seat *s)
 void keyboard_settle_focus(struct wayland_display *c)
 {
     for (struct seat *s = c->seats; s; s = s->next)
-        seat_settle_focus(s);
+        if (s->focus_leaving)
+            seat_settle_focus(s);
 }
 
 static void keyboard_leave(void *data, struct wl_keyboard *keyboard,