perf(gui): Present only changed pixels and cache the wallpaper

The compositor recomposited and re-uploaded the entire screen on every pass of
its loop. At the 1280x720 the desktop actually boots into, gui_gfx_present()
memcpy'd 3.6 MB into the framebuffer per frame, and the background underneath
it was recomputed pixel by pixel every time.

Present only the damage
- Keep a RAM shadow of what the framebuffer currently holds and diff the back
  buffer against it, pushing just the changed span of each row. Reading the
  shadow costs a RAM access; the framebuffer write it avoids costs far more,
  and the framebuffer is never read back.
- Each row is narrowed to its first and last differing pixel rather than
  tracked as a set of runs: one memcpy per row keeps the loop tight, and a row
  that changed in two distant places is rare enough that bridging the gap is
  cheaper than the bookkeeping.
- gui_gfx_invalidate() forces a full push for when something outside the
  compositor owns the screen, which the desktop calls on entry and on exit
  around the framebuffer console.

Cache the wallpaper
- The vertical gradient, the blended grid and the branding text never change
  between frames, so they are painted once into a cached buffer and blitted
  back per frame. The cache is keyed on resolution and the grid setting, so
  toggling the grid from the context menu still repaints it.

No drawing code changed: the damage is derived from the frame itself, so every
widget, window and app keeps rendering exactly as before.

Measured in QEMU at 1280x720: an idle desktop with two windows open went from
921600 pixels per frame to 3609, about 0.4% of a full blit. A screendump of
the running desktop is unchanged -- wallpaper, grid, icons, both windows,
taskbar and cursor all render identically.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0161Jh7uCEMLYEMNPsjC1LKf

gui/gui_gfx.c

@@ -6,18 +6,56 @@
 
 static uint32_t* g_frontbuffer = NULL;
 static uint32_t* g_backbuffer = NULL;
+/*
+ * Mirror of what the framebuffer currently holds.
+ *
+ * The front buffer lives in device memory, where a write costs far more than
+ * the same write to RAM and a read costs more still. Keeping a RAM copy of the
+ * last presented frame lets gui_gfx_present() work out which pixels actually
+ * changed without ever reading the framebuffer back, and push only those.
+ */
+static uint32_t* g_shadow = NULL;
+static bool g_shadow_valid = false;
 static int g_width = 800;
 static int g_height = 600;
+static uint64_t g_present_pixels = 0;
 
 
 void gui_gfx_init(uint32_t* fb, int width, int height) {
-    g_frontbuffer = fb;
-    g_width = (width > 0) ? width : 800;
-    g_height = (height > 0) ? height : 600;
+    int new_w = (width > 0) ? width : 800;
+    int new_h = (height > 0) ? height : 600;
 
-    if (!g_backbuffer) {
-        g_backbuffer = (uint32_t*)kzalloc(g_width * g_height * sizeof(uint32_t));
+    /* A resolution change invalidates both RAM buffers. */
+    if (g_backbuffer && (new_w != g_width || new_h != g_height)) {
+        kfree(g_backbuffer);
+        g_backbuffer = NULL;
+        if (g_shadow) {
+            kfree(g_shadow);
+            g_shadow = NULL;
+        }
     }
+
+    g_frontbuffer = fb;
+    g_width = new_w;
+    g_height = new_h;
+
+    size_t pixels = (size_t)g_width * (size_t)g_height;
+    if (!g_backbuffer) {
+        g_backbuffer = (uint32_t*)kzalloc(pixels * sizeof(uint32_t));
+    }
+    if (!g_shadow) {
+        /* Optional: without it present() just falls back to a full copy. */
+        g_shadow = (uint32_t*)kzalloc(pixels * sizeof(uint32_t));
+    }
+    g_shadow_valid = false;
+}
+
+void gui_gfx_invalidate(void) {
+    g_shadow_valid = false;
+}
+
+uint64_t gui_gfx_last_present_pixels(void) {
+    return g_present_pixels;
 }
 
 int gui_gfx_get_width(void) { return g_width; }
@@ -234,8 +272,55 @@ void gui_gfx_blit(int dx, int dy, int w, int h, const uint32_t* src_buf) {
     }
 }
 
+/*
+ * Push the back buffer to the screen, writing only what changed.
+ *
+ * Every frame is still composited in full, so the cheapest reliable way to
+ * find the damage is to diff against the shadow copy: a RAM read is orders of
+ * magnitude cheaper than the framebuffer write it avoids. A typical idle
+ * desktop frame only moves the cursor and the clock, which turns a 3.6 MB
+ * blit at 1280x720 into a few kilobytes.
+ *
+ * Each row is narrowed to the span between its first and last differing pixel
+ * rather than tracked as a set of runs: one memcpy per row keeps the inner
+ * loop tight, and a row that changed in two places is rare enough that
+ * bridging the gap costs less than the extra bookkeeping.
+ */
 void gui_gfx_present(void) {
-    if (g_frontbuffer && g_backbuffer) {
-        memcpy(g_frontbuffer, g_backbuffer, g_width * g_height * sizeof(uint32_t));
+    if (!g_frontbuffer || !g_backbuffer) return;
+
+    size_t pixels = (size_t)g_width * (size_t)g_height;
+
+    /* No shadow (allocation failed) or it does not describe the screen yet:
+     * push the whole frame and start tracking from there. */
+    if (!g_shadow || !g_shadow_valid) {
+        memcpy(g_frontbuffer, g_backbuffer, pixels * sizeof(uint32_t));
+        if (g_shadow) {
+            memcpy(g_shadow, g_backbuffer, pixels * sizeof(uint32_t));
+            g_shadow_valid = true;
+        }
+        g_present_pixels = pixels;
+        return;
     }
+
+    size_t pushed = 0;
+    for (int y = 0; y < g_height; y++) {
+        size_t row = (size_t)y * (size_t)g_width;
+        const uint32_t* back = g_backbuffer + row;
+        uint32_t* shadow = g_shadow + row;
+
+        int x0 = 0;
+        while (x0 < g_width && back[x0] == shadow[x0]) x0++;
+        if (x0 == g_width) continue;            /* row is untouched */
+
+        int x1 = g_width - 1;
+        while (x1 > x0 && back[x1] == shadow[x1]) x1--;
+
+        size_t span = (size_t)(x1 - x0 + 1);
+        memcpy(g_frontbuffer + row + x0, back + x0, span * sizeof(uint32_t));
+        memcpy(shadow + x0, back + x0, span * sizeof(uint32_t));
+        pushed += span;
+    }
+
+    g_present_pixels = pushed;
 }