Synchronize depth24 clears before immediate CPU reads

docs/GPU_GL_READBACK_REGIONS.md

@@ -23,7 +23,10 @@ CPU-authority mode retains its existing no-readback behavior.
 context. It checks the complete native CPU image against full hardware readback
 after ordered synthetic workloads. It covers single pixels, odd-width row
 stride, disjoint uploads, texture dependencies, wrapping fills, overlapping
-copies, masks, blending, clipping, precision margins and state restore.
+copies, masks, blending, clipping, precision margins and state restore. A depth24 exit followed by an immediate
+CPU read verifies that cleared movie pixels are visible and newer overlapping
+texture uploads survive. It tests this existing clear policy, not full movie
+playback or depth24 raster accuracy.
 The hardware rasterizer is shared by the comparison; this proves coherence,
 not independent raster accuracy. No retail assets are required.
 
@@ -32,8 +35,9 @@ Run `python runtime/tests/run_gl_readback_region.py --compiler-bin <mingw-bin>
 `sdl3-src/include` and `sdl3-build/libSDL3.a`. The Windows runner uses hidden
 windows. It tests
 native and4x scales. `--gl-source <old-gpu_gl_renderer.c> --expect-unbounded`
-checks the previous implementation: pixel comparisons pass, and only the small
-transfer assertion fails. Each command/result is saved in `receipt.json` under a fresh run directory inside the output root. Configure `runtime/` with `BUILD_TESTING=ON` and `PSX_GL_READBACK_SDL_ROOT=<deps>` on MinGW to register `gl_readback_region_test`; run it with `ctest --test-dir <build> -R gl_readback_region_test --output-on-failure`. Missing dependencies print a configure-time message and do not count as a hardware pass.
+accepts an unbounded implementation only when pixel comparisons pass and
+exactly the small-transfer assertion fails. An older implementation with a
+coherence defect (including the depth24 clear) is correctly rejected too. Each command/result is saved in `receipt.json` under a fresh run directory inside the output root. Configure `runtime/` with `BUILD_TESTING=ON` and `PSX_GL_READBACK_SDL_ROOT=<deps>` on MinGW to register `gl_readback_region_test`; run it with `ctest --test-dir <build> -R gl_readback_region_test --output-on-failure`. Missing dependencies print a configure-time message and do not count as a hardware pass.
 
 ## Prior art and scope
 

runtime/src/gpu_gl_renderer.c

@@ -2544,7 +2544,10 @@ static void depth24_clear_skipped_fb(void) {
     glDisable(GL_SCISSOR_TEST);
     p_glBindFramebuffer(PSXGL_FRAMEBUFFER, 0);
     rect_add(&s_pack_dirty, x0, y0, x1, y1);
-    if (!s_cpu_auth_dual) rect_add(&s_cpu_dirty, x0, y0, x1, y1);
+    if (!s_cpu_auth_dual) {
+        rect_add(&s_cpu_dirty, x0, y0, x1, y1);
+        s_gpu_dirty = 1; /* The clear must reach an immediate CPU read too. */
+    }
     present_dirty_rect(x0, y0, x1, y1, 1);
     rect_clear(&s_d24_skip_fb);
 }

runtime/tests/test_gl_readback_region.c

@@ -7,8 +7,9 @@ void gpu_vram_dirty_mark_row_impl(uint32_t y){}
 void gpu_vram_dirty_mark_rect(int x,int y,int w,int h){}
 void gpu_vram_dirty_mark_all(void){}
 int psx_netplay_active(void){return 0;}
-int gpu_display_is_depth24(void){return 0;}
-void gpu_get_display_info(GpuDisplayInfo *out){memset(out,0,sizeof(*out));}
+static int test_depth24;
+int gpu_display_is_depth24(void){return test_depth24;}
+void gpu_get_display_info(GpuDisplayInfo *out){memset(out,0,sizeof(*out));out->display_x=32;out->display_y=32;out->width=320;out->height=16;}
 int psx_ws_prim_in_backdrop(void){return 0;}
 int g_ws_tex_edge_pct=0;
 int psx_ws_prim_is_tagged(void){return 0;}
@@ -60,7 +61,17 @@ int main(int argc,char **argv){
  glb_set_draw_area(0,0,1023,511);glb_draw_flat_rect(320,320,8,8,0x4444);
  for(int i=0;i<1024*512;i++)image[i]=(uint16_t)((i*23)&0x7fff);
  gl_renderer_restage_vram_after_savestate();verify("state restage with pending draw");
- check(image[511*1024+1020]==oracle[511*1024+1020],"edge preserved");
+ /* Existing depth24 policy clears the skipped movie band on return to15-bit.
+  * That GPU write must become visible without waiting for another primitive. */
+ static uint16_t movie[480*16], texture[4]={0x3210,0x3210,0x3210,0x3210};
+ for(int i=0;i<480*16;i++)movie[i]=0x1234;
+ test_depth24=1;depth24_upload_policy();
+ glb_vram_transfer_in(32,32,480,16,movie);
+ glb_vram_transfer_in(33,33,2,2,texture);
+ test_depth24=0;depth24_upload_policy();
+ check(glb_vram_read(40,40)==0,"depth24 cleared band immediate CPU read");
+ check(glb_vram_read(33,33)==0x3210,"newer overlapping texture survives clear");
+ verify("depth24 leave coherence without subsequent primitive");
  printf("checks=%d failures=%d\n",checks,failures);
  gl_renderer_shutdown();SDL_DestroyWindow(win);SDL_Quit();return failures?1:0;
 }