Fix MPH HD and compute presentation regressions

THIRD_PARTY_ATTRIBUTION.md

@@ -79,7 +79,8 @@ remains a device model, not HLE).
   `GPU3D_Compute.cpp`, `GPU3D_Compute.h`, `GPU3D_Compute_shaders.h`
   (patch `0010`, adaptive width, polygon-ID attributes in the low
   resolution surface, and the internal-resolution accessors; patch `0014`,
-  correcting the compute memory-barrier bitfields), and
+  correcting the compute memory-barrier bitfields; patch `0015`, guarding
+  the texture-size uniform update to textured raster variants), and
   `GPU3D_Texcache.h`, `GPU3D_TexcacheOpenGL.cpp` (patch `0011`, routing
   decoded textures through the optional upscaler and sizing array storage
   and the per-array layer budget by its factor), and `OpenGLSupport.cpp`

runner/src/frontend.cpp

@@ -745,9 +745,12 @@ bool create_presentation(const NdsFrontendOptions& options,
     const bool direct_enabled =
         !direct_selection || !*direct_selection ||
         std::strcmp(direct_selection, "1") == 0;
+    const bool direct_top_requested =
+        (options.adaptive_screens & NDS_ADAPTIVE_TOP) != 0u ||
+        options.internal_resolution > 1u;
     presentation.gl_top = presentation.separate &&
         allow_gl_top &&
-        (options.adaptive_screens & NDS_ADAPTIVE_TOP) != 0u &&
+        direct_top_requested &&
         nds_gpu3d_renderer_prefers_compute() &&
         direct_enabled;
 #endif

runner/src/gpu3d.cpp

@@ -492,9 +492,11 @@ bool nds_gpu3d_use_compute_renderer() {
     // readback surface this bridge maps every frame is still RenderWidth x
     // 192 (the final pass point-samples it out of the scaled raster), so the
     // faithful 2D compositor and display capture see byte-identical input at
-    // every scale. Hi-res coordinates stay off: they change which samples the
-    // rasterizer produces, which the native readback would then observe.
-    renderer->SetRenderSettings(static_cast<int>(g_internal_scale), false);
+    // scale 1. Higher internal-resolution modes use melonDS's sub-native
+    // vertex coordinates so the extra samples reduce polygon/texture wobble
+    // rather than merely magnifying the native integer grid.
+    renderer->SetRenderSettings(
+        static_cast<int>(g_internal_scale), g_internal_scale > 1u);
     if (compute_gl_stage_failed("render settings")) return false;
     if (g_internal_scale > 1u)
         std::fprintf(stderr,

runner/src/melonds_compute/ComputeHost.cpp

@@ -576,7 +576,23 @@ bool nds_compute_host_present_top(const unsigned int* fallback_pixels,
     int drawable_width = 0;
     int drawable_height = 0;
     SDL_GL_GetDrawableSize(g_window, &drawable_width, &drawable_height);
-    glViewport(0, 0, drawable_width, drawable_height);
+    int viewport_x = 0;
+    int viewport_y = 0;
+    int viewport_width = drawable_width;
+    int viewport_height = drawable_height;
+    if (width == 256 && drawable_width > 0 && drawable_height > 0) {
+        const int target_width = drawable_height * 4 / 3;
+        if (target_width < drawable_width) {
+            viewport_width = target_width;
+            viewport_x = (drawable_width - viewport_width) / 2;
+        } else {
+            viewport_height = drawable_width * 3 / 4;
+            viewport_y = (drawable_height - viewport_height) / 2;
+        }
+    }
+    glClearColor(0.f, 0.f, 0.f, 1.f);
+    glClear(GL_COLOR_BUFFER_BIT);
+    glViewport(viewport_x, viewport_y, viewport_width, viewport_height);
     glUseProgram(g_present_program);
     glActiveTexture(GL_TEXTURE1);
     glBindTexture(GL_TEXTURE_2D, g_object_texture[buffer]);

runner/vendor/melonds/GPU3D_Compute.cpp

@@ -1033,7 +1033,10 @@ void ComputeRenderer::RenderFrame(GPU& gpu)
                 }
 
                 glUniform1ui(UniformIdxCurVariant, i);
-                glUniform2f(UniformIdxTextureSize, 1.f / variants[i].Width, 1.f / variants[i].Height);
+                if (variants[i].Texture != 0)
+                    glUniform2f(UniformIdxTextureSize,
+                        1.f / variants[i].Width,
+                        1.f / variants[i].Height);
                 glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, BinResultMemory);
                 glDispatchComputeIndirect(offsetof(BinResultHeader, VariantWorkCount) + i*4*4);
             }

runner/vendor/melonds/patches/0015-gpu3d-compute-texture-uniform-guard.patch

@@ -0,0 +1,16 @@
+diff --git a/src/GPU3D_Compute.cpp b/src/GPU3D_Compute.cpp
+index 0000000..0000000 100644
+--- a/src/GPU3D_Compute.cpp
++++ b/src/GPU3D_Compute.cpp
+@@ -1033,7 +1033,10 @@ void ComputeRenderer::RenderFrame(GPU& gpu)
+                 }
+
+                 glUniform1ui(UniformIdxCurVariant, i);
+-                glUniform2f(UniformIdxTextureSize, 1.f / variants[i].Width, 1.f / variants[i].Height);
++                if (variants[i].Texture != 0)
++                    glUniform2f(UniformIdxTextureSize,
++                        1.f / variants[i].Width,
++                        1.f / variants[i].Height);
+                 glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, BinResultMemory);
+                 glDispatchComputeIndirect(offsetof(BinResultHeader, VariantWorkCount) + i*4*4);
+             }

runner/vendor/melonds/patches/README.md

@@ -394,3 +394,12 @@ when the knob is unset.
 memory-barrier bitfield. The command barrier makes shader-written indirect
 dispatch parameters visible to `glDispatchComputeIndirect`; the correction
 matches upstream melonDS commit `77774538e56b118dcb5d64f08d784542ba77c72b`.
+
+## 0015-gpu3d-compute-texture-uniform-guard.patch
+
+`src/GPU3D_Compute.cpp`: updates `InvTextureSize` only for raster variants
+that actually bind a texture. On affected Intel OpenGL drivers, updating this
+uniform while the no-texture raster program is current can produce
+`GL_INVALID_OPERATION`, closing the runner after the first rendered frames.
+The shader only reads `InvTextureSize` inside `#ifdef UseTexture`, so this
+does not change textured variant sampling or no-texture rendering semantics.