GPU/HW: Avoid extra texture copy in Vulkan/GL

src/core/gpu_hw.cpp

@@ -2474,9 +2474,7 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
   const bool use_shader =
     (m_GPUSTAT.IsMaskingEnabled() || ((src_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
      ((src_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT || ((dst_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
-     ((dst_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT ||
-     Common::Rectangle<u32>::FromExtents(src_x, src_y, width, height)
-       .Intersects(Common::Rectangle<u32>::FromExtents(dst_x, dst_y, width, height)));
+     ((dst_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT);
 
   if (use_shader || IsUsingMultisampling())
   {
@@ -2526,14 +2524,15 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
     return;
   }
 
-  // We can't CopySubresourceRegion to the same resource. So use the shadow texture if we can, but that may need to be
-  // updated first. Copying to the same resource seemed to work on Windows 10, but breaks on Windows 7. But, it's
-  // against the API spec, so better to be safe than sorry.
-
-  // TODO: make this an optional feature, DX12 can do it
-
-  if (m_vram_dirty_rect.Intersects(Common::Rectangle<u32>::FromExtents(src_x, src_y, width, height)))
-    UpdateVRAMReadTexture();
+  GPUTexture* src_tex = m_vram_texture.get();
+  const bool overlaps_with_self = Common::Rectangle<u32>::FromExtents(src_x, src_y, width, height)
+                                    .Intersects(Common::Rectangle<u32>::FromExtents(dst_x, dst_y, width, height));
+  if (!g_gpu_device->GetFeatures().texture_copy_to_self || overlaps_with_self)
+  {
+    src_tex = m_vram_read_texture.get();
+    if (m_vram_dirty_rect.Intersects(Common::Rectangle<u32>::FromExtents(src_x, src_y, width, height)))
+      UpdateVRAMReadTexture();
+  }
 
   IncludeVRAMDirtyRectangle(
     Common::Rectangle<u32>::FromExtents(dst_x, dst_y, width, height).Clamped(0, 0, VRAM_WIDTH, VRAM_HEIGHT));
@@ -2545,9 +2544,10 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
   }
 
   g_gpu_device->CopyTextureRegion(m_vram_texture.get(), dst_x * m_resolution_scale, dst_y * m_resolution_scale, 0, 0,
-                                  m_vram_read_texture.get(), src_x * m_resolution_scale, src_y * m_resolution_scale, 0,
+                                  src_tex, src_x * m_resolution_scale, src_y * m_resolution_scale, 0,
                                   0, width * m_resolution_scale, height * m_resolution_scale);
-  m_vram_read_texture->MakeReadyForSampling();
+  if (src_tex != m_vram_texture.get())
+    m_vram_read_texture->MakeReadyForSampling();
 }
 
 void GPU_HW::DispatchRenderCommand()

src/util/d3d11_device.cpp

@@ -174,6 +174,7 @@ void D3D11Device::SetFeatures(FeatureMask disabled_features)
   m_features.framebuffer_fetch = false;
   m_features.per_sample_shading = (feature_level >= D3D_FEATURE_LEVEL_10_1);
   m_features.noperspective_interpolation = true;
+  m_features.texture_copy_to_self = false;
   m_features.supports_texture_buffers = !(disabled_features & FEATURE_MASK_TEXTURE_BUFFERS);
   m_features.texture_buffers_emulated_with_ssbo = false;
   m_features.geometry_shaders = !(disabled_features & FEATURE_MASK_GEOMETRY_SHADERS);

src/util/d3d12_device.cpp

@@ -116,7 +116,8 @@ D3D12Device::ComPtr<ID3D12RootSignature> D3D12Device::CreateRootSignature(const
   return rs;
 }
 
-bool D3D12Device::CreateDevice(const std::string_view& adapter, bool threaded_presentation, FeatureMask disabled_features)
+bool D3D12Device::CreateDevice(const std::string_view& adapter, bool threaded_presentation,
+                               FeatureMask disabled_features)
 {
   std::unique_lock lock(s_instance_mutex);
 
@@ -1176,6 +1177,8 @@ void D3D12Device::SetFeatures(FeatureMask disabled_features)
   m_features.framebuffer_fetch = false;
   m_features.per_sample_shading = true;
   m_features.noperspective_interpolation = true;
+  m_features.texture_copy_to_self =
+    /*!(disabled_features & FEATURE_MASK_TEXTURE_COPY_TO_SELF)*/ false; // TODO: Support with Enhanced Barriers
   m_features.supports_texture_buffers = !(disabled_features & FEATURE_MASK_TEXTURE_BUFFERS);
   m_features.texture_buffers_emulated_with_ssbo = false;
   m_features.geometry_shaders = !(disabled_features & FEATURE_MASK_GEOMETRY_SHADERS);

src/util/gpu_device.h

@@ -449,6 +449,7 @@ public:
     FEATURE_MASK_FRAMEBUFFER_FETCH = (1 << 1),
     FEATURE_MASK_TEXTURE_BUFFERS = (1 << 2),
     FEATURE_MASK_GEOMETRY_SHADERS = (1 << 3),
+    FEATURE_MASK_TEXTURE_COPY_TO_SELF = (1 << 4),
   };
 
   struct Features
@@ -457,6 +458,7 @@ public:
     bool framebuffer_fetch : 1;
     bool per_sample_shading : 1;
     bool noperspective_interpolation : 1;
+    bool texture_copy_to_self : 1;
     bool supports_texture_buffers : 1;
     bool texture_buffers_emulated_with_ssbo : 1;
     bool geometry_shaders : 1;

src/util/metal_device.mm

@@ -215,6 +215,7 @@ void MetalDevice::SetFeatures(FeatureMask disabled_features)
   m_features.framebuffer_fetch = !(disabled_features & FEATURE_MASK_FRAMEBUFFER_FETCH) && false; // TODO
   m_features.per_sample_shading = true;
   m_features.noperspective_interpolation = true;
+  m_features.texture_copy_to_self = !(disabled_features & FEATURE_MASK_TEXTURE_COPY_TO_SELF);
   m_features.supports_texture_buffers = !(disabled_features & FEATURE_MASK_TEXTURE_BUFFERS);
   m_features.texture_buffers_emulated_with_ssbo = true;
   m_features.geometry_shaders = false;

src/util/opengl_device.cpp

@@ -496,6 +496,8 @@ bool OpenGLDevice::CheckFeatures(bool* buggy_pbo, FeatureMask disabled_features)
   // noperspective is not supported in GLSL ES.
   m_features.noperspective_interpolation = !is_gles;
 
+  m_features.texture_copy_to_self = !(disabled_features & FEATURE_MASK_TEXTURE_COPY_TO_SELF);
+
   m_features.geometry_shaders =
     !(disabled_features & FEATURE_MASK_GEOMETRY_SHADERS) && (GLAD_GL_VERSION_3_2 || GLAD_GL_ES_VERSION_3_2);
 

src/util/vulkan_device.cpp

@@ -2202,6 +2202,7 @@ bool VulkanDevice::CheckFeatures(FeatureMask disabled_features)
     Log_WarningPrintf("Vulkan driver is missing dual-source blending. This will have an impact on performance.");
 
   m_features.noperspective_interpolation = true;
+  m_features.texture_copy_to_self = !(disabled_features & FEATURE_MASK_TEXTURE_COPY_TO_SELF);
   m_features.per_sample_shading = m_device_features.sampleRateShading;
   m_features.supports_texture_buffers = !(disabled_features & FEATURE_MASK_TEXTURE_BUFFERS);