GPU/HW: Elide buffer copy in scanout and fix flipped display

src/duckstation/d3d11_host_display.cpp

@@ -130,7 +130,7 @@ void D3D11HostDisplay::UpdateTexture(HostDisplayTexture* texture, u32 x, u32 y,
   }
 }
 
-void D3D11HostDisplay::SetDisplayTexture(void* texture, u32 offset_x, u32 offset_y, u32 width, u32 height,
+void D3D11HostDisplay::SetDisplayTexture(void* texture, s32 offset_x, s32 offset_y, s32 width, s32 height,
                                          u32 texture_width, u32 texture_height, float aspect_ratio)
 {
   m_display_srv = static_cast<ID3D11ShaderResourceView*>(texture);
@@ -249,29 +249,29 @@ bool D3D11HostDisplay::CreateSwapChainRTV()
 bool D3D11HostDisplay::CreateD3DResources()
 {
   static constexpr char fullscreen_quad_vertex_shader[] = R"(
-void main(in uint vertex_id : SV_VertexID,
-          out float2 v_tex0 : TEXCOORD0,
-          out float4 o_pos : SV_Position)
-{
-  v_tex0 = float2(float((vertex_id << 1) & 2u), float(vertex_id & 2u));
-  o_pos = float4(v_tex0 * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);
-}
-)";
-
-  static constexpr char display_pixel_shader[] = R"(
 cbuffer UBOBlock : register(b0)
 {
   float4 u_src_rect;
 };
 
+void main(in uint vertex_id : SV_VertexID,
+          out float2 v_tex0 : TEXCOORD0,
+          out float4 o_pos : SV_Position)
+{
+  float2 pos = float2(float((vertex_id << 1) & 2u), float(vertex_id & 2u));
+  v_tex0 = u_src_rect.xy + pos * u_src_rect.zw;
+  o_pos = float4(pos * float2(2.0f, -2.0f) + float2(-1.0f, 1.0f), 0.0f, 1.0f);
+}
+)";
+
+  static constexpr char display_pixel_shader[] = R"(
 Texture2D samp0 : register(t0);
 SamplerState samp0_ss : register(s0);
 
 void main(in float2 v_tex0 : TEXCOORD0,
           out float4 o_col0 : SV_Target)
 {
-  float2 coords = u_src_rect.xy + v_tex0 * u_src_rect.zw;
-  o_col0 = samp0.Sample(samp0_ss, coords);
+  o_col0 = samp0.Sample(samp0_ss, v_tex0);
 }
 )";
 
@@ -381,7 +381,7 @@ void D3D11HostDisplay::RenderDisplay()
   const auto map = m_display_uniform_buffer.Map(m_context.Get(), sizeof(uniforms), sizeof(uniforms));
   std::memcpy(map.pointer, uniforms, sizeof(uniforms));
   m_display_uniform_buffer.Unmap(m_context.Get(), sizeof(uniforms));
-  m_context->PSSetConstantBuffers(0, 1, m_display_uniform_buffer.GetD3DBufferArray());
+  m_context->VSSetConstantBuffers(0, 1, m_display_uniform_buffer.GetD3DBufferArray());
 
   const CD3D11_VIEWPORT vp(static_cast<float>(vp_left), static_cast<float>(vp_top), static_cast<float>(vp_width),
                            static_cast<float>(vp_height));

src/duckstation/d3d11_host_display.h

@@ -28,7 +28,7 @@ public:
   void UpdateTexture(HostDisplayTexture* texture, u32 x, u32 y, u32 width, u32 height, const void* data,
                      u32 data_stride) override;
 
-  void SetDisplayTexture(void* texture, u32 offset_x, u32 offset_y, u32 width, u32 height, u32 texture_width,
+  void SetDisplayTexture(void* texture, s32 offset_x, s32 offset_y, s32 width, s32 height, u32 texture_width,
                          u32 texture_height, float aspect_ratio) override;
   void SetDisplayLinearFiltering(bool enabled) override;
 
@@ -70,10 +70,10 @@ private:
   D3D11::StreamBuffer m_display_uniform_buffer;
 
   ID3D11ShaderResourceView* m_display_srv = nullptr;
-  u32 m_display_offset_x = 0;
-  u32 m_display_offset_y = 0;
-  u32 m_display_width = 0;
-  u32 m_display_height = 0;
+  s32 m_display_offset_x = 0;
+  s32 m_display_offset_y = 0;
+  s32 m_display_width = 0;
+  s32 m_display_height = 0;
   u32 m_display_texture_width = 0;
   u32 m_display_texture_height = 0;
   float m_display_aspect_ratio = 1.0f;

src/duckstation/opengl_host_display.cpp

@@ -28,7 +28,7 @@ public:
     glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_texture_binding);
 
     // TODO: Set pack width
-    Assert(initial_data_stride == (width * sizeof(u32)));
+    Assert(!initial_data || initial_data_stride == (width * sizeof(u32)));
 
     glBindTexture(GL_TEXTURE_2D, id);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, initial_data);
@@ -109,7 +109,7 @@ void OpenGLHostDisplay::UpdateTexture(HostDisplayTexture* texture, u32 x, u32 y,
   glBindTexture(GL_TEXTURE_2D, old_texture_binding);
 }
 
-void OpenGLHostDisplay::SetDisplayTexture(void* texture, u32 offset_x, u32 offset_y, u32 width, u32 height,
+void OpenGLHostDisplay::SetDisplayTexture(void* texture, s32 offset_x, s32 offset_y, s32 width, s32 height,
                                           u32 texture_width, u32 texture_height, float aspect_ratio)
 {
   m_display_texture_id = static_cast<GLuint>(reinterpret_cast<uintptr_t>(texture));
@@ -210,13 +210,14 @@ bool OpenGLHostDisplay::CreateGLResources()
   static constexpr char fullscreen_quad_vertex_shader[] = R"(
 #version 330 core
 
+uniform vec4 u_src_rect;
 out vec2 v_tex0;
 
 void main()
 {
-  v_tex0 = vec2(float((gl_VertexID << 1) & 2), float(gl_VertexID & 2));
-  gl_Position = vec4(v_tex0 * vec2(2.0f, -2.0f) + vec2(-1.0f, 1.0f), 0.0f, 1.0f);
-  gl_Position.y = -gl_Position.y;
+  vec2 pos = vec2(float((gl_VertexID << 1) & 2), float(gl_VertexID & 2));
+  v_tex0 = u_src_rect.xy + pos * u_src_rect.zw;
+  gl_Position = vec4(pos * vec2(2.0f, -2.0f) + vec2(-1.0f, 1.0f), 0.0f, 1.0f);
 }
 )";
 
@@ -224,15 +225,13 @@ void main()
 #version 330 core
 
 uniform sampler2D samp0;
-uniform vec4 u_src_rect;
 
 in vec2 v_tex0;
 out vec4 o_col0;
 
 void main()
 {
-  vec2 coords = u_src_rect.xy + v_tex0 * u_src_rect.zw;
-  o_col0 = texture(samp0, coords);
+  o_col0 = texture(samp0, v_tex0);
 }
 )";
 

src/duckstation/opengl_host_display.h

@@ -22,7 +22,7 @@ public:
   void UpdateTexture(HostDisplayTexture* texture, u32 x, u32 y, u32 width, u32 height, const void* data,
                      u32 data_stride) override;
 
-  void SetDisplayTexture(void* texture, u32 offset_x, u32 offset_y, u32 width, u32 height, u32 texture_width,
+  void SetDisplayTexture(void* texture, s32 offset_x, s32 offset_y, s32 width, s32 height, u32 texture_width,
                          u32 texture_height, float aspect_ratio) override;
   void SetDisplayLinearFiltering(bool enabled) override;
 
@@ -49,10 +49,10 @@ private:
   GL::Program m_display_program;
   GLuint m_display_vao = 0;
   GLuint m_display_texture_id = 0;
-  u32 m_display_offset_x = 0;
-  u32 m_display_offset_y = 0;
-  u32 m_display_width = 0;
-  u32 m_display_height = 0;
+  s32 m_display_offset_x = 0;
+  s32 m_display_offset_y = 0;
+  s32 m_display_width = 0;
+  s32 m_display_height = 0;
   u32 m_display_texture_width = 0;
   u32 m_display_texture_height = 0;
   float m_display_aspect_ratio = 1.0f;