From 16d83989e52d18e6f9bf80501c56b36f542784ba Mon Sep 17 00:00:00 2001
From: Connor McLaughlin <mclaughc@outlook.com>
Date: Tue, 5 Nov 2019 00:27:22 +1000
Subject: [PATCH] GPU/HW: Support line drawing at >1x scale in D3D11 via GS

---
 src/core/gpu_hw_d3d11.cpp     | 17 +++++++++++
 src/core/gpu_hw_d3d11.h       |  1 +
 src/core/gpu_hw_shadergen.cpp | 54 +++++++++++++++++++++++++++++++++++
 src/core/gpu_hw_shadergen.h   |  1 +
 4 files changed, 73 insertions(+)

diff --git a/src/core/gpu_hw_d3d11.cpp b/src/core/gpu_hw_d3d11.cpp
index 403c6644e..b915a3de4 100644
--- a/src/core/gpu_hw_d3d11.cpp
+++ b/src/core/gpu_hw_d3d11.cpp
@@ -92,6 +92,8 @@ void GPU_HW_D3D11::Reset()
 void GPU_HW_D3D11::ResetGraphicsAPIState()
 {
   GPU_HW::ResetGraphicsAPIState();
+
+  m_context->GSSetShader(nullptr, nullptr, 0);
 }
 
 void GPU_HW_D3D11::RestoreGraphicsAPIState()
@@ -403,6 +405,15 @@ bool GPU_HW_D3D11::CompileShaders()
     }
   }
 
+  m_batch_line_expand_geometry_shader.Reset();
+  if (m_resolution_scale > 1)
+  {
+    m_batch_line_expand_geometry_shader = D3D11::ShaderCompiler::CompileAndCreateGeometryShader(
+      m_device.Get(), shadergen.GenerateBatchLineExpandGeometryShader(), debug);
+    if (!m_batch_line_expand_geometry_shader)
+      return false;
+  }
+
   m_copy_pixel_shader =
     D3D11::ShaderCompiler::CompileAndCreatePixelShader(m_device.Get(), shadergen.GenerateCopyFragmentShader(), debug);
   if (!m_copy_pixel_shader)
@@ -490,6 +501,7 @@ void GPU_HW_D3D11::DrawUtilityShader(ID3D11PixelShader* shader, const void* unif
 
   m_context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
   m_context->VSSetShader(m_screen_quad_vertex_shader.Get(), nullptr, 0);
+  m_context->GSSetShader(nullptr, nullptr, 0);
   m_context->PSSetShader(shader, nullptr, 0);
   m_context->OMSetBlendState(m_blend_disabled_state.Get(), nullptr, 0xFFFFFFFFu);
 
@@ -507,6 +519,11 @@ void GPU_HW_D3D11::SetDrawState(BatchRenderMode render_mode)
 
   m_context->VSSetShader(m_batch_vertex_shaders[BoolToUInt8(textured)].Get(), nullptr, 0);
 
+  m_context->GSSetShader((m_batch.primitive < GPU_HW::BatchPrimitive::Triangles && m_resolution_scale > 1) ?
+                           m_batch_line_expand_geometry_shader.Get() :
+                           nullptr,
+                         nullptr, 0);
+
   m_context->PSSetShader(m_batch_pixel_shaders[static_cast<u8>(render_mode)][static_cast<u8>(m_batch.texture_mode)]
                                               [BoolToUInt8(m_batch.dithering)]
                                                 .Get(),
diff --git a/src/core/gpu_hw_d3d11.h b/src/core/gpu_hw_d3d11.h
index a4d082b72..03a7bb4b0 100644
--- a/src/core/gpu_hw_d3d11.h
+++ b/src/core/gpu_hw_d3d11.h
@@ -106,6 +106,7 @@ private:
   std::array<ComPtr<ID3D11VertexShader>, 2> m_batch_vertex_shaders; // [textured]
   std::array<std::array<std::array<ComPtr<ID3D11PixelShader>, 2>, 9>, 4>
     m_batch_pixel_shaders; // [render_mode][texture_mode][dithering]
+  ComPtr<ID3D11GeometryShader> m_batch_line_expand_geometry_shader;
 
   ComPtr<ID3D11VertexShader> m_screen_quad_vertex_shader;
   ComPtr<ID3D11PixelShader> m_copy_pixel_shader;
diff --git a/src/core/gpu_hw_shadergen.cpp b/src/core/gpu_hw_shadergen.cpp
index 85543441d..7105a3ca9 100644
--- a/src/core/gpu_hw_shadergen.cpp
+++ b/src/core/gpu_hw_shadergen.cpp
@@ -502,6 +502,60 @@ int4 SampleFromVRAM(int4 texpage, float2 coord)
   return ss.str();
 }
 
+std::string GPU_HW_ShaderGen::GenerateBatchLineExpandGeometryShader()
+{
+  std::stringstream ss;
+  WriteHeader(ss);
+  WriteCommonFunctions(ss);
+
+  // GS is a pain, too different between HLSL and GLSL...
+  if (m_glsl)
+  {
+  }
+  else
+  {
+    ss << R"(
+CONSTANT float2 OFFSET = (1.0 / float2(VRAM_SIZE)) * float2(RESOLUTION_SCALE, RESOLUTION_SCALE);
+
+struct Vertex
+{
+  float4 col0 : COLOR0;
+  float4 pos : SV_Position;
+};
+
+[maxvertexcount(4)]
+void main(line Vertex input[2], inout TriangleStream<Vertex> output)
+{
+  Vertex v;
+
+  // top-left
+  v.col0 = input[0].col0;
+  v.pos = input[0].pos + float4(-OFFSET.x, +OFFSET.y, 0.0, 0.0);
+  output.Append(v);
+
+  // top-right
+  v.col0 = input[0].col0;
+  v.pos = input[0].pos + float4(+OFFSET.x, +OFFSET.y, 0.0, 0.0);
+  output.Append(v);
+
+  // bottom-left
+  v.col0 = input[1].col0;
+  v.pos = input[1].pos + float4(-OFFSET.x, -OFFSET.y, 0.0, 0.0);
+  output.Append(v);
+
+  // bottom-right
+  v.col0 = input[1].col0;
+  v.pos = input[1].pos + float4(+OFFSET.x, -OFFSET.y, 0.0, 0.0);
+  output.Append(v);
+
+  output.RestartStrip();
+}
+)";
+  }
+
+  return ss.str();
+}
+
 std::string GPU_HW_ShaderGen::GenerateScreenQuadVertexShader()
 {
   std::stringstream ss;
diff --git a/src/core/gpu_hw_shadergen.h b/src/core/gpu_hw_shadergen.h
index 547ec1259..d21cb165b 100644
--- a/src/core/gpu_hw_shadergen.h
+++ b/src/core/gpu_hw_shadergen.h
@@ -21,6 +21,7 @@ public:
   std::string GenerateBatchVertexShader(bool textured);
   std::string GenerateBatchFragmentShader(GPU_HW::BatchRenderMode transparency, GPU::TextureMode texture_mode,
                                           bool dithering);
+  std::string GenerateBatchLineExpandGeometryShader();
   std::string GenerateScreenQuadVertexShader();
   std::string GenerateFillFragmentShader();
   std::string GenerateCopyFragmentShader();