Merge pull request #12324 from Pokechu22/vertex-loader-dummy-tex-coords

VertexLoader: Fixes and cleanup related to skipped components

Source/Core/VideoCommon/CPMemory.h

@@ -575,6 +575,102 @@ struct VAT
       return 0;
     }
   }
+  void SetTexElements(size_t idx, TexComponentCount value)
+  {
+    switch (idx)
+    {
+    case 0:
+      g0.Tex0CoordElements = value;
+      return;
+    case 1:
+      g1.Tex1CoordElements = value;
+      return;
+    case 2:
+      g1.Tex2CoordElements = value;
+      return;
+    case 3:
+      g1.Tex3CoordElements = value;
+      return;
+    case 4:
+      g1.Tex4CoordElements = value;
+      return;
+    case 5:
+      g2.Tex5CoordElements = value;
+      return;
+    case 6:
+      g2.Tex6CoordElements = value;
+      return;
+    case 7:
+      g2.Tex7CoordElements = value;
+      return;
+    default:
+      PanicAlertFmt("Invalid tex coord index {}", idx);
+    }
+  }
+  void SetTexFormat(size_t idx, ComponentFormat value)
+  {
+    switch (idx)
+    {
+    case 0:
+      g0.Tex0CoordFormat = value;
+      return;
+    case 1:
+      g1.Tex1CoordFormat = value;
+      return;
+    case 2:
+      g1.Tex2CoordFormat = value;
+      return;
+    case 3:
+      g1.Tex3CoordFormat = value;
+      return;
+    case 4:
+      g1.Tex4CoordFormat = value;
+      return;
+    case 5:
+      g2.Tex5CoordFormat = value;
+      return;
+    case 6:
+      g2.Tex6CoordFormat = value;
+      return;
+    case 7:
+      g2.Tex7CoordFormat = value;
+      return;
+    default:
+      PanicAlertFmt("Invalid tex coord index {}", idx);
+    }
+  }
+  void SetTexFrac(size_t idx, u8 value)
+  {
+    switch (idx)
+    {
+    case 0:
+      g0.Tex0Frac = value;
+      return;
+    case 1:
+      g1.Tex1Frac = value;
+      return;
+    case 2:
+      g1.Tex2Frac = value;
+      return;
+    case 3:
+      g1.Tex3Frac = value;
+      return;
+    case 4:
+      g2.Tex4Frac = value;
+      return;
+    case 5:
+      g2.Tex5Frac = value;
+      return;
+    case 6:
+      g2.Tex6Frac = value;
+      return;
+    case 7:
+      g2.Tex7Frac = value;
+      return;
+    default:
+      PanicAlertFmt("Invalid tex coord index {}", idx);
+    }
+  }
 };
 template <>
 struct fmt::formatter<VAT>

Source/Core/VideoCommon/VertexLoader.cpp

@@ -77,9 +77,6 @@ VertexLoader::VertexLoader(const TVtxDesc& vtx_desc, const VAT& vtx_attr)
 
 void VertexLoader::CompileVertexTranslator()
 {
-  // Reset pipeline
-  m_numPipelineStages = 0;
-
   // Position in pc vertex format.
   int nat_offset = 0;
 
@@ -149,9 +146,16 @@ void VertexLoader::CompileVertexTranslator()
         VertexLoader_Color::GetFunction(m_VtxDesc.low.Color[i], m_VtxAttr.GetColorFormat(i));
 
     if (pFunc != nullptr)
+    {
       WriteCall(pFunc);
+    }
     else
+    {
       ASSERT(m_VtxDesc.low.Color[i] == VertexComponentFormat::NotPresent);
+      // Keep colIndex in sync if color 0 is absent but color 1 is present
+      if (i == 0 && m_VtxDesc.low.Color[1] != VertexComponentFormat::NotPresent)
+        WriteCall(VertexLoader_Color::GetDummyFunction());
+    }
 
     if (m_VtxDesc.low.Color[i] != VertexComponentFormat::NotPresent)
     {
@@ -213,12 +217,13 @@ void VertexLoader::CompileVertexTranslator()
     {
       // if there's more tex coords later, have to write a dummy call
       bool has_more = false;
-      for (size_t j = 0; j < m_VtxDesc.high.TexCoord.Size(); ++j)
+      for (size_t j = i + 1; j < m_VtxDesc.high.TexCoord.Size(); ++j)
       {
         if (m_VtxDesc.high.TexCoord[j] != VertexComponentFormat::NotPresent)
         {
           has_more = true;
-          WriteCall(VertexLoader_TextCoord::GetDummyFunction());  // important to get indices right!
+          // Keep tcIndex in sync so that the correct array is used later
+          WriteCall(VertexLoader_TextCoord::GetDummyFunction());
           break;
         }
         else if (m_VtxDesc.low.TexMatIdx[j])
@@ -245,7 +250,7 @@ void VertexLoader::CompileVertexTranslator()
 
 void VertexLoader::WriteCall(TPipelineFunction func)
 {
-  m_PipelineStages[m_numPipelineStages++] = func;
+  m_PipelineStages.push_back(func);
 }
 
 int VertexLoader::RunVertices(const u8* src, u8* dst, int count)
@@ -261,8 +266,8 @@ int VertexLoader::RunVertices(const u8* src, u8* dst, int count)
     m_tcIndex = 0;
     m_colIndex = 0;
     m_texmtxwrite = m_texmtxread = 0;
-    for (int i = 0; i < m_numPipelineStages; i++)
-      m_PipelineStages[i](this);
+    for (TPipelineFunction& func : m_PipelineStages)
+      func(this);
     PRIM_LOG("\n");
   }
 

Source/Core/VideoCommon/VertexLoader.h

@@ -9,6 +9,7 @@
 #include <string>
 
 #include "Common/CommonTypes.h"
+#include "Common/SmallVector.h"
 #include "VideoCommon/VertexLoaderBase.h"
 
 class VertexLoader;
@@ -38,8 +39,11 @@ public:
 
 private:
   // Pipeline.
-  TPipelineFunction m_PipelineStages[64];  // TODO - figure out real max. it's lower.
-  int m_numPipelineStages;
+  // 1 pos matrix + 8 texture matrices + 1 position + 1 normal or normal/binormal/tangent
+  // + 2 colors + 8 texture coordinates or dummy texture coordinates + 8 texture matrices
+  // merged into texture coordinates + 1 skip gives a maximum of 30
+  // (Tested by VertexLoaderTest.LargeFloatVertexSpeed)
+  Common::SmallVector<TPipelineFunction, 30> m_PipelineStages;
 
   void CompileVertexTranslator();
 

Source/Core/VideoCommon/VertexLoader_Color.cpp

@@ -16,6 +16,11 @@
 
 namespace
 {
+void Color_Read_Dummy(VertexLoader* loader)
+{
+  loader->m_colIndex++;
+}
+
 constexpr u32 alpha_mask = 0xFF000000;
 
 void SetCol(VertexLoader* loader, u32 val)
@@ -201,3 +206,8 @@ TPipelineFunction VertexLoader_Color::GetFunction(VertexComponentFormat type, Co
   }
   return s_table_read_color[type][format];
 }
+
+TPipelineFunction VertexLoader_Color::GetDummyFunction()
+{
+  return Color_Read_Dummy;
+}

Source/Core/VideoCommon/VertexLoader_Color.h

@@ -25,6 +25,9 @@ public:
 
   static TPipelineFunction GetFunction(VertexComponentFormat type, ColorFormat format);
 
+  // It is important to synchronize colIndex, or else the wrong color array will be used
+  static TPipelineFunction GetDummyFunction();
+
 private:
   template <typename T, auto last_member>
   using EnumMap = typename Common::EnumMap<T, last_member>;

Source/Core/VideoCommon/VertexLoader_TextCoord.h

@@ -22,7 +22,7 @@ public:
   static TPipelineFunction GetFunction(VertexComponentFormat type, ComponentFormat format,
                                        TexComponentCount elements);
 
-  // It is important to synchronize tcIndex.
+  // It is important to synchronize tcIndex, or else the wrong texture coordinate array will be used
   static TPipelineFunction GetDummyFunction();
 
 private:

Source/UnitTests/VideoCommon/VertexLoaderTest.cpp

@@ -759,6 +759,284 @@ TEST_P(VertexLoaderNormalTest, NormalAll)
   }
 }
 
+class VertexLoaderSkippedColorsTest : public VertexLoaderTest,
+                                      public ::testing::WithParamInterface<std::tuple<bool, bool>>
+{
+};
+INSTANTIATE_TEST_SUITE_P(AllCombinations, VertexLoaderSkippedColorsTest,
+                         ::testing::Combine(::testing::Values(false, true),
+                                            ::testing::Values(false, true)));
+
+TEST_P(VertexLoaderSkippedColorsTest, SkippedColors)
+{
+  bool enable_color_0, enable_color_1;
+  std::tie(enable_color_0, enable_color_1) = GetParam();
+
+  size_t input_size = 1;
+  size_t output_size = 3 * sizeof(float);
+  size_t color_0_offset = 0;
+  size_t color_1_offset = 0;
+
+  m_vtx_desc.low.Position = VertexComponentFormat::Index8;
+  if (enable_color_0)
+  {
+    m_vtx_desc.low.Color0 = VertexComponentFormat::Index8;
+    input_size++;
+    color_0_offset = output_size;
+    output_size += sizeof(u32);
+  }
+  if (enable_color_1)
+  {
+    m_vtx_desc.low.Color1 = VertexComponentFormat::Index8;
+    input_size++;
+    color_1_offset = output_size;
+    output_size += sizeof(u32);
+  }
+
+  m_vtx_attr.g0.PosElements = CoordComponentCount::XYZ;
+  m_vtx_attr.g0.PosFormat = ComponentFormat::Float;
+  m_vtx_attr.g0.Color0Elements = ColorComponentCount::RGBA;
+  m_vtx_attr.g0.Color0Comp = ColorFormat::RGBA8888;
+  m_vtx_attr.g0.Color1Elements = ColorComponentCount::RGBA;
+  m_vtx_attr.g0.Color1Comp = ColorFormat::RGBA8888;
+
+  CreateAndCheckSizes(input_size, output_size);
+
+  // Vertex 0
+  Input<u8>(1);
+  if (enable_color_0)
+    Input<u8>(1);
+  if (enable_color_1)
+    Input<u8>(1);
+  // Vertex 1
+  Input<u8>(0);
+  if (enable_color_0)
+    Input<u8>(0);
+  if (enable_color_1)
+    Input<u8>(0);
+  // Position array
+  VertexLoaderManager::cached_arraybases[CPArray::Position] = m_src.GetPointer();
+  g_main_cp_state.array_strides[CPArray::Position] =
+      sizeof(float);  // so 1, 2, 3 for index 0; 2, 3, 4 for index 1
+  Input(1.f);
+  Input(2.f);
+  Input(3.f);
+  Input(4.f);
+  // Color array 0
+  VertexLoaderManager::cached_arraybases[CPArray::Color0] = m_src.GetPointer();
+  g_main_cp_state.array_strides[CPArray::Color0] = sizeof(u32);
+  Input<u32>(0x00010203u);
+  Input<u32>(0x04050607u);
+  // Color array 1
+  VertexLoaderManager::cached_arraybases[CPArray::Color1] = m_src.GetPointer();
+  g_main_cp_state.array_strides[CPArray::Color1] = sizeof(u32);
+  Input<u32>(0x08090a0bu);
+  Input<u32>(0x0c0d0e0fu);
+
+  ASSERT_EQ(m_loader->m_native_vtx_decl.colors[0].enable, enable_color_0);
+  if (enable_color_0)
+    ASSERT_EQ(m_loader->m_native_vtx_decl.colors[0].offset, color_0_offset);
+  ASSERT_EQ(m_loader->m_native_vtx_decl.colors[1].enable, enable_color_1);
+  if (enable_color_1)
+    ASSERT_EQ(m_loader->m_native_vtx_decl.colors[1].offset, color_1_offset);
+
+  RunVertices(2);
+  // Vertex 0
+  ExpectOut(2);
+  ExpectOut(3);
+  ExpectOut(4);
+  if (enable_color_0)
+    EXPECT_EQ((m_dst.Read<u32, true>()), 0x04050607u);
+  if (enable_color_1)
+    EXPECT_EQ((m_dst.Read<u32, true>()), 0x0c0d0e0fu);
+  // Vertex 1
+  ExpectOut(1);
+  ExpectOut(2);
+  ExpectOut(3);
+  if (enable_color_0)
+    EXPECT_EQ((m_dst.Read<u32, true>()), 0x00010203u);
+  if (enable_color_1)
+    EXPECT_EQ((m_dst.Read<u32, true>()), 0x08090a0bu);
+}
+
+class VertexLoaderSkippedTexCoordsTest : public VertexLoaderTest,
+                                         public ::testing::WithParamInterface<u32>
+{
+public:
+  static constexpr u32 NUM_COMPONENTS_TO_TEST = 3;
+  static constexpr u32 NUM_PARAMETERS_PER_COMPONENT = 3;
+  static constexpr u32 NUM_COMBINATIONS =
+      1 << (NUM_COMPONENTS_TO_TEST * NUM_PARAMETERS_PER_COMPONENT);
+};
+INSTANTIATE_TEST_SUITE_P(AllCombinations, VertexLoaderSkippedTexCoordsTest,
+                         ::testing::Range(0u, VertexLoaderSkippedTexCoordsTest::NUM_COMBINATIONS));
+
+TEST_P(VertexLoaderSkippedTexCoordsTest, SkippedTextures)
+{
+  std::array<bool, NUM_COMPONENTS_TO_TEST> enable_tex, enable_matrix, use_st;
+  const u32 param = GetParam();
+  for (u32 component = 0; component < NUM_COMPONENTS_TO_TEST; component++)
+  {
+    const u32 bits = param >> (component * NUM_PARAMETERS_PER_COMPONENT);
+    enable_tex[component] = (bits & 1);
+    enable_matrix[component] = (bits & 2);
+    use_st[component] = (bits & 4);
+  }
+
+  size_t input_size = 1;
+  size_t output_size = 3 * sizeof(float);
+
+  std::array<bool, NUM_COMPONENTS_TO_TEST> component_enabled{};
+  std::array<size_t, NUM_COMPONENTS_TO_TEST> component_offset{};
+
+  m_vtx_desc.low.Position = VertexComponentFormat::Index8;
+  m_vtx_attr.g0.PosElements = CoordComponentCount::XYZ;
+  m_vtx_attr.g0.PosFormat = ComponentFormat::Float;
+
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    if (enable_matrix[i] || enable_tex[i])
+    {
+      component_enabled[i] = true;
+      component_offset[i] = output_size;
+      if (enable_matrix[i])
+      {
+        output_size += 3 * sizeof(float);
+      }
+      else
+      {
+        if (use_st[i])
+        {
+          output_size += 2 * sizeof(float);
+        }
+        else
+        {
+          output_size += sizeof(float);
+        }
+      }
+    }
+    if (enable_matrix[i])
+    {
+      m_vtx_desc.low.TexMatIdx[i] = enable_matrix[i];
+      input_size++;
+    }
+    if (enable_tex[i])
+    {
+      m_vtx_desc.high.TexCoord[i] = VertexComponentFormat::Index8;
+      input_size++;
+    }
+
+    m_vtx_attr.SetTexElements(i, use_st[i] ? TexComponentCount::ST : TexComponentCount::S);
+    m_vtx_attr.SetTexFormat(i, ComponentFormat::Float);
+    m_vtx_attr.SetTexFrac(i, 0);
+  }
+
+  CreateAndCheckSizes(input_size, output_size);
+
+  // Vertex 0
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    if (enable_matrix[i])
+      Input<u8>(u8(20 + i));
+  }
+  Input<u8>(1);  // Position
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    if (enable_tex[i])
+      Input<u8>(1);
+  }
+  // Vertex 1
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    if (enable_matrix[i])
+      Input<u8>(u8(10 + i));
+  }
+  Input<u8>(0);  // Position
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    if (enable_tex[i])
+      Input<u8>(0);
+  }
+  // Position array
+  VertexLoaderManager::cached_arraybases[CPArray::Position] = m_src.GetPointer();
+  g_main_cp_state.array_strides[CPArray::Position] =
+      sizeof(float);  // so 1, 2, 3 for index 0; 2, 3, 4 for index 1
+  Input(1.f);
+  Input(2.f);
+  Input(3.f);
+  Input(4.f);
+  // Texture coord arrays
+  for (u8 i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    VertexLoaderManager::cached_arraybases[CPArray::TexCoord0 + i] = m_src.GetPointer();
+    g_main_cp_state.array_strides[CPArray::TexCoord0 + i] = 2 * sizeof(float);
+    Input<float>(i * 100 + 11);
+    Input<float>(i * 100 + 12);
+    Input<float>(i * 100 + 21);
+    Input<float>(i * 100 + 22);
+  }
+
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    ASSERT_EQ(m_loader->m_native_vtx_decl.texcoords[i].enable, component_enabled[i]);
+    if (component_enabled[i])
+      ASSERT_EQ(m_loader->m_native_vtx_decl.texcoords[i].offset, component_offset[i]);
+  }
+
+  RunVertices(2);
+
+  // Vertex 0
+  ExpectOut(2);
+  ExpectOut(3);
+  ExpectOut(4);
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    size_t num_read = 0;
+    if (enable_tex[i])
+    {
+      ExpectOut(i * 100 + 21);
+      num_read++;
+      if (use_st[i])
+      {
+        ExpectOut(i * 100 + 22);
+        num_read++;
+      }
+    }
+    if (enable_matrix[i])
+    {
+      // With a matrix there are always 3 components; otherwise-unused components should be 0
+      while (num_read++ < 2)
+        ExpectOut(0);
+      ExpectOut(20 + i);
+    }
+  }
+  // Vertex 1
+  ExpectOut(1);
+  ExpectOut(2);
+  ExpectOut(3);
+  for (size_t i = 0; i < NUM_COMPONENTS_TO_TEST; i++)
+  {
+    size_t num_read = 0;
+    if (enable_tex[i])
+    {
+      ExpectOut(i * 100 + 11);
+      num_read++;
+      if (use_st[i])
+      {
+        ExpectOut(i * 100 + 12);
+        num_read++;
+      }
+    }
+    if (enable_matrix[i])
+    {
+      // With a matrix there are always 3 components; otherwise-unused components should be 0
+      while (num_read++ < 2)
+        ExpectOut(0);
+      ExpectOut(10 + i);
+    }
+  }
+}
+
 // For gtest, which doesn't know about our fmt::formatters by default
 static void PrintTo(const VertexComponentFormat& t, std::ostream* os)
 {