7 files changed, 126 insertions, 57 deletions

diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 0506ac8fe..1b30ce018 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -93,6 +93,7 @@ public:
 
         struct VertexAttribute {
             enum class Size : u32 {
+                Invalid = 0x0,
                 Size_32_32_32_32 = 0x01,
                 Size_32_32_32 = 0x02,
                 Size_16_16_16_16 = 0x03,
@@ -257,6 +258,10 @@ public:
             bool IsNormalized() const {
                 return (type == Type::SignedNorm) || (type == Type::UnsignedNorm);
             }
+
+            bool IsValid() const {
+                return size != Size::Invalid;
+            }
         };
 
         enum class PrimitiveTopology : u32 {
@@ -352,6 +357,27 @@ public:
                 OneMinusConstantColor = 0x62,
                 ConstantAlpha = 0x63,
                 OneMinusConstantAlpha = 0x64,
+
+                // These values are used by Nouveau and some games.
+                ZeroGL = 0x4000,
+                OneGL = 0x4001,
+                SourceColorGL = 0x4300,
+                OneMinusSourceColorGL = 0x4301,
+                SourceAlphaGL = 0x4302,
+                OneMinusSourceAlphaGL = 0x4303,
+                DestAlphaGL = 0x4304,
+                OneMinusDestAlphaGL = 0x4305,
+                DestColorGL = 0x4306,
+                OneMinusDestColorGL = 0x4307,
+                SourceAlphaSaturateGL = 0x4308,
+                ConstantColorGL = 0xc001,
+                OneMinusConstantColorGL = 0xc002,
+                ConstantAlphaGL = 0xc003,
+                OneMinusConstantAlphaGL = 0xc004,
+                Source1ColorGL = 0xc900,
+                OneMinusSource1ColorGL = 0xc901,
+                Source1AlphaGL = 0xc902,
+                OneMinusSource1AlphaGL = 0xc903,
             };
 
             u32 separate_alpha;
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 834940b83..19e7f1161 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -70,6 +70,7 @@ u32 RenderTargetBytesPerPixel(RenderTargetFormat format) {
     case RenderTargetFormat::RG8_SNORM:
         return 2;
     case RenderTargetFormat::R8_UNORM:
+    case RenderTargetFormat::R8_UINT:
         return 1;
     default:
         UNIMPLEMENTED_MSG("Unimplemented render target format {}", static_cast<u32>(format));
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index de5b037be..e008d8f26 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -42,6 +42,7 @@ enum class RenderTargetFormat : u32 {
     R16_UINT = 0xF1,
     R16_FLOAT = 0xF2,
     R8_UNORM = 0xF3,
+    R8_UINT = 0xF6,
 };
 
 enum class DepthFormat : u32 {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index b97e892e6..38a7b1413 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -161,11 +161,16 @@ std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
     // assume every shader uses them all.
     for (unsigned index = 0; index < 16; ++index) {
         auto& attrib = regs.vertex_attrib_format[index];
+
+        // Ignore invalid attributes.
+        if (!attrib.IsValid())
+            continue;
+
+        auto& buffer = regs.vertex_array[attrib.buffer];
         LOG_TRACE(HW_GPU, "vertex attrib {}, count={}, size={}, type={}, offset={}, normalize={}",
                   index, attrib.ComponentCount(), attrib.SizeString(), attrib.TypeString(),
                   attrib.offset.Value(), attrib.IsNormalized());
 
-        auto& buffer = regs.vertex_array[attrib.buffer];
         ASSERT(buffer.IsEnabled());
 
         glEnableVertexAttribArray(index);
@@ -211,7 +216,7 @@ void RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
 
     // Next available bindpoints to use when uploading the const buffers and textures to the GLSL
     // shaders. The constbuffer bindpoint starts after the shader stage configuration bind points.
-    u32 current_constbuffer_bindpoint = uniform_buffers.size();
+    u32 current_constbuffer_bindpoint = static_cast<u32>(uniform_buffers.size());
     u32 current_texture_bindpoint = 0;
 
     for (size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index d055b1dfa..84c250c63 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -99,6 +99,7 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
      false}, // A2B10G10R10
     {GL_RGB5_A1, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, ComponentType::UNorm, false}, // A1B5G5R5
     {GL_R8, GL_RED, GL_UNSIGNED_BYTE, ComponentType::UNorm, false},                    // R8
+    {GL_R8UI, GL_RED_INTEGER, GL_UNSIGNED_BYTE, ComponentType::UInt, false},           // R8UI
     {GL_RGBA16F, GL_RGBA, GL_HALF_FLOAT, ComponentType::Float, false},                 // RGBA16F
     {GL_R11F_G11F_B10F, GL_RGB, GL_UNSIGNED_INT_10F_11F_11F_REV, ComponentType::Float,
      false},                                                                     // R11FG11FB10F
@@ -233,26 +234,27 @@ void MortonCopy(u32 stride, u32 block_height, u32 height, std::vector<u8>& gl_bu
 static constexpr std::array<void (*)(u32, u32, u32, std::vector<u8>&, Tegra::GPUVAddr),
                             SurfaceParams::MaxPixelFormat>
     morton_to_gl_fns = {
-        MortonCopy<true, PixelFormat::ABGR8U>,      MortonCopy<true, PixelFormat::ABGR8S>,
-        MortonCopy<true, PixelFormat::B5G6R5>,      MortonCopy<true, PixelFormat::A2B10G10R10>,
-        MortonCopy<true, PixelFormat::A1B5G5R5>,    MortonCopy<true, PixelFormat::R8>,
-        MortonCopy<true, PixelFormat::RGBA16F>,     MortonCopy<true, PixelFormat::R11FG11FB10F>,
-        MortonCopy<true, PixelFormat::RGBA32UI>,    MortonCopy<true, PixelFormat::DXT1>,
-        MortonCopy<true, PixelFormat::DXT23>,       MortonCopy<true, PixelFormat::DXT45>,
-        MortonCopy<true, PixelFormat::DXN1>,        MortonCopy<true, PixelFormat::DXN2UNORM>,
-        MortonCopy<true, PixelFormat::DXN2SNORM>,   MortonCopy<true, PixelFormat::BC7U>,
-        MortonCopy<true, PixelFormat::ASTC_2D_4X4>, MortonCopy<true, PixelFormat::G8R8>,
-        MortonCopy<true, PixelFormat::BGRA8>,       MortonCopy<true, PixelFormat::RGBA32F>,
-        MortonCopy<true, PixelFormat::RG32F>,       MortonCopy<true, PixelFormat::R32F>,
-        MortonCopy<true, PixelFormat::R16F>,        MortonCopy<true, PixelFormat::R16UNORM>,
-        MortonCopy<true, PixelFormat::R16S>,        MortonCopy<true, PixelFormat::R16UI>,
-        MortonCopy<true, PixelFormat::R16I>,        MortonCopy<true, PixelFormat::RG16>,
-        MortonCopy<true, PixelFormat::RG16F>,       MortonCopy<true, PixelFormat::RG16UI>,
-        MortonCopy<true, PixelFormat::RG16I>,       MortonCopy<true, PixelFormat::RG16S>,
-        MortonCopy<true, PixelFormat::RGB32F>,      MortonCopy<true, PixelFormat::SRGBA8>,
-        MortonCopy<true, PixelFormat::RG8S>,        MortonCopy<true, PixelFormat::Z24S8>,
-        MortonCopy<true, PixelFormat::S8Z24>,       MortonCopy<true, PixelFormat::Z32F>,
-        MortonCopy<true, PixelFormat::Z16>,         MortonCopy<true, PixelFormat::Z32FS8>,
+        MortonCopy<true, PixelFormat::ABGR8U>,       MortonCopy<true, PixelFormat::ABGR8S>,
+        MortonCopy<true, PixelFormat::B5G6R5>,       MortonCopy<true, PixelFormat::A2B10G10R10>,
+        MortonCopy<true, PixelFormat::A1B5G5R5>,     MortonCopy<true, PixelFormat::R8>,
+        MortonCopy<true, PixelFormat::R8UI>,         MortonCopy<true, PixelFormat::RGBA16F>,
+        MortonCopy<true, PixelFormat::R11FG11FB10F>, MortonCopy<true, PixelFormat::RGBA32UI>,
+        MortonCopy<true, PixelFormat::DXT1>,         MortonCopy<true, PixelFormat::DXT23>,
+        MortonCopy<true, PixelFormat::DXT45>,        MortonCopy<true, PixelFormat::DXN1>,
+        MortonCopy<true, PixelFormat::DXN2UNORM>,    MortonCopy<true, PixelFormat::DXN2SNORM>,
+        MortonCopy<true, PixelFormat::BC7U>,         MortonCopy<true, PixelFormat::ASTC_2D_4X4>,
+        MortonCopy<true, PixelFormat::G8R8>,         MortonCopy<true, PixelFormat::BGRA8>,
+        MortonCopy<true, PixelFormat::RGBA32F>,      MortonCopy<true, PixelFormat::RG32F>,
+        MortonCopy<true, PixelFormat::R32F>,         MortonCopy<true, PixelFormat::R16F>,
+        MortonCopy<true, PixelFormat::R16UNORM>,     MortonCopy<true, PixelFormat::R16S>,
+        MortonCopy<true, PixelFormat::R16UI>,        MortonCopy<true, PixelFormat::R16I>,
+        MortonCopy<true, PixelFormat::RG16>,         MortonCopy<true, PixelFormat::RG16F>,
+        MortonCopy<true, PixelFormat::RG16UI>,       MortonCopy<true, PixelFormat::RG16I>,
+        MortonCopy<true, PixelFormat::RG16S>,        MortonCopy<true, PixelFormat::RGB32F>,
+        MortonCopy<true, PixelFormat::SRGBA8>,       MortonCopy<true, PixelFormat::RG8S>,
+        MortonCopy<true, PixelFormat::Z24S8>,        MortonCopy<true, PixelFormat::S8Z24>,
+        MortonCopy<true, PixelFormat::Z32F>,         MortonCopy<true, PixelFormat::Z16>,
+        MortonCopy<true, PixelFormat::Z32FS8>,
 };
 
 static constexpr std::array<void (*)(u32, u32, u32, std::vector<u8>&, Tegra::GPUVAddr),
@@ -264,6 +266,7 @@ static constexpr std::array<void (*)(u32, u32, u32, std::vector<u8>&, Tegra::GPU
         MortonCopy<false, PixelFormat::A2B10G10R10>,
         MortonCopy<false, PixelFormat::A1B5G5R5>,
         MortonCopy<false, PixelFormat::R8>,
+        MortonCopy<false, PixelFormat::R8UI>,
         MortonCopy<false, PixelFormat::RGBA16F>,
         MortonCopy<false, PixelFormat::R11FG11FB10F>,
         MortonCopy<false, PixelFormat::RGBA32UI>,
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index d7a43652e..202257b58 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -29,44 +29,45 @@ struct SurfaceParams {
         A2B10G10R10 = 3,
         A1B5G5R5 = 4,
         R8 = 5,
-        RGBA16F = 6,
-        R11FG11FB10F = 7,
-        RGBA32UI = 8,
-        DXT1 = 9,
-        DXT23 = 10,
-        DXT45 = 11,
-        DXN1 = 12, // This is also known as BC4
-        DXN2UNORM = 13,
-        DXN2SNORM = 14,
-        BC7U = 15,
-        ASTC_2D_4X4 = 16,
-        G8R8 = 17,
-        BGRA8 = 18,
-        RGBA32F = 19,
-        RG32F = 20,
-        R32F = 21,
-        R16F = 22,
-        R16UNORM = 23,
-        R16S = 24,
-        R16UI = 25,
-        R16I = 26,
-        RG16 = 27,
-        RG16F = 28,
-        RG16UI = 29,
-        RG16I = 30,
-        RG16S = 31,
-        RGB32F = 32,
-        SRGBA8 = 33,
-        RG8S = 34,
+        R8UI = 6,
+        RGBA16F = 7,
+        R11FG11FB10F = 8,
+        RGBA32UI = 9,
+        DXT1 = 10,
+        DXT23 = 11,
+        DXT45 = 12,
+        DXN1 = 13, // This is also known as BC4
+        DXN2UNORM = 14,
+        DXN2SNORM = 15,
+        BC7U = 16,
+        ASTC_2D_4X4 = 17,
+        G8R8 = 18,
+        BGRA8 = 19,
+        RGBA32F = 20,
+        RG32F = 21,
+        R32F = 22,
+        R16F = 23,
+        R16UNORM = 24,
+        R16S = 25,
+        R16UI = 26,
+        R16I = 27,
+        RG16 = 28,
+        RG16F = 29,
+        RG16UI = 30,
+        RG16I = 31,
+        RG16S = 32,
+        RGB32F = 33,
+        SRGBA8 = 34,
+        RG8S = 35,
 
         MaxColorFormat,
 
         // DepthStencil formats
-        Z24S8 = 35,
-        S8Z24 = 36,
-        Z32F = 37,
-        Z16 = 38,
-        Z32FS8 = 39,
+        Z24S8 = 36,
+        S8Z24 = 37,
+        Z32F = 38,
+        Z16 = 39,
+        Z32FS8 = 40,
 
         MaxDepthStencilFormat,
 
@@ -110,6 +111,7 @@ struct SurfaceParams {
             1, // A2B10G10R10
             1, // A1B5G5R5
             1, // R8
+            1, // R8UI
             1, // RGBA16F
             1, // R11FG11FB10F
             1, // RGBA32UI
@@ -161,6 +163,7 @@ struct SurfaceParams {
             32,  // A2B10G10R10
             16,  // A1B5G5R5
             8,   // R8
+            8,   // R8UI
             64,  // RGBA16F
             32,  // R11FG11FB10F
             128, // RGBA32UI
@@ -250,6 +253,8 @@ struct SurfaceParams {
             return PixelFormat::RGBA32UI;
         case Tegra::RenderTargetFormat::R8_UNORM:
             return PixelFormat::R8;
+        case Tegra::RenderTargetFormat::R8_UINT:
+            return PixelFormat::R8UI;
         case Tegra::RenderTargetFormat::RG16_FLOAT:
             return PixelFormat::RG16F;
         case Tegra::RenderTargetFormat::RG16_UINT:
@@ -301,7 +306,15 @@ struct SurfaceParams {
         case Tegra::Texture::TextureFormat::A1B5G5R5:
             return PixelFormat::A1B5G5R5;
         case Tegra::Texture::TextureFormat::R8:
-            return PixelFormat::R8;
+            switch (component_type) {
+            case Tegra::Texture::ComponentType::UNORM:
+                return PixelFormat::R8;
+            case Tegra::Texture::ComponentType::UINT:
+                return PixelFormat::R8UI;
+            }
+            LOG_CRITICAL(HW_GPU, "Unimplemented component_type={}",
+                         static_cast<u32>(component_type));
+            UNREACHABLE();
         case Tegra::Texture::TextureFormat::G8R8:
             return PixelFormat::G8R8;
         case Tegra::Texture::TextureFormat::R16_G16_B16_A16:
@@ -435,6 +448,7 @@ struct SurfaceParams {
             return ComponentType::Float;
         case Tegra::RenderTargetFormat::RGBA32_UINT:
         case Tegra::RenderTargetFormat::RG16_UINT:
+        case Tegra::RenderTargetFormat::R8_UINT:
         case Tegra::RenderTargetFormat::R16_UINT:
             return ComponentType::UInt;
         case Tegra::RenderTargetFormat::RG16_SINT:
diff --git a/src/video_core/renderer_opengl/maxwell_to_gl.h b/src/video_core/renderer_opengl/maxwell_to_gl.h
index c439446b1..5afd20dbe 100644
--- a/src/video_core/renderer_opengl/maxwell_to_gl.h
+++ b/src/video_core/renderer_opengl/maxwell_to_gl.h
@@ -156,42 +156,61 @@ inline GLenum BlendEquation(Maxwell::Blend::Equation equation) {
 inline GLenum BlendFunc(Maxwell::Blend::Factor factor) {
     switch (factor) {
     case Maxwell::Blend::Factor::Zero:
+    case Maxwell::Blend::Factor::ZeroGL:
         return GL_ZERO;
     case Maxwell::Blend::Factor::One:
+    case Maxwell::Blend::Factor::OneGL:
         return GL_ONE;
     case Maxwell::Blend::Factor::SourceColor:
+    case Maxwell::Blend::Factor::SourceColorGL:
         return GL_SRC_COLOR;
     case Maxwell::Blend::Factor::OneMinusSourceColor:
+    case Maxwell::Blend::Factor::OneMinusSourceColorGL:
         return GL_ONE_MINUS_SRC_COLOR;
     case Maxwell::Blend::Factor::SourceAlpha:
+    case Maxwell::Blend::Factor::SourceAlphaGL:
         return GL_SRC_ALPHA;
     case Maxwell::Blend::Factor::OneMinusSourceAlpha:
+    case Maxwell::Blend::Factor::OneMinusSourceAlphaGL:
         return GL_ONE_MINUS_SRC_ALPHA;
     case Maxwell::Blend::Factor::DestAlpha:
+    case Maxwell::Blend::Factor::DestAlphaGL:
         return GL_DST_ALPHA;
     case Maxwell::Blend::Factor::OneMinusDestAlpha:
+    case Maxwell::Blend::Factor::OneMinusDestAlphaGL:
         return GL_ONE_MINUS_DST_ALPHA;
     case Maxwell::Blend::Factor::DestColor:
+    case Maxwell::Blend::Factor::DestColorGL:
         return GL_DST_COLOR;
     case Maxwell::Blend::Factor::OneMinusDestColor:
+    case Maxwell::Blend::Factor::OneMinusDestColorGL:
         return GL_ONE_MINUS_DST_COLOR;
     case Maxwell::Blend::Factor::SourceAlphaSaturate:
+    case Maxwell::Blend::Factor::SourceAlphaSaturateGL:
         return GL_SRC_ALPHA_SATURATE;
     case Maxwell::Blend::Factor::Source1Color:
+    case Maxwell::Blend::Factor::Source1ColorGL:
         return GL_SRC1_COLOR;
     case Maxwell::Blend::Factor::OneMinusSource1Color:
+    case Maxwell::Blend::Factor::OneMinusSource1ColorGL:
         return GL_ONE_MINUS_SRC1_COLOR;
     case Maxwell::Blend::Factor::Source1Alpha:
+    case Maxwell::Blend::Factor::Source1AlphaGL:
         return GL_SRC1_ALPHA;
     case Maxwell::Blend::Factor::OneMinusSource1Alpha:
+    case Maxwell::Blend::Factor::OneMinusSource1AlphaGL:
         return GL_ONE_MINUS_SRC1_ALPHA;
     case Maxwell::Blend::Factor::ConstantColor:
+    case Maxwell::Blend::Factor::ConstantColorGL:
         return GL_CONSTANT_COLOR;
     case Maxwell::Blend::Factor::OneMinusConstantColor:
+    case Maxwell::Blend::Factor::OneMinusConstantColorGL:
         return GL_ONE_MINUS_CONSTANT_COLOR;
     case Maxwell::Blend::Factor::ConstantAlpha:
+    case Maxwell::Blend::Factor::ConstantAlphaGL:
         return GL_CONSTANT_ALPHA;
     case Maxwell::Blend::Factor::OneMinusConstantAlpha:
+    case Maxwell::Blend::Factor::OneMinusConstantAlphaGL:
         return GL_ONE_MINUS_CONSTANT_ALPHA;
     }
     LOG_CRITICAL(Render_OpenGL, "Unimplemented blend factor={}", static_cast<u32>(factor));