Merge pull request #528 from Subv/rg11b10f

GPU: Implemented the R11FG11FB10F texture and rendertarget formats.

src/video_core/gpu.h

@@ -19,6 +19,7 @@ enum class RenderTargetFormat : u32 {
     RGB10_A2_UNORM = 0xD1,
     RGBA8_UNORM = 0xD5,
     RGBA8_SRGB = 0xD6,
+    R11G11B10_FLOAT = 0xE0,
 };
 
 /// Returns the number of bytes per pixel of each rendertarget format.

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -50,6 +50,7 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {GL_RGB5_A1, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, false},                // A1B5G5R5
     {GL_R8, GL_RED, GL_UNSIGNED_BYTE, false},                                   // R8
     {GL_RGBA16F, GL_RGBA, GL_HALF_FLOAT, false},                                // RGBA16F
+    {GL_R11F_G11F_B10F, GL_RGB, GL_UNSIGNED_INT_10F_11F_11F_REV, false},        // R11FG11FB10F
     {GL_COMPRESSED_RGB_S3TC_DXT1_EXT, GL_RGB, GL_UNSIGNED_INT_8_8_8_8, true},   // DXT1
     {GL_COMPRESSED_RGBA_S3TC_DXT3_EXT, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8, true}, // DXT23
     {GL_COMPRESSED_RGBA_S3TC_DXT5_EXT, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8, true}, // DXT45
@@ -60,8 +61,10 @@ static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType
     const SurfaceType type = SurfaceParams::GetFormatType(pixel_format);
     if (type == SurfaceType::ColorTexture) {
         ASSERT(static_cast<size_t>(pixel_format) < tex_format_tuples.size());
-        // For now only UNORM components are supported, or RGBA16F which is type FLOAT
+        // For now only UNORM components are supported, or either R11FG11FB10F or RGBA16F which are
-        ASSERT(component_type == ComponentType::UNorm || pixel_format == PixelFormat::RGBA16F);
+        // type FLOAT
+        ASSERT(component_type == ComponentType::UNorm || pixel_format == PixelFormat::RGBA16F ||
+               pixel_format == PixelFormat::R11FG11FB10F);
         return tex_format_tuples[static_cast<unsigned int>(pixel_format)];
     } else if (type == SurfaceType::Depth || type == SurfaceType::DepthStencil) {
         // TODO(Subv): Implement depth formats
@@ -113,8 +116,9 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
         MortonCopy<true, PixelFormat::ABGR8>,        MortonCopy<true, PixelFormat::B5G6R5>,
         MortonCopy<true, PixelFormat::A2B10G10R10>,  MortonCopy<true, PixelFormat::A1B5G5R5>,
         MortonCopy<true, PixelFormat::R8>,           MortonCopy<true, PixelFormat::RGBA16F>,
-        MortonCopy<true, PixelFormat::DXT1>,        MortonCopy<true, PixelFormat::DXT23>,
+        MortonCopy<true, PixelFormat::R11FG11FB10F>, MortonCopy<true, PixelFormat::DXT1>,
-        MortonCopy<true, PixelFormat::DXT45>,       MortonCopy<true, PixelFormat::DXN1>,
+        MortonCopy<true, PixelFormat::DXT23>,        MortonCopy<true, PixelFormat::DXT45>,
+        MortonCopy<true, PixelFormat::DXN1>,
 };
 
 static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra::GPUVAddr,
@@ -127,6 +131,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
         MortonCopy<false, PixelFormat::A1B5G5R5>,
         MortonCopy<false, PixelFormat::R8>,
         MortonCopy<false, PixelFormat::RGBA16F>,
+        MortonCopy<false, PixelFormat::R11FG11FB10F>,
         // TODO(Subv): Swizzling the DXT1/DXT23/DXT45/DXN1 formats is not yet supported
         nullptr,
         nullptr,

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -60,10 +60,11 @@ struct SurfaceParams {
         A1B5G5R5 = 3,
         R8 = 4,
         RGBA16F = 5,
-        DXT1 = 6,
+        R11FG11FB10F = 6,
-        DXT23 = 7,
+        DXT1 = 7,
-        DXT45 = 8,
+        DXT23 = 8,
-        DXN1 = 9, // This is also known as BC4
+        DXT45 = 9,
+        DXN1 = 10, // This is also known as BC4
 
         Max,
         Invalid = 255,
@@ -104,7 +105,8 @@ struct SurfaceParams {
             1, // A2B10G10R10
             1, // A1B5G5R5
             1, // R8
-            2, // RGBA16F
+            1, // RGBA16F
+            1, // R11FG11FB10F
             4, // DXT1
             4, // DXT23
             4, // DXT45
@@ -129,6 +131,7 @@ struct SurfaceParams {
             16,  // A1B5G5R5
             8,   // R8
             64,  // RGBA16F
+            32,  // R11FG11FB10F
             64,  // DXT1
             128, // DXT23
             128, // DXT45
@@ -151,6 +154,8 @@ struct SurfaceParams {
             return PixelFormat::A2B10G10R10;
         case Tegra::RenderTargetFormat::RGBA16_FLOAT:
             return PixelFormat::RGBA16F;
+        case Tegra::RenderTargetFormat::R11G11B10_FLOAT:
+            return PixelFormat::R11FG11FB10F;
         default:
             NGLOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
@@ -182,6 +187,8 @@ struct SurfaceParams {
             return PixelFormat::R8;
         case Tegra::Texture::TextureFormat::R16_G16_B16_A16:
             return PixelFormat::RGBA16F;
+        case Tegra::Texture::TextureFormat::BF10GF11RF11:
+            return PixelFormat::R11FG11FB10F;
         case Tegra::Texture::TextureFormat::DXT1:
             return PixelFormat::DXT1;
         case Tegra::Texture::TextureFormat::DXT23:
@@ -211,6 +218,8 @@ struct SurfaceParams {
             return Tegra::Texture::TextureFormat::R8;
         case PixelFormat::RGBA16F:
             return Tegra::Texture::TextureFormat::R16_G16_B16_A16;
+        case PixelFormat::R11FG11FB10F:
+            return Tegra::Texture::TextureFormat::BF10GF11RF11;
         case PixelFormat::DXT1:
             return Tegra::Texture::TextureFormat::DXT1;
         case PixelFormat::DXT23:
@@ -243,6 +252,7 @@ struct SurfaceParams {
         case Tegra::RenderTargetFormat::RGB10_A2_UNORM:
             return ComponentType::UNorm;
         case Tegra::RenderTargetFormat::RGBA16_FLOAT:
+        case Tegra::RenderTargetFormat::R11G11B10_FLOAT:
             return ComponentType::Float;
         default:
             NGLOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));

src/video_core/textures/decoders.cpp

@@ -55,6 +55,7 @@ u32 BytesPerPixel(TextureFormat format) {
         return 16;
     case TextureFormat::A8R8G8B8:
     case TextureFormat::A2B10G10R10:
+    case TextureFormat::BF10GF11RF11:
         return 4;
     case TextureFormat::A1B5G5R5:
     case TextureFormat::B5G6R5:
@@ -92,6 +93,7 @@ std::vector<u8> UnswizzleTexture(VAddr address, TextureFormat format, u32 width,
     case TextureFormat::B5G6R5:
     case TextureFormat::R8:
     case TextureFormat::R16_G16_B16_A16:
+    case TextureFormat::BF10GF11RF11:
         CopySwizzledData(width, height, bytes_per_pixel, bytes_per_pixel, data,
                          unswizzled_data.data(), true, block_height);
         break;
@@ -118,6 +120,7 @@ std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat
     case TextureFormat::A1B5G5R5:
     case TextureFormat::B5G6R5:
     case TextureFormat::R8:
+    case TextureFormat::BF10GF11RF11:
         // TODO(Subv): For the time being just forward the same data without any decoding.
         rgba_data = texture_data;
         break;