Merge pull request #497 from Subv/dxn1

GPU: Implemented the DXN1 (BC4) texture format.

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -53,6 +53,7 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {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
+    {GL_COMPRESSED_RED_RGTC1, GL_RED, GL_UNSIGNED_INT_8_8_8_8, true},           // DXN1
 }};
 
 static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType component_type) {
@@ -113,7 +114,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
         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::DXT45>,
+        MortonCopy<true, PixelFormat::DXT45>,       MortonCopy<true, PixelFormat::DXN1>,
 };
 
 static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra::GPUVAddr,
@@ -126,7 +127,8 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
         MortonCopy<false, PixelFormat::A1B5G5R5>,
         MortonCopy<false, PixelFormat::R8>,
         MortonCopy<false, PixelFormat::RGBA16F>,
-        // TODO(Subv): Swizzling the DXT1/DXT23/DXT45 formats is not yet supported
+        // TODO(Subv): Swizzling the DXT1/DXT23/DXT45/DXN1 formats is not yet supported
+        nullptr,
         nullptr,
         nullptr,
         nullptr,

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -63,6 +63,7 @@ struct SurfaceParams {
         DXT1 = 6,
         DXT23 = 7,
         DXT45 = 8,
+        DXN1 = 9, // This is also known as BC4
 
         Max,
         Invalid = 255,
@@ -107,6 +108,7 @@ struct SurfaceParams {
             4, // DXT1
             4, // DXT23
             4, // DXT45
+            4, // DXN1
         }};
 
         ASSERT(static_cast<size_t>(format) < compression_factor_table.size());
@@ -130,6 +132,7 @@ struct SurfaceParams {
             64,  // DXT1
             128, // DXT23
             128, // DXT45
+            64,  // DXN1
         }};
 
         ASSERT(static_cast<size_t>(format) < bpp_table.size());
@@ -185,6 +188,8 @@ struct SurfaceParams {
             return PixelFormat::DXT23;
         case Tegra::Texture::TextureFormat::DXT45:
             return PixelFormat::DXT45;
+        case Tegra::Texture::TextureFormat::DXN1:
+            return PixelFormat::DXN1;
         default:
             NGLOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
@@ -212,6 +217,8 @@ struct SurfaceParams {
             return Tegra::Texture::TextureFormat::DXT23;
         case PixelFormat::DXT45:
             return Tegra::Texture::TextureFormat::DXT45;
+        case PixelFormat::DXN1:
+            return Tegra::Texture::TextureFormat::DXN1;
         default:
             UNREACHABLE();
         }

src/video_core/textures/decoders.cpp

@@ -46,6 +46,7 @@ void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_
 u32 BytesPerPixel(TextureFormat format) {
     switch (format) {
     case TextureFormat::DXT1:
+    case TextureFormat::DXN1:
         // In this case a 'pixel' actually refers to a 4x4 tile.
         return 8;
     case TextureFormat::DXT23:
@@ -79,7 +80,9 @@ std::vector<u8> UnswizzleTexture(VAddr address, TextureFormat format, u32 width,
     case TextureFormat::DXT1:
     case TextureFormat::DXT23:
     case TextureFormat::DXT45:
-        // In the DXT formats, each 4x4 tile is swizzled instead of just individual pixel values.
+    case TextureFormat::DXN1:
+        // In the DXT and DXN formats, each 4x4 tile is swizzled instead of just individual pixel
+        // values.
         CopySwizzledData(width / 4, height / 4, bytes_per_pixel, bytes_per_pixel, data,
                          unswizzled_data.data(), true, block_height);
         break;
@@ -109,6 +112,7 @@ std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat
     case TextureFormat::DXT1:
     case TextureFormat::DXT23:
     case TextureFormat::DXT45:
+    case TextureFormat::DXN1:
     case TextureFormat::A8R8G8B8:
     case TextureFormat::A2B10G10R10:
     case TextureFormat::A1B5G5R5: