Merge pull request #491 from bunnei/rgba16f

gl_rasterizer_cache: Implement PixelFormat RGBA16F.

3 files changed, 24 insertions, 7 deletions

diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index ac3e4bf27..10a460a82 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -49,6 +49,7 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {GL_RGB10_A2, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, false},              // A2B10G10R10
     {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_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
@@ -58,8 +59,8 @@ 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
-        ASSERT(component_type == ComponentType::UNorm);
+        // For now only UNORM components are supported, or RGBA16F which is type FLOAT
+        ASSERT(component_type == ComponentType::UNorm || pixel_format == PixelFormat::RGBA16F);
         return tex_format_tuples[static_cast<unsigned int>(pixel_format)];
     } else if (type == SurfaceType::Depth || type == SurfaceType::DepthStencil) {
         // TODO(Subv): Implement depth formats
@@ -110,8 +111,9 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
     morton_to_gl_fns = {
         MortonCopy<true, PixelFormat::ABGR8>,       MortonCopy<true, PixelFormat::B5G6R5>,
         MortonCopy<true, PixelFormat::A2B10G10R10>, MortonCopy<true, PixelFormat::A1B5G5R5>,
-        MortonCopy<true, PixelFormat::R8>,          MortonCopy<true, PixelFormat::DXT1>,
-        MortonCopy<true, PixelFormat::DXT23>,       MortonCopy<true, PixelFormat::DXT45>,
+        MortonCopy<true, PixelFormat::R8>,          MortonCopy<true, PixelFormat::RGBA16F>,
+        MortonCopy<true, PixelFormat::DXT1>,        MortonCopy<true, PixelFormat::DXT23>,
+        MortonCopy<true, PixelFormat::DXT45>,
 };
 
 static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra::GPUVAddr,
@@ -123,6 +125,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr, Tegra:
         MortonCopy<false, PixelFormat::A2B10G10R10>,
         MortonCopy<false, PixelFormat::A1B5G5R5>,
         MortonCopy<false, PixelFormat::R8>,
+        MortonCopy<false, PixelFormat::RGBA16F>,
         // TODO(Subv): Swizzling the DXT1/DXT23/DXT45 formats is not yet supported
         nullptr,
         nullptr,
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index fc09f108c..07461017d 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -59,9 +59,10 @@ struct SurfaceParams {
         A2B10G10R10 = 2,
         A1B5G5R5 = 3,
         R8 = 4,
-        DXT1 = 5,
-        DXT23 = 6,
-        DXT45 = 7,
+        RGBA16F = 5,
+        DXT1 = 6,
+        DXT23 = 7,
+        DXT45 = 8,
 
         Max,
         Invalid = 255,
@@ -102,6 +103,7 @@ struct SurfaceParams {
             1, // A2B10G10R10
             1, // A1B5G5R5
             1, // R8
+            2, // RGBA16F
             4, // DXT1
             4, // DXT23
             4, // DXT45
@@ -124,6 +126,7 @@ struct SurfaceParams {
             32,  // A2B10G10R10
             16,  // A1B5G5R5
             8,   // R8
+            64,  // RGBA16F
             64,  // DXT1
             128, // DXT23
             128, // DXT45
@@ -143,6 +146,8 @@ struct SurfaceParams {
             return PixelFormat::ABGR8;
         case Tegra::RenderTargetFormat::RGB10_A2_UNORM:
             return PixelFormat::A2B10G10R10;
+        case Tegra::RenderTargetFormat::RGBA16_FLOAT:
+            return PixelFormat::RGBA16F;
         default:
             NGLOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
@@ -172,6 +177,8 @@ struct SurfaceParams {
             return PixelFormat::A1B5G5R5;
         case Tegra::Texture::TextureFormat::R8:
             return PixelFormat::R8;
+        case Tegra::Texture::TextureFormat::R16_G16_B16_A16:
+            return PixelFormat::RGBA16F;
         case Tegra::Texture::TextureFormat::DXT1:
             return PixelFormat::DXT1;
         case Tegra::Texture::TextureFormat::DXT23:
@@ -197,6 +204,8 @@ struct SurfaceParams {
             return Tegra::Texture::TextureFormat::A1B5G5R5;
         case PixelFormat::R8:
             return Tegra::Texture::TextureFormat::R8;
+        case PixelFormat::RGBA16F:
+            return Tegra::Texture::TextureFormat::R16_G16_B16_A16;
         case PixelFormat::DXT1:
             return Tegra::Texture::TextureFormat::DXT1;
         case PixelFormat::DXT23:
@@ -226,6 +235,8 @@ struct SurfaceParams {
         case Tegra::RenderTargetFormat::RGBA8_SRGB:
         case Tegra::RenderTargetFormat::RGB10_A2_UNORM:
             return ComponentType::UNorm;
+        case Tegra::RenderTargetFormat::RGBA16_FLOAT:
+            return ComponentType::Float;
         default:
             NGLOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 9d7b73b73..0179663e8 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -60,6 +60,8 @@ u32 BytesPerPixel(TextureFormat format) {
         return 2;
     case TextureFormat::R8:
         return 1;
+    case TextureFormat::R16_G16_B16_A16:
+        return 8;
     default:
         UNIMPLEMENTED_MSG("Format not implemented");
         break;
@@ -86,6 +88,7 @@ std::vector<u8> UnswizzleTexture(VAddr address, TextureFormat format, u32 width,
     case TextureFormat::A1B5G5R5:
     case TextureFormat::B5G6R5:
     case TextureFormat::R8:
+    case TextureFormat::R16_G16_B16_A16:
         CopySwizzledData(width, height, bytes_per_pixel, bytes_per_pixel, data,
                          unswizzled_data.data(), true, block_height);
         break;