From 75e12a33ae0479fed08d4f7dfe0ec95bf222084a Mon Sep 17 00:00:00 2001
From: Khangaroo <khang06@users.noreply.github.com>
Date: Thu, 9 Aug 2018 19:15:32 -0400
Subject: Implement SNORM for BC5/DXN2 (#998)

* Implement BC5/DXN2 (#996)

- Used by Kirby Star Allies.

* Implement BC5/DXN2 SNORM

UNORM for Kirby Star Allies
SNORM for Super Mario Odyssey
---
 .../renderer_opengl/gl_rasterizer_cache.cpp        | 29 ++++++----
 .../renderer_opengl/gl_rasterizer_cache.h          | 64 +++++++++++++---------
 2 files changed, 55 insertions(+), 38 deletions(-)

(limited to 'src/video_core/renderer_opengl')

diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index c447e999c..f6efce818 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -109,7 +109,9 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {GL_COMPRESSED_RGBA_S3TC_DXT5_EXT, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8, ComponentType::UNorm,
      true},                                                                                 // DXT45
     {GL_COMPRESSED_RED_RGTC1, GL_RED, GL_UNSIGNED_INT_8_8_8_8, ComponentType::UNorm, true}, // DXN1
-    {GL_COMPRESSED_RG_RGTC2, GL_RG, GL_UNSIGNED_INT_8_8_8_8, ComponentType::UNorm, true},   // DXN2
+    {GL_COMPRESSED_RG_RGTC2, GL_RG, GL_UNSIGNED_INT_8_8_8_8, ComponentType::UNorm,
+     true},                                                                     // DXN2UNORM
+    {GL_COMPRESSED_SIGNED_RG_RGTC2, GL_RG, GL_INT, ComponentType::SNorm, true}, // DXN2SNORM
     {GL_COMPRESSED_RGBA_BPTC_UNORM_ARB, GL_RGB, GL_UNSIGNED_INT_8_8_8_8, ComponentType::UNorm,
      true},                                                                    // BC7U
     {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false},        // ASTC_2D_4X4
@@ -219,17 +221,18 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
         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::DXN2>,         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::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::Z24S8>,
-        MortonCopy<true, PixelFormat::S8Z24>,        MortonCopy<true, PixelFormat::Z32F>,
-        MortonCopy<true, PixelFormat::Z16>,          MortonCopy<true, PixelFormat::Z32FS8>,
+        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::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::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, u8*, Tegra::GPUVAddr),
@@ -251,6 +254,8 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
         nullptr,
         nullptr,
         nullptr,
+        nullptr,
+        nullptr,
         MortonCopy<false, PixelFormat::G8R8>,
         MortonCopy<false, PixelFormat::BGRA8>,
         MortonCopy<false, PixelFormat::RGBA32F>,
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index 6f01b2bf0..26e2ee203 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -35,32 +35,33 @@ struct SurfaceParams {
         DXT23 = 9,
         DXT45 = 10,
         DXN1 = 11, // This is also known as BC4
-        DXN2 = 12, // This is also known as BC5
-        BC7U = 13,
-        ASTC_2D_4X4 = 14,
-        G8R8 = 15,
-        BGRA8 = 16,
-        RGBA32F = 17,
-        RG32F = 18,
-        R32F = 19,
-        R16F = 20,
-        R16UNORM = 21,
-        RG16 = 22,
-        RG16F = 23,
-        RG16UI = 24,
-        RG16I = 25,
-        RG16S = 26,
-        RGB32F = 27,
-        SRGBA8 = 28,
+        DXN2UNORM = 12,
+        DXN2SNORM = 13,
+        BC7U = 14,
+        ASTC_2D_4X4 = 15,
+        G8R8 = 16,
+        BGRA8 = 17,
+        RGBA32F = 18,
+        RG32F = 19,
+        R32F = 20,
+        R16F = 21,
+        R16UNORM = 22,
+        RG16 = 23,
+        RG16F = 24,
+        RG16UI = 25,
+        RG16I = 26,
+        RG16S = 27,
+        RGB32F = 28,
+        SRGBA8 = 29,
 
         MaxColorFormat,
 
         // DepthStencil formats
-        Z24S8 = 29,
-        S8Z24 = 30,
-        Z32F = 31,
-        Z16 = 32,
-        Z32FS8 = 33,
+        Z24S8 = 30,
+        S8Z24 = 31,
+        Z32F = 32,
+        Z16 = 33,
+        Z32FS8 = 34,
 
         MaxDepthStencilFormat,
 
@@ -110,7 +111,8 @@ struct SurfaceParams {
             4, // DXT23
             4, // DXT45
             4, // DXN1
-            4, // DXN2
+            4, // DXN2UNORM
+            4, // DXN2SNORM
             4, // BC7U
             4, // ASTC_2D_4X4
             1, // G8R8
@@ -155,7 +157,8 @@ struct SurfaceParams {
             128, // DXT23
             128, // DXT45
             64,  // DXN1
-            128, // DXN2
+            128, // DXN2UNORM
+            128, // DXN2SNORM
             128, // BC7U
             32,  // ASTC_2D_4X4
             16,  // G8R8
@@ -309,7 +312,15 @@ struct SurfaceParams {
         case Tegra::Texture::TextureFormat::DXN1:
             return PixelFormat::DXN1;
         case Tegra::Texture::TextureFormat::DXN2:
-            return PixelFormat::DXN2;
+            switch (component_type) {
+            case Tegra::Texture::ComponentType::UNORM:
+                return PixelFormat::DXN2UNORM;
+            case Tegra::Texture::ComponentType::SNORM:
+                return PixelFormat::DXN2SNORM;
+            }
+            LOG_CRITICAL(HW_GPU, "Unimplemented component_type={}",
+                         static_cast<u32>(component_type));
+            UNREACHABLE();
         case Tegra::Texture::TextureFormat::BC7U:
             return PixelFormat::BC7U;
         case Tegra::Texture::TextureFormat::ASTC_2D_4X4:
@@ -367,7 +378,8 @@ struct SurfaceParams {
             return Tegra::Texture::TextureFormat::DXT45;
         case PixelFormat::DXN1:
             return Tegra::Texture::TextureFormat::DXN1;
-        case PixelFormat::DXN2:
+        case PixelFormat::DXN2UNORM:
+        case PixelFormat::DXN2SNORM:
             return Tegra::Texture::TextureFormat::DXN2;
         case PixelFormat::BC7U:
             return Tegra::Texture::TextureFormat::BC7U;
-- 
cgit v1.2.3