Merge pull request #1625 from FernandoS27/astc

Implement ASTC Textures 5x5 and fix a bunch of ASTC texture problems

9 files changed, 154 insertions, 75 deletions

diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index f194a7687..0e11557c4 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -58,16 +58,14 @@ void SurfaceParams::InitCacheParameters(Tegra::GPUVAddr gpu_addr_) {
 
 std::size_t SurfaceParams::InnerMipmapMemorySize(u32 mip_level, bool force_gl, bool layer_only,
                                                  bool uncompressed) const {
-    const u32 compression_factor{GetCompressionFactor(pixel_format)};
+    const u32 tile_x{GetDefaultBlockWidth(pixel_format)};
+    const u32 tile_y{GetDefaultBlockHeight(pixel_format)};
     const u32 bytes_per_pixel{GetBytesPerPixel(pixel_format)};
     u32 m_depth = (layer_only ? 1U : depth);
     u32 m_width = MipWidth(mip_level);
     u32 m_height = MipHeight(mip_level);
-    m_width = uncompressed ? m_width
-                           : std::max(1U, (m_width + compression_factor - 1) / compression_factor);
-    m_height = uncompressed
-                   ? m_height
-                   : std::max(1U, (m_height + compression_factor - 1) / compression_factor);
+    m_width = uncompressed ? m_width : std::max(1U, (m_width + tile_x - 1) / tile_x);
+    m_height = uncompressed ? m_height : std::max(1U, (m_height + tile_y - 1) / tile_y);
     m_depth = std::max(1U, m_depth >> mip_level);
     u32 m_block_height = MipBlockHeight(mip_level);
     u32 m_block_depth = MipBlockDepth(mip_level);
@@ -305,6 +303,8 @@ static constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> tex
     {GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_8X8_SRGB
     {GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_8X5_SRGB
     {GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_5X4_SRGB
+    {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false},        // ASTC_2D_5X5
+    {GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_5X5_SRGB
 
     // Depth formats
     {GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
@@ -364,15 +364,18 @@ void MortonCopy(u32 stride, u32 block_height, u32 height, u32 block_depth, u32 d
 
     // With the BCn formats (DXT and DXN), each 4x4 tile is swizzled instead of just individual
     // pixel values.
-    const u32 tile_size{IsFormatBCn(format) ? 4U : 1U};
+    const u32 tile_size_x{GetDefaultBlockWidth(format)};
+    const u32 tile_size_y{GetDefaultBlockHeight(format)};
 
     if (morton_to_gl) {
-        const std::vector<u8> data = Tegra::Texture::UnswizzleTexture(
-            addr, tile_size, bytes_per_pixel, stride, height, depth, block_height, block_depth);
+        const std::vector<u8> data =
+            Tegra::Texture::UnswizzleTexture(addr, tile_size_x, tile_size_y, bytes_per_pixel,
+                                             stride, height, depth, block_height, block_depth);
         const std::size_t size_to_copy{std::min(gl_buffer_size, data.size())};
         memcpy(gl_buffer, data.data(), size_to_copy);
     } else {
-        Tegra::Texture::CopySwizzledData(stride / tile_size, height / tile_size, depth,
+        Tegra::Texture::CopySwizzledData((stride + tile_size_x - 1) / tile_size_x,
+                                         (height + tile_size_y - 1) / tile_size_y, depth,
                                          bytes_per_pixel, bytes_per_pixel, Memory::GetPointer(addr),
                                          gl_buffer, false, block_height, block_depth);
     }
@@ -440,6 +443,8 @@ static constexpr GLConversionArray morton_to_gl_fns = {
         MortonCopy<true, PixelFormat::ASTC_2D_8X8_SRGB>,
         MortonCopy<true, PixelFormat::ASTC_2D_8X5_SRGB>,
         MortonCopy<true, PixelFormat::ASTC_2D_5X4_SRGB>,
+        MortonCopy<true, PixelFormat::ASTC_2D_5X5>,
+        MortonCopy<true, PixelFormat::ASTC_2D_5X5_SRGB>,
         MortonCopy<true, PixelFormat::Z32F>,
         MortonCopy<true, PixelFormat::Z16>,
         MortonCopy<true, PixelFormat::Z24S8>,
@@ -508,6 +513,8 @@ static constexpr GLConversionArray gl_to_morton_fns = {
         nullptr,
         nullptr,
         nullptr,
+        nullptr,
+        nullptr,
         MortonCopy<false, PixelFormat::Z32F>,
         MortonCopy<false, PixelFormat::Z16>,
         MortonCopy<false, PixelFormat::Z24S8>,
@@ -897,21 +904,24 @@ static void ConvertG8R8ToR8G8(std::vector<u8>& data, u32 width, u32 height) {
  * typical desktop GPUs.
  */
 static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelFormat pixel_format,
-                                               u32 width, u32 height) {
+                                               u32 width, u32 height, u32 depth) {
     switch (pixel_format) {
     case PixelFormat::ASTC_2D_4X4:
     case PixelFormat::ASTC_2D_8X8:
     case PixelFormat::ASTC_2D_8X5:
     case PixelFormat::ASTC_2D_5X4:
+    case PixelFormat::ASTC_2D_5X5:
     case PixelFormat::ASTC_2D_4X4_SRGB:
     case PixelFormat::ASTC_2D_8X8_SRGB:
     case PixelFormat::ASTC_2D_8X5_SRGB:
-    case PixelFormat::ASTC_2D_5X4_SRGB: {
+    case PixelFormat::ASTC_2D_5X4_SRGB:
+    case PixelFormat::ASTC_2D_5X5_SRGB: {
         // Convert ASTC pixel formats to RGBA8, as most desktop GPUs do not support ASTC.
         u32 block_width{};
         u32 block_height{};
         std::tie(block_width, block_height) = GetASTCBlockSize(pixel_format);
-        data = Tegra::Texture::ASTC::Decompress(data, width, height, block_width, block_height);
+        data =
+            Tegra::Texture::ASTC::Decompress(data, width, height, depth, block_width, block_height);
         break;
     }
     case PixelFormat::S8Z24:
@@ -971,7 +981,7 @@ void CachedSurface::LoadGLBuffer() {
     }
     for (u32 i = 0; i < params.max_mip_level; i++)
         ConvertFormatAsNeeded_LoadGLBuffer(gl_buffer[i], params.pixel_format, params.MipWidth(i),
-                                           params.MipHeight(i));
+                                           params.MipHeight(i), params.MipDepth(i));
 }
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64));
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index f255f4419..46ad37897 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -139,7 +139,7 @@ struct SurfaceParams {
     }
 
     u32 MipDepth(u32 mip_level) const {
-        return std::max(1U, depth >> mip_level);
+        return is_layered ? depth : std::max(1U, depth >> mip_level);
     }
 
     // Auto block resizing algorithm from:
diff --git a/src/video_core/surface.cpp b/src/video_core/surface.cpp
index d9a97e30b..e284a4604 100644
--- a/src/video_core/surface.cpp
+++ b/src/video_core/surface.cpp
@@ -297,6 +297,8 @@ PixelFormat PixelFormatFromTextureFormat(Tegra::Texture::TextureFormat format,
         return is_srgb ? PixelFormat::ASTC_2D_4X4_SRGB : PixelFormat::ASTC_2D_4X4;
     case Tegra::Texture::TextureFormat::ASTC_2D_5X4:
         return is_srgb ? PixelFormat::ASTC_2D_5X4_SRGB : PixelFormat::ASTC_2D_5X4;
+    case Tegra::Texture::TextureFormat::ASTC_2D_5X5:
+        return is_srgb ? PixelFormat::ASTC_2D_5X5_SRGB : PixelFormat::ASTC_2D_5X5;
     case Tegra::Texture::TextureFormat::ASTC_2D_8X8:
         return is_srgb ? PixelFormat::ASTC_2D_8X8_SRGB : PixelFormat::ASTC_2D_8X8;
     case Tegra::Texture::TextureFormat::ASTC_2D_8X5:
@@ -440,10 +442,12 @@ bool IsPixelFormatASTC(PixelFormat format) {
     switch (format) {
     case PixelFormat::ASTC_2D_4X4:
     case PixelFormat::ASTC_2D_5X4:
+    case PixelFormat::ASTC_2D_5X5:
     case PixelFormat::ASTC_2D_8X8:
     case PixelFormat::ASTC_2D_8X5:
     case PixelFormat::ASTC_2D_4X4_SRGB:
     case PixelFormat::ASTC_2D_5X4_SRGB:
+    case PixelFormat::ASTC_2D_5X5_SRGB:
     case PixelFormat::ASTC_2D_8X8_SRGB:
     case PixelFormat::ASTC_2D_8X5_SRGB:
         return true;
@@ -453,27 +457,7 @@ bool IsPixelFormatASTC(PixelFormat format) {
 }
 
 std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) {
-    switch (format) {
-    case PixelFormat::ASTC_2D_4X4:
-        return {4, 4};
-    case PixelFormat::ASTC_2D_5X4:
-        return {5, 4};
-    case PixelFormat::ASTC_2D_8X8:
-        return {8, 8};
-    case PixelFormat::ASTC_2D_8X5:
-        return {8, 5};
-    case PixelFormat::ASTC_2D_4X4_SRGB:
-        return {4, 4};
-    case PixelFormat::ASTC_2D_5X4_SRGB:
-        return {5, 4};
-    case PixelFormat::ASTC_2D_8X8_SRGB:
-        return {8, 8};
-    case PixelFormat::ASTC_2D_8X5_SRGB:
-        return {8, 5};
-    default:
-        LOG_CRITICAL(HW_GPU, "Unhandled format: {}", static_cast<u32>(format));
-        UNREACHABLE();
-    }
+    return {GetDefaultBlockWidth(format), GetDefaultBlockHeight(format)};
 }
 
 bool IsFormatBCn(PixelFormat format) {
diff --git a/src/video_core/surface.h b/src/video_core/surface.h
index 3232e437f..0ef7849a4 100644
--- a/src/video_core/surface.h
+++ b/src/video_core/surface.h
@@ -72,19 +72,21 @@ enum class PixelFormat {
     ASTC_2D_8X8_SRGB = 54,
     ASTC_2D_8X5_SRGB = 55,
     ASTC_2D_5X4_SRGB = 56,
+    ASTC_2D_5X5 = 57,
+    ASTC_2D_5X5_SRGB = 58,
 
     MaxColorFormat,
 
     // Depth formats
-    Z32F = 57,
-    Z16 = 58,
+    Z32F = 59,
+    Z16 = 60,
 
     MaxDepthFormat,
 
     // DepthStencil formats
-    Z24S8 = 59,
-    S8Z24 = 60,
-    Z32FS8 = 61,
+    Z24S8 = 61,
+    S8Z24 = 62,
+    Z32FS8 = 63,
 
     MaxDepthStencilFormat,
 
@@ -188,6 +190,8 @@ static constexpr u32 GetCompressionFactor(PixelFormat format) {
         4, // ASTC_2D_8X8_SRGB
         4, // ASTC_2D_8X5_SRGB
         4, // ASTC_2D_5X4_SRGB
+        4, // ASTC_2D_5X5
+        4, // ASTC_2D_5X5_SRGB
         1, // Z32F
         1, // Z16
         1, // Z24S8
@@ -199,6 +203,79 @@ static constexpr u32 GetCompressionFactor(PixelFormat format) {
     return compression_factor_table[static_cast<std::size_t>(format)];
 }
 
+static constexpr u32 GetDefaultBlockWidth(PixelFormat format) {
+    if (format == PixelFormat::Invalid)
+        return 0;
+    constexpr std::array<u32, MaxPixelFormat> block_width_table = {{
+        1, // ABGR8U
+        1, // ABGR8S
+        1, // ABGR8UI
+        1, // B5G6R5U
+        1, // A2B10G10R10U
+        1, // A1B5G5R5U
+        1, // R8U
+        1, // R8UI
+        1, // RGBA16F
+        1, // RGBA16U
+        1, // RGBA16UI
+        1, // R11FG11FB10F
+        1, // RGBA32UI
+        4, // DXT1
+        4, // DXT23
+        4, // DXT45
+        4, // DXN1
+        4, // DXN2UNORM
+        4, // DXN2SNORM
+        4, // BC7U
+        4, // BC6H_UF16
+        4, // BC6H_SF16
+        4, // ASTC_2D_4X4
+        1, // G8R8U
+        1, // G8R8S
+        1, // BGRA8
+        1, // RGBA32F
+        1, // RG32F
+        1, // R32F
+        1, // R16F
+        1, // R16U
+        1, // R16S
+        1, // R16UI
+        1, // R16I
+        1, // RG16
+        1, // RG16F
+        1, // RG16UI
+        1, // RG16I
+        1, // RG16S
+        1, // RGB32F
+        1, // RGBA8_SRGB
+        1, // RG8U
+        1, // RG8S
+        1, // RG32UI
+        1, // R32UI
+        8, // ASTC_2D_8X8
+        8, // ASTC_2D_8X5
+        5, // ASTC_2D_5X4
+        1, // BGRA8_SRGB
+        4, // DXT1_SRGB
+        4, // DXT23_SRGB
+        4, // DXT45_SRGB
+        4, // BC7U_SRGB
+        4, // ASTC_2D_4X4_SRGB
+        8, // ASTC_2D_8X8_SRGB
+        8, // ASTC_2D_8X5_SRGB
+        5, // ASTC_2D_5X4_SRGB
+        5, // ASTC_2D_5X5
+        5, // ASTC_2D_5X5_SRGB
+        1, // Z32F
+        1, // Z16
+        1, // Z24S8
+        1, // S8Z24
+        1, // Z32FS8
+    }};
+    ASSERT(static_cast<std::size_t>(format) < block_width_table.size());
+    return block_width_table[static_cast<std::size_t>(format)];
+}
+
 static constexpr u32 GetDefaultBlockHeight(PixelFormat format) {
     if (format == PixelFormat::Invalid)
         return 0;
@@ -261,6 +338,8 @@ static constexpr u32 GetDefaultBlockHeight(PixelFormat format) {
         8, // ASTC_2D_8X8_SRGB
         5, // ASTC_2D_8X5_SRGB
         4, // ASTC_2D_5X4_SRGB
+        5, // ASTC_2D_5X5
+        5, // ASTC_2D_5X5_SRGB
         1, // Z32F
         1, // Z16
         1, // Z24S8
@@ -299,7 +378,7 @@ static constexpr u32 GetFormatBpp(PixelFormat format) {
         128, // BC7U
         128, // BC6H_UF16
         128, // BC6H_SF16
-        32,  // ASTC_2D_4X4
+        128, // ASTC_2D_4X4
         16,  // G8R8U
         16,  // G8R8S
         32,  // BGRA8
@@ -322,18 +401,20 @@ static constexpr u32 GetFormatBpp(PixelFormat format) {
         16,  // RG8S
         64,  // RG32UI
         32,  // R32UI
-        16,  // ASTC_2D_8X8
-        16,  // ASTC_2D_8X5
-        32,  // ASTC_2D_5X4
+        128, // ASTC_2D_8X8
+        128, // ASTC_2D_8X5
+        128, // ASTC_2D_5X4
         32,  // BGRA8_SRGB
         64,  // DXT1_SRGB
         128, // DXT23_SRGB
         128, // DXT45_SRGB
         128, // BC7U
-        32,  // ASTC_2D_4X4_SRGB
-        16,  // ASTC_2D_8X8_SRGB
-        16,  // ASTC_2D_8X5_SRGB
-        32,  // ASTC_2D_5X4_SRGB
+        128, // ASTC_2D_4X4_SRGB
+        128, // ASTC_2D_8X8_SRGB
+        128, // ASTC_2D_8X5_SRGB
+        128, // ASTC_2D_5X4_SRGB
+        128, // ASTC_2D_5X5
+        128, // ASTC_2D_5X5_SRGB
         32,  // Z32F
         16,  // Z16
         32,  // Z24S8
diff --git a/src/video_core/textures/astc.cpp b/src/video_core/textures/astc.cpp
index b1feacae9..bc50a4876 100644
--- a/src/video_core/textures/astc.cpp
+++ b/src/video_core/textures/astc.cpp
@@ -1598,27 +1598,29 @@ static void DecompressBlock(uint8_t inBuf[16], const uint32_t blockWidth,
 namespace Tegra::Texture::ASTC {
 
 std::vector<uint8_t> Decompress(std::vector<uint8_t>& data, uint32_t width, uint32_t height,
-                                uint32_t block_width, uint32_t block_height) {
+                                uint32_t depth, uint32_t block_width, uint32_t block_height) {
     uint32_t blockIdx = 0;
-    std::vector<uint8_t> outData(height * width * 4);
-    for (uint32_t j = 0; j < height; j += block_height) {
-        for (uint32_t i = 0; i < width; i += block_width) {
+    std::vector<uint8_t> outData(height * width * depth * 4);
+    for (uint32_t k = 0; k < depth; k++) {
+        for (uint32_t j = 0; j < height; j += block_height) {
+            for (uint32_t i = 0; i < width; i += block_width) {
 
-            uint8_t* blockPtr = data.data() + blockIdx * 16;
+                uint8_t* blockPtr = data.data() + blockIdx * 16;
 
-            // Blocks can be at most 12x12
-            uint32_t uncompData[144];
-            ASTCC::DecompressBlock(blockPtr, block_width, block_height, uncompData);
+                // Blocks can be at most 12x12
+                uint32_t uncompData[144];
+                ASTCC::DecompressBlock(blockPtr, block_width, block_height, uncompData);
 
-            uint32_t decompWidth = std::min(block_width, width - i);
-            uint32_t decompHeight = std::min(block_height, height - j);
+                uint32_t decompWidth = std::min(block_width, width - i);
+                uint32_t decompHeight = std::min(block_height, height - j);
 
-            uint8_t* outRow = outData.data() + (j * width + i) * 4;
-            for (uint32_t jj = 0; jj < decompHeight; jj++) {
-                memcpy(outRow + jj * width * 4, uncompData + jj * block_width, decompWidth * 4);
-            }
+                uint8_t* outRow = outData.data() + (j * width + i) * 4;
+                for (uint32_t jj = 0; jj < decompHeight; jj++) {
+                    memcpy(outRow + jj * width * 4, uncompData + jj * block_width, decompWidth * 4);
+                }
 
-            blockIdx++;
+                blockIdx++;
+            }
         }
     }
 
diff --git a/src/video_core/textures/astc.h b/src/video_core/textures/astc.h
index f0d7c0e56..d419dd025 100644
--- a/src/video_core/textures/astc.h
+++ b/src/video_core/textures/astc.h
@@ -10,6 +10,6 @@
 namespace Tegra::Texture::ASTC {
 
 std::vector<uint8_t> Decompress(std::vector<uint8_t>& data, uint32_t width, uint32_t height,
-                                uint32_t block_width, uint32_t block_height);
+                                uint32_t depth, uint32_t block_width, uint32_t block_height);
 
 } // namespace Tegra::Texture::ASTC
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 550ca856c..3066abf61 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -227,12 +227,14 @@ u32 BytesPerPixel(TextureFormat format) {
     }
 }
 
-std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width,
-                                 u32 height, u32 depth, u32 block_height, u32 block_depth) {
+std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size_x, u32 tile_size_y,
+                                 u32 bytes_per_pixel, u32 width, u32 height, u32 depth,
+                                 u32 block_height, u32 block_depth) {
     std::vector<u8> unswizzled_data(width * height * depth * bytes_per_pixel);
-    CopySwizzledData(width / tile_size, height / tile_size, depth, bytes_per_pixel, bytes_per_pixel,
-                     Memory::GetPointer(address), unswizzled_data.data(), true, block_height,
-                     block_depth);
+    CopySwizzledData((width + tile_size_x - 1) / tile_size_x,
+                     (height + tile_size_y - 1) / tile_size_y, depth, bytes_per_pixel,
+                     bytes_per_pixel, Memory::GetPointer(address), unswizzled_data.data(), true,
+                     block_height, block_depth);
     return unswizzled_data;
 }
 
diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h
index b390219e4..ba065510b 100644
--- a/src/video_core/textures/decoders.h
+++ b/src/video_core/textures/decoders.h
@@ -19,8 +19,8 @@ inline std::size_t GetGOBSize() {
 /**
  * Unswizzles a swizzled texture without changing its format.
  */
-std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width,
-                                 u32 height, u32 depth,
+std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size_x, u32 tile_size_y,
+                                 u32 bytes_per_pixel, u32 width, u32 height, u32 depth,
                                  u32 block_height = TICEntry::DefaultBlockHeight,
                                  u32 block_depth = TICEntry::DefaultBlockHeight);
 
diff --git a/src/yuzu/debugger/graphics/graphics_surface.cpp b/src/yuzu/debugger/graphics/graphics_surface.cpp
index 0adbab27d..707747422 100644
--- a/src/yuzu/debugger/graphics/graphics_surface.cpp
+++ b/src/yuzu/debugger/graphics/graphics_surface.cpp
@@ -386,9 +386,9 @@ void GraphicsSurfaceWidget::OnUpdate() {
 
     // TODO(bunnei): Will not work with BCn formats that swizzle 4x4 tiles.
     // Needs to be fixed if we plan to use this feature more, otherwise we may remove it.
-    auto unswizzled_data =
-        Tegra::Texture::UnswizzleTexture(*address, 1, Tegra::Texture::BytesPerPixel(surface_format),
-                                         surface_width, surface_height, 1U);
+    auto unswizzled_data = Tegra::Texture::UnswizzleTexture(
+        *address, 1, 1, Tegra::Texture::BytesPerPixel(surface_format), surface_width,
+        surface_height, 1U);
 
     auto texture_data = Tegra::Texture::DecodeTexture(unswizzled_data, surface_format,
                                                       surface_width, surface_height);