3 files changed, 32 insertions, 7 deletions

diff --git a/src/video_core/textures/astc.cpp b/src/video_core/textures/astc.cpp
index b508d64e9..a9b8f69af 100644
--- a/src/video_core/textures/astc.cpp
+++ b/src/video_core/textures/astc.cpp
@@ -25,8 +25,8 @@
 
 class InputBitStream {
 public:
-    explicit InputBitStream(const unsigned char* ptr, int nBits = 0, int start_offset = 0)
-        : m_NumBits(nBits), m_CurByte(ptr), m_NextBit(start_offset % 8) {}
+    explicit InputBitStream(const unsigned char* ptr, int start_offset = 0)
+        : m_CurByte(ptr), m_NextBit(start_offset % 8) {}
 
     ~InputBitStream() = default;
 
@@ -55,12 +55,9 @@ public:
     }
 
 private:
-    const int m_NumBits;
     const unsigned char* m_CurByte;
     int m_NextBit = 0;
     int m_BitsRead = 0;
-
-    bool done = false;
 };
 
 class OutputBitStream {
@@ -114,7 +111,6 @@ private:
     const int m_NumBits;
     unsigned char* m_CurByte;
     int m_NextBit = 0;
-    int m_BitsRead = 0;
 
     bool done = false;
 };
@@ -1616,6 +1612,7 @@ namespace Tegra::Texture::ASTC {
 std::vector<uint8_t> Decompress(const uint8_t* data, uint32_t width, uint32_t height,
                                 uint32_t depth, uint32_t block_width, uint32_t block_height) {
     uint32_t blockIdx = 0;
+    std::size_t depth_offset = 0;
     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) {
@@ -1630,7 +1627,7 @@ std::vector<uint8_t> Decompress(const uint8_t* data, uint32_t width, uint32_t he
                 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;
+                uint8_t* outRow = depth_offset + 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);
                 }
@@ -1638,6 +1635,7 @@ std::vector<uint8_t> Decompress(const uint8_t* data, uint32_t width, uint32_t he
                 blockIdx++;
             }
         }
+        depth_offset += height * width * 4;
     }
 
     return outData;
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 995d0e068..217805386 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -288,6 +288,29 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
     }
 }
 
+void SwizzleKepler(const u32 width, const u32 height, const u32 dst_x, const u32 dst_y,
+                   const u32 block_height, const std::size_t copy_size, const u8* source_data,
+                   u8* swizzle_data) {
+    const u32 image_width_in_gobs{(width + gob_size_x - 1) / gob_size_x};
+    std::size_t count = 0;
+    for (std::size_t y = dst_y; y < height && count < copy_size; ++y) {
+        const std::size_t gob_address_y =
+            (y / (gob_size_y * block_height)) * gob_size * block_height * image_width_in_gobs +
+            ((y % (gob_size_y * block_height)) / gob_size_y) * gob_size;
+        const auto& table = legacy_swizzle_table[y % gob_size_y];
+        for (std::size_t x = dst_x; x < width && count < copy_size; ++x) {
+            const std::size_t gob_address =
+                gob_address_y + (x / gob_size_x) * gob_size * block_height;
+            const std::size_t swizzled_offset = gob_address + table[x % gob_size_x];
+            const u8* source_line = source_data + count;
+            u8* dest_addr = swizzle_data + swizzled_offset;
+            count++;
+
+            std::memcpy(dest_addr, source_line, 1);
+        }
+    }
+}
+
 std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat format, u32 width,
                               u32 height) {
     std::vector<u8> rgba_data;
diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h
index e078fa274..e072d8401 100644
--- a/src/video_core/textures/decoders.h
+++ b/src/video_core/textures/decoders.h
@@ -51,4 +51,8 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
                       u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, u32 block_height,
                       u32 offset_x, u32 offset_y);
 
+void SwizzleKepler(const u32 width, const u32 height, const u32 dst_x, const u32 dst_y,
+                   const u32 block_height, const std::size_t copy_size, const u8* source_data,
+                   u8* swizzle_data);
+
 } // namespace Tegra::Texture