GPU: Added a method to unswizzle a texture without decoding it.

Allow unswizzling of DXT1 textures.

4 files changed, 95 insertions, 5 deletions

diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index ca1b150a7..d1edfe09a 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -196,7 +196,7 @@ void Maxwell3D::DrawArrays() {
 
             auto format = tic_entry.format.Value();
 
-            auto texture = Texture::DecodeTexture(
+            auto texture = Texture::UnswizzleTexture(
                 memory_manager.PhysicalToVirtualAddress(tic_entry.Address()),
                 tic_entry.format.Value(), tic_entry.Width(), tic_entry.Height());
 
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 705e2e066..300267209 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -2,13 +2,93 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <cstring>
+#include "common/assert.h"
 #include "video_core/textures/decoders.h"
+#include "video_core/textures/texture.h"
 
 namespace Tegra {
 namespace Texture {
 
-std::vector<u8> DecodeTexture(VAddr address, TextureFormat format, u32 width, u32 height) {
-    return {};
+/**
+ * Calculates the offset of an (x, y) position within a swizzled texture.
+ * Taken from the Tegra X1 TRM.
+ */
+static u32 GetSwizzleOffset(u32 x, u32 y, u32 image_width, u32 bytes_per_pixel, u32 block_height) {
+    u32 image_width_in_gobs = image_width * bytes_per_pixel / 64;
+    u32 GOB_address = 0 + (y / (8 * block_height)) * 512 * block_height * image_width_in_gobs +
+                      (x * bytes_per_pixel / 64) * 512 * block_height +
+                      (y % (8 * block_height) / 8) * 512;
+    x *= bytes_per_pixel;
+    u32 address = GOB_address + ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
+                  (y % 2) * 16 + (x % 16);
+
+    return address;
 }
+
+static void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel,
+                             u8* swizzled_data, u8* unswizzled_data, bool unswizzle,
+                             u32 block_height) {
+    u8* data_ptrs[2];
+    for (unsigned y = 0; y < height; ++y) {
+        for (unsigned x = 0; x < width; ++x) {
+            u32 swizzle_offset = GetSwizzleOffset(x, y, width, bytes_per_pixel, block_height);
+            u32 pixel_index = (x + y * width) * out_bytes_per_pixel;
+
+            data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
+            data_ptrs[!unswizzle] = &unswizzled_data[pixel_index];
+
+            std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
+        }
+    }
 }
+
+u32 BytesPerPixel(TextureFormat format) {
+    switch (format) {
+    case TextureFormat::DXT1:
+        // In this case a 'pixel' actually refers to a 4x4 tile.
+        return 8;
+    default:
+        UNIMPLEMENTED_MSG("Format not implemented");
+        break;
+    }
+}
+
+std::vector<u8> UnswizzleTexture(VAddr address, TextureFormat format, u32 width, u32 height) {
+    u8* data = Memory::GetPointer(address);
+    u32 bytes_per_pixel = BytesPerPixel(format);
+
+    static constexpr u32 DefaultBlockHeight = 16;
+
+    std::vector<u8> unswizzled_data(width * height * bytes_per_pixel);
+
+    switch (format) {
+    case TextureFormat::DXT1:
+        // In the DXT1 format, 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, DefaultBlockHeight);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Format not implemented");
+        break;
+    }
+
+    return unswizzled_data;
+}
+
+std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat format, u32 width,
+                              u32 height) {
+    std::vector<u8> rgba_data;
+
+    // TODO(Subv): Implement.
+    switch (format) {
+    default:
+        UNIMPLEMENTED_MSG("Format not implemented");
+        break;
+    }
+
+    return rgba_data;
+}
+
+} // namespace Texture
 } // namespace Tegra
diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h
index e0d55600e..0c21694ff 100644
--- a/src/video_core/textures/decoders.h
+++ b/src/video_core/textures/decoders.h
@@ -12,9 +12,15 @@ namespace Tegra {
 namespace Texture {
 
 /**
- * Decodes a swizzled texture into a RGBA8888 texture.
+ * Unswizzles a swizzled texture without changing its format.
  */
-std::vector<u8> DecodeTexture(VAddr address, TextureFormat format, u32 width, u32 height);
+std::vector<u8> UnswizzleTexture(VAddr address, TextureFormat format, u32 width, u32 height);
+
+/**
+ * Decodes an unswizzled texture into a A8R8G8B8 texture.
+ */
+std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat format, u32 width,
+                              u32 height);
 
 } // namespace Texture
 } // namespace Tegra
diff --git a/src/video_core/textures/texture.h b/src/video_core/textures/texture.h
index 3306d2ab2..d969bcdd9 100644
--- a/src/video_core/textures/texture.h
+++ b/src/video_core/textures/texture.h
@@ -13,6 +13,7 @@ namespace Tegra {
 namespace Texture {
 
 enum class TextureFormat : u32 {
+    A8R8G8B8 = 8,
     DXT1 = 0x24,
 };
 
@@ -53,5 +54,8 @@ struct TICEntry {
 };
 static_assert(sizeof(TICEntry) == 0x20, "TICEntry has wrong size");
 
+/// Returns the number of bytes per pixel of the input texture format.
+u32 BytesPerPixel(TextureFormat format);
+
 } // namespace Texture
 } // namespace Tegra