TinyBCDec is a tiny library for decoding block compressed texture formats. It's written in pure Java (11+), without any dependencies, with a focus on speed and accuracy.

Currently, the following formats are supported:

• BC1 (DXT1)
• BC2 (DXT3)
• BC3 (DXT5)
• BC4 (ATI1)
• BC5 (ATI2)
• BC6H
• BC7

For ease of use, the library provides functionality to reorder the output channels, so the decoded data can be used directly AWT, JavaFX, or even OpenGL or Vulkan.

Using the library is very easy, all you need is the compressed data, and the format of the data.

The following features are present:

• Pixel order with alpha filling: If the pixel order contains an alpha channel, and the source does not, the alpha channel will be set to 255 or 1.0 in the case of BC6H.
• Partial decodes: The width and height do not need to be a multiple of the block size (4 in this case).
• Z reconstruction ("Blue" normal maps): BC5 uses two channels to store a normal map, the library can reconstruct the Z value from the two channels. Clamping is done to prevent negative values or values above 255.
• BC6H: BC6H is decoded to a Little-Endian Half-Float buffer. The library does not provide a way to convert this to full float. This can be done with Float.float16toFloat in Java 21, or with a library. This also means that the output buffer is twice as big.

The library provides the BlockDecoder class, which can be used to decode, by specifying the format and the pixel order. You can let the library create a new buffer, or pass an existing one to save allocations.

Given src is the compressed data, starting at offset srcPos, the following code snippet shows how to decode a BC1 texture.

import be.twofold.tinybcdec.*;

BlockDecoder decoder = BlockDecoder.create(BlockFormat.BC1);
byte[] result = decoder.decode(256, 256, src, srcPos);

If you want to pass an existing buffer, you can pass it as the last two arguments dst and dstPos. There will be no return value.

decoder.decode(256, 256, src, srcPos, dst, dstPos);

I've done some performance testing, and the library is quite fast. I've run some benchmarks on my machine (AMD 7840U).

Some quick benchmarks, tested on a Ryzen 5950X:

• BC1: ~1050MP/s
• BC2: ~650MP/s
• BC3: ~500MP/s
• BC4: ~550MP/s
• BC5: ~425MP/s
• BC6: ~130MP/s
• BC7: ~150MP/s

These numbers are just a rough estimate, and can vary depending on the hardware and the JVM.

To give you an idea, this means about 30ms to decode a 4K texture in BC1, and about 120ms for BC6 or BC7.

A final note on accuracy, the library is tested against the output of DirectXTex. I generated images, encoded them, decoded them again and compared the output. The output is identical, except when reconstructing Z in BC5, where the library uses a different method.

This is done by doing a full float implementation of BC1 (and by extension BC2 and BC3). The other formats have bit-exact implementations. Z reconstruction uses a lookup table that is generated at runtime, and is also full float accurate.