BasisBlockEncoder — managed BCn texture compression

A fully-managed, hardware-accelerated, zero-allocation CPU block compressor for BC1, BC3, BC4, BC5, BC6H, and BC7 based on BinomialLLC's Transcoder Internals Analytical Real Time Encoders.

Parity with transcoder/basisu_transcoder.cpp

• Bit-for-bit deterministic across architectures. The same input yields the same output bytes on x86 and ARM, regardless of SIMD width
• Reference-grade quality. BC4/BC5 are byte-exact against the C++ reference; BC6H meets or exceeds its PSNR; BC7 reproduces the reference's mode decisions with PSNR within a rounding step

Install

dotnet add package BasisBlockEncoder

Targets net7.0 through net10.0, no third-party runtime dependencies.

Quick start

using BasisBlockEncoder;

// RGBA8 image, tightly packed (strideBytes = width * 4) -> BC7.
int width = 1024, height = 1024;
byte[] rgba = LoadRgba8(width, height);

byte[] dst = new byte[BlockEncoder.EncodedSize(BcFormat.Bc7, width, height)];
BlockEncoder.EncodeBc7(rgba, width, height, width * 4, dst);   // defaults to Bc7Flags.Default

Choose a preset, or any other format:

BlockEncoder.EncodeBc7(rgba, w, h, w * 4, dst, Bc7Flags.Fast);            // faster, slightly lower quality
BlockEncoder.EncodeBc1(rgba, w, h, w * 4, dst, Bc1Quality.HighQuality);   // RGB (+1-bit alpha)
BlockEncoder.EncodeBc3(rgba, w, h, w * 4, dst);                           // RGBA: BC4 alpha + BC1 color
BlockEncoder.EncodeBc4(rgba, w, h, w * 4, dst, channel: 0);              // single channel (R)
BlockEncoder.EncodeBc5(rgba, w, h, w * 4, dst, channel0: 0, channel1: 1); // two channels (RG) — normal maps

HDR (BC6H) takes RGB half-floats:

ReadOnlySpan<Half> rgbHalf = /* width * height * 3 Half values */;
byte[] dst = new byte[BlockEncoder.EncodedSize(BcFormat.Bc6h, w, h)];
BlockEncoder.EncodeBc6h(rgbHalf, w, h, w * 3 * sizeof(ushort), dst, Bc6hQuality.Default);

If you prefer to drive it generically, Encode dispatches by format:

BlockEncoder.Encode(BcFormat.Bc7, rgba, w, h, w * 4, dst, flags: (uint)Bc7Flags.Default);

Supported formats

Format	Input	Block size	Notes
BC1	RGBA8	8 B	RGB + 1-bit alpha; always 4-color (GPU-safe) — never 3-color/punch-through
BC3	RGBA8	16 B	BC4 alpha + BC1 color
BC4	1 channel	8 B	Byte-exact vs the reference
BC5	2 channels	16 B	Two BC4 halves, ideal for tangent-space normal maps
BC6H	RGB FP16	16 B	Unsigned HDR
BC7	RGBA8	16 B	Highest-quality LDR; analytical mode search

BC2 is intentionally omitted, matching the C++ implementation.

Performance

This is a real-time analytical encoder: it computes each block's best fit directly rather than brute-forcing the mode space, and it vectorizes that math with a width-adaptive path — AVX-512 → AVX2 → NEON → scalar — selected at JIT time behind hardware-acceleration guards. The per-block path is allocation-free.

Each block is encoded on a single thread by design, but blocks are independent: fan the per-block calls across cores ( for example, Parallel.For over block rows) and throughput scales with them.

Quality knobs

• Bc7Flags — Fastest, Faster, Fast, Default (recommended), HighQuality. Higher presets enable more mode families (two-/three-subset, dual-plane) and p-bit optimization; HighQuality runs the full analytical search. The optional non-analytical brute-force passes are accepted for source compatibility but not performed — the analytical search is what delivers the quality.
• Bc1Quality — Fast or HighQuality, for the BC1/BC3 color block.
• Bc6hQuality — Fast, Default, HighQuality.

Streaming

Encode in horizontal bands instead of materializing the whole surface: EncodeRows / EncodeBc6hRows (bands must be a multiple of 4 rows, except the last), or the BlockRowStreamWriter / Bc6hBlockRowStreamWriter helpers.

License

This package is MIT. The encoder is a managed port of Basis Universal's block encoders, so retain the Apache-2.0 Basis Universal attribution/NOTICE when redistributing.