FZGPUModules

GPU-accelerated graph composable compression pipeline builder for analytical workflows.

Overview

FZGPUModules is a CUDA library for building composable, high-throughput compression pipelines. Each pipeline is a directed acyclic graph (DAG) of stages - coders, predictors, quantizers, shufflers, transforms, fused stages, and external stages - connected and executed entirely on the GPU with stream-ordered memory management.

Key properties:

• Modular — mix and match stages (Lorenzo, G-Interp, Quantizer, ADM, RLE, RZE, RRE, Bitshuffle, Huffman, ANS, …)
• High throughput — parallel level execution, persistent scratch, CUDA Graph support
• Memory-efficient — MINIMAL and PREALLOCATE strategies; buffer coloring to alias non-overlapping allocations
• File format — FZM format with CRC32 checksums and full stage config serialization

---

Requirements

Requirement	Minimum	Notes
CUDA Toolkit	11.2+	Stream-ordered allocator required
Host Compiler	GCC 7+ or Clang 5+	Upper bound set by CUDA version — see NVIDIA release notes; NVHPC 23.11 tested in CI
C++ Standard	C++17
CMake	3.24+
Host byte order	Little-endian

Note: using a vGPU will result in the CUDA mempool creation to fail, resulting in an automatic fallback allocation using cudaMalloc. This will work correctly but without the performance benefits of the stream-ordered allocator. For perfomance critical workloads avoid vGPU setups. The lack of stream-ordered allocator support also prevents CUDA Graph capture on vGPUs so this feature is unavailable in those environments.

---

Quick Start

Building from Source

For full build options (presets, examples/tests, install), see the Building from Source page.

git clone https://github.com/szcompressor/FZGPUModules.git
git submodule update --init --recursive
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

C++ API Usage

#include "fzgpumodules.h"
 
// 1. Build a pipeline
fz::Pipeline pipeline(input_bytes);
 
auto* lrz = pipeline.addStage<fz::LorenzoQuantStage<float, uint16_t>>(
    fz::LorenzoQuantStage<float, uint16_t>::Config{1e-4f});
auto* rle = pipeline.addStage<fz::RLEStage<uint16_t>>();
 
pipeline.connect(rle, lrz, "codes");
pipeline.finalize();
 
// 2. Compress
void* d_compressed = nullptr;
size_t compressed_size = 0;
pipeline.compress(d_input, n * sizeof(float), &d_compressed, &compressed_size, stream);
 
// 3. Decompress
void* d_output = nullptr;
size_t output_size = 0;
pipeline.decompress(d_compressed, compressed_size, &d_output, &output_size, stream);
cudaStreamSynchronize(stream);
// d_output is pool-owned by default; do not cudaFree it.
// Call pipeline.setPoolManagedDecompOutput(false) for caller-owned output.

See examples/ for more usage patterns including multi-branch pipelines, CUDA Graph capture, and the low-level DAG API.

---

Available Stages

For detailed per-stage documentation — constraints, behavioral rules, and extended usage notes — see the Stage Reference.

Stage	Header	Description
LorenzoQuantStage<TInput, TCode>	modules/fused/lorenzo_quant/lorenzo_quant.h	Fused float predictor + quantizer (lossy)
LorenzoStage<T>	modules/predictors/lorenzo/lorenzo_stage.h	Plain integer Lorenzo predictor (lossless)
TiledLorenzoStage<T>	modules/predictors/tiled_lorenzo/tiled_lorenzo_stage.h	Dimension-aware (tiled separable) Lorenzo predictor (lossless, 2D/3D, cuSZp3 delta)
GInterpStage<TInput, TCode>	modules/fused/ginterp/ginterp_stage.h	Multi-level spline interpolation predictor + quantizer (lossy, 3D, cuSZ-Hi port)
QuantizerStage<TInput, TCode>	modules/quantizers/quantizer/quantizer.h	Direct-value quantizer (ABS/REL/NOA)
RLEStage<T>	modules/coders/rle/rle.h	Run-length encoding
DifferenceStage<T, TOut>	modules/predictors/diff/diff.h	First-order difference / cumulative-sum coding
ADMStage	modules/transforms/adm/adm_stage.h	Adaptive data mapping — uint16/32 → 8-bit symbol domain (MANS port)
BitshuffleStage	modules/shufflers/bitshuffle/bitshuffle_stage.h	Bit-matrix transpose
RZEStage	modules/coders/rze/rze_stage.h	Recursive zero-byte elimination
RREStage	modules/coders/rre/rre_stage.h	Repetition-reduction encoding (LC component)
ZigzagStage<TIn, TOut>	modules/transforms/zigzag/zigzag_stage.h	Zigzag encode/decode
NegabinaryStage<TIn, TOut>	modules/transforms/negabinary/negabinary_stage.h	Negabinary encode/decode
BitpackStage<T>	modules/coders/bitpack/bitpack_stage.h	Pack/unpack power-of-two value streams
AdaptiveBitpackStage<T>	modules/coders/adaptive_bitpack/adaptive_bitpack_stage.h	Per-block adaptive fixed-rate bit-plane coding (cuSZp/cuSZp2 port)
HuffmanStage<T>	modules/coders/huffman/huffman_stage.h	GPU Huffman entropy coding (PHF, cuSZ port)
ANSStage	modules/coders/ans/ans_stage.h	GPU rANS entropy coding (dietGPU port)
BitplaneRZEStage	modules/fused/bitplane_rze/bitplane_rze_stage.h	Fused bitplane transpose + zero-group RZE lossless encoder (FZ-GPU port)
MergeStage	modules/structural/merge/merge_stage.h	Concatenate N producer ports into one buffer / split back (structural)

Memory Strategies

Strategy	Description
MINIMAL	Allocate on demand, free at last consumer. Lowest peak GPU memory.
PREALLOCATE	Allocate everything at finalize(). Required for CUDA Graph capture. Enables buffer coloring for efficient buffer reuse.

---

Caller-Allocated Output

If you want full memory control, use the caller-allocated overloads. This mirrors nvcomp-style APIs: you pre-allocate an output buffer and pass its capacity; the API returns the actual size.

// After finalize()
size_t comp_capacity = pipeline.getMaxCompressedSize(input_bytes);
void* d_comp_user = nullptr;
cudaMalloc(&d_comp_user, comp_capacity);
 
size_t comp_size = 0;
pipeline.compress(d_input, input_bytes,
                  d_comp_user, comp_capacity,
                  &comp_size, stream);

For decompression, size the output from the original input or from the FZM header:

auto header = fz::Pipeline::readHeader("output.fzm");
size_t decomp_capacity = header.core.uncompressed_size;
 
void* d_decomp_user = nullptr;
cudaMalloc(&d_decomp_user, decomp_capacity);
 
size_t decomp_size = 0;
pipeline.decompress(d_comp_user, comp_size,
                    d_decomp_user, decomp_capacity,
                    &decomp_size, stream);

See examples/ownership_example.cpp for a minimal end-to-end example.

---

CUDA Graph Support

For throughput-critical workloads, enable CUDA Graph capture to eliminate CPU-side kernel launch overhead on repeated compress calls:

fz::Pipeline pipeline(input_bytes, fz::MemoryStrategy::PREALLOCATE, 2.0f);
// ... addStage, connect ...
pipeline.enableGraphMode(true);
pipeline.finalize();
pipeline.warmup(stream);      // JIT-compiles all kernels once
pipeline.captureGraph(stream);
 
// subsequent compress() calls replay the captured graph
pipeline.compress(d_input, input_bytes, &d_compressed, &compressed_sz, stream);

Call compress() only after captureGraph(); use the same stream for capture and replay.

---

Compressor Config File

For complex pipelines, you can also load the stage graph from a TOML config file:

fzgmod-cli -z -i data.f32 -c examples/presets/pfpl.toml -o compressed.fzm --report

You can also use the Pipeline::loadFromConfig() API to load a config file from C++. The config schema supports arbitrary DAGs.

See examples/presets/ for reference and pre-built pipeline configurations and the Config File Reference for the full config schema.

---

File I/O

// Write to file after compressing
pipeline.writeToFile("output.fzm", stream);
 
// Decompress directly from file (no pipeline setup needed)
void* d_out = nullptr;
size_t out_size = 0;
fz::Pipeline::decompressFromFile("output.fzm", &d_out, &out_size, stream);
cudaStreamSynchronize(stream);
cudaFree(d_out);

FZM files embed the full stage configuration and compressed payload with CRC32 checksums. See the FZM File Format page for the full specification.

---

Decode-only pipelines (no warmup compress)

For streaming decode loops that only ever decompress blobs produced elsewhere, a decode-only pipeline can decode an in-memory blob with no prior compress() by carrying a small metadata header:

// Producer (after compress()): store the header alongside the blob.
std::vector<uint8_t> header = producer.serializeHeaderToMemory();   // ~1 KB, no payload
 
// Consumer (fresh, finalized, same topology, never compress()ed): one call per blob.
slot.decompressFromMemory(header.data(), header.size(),
                          d_blob, blob_size, &d_out, &out_size, stream);

The header carries the data-dependent inverse metadata that is not in the raw blob (Huffman symbol count, quantizer outlier count). See examples/decode_only_slots.cpp and the API Reference.

---

Thread Safety

Each Pipeline must be used from a single host thread. There is no internal locking.

Safe — run one independent pipeline per thread:

std::thread t1([&] {
    fz::Pipeline p1(input_size);
    // build, finalize, compress, decompress ...
});
std::thread t2([&] {
    fz::Pipeline p2(input_size);
    // build, finalize, compress, decompress ...
});
t1.join(); t2.join();

Not safe — two threads sharing one pipeline:

fz::Pipeline shared;
std::thread t1([&] { shared.compress(...); });  // data race
std::thread t2([&] { shared.compress(...); });  // data race

The library has no global mutable state. The FZ_LOG logger singleton is set once at startup; do not change log level or callback while pipelines are running on other threads.

---

Coding with AI agents

If you're using Claude Code (or another agent-driven IDE assistant), point it at CLAUDE.md at the repo root for an LLM-tailored overview of the pipeline model, the stage catalog, build/test recipes, and the ownership rules and gotchas that are easy to get wrong from source alone.

---

Citation

If you reference this work, please cite:

Note: this paper describes the 1.0 release of the library; the 2.0 API and documentation may differ.

‍[DRBSD-11] FZModules: A Heterogeneous Computing Framework for Customizable Scientific Data Compression Pipelines

@inproceedings{ruiter2025fzmodules,
    author = {Ruiter, Skyler and Tian, Jiannan and Song, Fengguang},
    title = {FZModules: A Heterogeneous Computing Framework for Customizable Scientific Data Compression Pipelines},
    year = {2025},
    url = {https://doi.org/10.1145/3731599.3767376},
    booktitle = {Proceedings of the SC '25 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis},
    pages = {332-338},
    series = {SC Workshops '25}
}