Unitbench Openzl Scenarios Agent Skill

unitBench Scenario Creation

Create benchmark scenarios for openzl codec nodes. Two benchmark types exist: kernel (test encode/decode kernel functions directly) and graph (test a node within a full compress/decompress pipeline).

Deciding What to Benchmark

Before creating scenarios, ask the user:

• Does this node have a standalone kernel function? (e.g., ZL_bitSplitEncode, ZL_bitSplitDecode)
• If yes: kernel benchmarks are an option - test the encode/decode functions directly with minimal overhead.
• If no: the node must be tested as part of a graph.

• Should the node be tested in a graph?
• Graph benchmarks test the full pipeline: tokenization -> node -> downstream processing -> round-trip decompression.
• Useful for measuring real-world overhead vs kernel-only performance.

• What data types/widths does the node operate on?
• Determines element widths, bit layouts, and which tokenizer node to use in graphs.
• Ask the user for the specific parameters (bitWidths, element widths, etc.).

File Locations

| What | Where |

|------|-------|

| Kernel benchmarks | benchmark/unitBench/scenarios/codecs/<name>.c and .h |

| Graph benchmarks | benchmark/unitBench/scenarios/<name>_graph.c and .h |

| Scenario registration | benchmark/unitBench/benchList.h |

| BUCK file | benchmark/unitBench/BUCK |

| Test data | /tmp/ (use dd if=/dev/urandom) |

All paths relative to the openzl dev root.

Kernel Benchmark

Test encode/decode kernel functions directly. Requires a standalone kernel API.

Header (.h)

Add declarations to existing scenarios/codecs/<codec>.h or create a new one:

// Decode
size_t <codec>Decode_<type>_prep(void* src, size_t srcSize, const BenchPayload* bp);
size_t <codec>Decode_<type>_outSize(const void* src, size_t srcSize);
size_t <codec>Decode_<type>_wrapper(const void* src, size_t srcSize, void* dst, size_t dstCapacity, void* customPayload);

// Encode
size_t <codec>Encode_<type>_prep(void* src, size_t srcSize, const BenchPayload* bp);
size_t <codec>Encode_<type>_outSize(const void* src, size_t srcSize);
size_t <codec>Encode_<type>_wrapper(const void* src, size_t srcSize, void* dst, size_t dstCapacity, void* customPayload);

Source (.c)

Decode scenario: prep packs split streams contiguously into src, wrapper recomputes pointers and calls the decode kernel, outSize returns (srcSize / sumSrcElt) * dstEltWidth.

Encode scenario: prep fills src with random values, wrapper calls the encode kernel writing streams contiguously into dst, outSize returns (srcSize / srcEltWidth) * sumDstElt.

Reference implementation: See scenarios/codecs/bitSplit.c for the complete pattern with multiple data type examples.

Graph Benchmark

Test a node within a full compress/decompress graph. Required when no standalone kernel exists. Also useful alongside kernel benchmarks to measure graph overhead.

Header (.h)

// Copyright (c) Meta Platforms, Inc. and affiliates.

#ifndef GUARD_MACRO_H
#define GUARD_MACRO_H

#include "openzl/shared/portability.h"
#include "openzl/zl_compressor.h"

ZL_BEGIN_C_DECLS

ZL_GraphID <name>_graph(ZL_Compressor* cgraph);

ZL_END_C_DECLS

#endif

Source (.c)

Build the graph using ZL_Compressor_registerStaticGraph_fromNode1o. Typical pattern: tokenize input -> apply node -> downstream graph.

#include "openzl/codecs/zl_<codec>.h"   // ZL_NODE_<YOUR_NODE>
#include "openzl/zl_compressor.h"
#include "openzl/zl_public_nodes.h"      // ZL_NODE_INTERPRET_AS_LE*

ZL_GraphID my_graph(ZL_Compressor* cgraph)
{
    if (ZL_isError(ZL_Compressor_setParameter(
                cgraph, ZL_CParam_formatVersion, ZL_MAX_FORMAT_VERSION))) {
        abort();
    }
    if (ZL_isError(ZL_Compressor_setParameter(
                cgraph, ZL_CParam_compressionLevel, 1))) {
        abort();
    }

    return ZL_Compressor_registerStaticGraph_fromNode1o(
            cgraph,
            ZL_NODE_INTERPRET_AS_LE64,  // tokenizer matching element width
            ZL_Compressor_registerStaticGraph_fromNode1o(
                    cgraph, ZL_NODE_YOUR_NODE, ZL_GRAPH_STORE));
                    // ZL_GRAPH_STORE to benchmark node in isolation
                    // ZL_GRAPH_ZSTD to benchmark with compression
}

Tokenizer node must match element width: ZL_NODE_INTERPRET_AS_LE16 (2 bytes), ZL_NODE_INTERPRET_AS_LE32 (4 bytes), ZL_NODE_INTERPRET_AS_LE64 (8 bytes).

Reference: See scenarios/sao_graph.c for a complex multi-stream graph example.

Registration

benchList.h

• Add include for graph header (near other graph includes around line 171):

#include "benchmark/unitBench/scenarios/<name>_graph.h"

• Add entries to scenarioList[] array (maintain alphabetical order):

// Kernel scenarios (set .func via first positional arg)
{ "<codec>Decode_<type>", <codec>Decode_<type>_wrapper, .prep = <codec>Decode_<type>_prep, .outSize = <codec>Decode_<type>_outSize },
{ "<codec>Encode_<type>", <codec>Encode_<type>_wrapper, .prep = <codec>Encode_<type>_prep, .outSize = <codec>Encode_<type>_outSize },

// Graph scenario (set .graphF - harness auto-wires init and compression)
{ "<graphName>", .graphF = <name>_graph },

BUCK

Add a library target for graph benchmarks (following sao_graph pattern):

zs_library(
    name = "<name>_graph",
    srcs = ["scenarios/<name>_graph.c"],
    headers = ["scenarios/<name>_graph.h"],
    deps = [
        "../..:zstronglib",
    ],
)

Kernel .c/.h files are auto-included by the unitBench binary's glob(["**/*.c"]).

Test Data and Running

Test data size must be a multiple of the element width for the codec/node being tested. For example, fp64 (8-byte elements) needs a file size divisible by 8. Using standard sizes like 1MB/10MB works for all common element widths.

# Generate test data (use sizes that are multiples of element width)
dd if=/dev/urandom of=/tmp/openzl_bench/test_1MB.bin bs=1M count=1
dd if=/dev/urandom of=/tmp/openzl_bench/test_10MB.bin bs=1M count=10

# Build with BUCK in opt mode (best practice - optimized, no ASAN)
buck build @//mode/opt //openzl/dev/benchmark/unitBench:unitBench

# Run benchmark via buck run
buck run @//mode/opt //openzl/dev/benchmark/unitBench:unitBench -- <scenarioName> /tmp/openzl_bench/test_10MB.bin

# Useful options (after the -- separator)
#   -i <seconds>    benchmark duration (default ~2s)
#   -B <bytes>      split input into blocks
#   --csv           CSV output for parsing
#   -z              compression only (skip decompression round-trip)

# List all scenarios
buck run @//mode/opt //openzl/dev/benchmark/unitBench:unitBench -- --list

Always use buck build/run @//mode/opt for benchmarking. If buck is not available, fall back to make unitBench (from the openzl dev root).

Common Mistakes

• Forgetting prep function: Kernel decode benchmarks need prep to fill split streams. Encode benchmarks need prep to fill random source values.
• Wrong outSize: Decode: (srcSize / sumSrcElt) * dstEltWidth. Encode: (srcSize / srcEltWidth) * sumDstElt.
• Graph not setting formatVersion: Must set ZL_CParam_formatVersion to ZL_MAX_FORMAT_VERSION for newer nodes.
• scenarioList not alphabetical: Entries must be in alphabetical order by name.
• Test data in repo: Put test data in /tmp/, not in the source tree.