Add AOCL-Compression 5.3.0 (AMD-optimized compression)

[BiplabRaut]

AOCL-Compression provides AMD-optimized implementations of widely used lossless compression algorithms. This integration exposes them alongside lzbench's native implementations for direct comparison.

• License: BSD 3-Clause / MIT / NCSA
• Repository: https://github.com/amd/aocl-compression
• Version: AOCL-Compression 5.3.0, pinned to upstream tag 5.3

Registered variants

• aocl-lz4 (level 0)
• aocl-lz4hc (levels 1-12)
• aocl-lzma (levels 0-9)
• aocl-bzip2 (levels 1-9)
• aocl-snappy (level 0)
• aocl-zlib (levels 1-9)
• aocl-zstd (levels 1-22)

The AOCL alias runs representative levels of all seven codecs.

Performance

Native and AOCL implementations were measured under identical settings:

• Dataset: Silesia corpus as an uncompressed tar (211,957,760 bytes)
• CPU: AMD Ryzen AI 9 HX PRO 370
• Compiler: GCC 15.2.0
• Execution: single core, pinned to CPU 0
• Block size: lzbench default (whole input)
• CPU boost: disabled
• CPU governor: performance
• Runs: five independent runs
• Per-run settings: -t3,3 -i3,3 -p1
• Reported throughput: geometric mean across five runs

Results summary

AOCL delivers substantial throughput gains on zlib (up to 3.7× compression, 2× decompression),
bzip2 (2.1–2.3× compression, +60–69% decompression), snappy (+20% compression, 2× decompression),
and moderate gains on zstd (+14–21% compression, +8–10% decompression), lz4hc (+18% compression),
lzma (+12% compression, +17% decompression), and lz4 (+5% compression, +2% decompression) with
minor compression ratio trade-offs.

Detailed Results

Compressor	Compression	Decompression	Ratio
lz4 1.10.0	381.25 MB/s	2499.41 MB/s	47.59%
aocl-lz4 5.3.0	400.32 MB/s	2550.71 MB/s	48.00%
snappy 1.2.2	342.91 MB/s	719.19 MB/s	47.85%
aocl-snappy 5.3.0	413.03 MB/s	1433.58 MB/s	48.55%
zlib 1.3.2 -1	61.10 MB/s	199.93 MB/s	36.45%
aocl-zlib 5.3.0 -1	82.78 MB/s	414.80 MB/s	34.86%
zlib 1.3.2 -6	18.16 MB/s	210.76 MB/s	32.19%
aocl-zlib 5.3.0 -6	46.70 MB/s	446.00 MB/s	32.52%
zlib 1.3.2 -9	7.37 MB/s	212.55 MB/s	31.91%
aocl-zlib 5.3.0 -9	27.47 MB/s	439.38 MB/s	32.21%
bzip2 1.0.8 -1	9.55 MB/s	27.54 MB/s	28.54%
aocl-bzip2 5.3.0 -1	18.13 MB/s	46.47 MB/s	28.56%
bzip2 1.0.8 -5	9.36 MB/s	24.46 MB/s	26.29%
aocl-bzip2 5.3.0 -5	20.31 MB/s	39.21 MB/s	26.31%
bzip2 1.0.8 -9	8.92 MB/s	23.16 MB/s	25.75%
aocl-bzip2 5.3.0 -9	20.58 MB/s	37.65 MB/s	25.77%
zstd 1.5.7 -3	169.38 MB/s	871.35 MB/s	31.20%
aocl-zstd 5.3.0 -3	204.97 MB/s	959.04 MB/s	31.94%
zstd 1.5.7 -9	44.15 MB/s	920.85 MB/s	27.87%
aocl-zstd 5.3.0 -9	50.33 MB/s	994.89 MB/s	28.14%
lzma 26.01 -9	1.71 MB/s	52.42 MB/s	22.97%
aocl-lzma 5.3.0 -9	1.91 MB/s	61.51 MB/s	23.01%
lz4hc 1.10.0 -9	22.20 MB/s	2412.30 MB/s	36.75%
aocl-lz4hc 5.3.0 -9	26.19 MB/s	2436.0 MB/s	36.78%

Benchmark test was executed with the following command (5 iterations):

taskset -c 0 ./lzbench \
  -elz4/aocl-lz4/snappy/aocl-snappy/zlib,1/aocl-zlib,1/zlib,6/aocl-zlib,6/zlib,9/aocl-zlib,9/bzip2,1/aocl-bzip2,1/bzip2,5/aocl-bzip2,5/bzip2,9/aocl-bzip2,9/zstd,3/aocl-zstd,3/zstd,9/aocl-zstd,9/lzma,9/aocl-lzma,9/lz4hc,9/aocl-lz4hc,9 \
  -t3,3 -i3,3 -p1 -o4 silesia.tar

Requirements

• Linux on x86-64
• GNU Make
• GNU nm and objcopy
• AOCL-Compression sources initialized through the git submodule

Unsupported configurations automatically define BENCH_REMOVE_AOCL. The integration can be disabled explicitly with:

make DONT_BUILD_AOCL=1

Integration

• Added AOCL-Compression as the lz/aocl-compression git submodule, pinned to upstream tag 5.3.
• Added wrappers using AOCL's unified aocl_llc_* API.
• Registered all seven AOCL codecs and their supported levels.
• Added the AOCL alias for representative levels.
• Builds AOCL as a static archive using its bundled GNU Make build.
• Prefixes AOCL's internal symbols with AOCLLZB_ using objcopy to prevent collisions with lzbench's native codec implementations.
• Preserves AOCL's public aocl_llc_* API symbols.
• Updated README.md, BUILD.md, and CHANGELOG.