FPNG Java¶

FPNG-Java provides Java wrappers around three native, SIMD-optimised PNG encoders — FPNG (SSE2), FPNGE (AVX2), and ZPNG (zlib-ng). All produce fully compliant PNG files at significantly higher throughput than javax.imageio.

The three encoders solve different problems:

FPNG and FPNGE are tuned for photographic content and raw throughput. They use minimal-effort deflate via lookup tables, producing large files very fast.
ZPNG uses zlib-ng’s full deflate implementation and is tuned for screen content (UI screenshots, dashboards, remote-desktop tiles, anything with sharp edges and large flat-colour regions). It produces files about half the size of FPNG/FPNGE on this kind of input, at roughly 2-3× the encoding cost. For bandwidth-constrained delivery this is the right trade.

The JDK22+ version of the encoders uses Foreign Function & Memory API and is the fastest. Otherwise JNI is used via JNA, which comes with some overhead.

Latest stable release: fpng-java-2.1.jar

GitHub Repository: https://github.com/manticore-projects/fpng-java

Features¶

Three encoders, each tuned for a different workload
- FPNG — SSE2/SSE4.1/PCLMUL, works on any x86-64 CPU; raw throughput
- FPNGE — AVX2, used automatically on capable CPUs; faster than FPNG
- ZPNG — zlib-ng backed, NEON/AVX2 accelerated; best file size for screen content
Runtime hasAVX2() CPUID probe — no AVX2 library is loaded on unsupported hardware
FFM for JDK22+, JNA for older JDKs
Encodes BufferedImage directly — 3-channel (RGB) and 4-channel (RGBA)
Handles Java’s native TYPE_3BYTE_BGR / TYPE_4BYTE_ABGR channel order in C via SIMD byte shuffles (NEON on ARM64, SSE/AVX2 on x86)
Interleaved Adler32 checksum — no redundant full-image pass (FPNG/FPNGE)
PCLMUL-accelerated CRC32
Supports Java 8, Java 11, Java 17, Java 21, and Java 23
Platforms: Linux x86-64 / ARM64, macOS x86-64 / ARM64 (Apple Silicon), Windows x86 / x86-64
All binaries built and tested natively on each target platform via GitHub Runners

Benchmarks¶

Run the JMH benchmarks with:

./gradlew jmh

Results are written to benchmark/build/results/jmh/results.json.

Speed¶

JMH PNG Encoder Benchmark Report

PNG Encoder Benchmark Report

Generated by Manticore Projects · JMH ms/op · lower is better

3 ch text_test.png

Encoder	Score (ms/op)	Error	vs Winner
FPNGE (Java 23 FFM)	2.291	±NaN	baseline
FPNGE (Java 23 FFM)	2.591	±NaN	+13.1%
FPNGE (JNI)	2.657	±NaN	+16.0%
FPNGE (JNI)	2.972	±NaN	+29.8%
FPNGE (Java 23 FFM)	3.350	±NaN	+46.2%
FPNGE (JNI)	3.763	±NaN	+64.3%
FPNGE (Java 23 FFM)	4.625	±NaN	+101.9%
FPNG Encoder (Java 23)	4.695	±NaN	+104.9%
FPNGE (JNI)	5.025	±NaN	+119.4%
FPNGE (Java 23 FFM)	5.104	±NaN	+122.8%
ZPNG23Benchmark	5.517	±NaN	+140.8%
FPNGE (JNI)	5.531	±NaN	+141.4%
FPNG Encoder (Java 23)	5.871	±NaN	+156.3%
FPNG Encoder (Java 23)	5.902	±NaN	+157.7%
PNGEncoder (fastest)	6.984	±NaN	+204.9%
ZPNG23Benchmark	7.307	±NaN	+219.0%
FPNG Encoder (JNI)	7.383	±NaN	+222.3%
FPNG Encoder (Java 23)	7.792	±NaN	+240.1%
FPNG Encoder (Java 23)	7.914	±NaN	+245.5%
PNGEncoder (fastest)	8.323	±NaN	+263.3%
FPNG Encoder (JNI)	8.579	±NaN	+274.5%
FPNG Encoder (JNI)	8.591	±NaN	+275.0%
FPNG Encoder (JNI)	9.835	±NaN	+329.4%
FPNG Encoder (JNI)	10.120	±NaN	+341.8%
PNGEncoder (fastest)	10.698	±NaN	+367.0%
PNGEncoder (fastest)	12.615	±NaN	+450.7%
PNGEncoder (fastest)	15.023	±NaN	+555.8%
ZPNG23Benchmark	22.746	±NaN	+892.9%
ZPNG23Benchmark	23.545	±NaN	+927.8%
ZPNG23Benchmark	25.026	±NaN	+992.5%
ImageIO	36.859	±NaN	+1509.0%
ImageIO	36.895	±NaN	+1510.6%
ImageIO	36.928	±NaN	+1512.1%
ImageIO	38.221	±NaN	+1568.5%
ImageIO	38.263	±NaN	+1570.4%
ObjectPlanet PNG	53.660	±NaN	+2242.5%
ObjectPlanet PNG	55.038	±NaN	+2302.6%
ObjectPlanet PNG	57.271	±NaN	+2400.1%
PNGEncoder	57.828	±NaN	+2424.4%
ObjectPlanet PNG	59.245	±NaN	+2486.3%
ObjectPlanet PNG	61.444	±NaN	+2582.3%
PNGEncoder	61.974	±NaN	+2605.4%
PNGEncoder	62.756	±NaN	+2639.6%
PNGEncoder	66.641	±NaN	+2809.2%
PNGEncoder	69.236	±NaN	+2922.4%

Efficiency¶

JMH PNG Encoder Pareto Report

PNG Encoder Pareto Report

Generated by Manticore Projects · encode time vs. file size · lower-left is better

3 ch text_test.png

overall winner

per-encoder knee

pareto-optimal

dominated

FPNG Encoder (JNI)

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	dominated	10.12 ±NaN	339.3	4.2%
2	dominated	8.58 ±NaN	8155.5	100.0%
3	dominated	8.59 ±NaN	8155.5	100.0%
4	PARETO	7.38 ±NaN	668.2	8.2%
5	PARETO	9.84 ±NaN	339.3	4.2%

FPNG Encoder (Java 23)

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	PARETO	7.79 ±NaN	339.3	4.2%
2	dominated	5.90 ±NaN	8155.5	100.0%
3	dominated	5.87 ±NaN	8155.5	100.0%
4	PARETO	4.69 ±NaN	668.2	8.2%
5	dominated	7.91 ±NaN	339.3	4.2%

FPNGE (JNI)

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	PARETO	2.66 ±NaN	741.6	94.7%
2	dominated	2.97 ±NaN	783.3	100.0%
3	PARETO	3.76 ±NaN	685.4	87.5%
4	KNEE	5.02 ±NaN	684.7	87.4%
5	PARETO	5.53 ±NaN	525.9	67.1%

FPNGE (Java 23 FFM)

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	PARETO	2.29 ±NaN	736.1	94.0%
2	dominated	2.59 ±NaN	783.3	100.0%
3	KNEE	3.35 ±NaN	679.9	86.8%
4	dominated	4.63 ±NaN	684.7	87.4%
5	PARETO	5.10 ±NaN	525.9	67.1%

ImageIO

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	dominated	38.26 ±NaN	152.1	100.0%
2	PARETO	36.86 ±NaN	152.1	100.0%
3	dominated	38.22 ±NaN	152.1	100.0%
4	dominated	36.89 ±NaN	152.1	100.0%
5	dominated	36.93 ±NaN	152.1	100.0%

ObjectPlanet PNG

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	PARETO	53.66 ±NaN	336.4	100.0%
2	KNEE	55.04 ±NaN	187.0	55.6%
3	PARETO	57.27 ±NaN	169.0	50.2%
4	PARETO	59.25 ±NaN	152.0	45.2%
5	PARETO	61.44 ±NaN	144.6	43.0%

PNGEncoder

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	dominated	61.97 ±NaN	336.5	100.0%
1	PARETO	6.98 ±NaN	336.5	100.0%
2	dominated	62.76 ±NaN	187.1	55.6%
2	KNEE	8.32 ±NaN	187.1	55.6%
3	dominated	57.83 ±NaN	169.1	50.2%
3	PARETO	10.70 ±NaN	169.1	50.2%
4	dominated	66.64 ±NaN	152.1	45.2%
4	PARETO	12.62 ±NaN	152.1	45.2%
5	dominated	69.24 ±NaN	144.6	43.0%
5	PARETO	15.02 ±NaN	144.6	43.0%

ZPNG23Benchmark

Lvl	Status	Time (ms)	Size (KB)	Ratio
1	PARETO	5.52 ±NaN	338.8	100.0%
2	WINNER	7.31 ±NaN	187.0	55.2%
3	PARETO	22.75 ±NaN	168.2	49.6%
4	PARETO	23.55 ±NaN	151.7	44.8%
5	PARETO	25.03 ±NaN	144.2	42.6%