Very fast C++ .PNG writer for 24/32bpp images. fpng.cpp was written to see just how fast you can write .PNG's without sacrificing too much compression, and to determine how slow the existing .PNG writers are. (Turns out they are quite slow.)
On average, fpng.cpp produces files roughly 1-15% smaller than QOI but is 1.4-1.5x slower on 24bpp images.
fpng.cpp compared to stb_image_write.h: fpng.cpp is over 10x faster with roughly 16-18% smaller files. (I was surprised by this. stb_image_write.h does a decent job choosing the PNG filter method and has a stronger byte-oriented parser than fpng's, but its Deflate compressor is crippled by being limited to only static Huffman tables.)
fpng.cpp compared to lodepng.cpp: fpng.cpp is 12-16x faster with roughly 15-18% larger files.
On 32bpp images with correlated alpha signals (a common case in video games) qoi outputs files 1.7x larger vs. fpng and 1.98x larger vs. lodepng.
Benchmark using the 184 QOI test images:
qoi: 4.69 secs 359.55 MB
fpng: 6.66 secs 353.50 MB
lodepng: 110.72 secs 300.14 MB
stb_image_write: 68.07 secs 425.64 MB
Benchmark using my 303 test images:
qoi: 2.25 secs 300.84 MB
fpng: 3.40 secs 253.26 MB
lodepng: 41.68 secs 220.40 MB
stb_image_write: 37.23 secs 311.41 MB
Benchmark using the 184 QOI test images, but with the green channel copied into alpha (to create a correlated alpha channel):
qoi: 2.75 secs 697.2 MB
fpng: 7.48 secs 404.0 MB
lodepng: 151.57 secs 352.1 MB
stbi: 81.58 secs 486.4 MB
182 of the qoi test images don't have alpha channels, so I'm guessing it's not being heavily tested with alpha images much yet.
fpng.cpp uses a custom image aware pixel-wise Deflate compressor which was optimized for simplicity over high ratios. It uses a simple LZRW1-style parser, all literals (except the PNG filter bytes) are output in groups of 3 or 4, all matches are multiples of 3/4 bytes, and it only utilizes a single dynamic Huffman block with one PNG IDAT block. It utilizes 64-bit registers and exploits unaligned little endian reads/writes. (On big endian CPU's it'll use 32/64bpp byteswaps.)
Passes over the input image and dynamic allocations are minimized, although it does use std::vector internally. The first scanline always uses filter #0, and the rest use filter #2 (previous scanline). It uses the fast CRC-32 code described by Brumme here. The original high-level PNG function (that code that writes the headers) was written by Alex Evans.
lodepng v20210627 fetched 12/18/2021
stb_image_write.h v1.16 fetched 12/18/2021
qoi.h fetched 12/18/2021
To build, compile from the included .SLN with Visual Studio 2019/2022 or use cmake to generate a .SLN file. For Linux/OSX, use "cmake ." then "make". Tested with MSVC 2019/gcc/clang.
I have only tested fpng.cpp on little endian systems. The code is there for big endian, and it should work, but it needs testing.
From the "bin" directory, run "fpng_test.exe" or "./fpng_test" like this:
fpng_test.exe <image_filename.png>
To generate .CSV output only:
fpng_test.exe -c <image_filename.png>
There will be several output files written to the current directory: stbi.png, lodepng.png, qoi.qoi, and fpng.png. Statistics or .CSV data will be printed to stdout, and errors to stderr.
The test app decompresses fpng's output using lodepng and stb_image to validate the compressed data. The compressed output has also been validated using pngcheck.
To use fpng.cpp in other programs, copy fpng.cpp/.h and Crc32.cpp/.h into your project. No other configuration or files are needed. Computing the CRC-32 of buffers is a substantial proportion of overall compression time in fpng, so if you have a faster CRC-32 function you can modify fpng_crc() in fpng.cpp to call that instead. The one included in Crc32.cpp doesn't utilize any special CPU instruction sets, so it could be faster.
#include "fpng.h" then call one of these C-style functions in the "fpng" namespace:
bool fpng_encode_image_to_memory(const void* pImage, int w, int h, int num_chans, bool flip, std::vector<uint8_t>& out_buf);
bool fpng_encode_image_to_file(const char* pFilename, const void* pImage, int w, int h, int num_chans, bool flip);
num_chans must be 3 or 4. There must be w*3*h or w*4*h bytes pointed to by pImage. The image row pitch is always w*3 or w*4 bytes. There is no automatic determination if the image actually uses an alpha channel, so if you call it with 4 you will always get a 32bpp .PNG file.
Note the adler32 function in fpng.cpp could be vectorized for higher performance.
I've started to fuzz fpng.cpp, but it will take several days to complete. So far it's survived several different random fuzz methods.
fpng.cpp/.h: Public Domain or zlib (your choice). See the end of fpng.cpp.
Crc32.cpp/.h: Copyright (c) 2011-2019 Stephan Brumme. zlib license: