Ignore checksums more often

[nigeltao]

Adler-32 and CRC-32 computations are fast with SIMD, but if you just want the fastest possible PNG decoder (e.g. to compare to QOI's speed), ignoring the checksums can be even faster.

For decode (not encode, obviously), IIUC fpng already skips computing/verifying Adler-32 always and CRC-32 sometimes (depending on FPNG_DISABLE_DECODE_CRC32_CHECKS).

Some of the other libraries (lodepng and stb_image) also do this automatically. Wuffs needs to opt in, with a one-liner patch:

diff --git a/src/fpng_test.cpp b/src/fpng_test.cpp
index afc3f77..7302146 100644
--- a/src/fpng_test.cpp
+++ b/src/fpng_test.cpp
@@ -683,6 +683,7 @@ static void* wuffs_decode(void* pData, size_t data_len, uint32_t &width, uint32_
        wuffs_png__decoder* pDec = wuffs_png__decoder__alloc();
        if (!pDec) 
                return nullptr;
+       wuffs_png__decoder__set_quirk_enabled(pDec, WUFFS_BASE__QUIRK_IGNORE_CHECKSUM, true);

        wuffs_base__image_config ic;
        wuffs_base__io_buffer src = wuffs_base__ptr_u8__reader((uint8_t *)pData, data_len, true);

[nigeltao]

Tangentially, if you're using Wuffs from C++ (instead of C), you can replace

wuffs_png__decoder__set_quirk_enabled(pDec, etc);

with

pDec->set_quirk_enabled(etc);

and similarly for other Wuffs' C API calls.

[richgel999]

Cool- I will hook this up to see how wuffs compares.

[richgel999]

I disabled Wuffs checksum checking. Also I updated the README. I didn't conduct exhaustive benchmarking yet, but I will. So far Wuffs's PNG decoder is extremely fast.

fpng seems a bit faster (within 10-20%, depending on whether you disable Wuffs checksums), but it's not a general purpose decoder. FPNG is really just existence proof and a push to others to make their PNG libraries faster. The real competition is QOI.

[nigeltao]

The real competition is QOI.

Also tangential, but if you haven't already seen it, https://github.com/nigeltao/qoi2-bikeshed/issues/28 has some numbers for unofficial, experimental QOI variants. The highlights:

• libpng is... libpng 1.6.37.
• qoi-mas-c04 is "Official QOI": commit c04a975 from the phoboslab/master branch (2021-12-24).
• qoi-opt-c04 is the same implementation (and file format), but the decoder was optimized. The encoder is unchanged.
• qoi-lz4-d28 tweaks the QOI opcodes and also introduces LZ4 compression.

qoibench.c numbers on two test suites:

        decode ms   encode ms   decode mpps   encode mpps   size kb    rate

# Grand total for images
libpng:      13.5       153.7         34.44          3.02       398   24.2%
qoi-mas-c04:  4.7         5.6         99.25         82.46       463   28.2%
qoi-opt-c04:  4.2         5.7        111.40         81.60       463   28.2%
qoi-lz4-d28:  4.2         6.0        110.38         77.11       416   25.4%

# Grand total for images-lance
libpng:     103.5       925.0         39.25          4.39      1395    9.0%
qoi-mas-c04: 35.2        32.5        115.26        124.91      2109   13.6%
qoi-opt-c04: 29.0        32.5        139.87        124.88      2109   13.6%
qoi-lz4-d28: 28.7        36.2        141.62        112.13      1530    9.8%

It looks like QOI can get a 1.1x - 1.2x decode speed improvement above the current official implementation, without any file format changes. With file format changes, something QOI-like can keep that higher speed but also improve the compression ratio, closing the gap on PNG.

https://github.com/nigeltao/qoi2-bikeshed/issues/31 and https://github.com/nigeltao/qoi2-bikeshed/issues/34 also suggests that there may be further compression-ratio gains from tweaking a QOI-like file format.

[nigeltao]

The real competition is QOI.

You (@richgel999) are already mentioned in the tweet but for anyone else coming here: @veluca93 and @jonsneyers have a standalone JpegXL lossless encoder that's about 1000 lines of code (https://github.com/libjxl/libjxl/tree/f413c429cfd5f2e1f65b5ebfc0c969b67ddbcbaf/experimental/fast_lossless) and https://twitter.com/jonsneyers/status/1472959101416226823 suggests that it's 2x faster than QOI (for roughly equivalent compression ratio).

[nigeltao]

See also: https://github.com/veluca93/fpnge

Anyway, happy to see healthy competition making everyone faster. :-)