How to set optimized options?

[ebiggers]

How to set optimized options?

[ebiggers]

It's optimized by default. Please give details. What CPU? How did you measure the speed, exactly? What input sizes did you test? Is there any way that I could reproduce your results? Are you sure it was a controlled test?

[kanryu]

I first built it with msvc, but it seems that the optimized function is not included in lib and it is not optimized by this build method. On the other hand, when building with gcc, I confirmed that it contains optimized functions without any special handling.

As for the measurement method of processing, I use lodepng which is one of the OSS implementation of the PNG library, and confirmed it by replacing the expansion part of deflate with libdeflate. Expanding images is completed in half the time from lodepng origin, but it is almost the same as libpng.

[ebiggers]

Yes, due to limited resources to develop and test with MSVC, libdeflate is only properly optimized when built with gcc or clang.

Anyway, if I understand you correctly, your results were:

• lobepng with its own DEFLATE implementation is slow.
• lobepng with libdeflate is fast.
• libpng with zlib is fast.

I don't see where you actually compared libdeflate to zlib directly. It could be that libpng is faster for other reasons. Can you please clarify whether you actually did a controlled test that compared libdeflate to zlib?

[jibsen]

If you are compiling with GCC for Windows x86 (and not x64), you might also want to add -msse2, unless you specifically need to support old machines. Compiling for x86 with SSE2 enabled has been the default in MSVC since 2012.

It is unlikely to affect the decompression speed, since that is handled by inspecting CPU features, but it does appear to speed up compression (as a quick test, the included benchmark tool reports 63 MB/s with the default x86 options, and 76 MB/s with -msse2 for compression level 1 on silesia.tar).

[ebiggers]

Correct, I haven't made the matchfinder optimizations detect CPU features at runtime yet. So adding
-msse2 (for x86) or -mavx2 (for x86 and x64) will help a bit, if you know the code will only be run on a CPU with those features. However, that only affects compression, whereas the original question here was about decompression.

[kanryu]

Hi,
As a result of profiling lodepng from last time, I noticed that adler32 and unfilter processing is late by lodepng 's default (PNG image has filters). Replacing the zlib module with libdeflate instead of inflate and replacing it with unfilter in libpng confirmed that it can be expanded at 1.35 times faster than normal libpng.

It seems necessary to modify the processing of lodepng itself in order to further increase the speed (some unnecessary processing is found).

I added -O3, -Ofast, -mtune etc when adding libdeflate with gcc of msys2 but there was no difference in processing performance.

[ebiggers]

Okay, so you're saying that libdeflate is faster after all, in a proper comparison?

Is there any remaining issue here?

[kanryu]

No.
Thank you very much for solving my first question :)