-
Notifications
You must be signed in to change notification settings - Fork 2.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
zlib: Add support for building with Chromium's zlib implementation #5748
Conversation
fe9ab62
to
e0013cc
Compare
Wow, I'm very much torn by this. On one hand, speed! One the other, submodules. And two different bundled zlibs. It feels like not bundling this and letting the caller deal with it would be the easiest solution for us. But it also looks like that's actually quite difficult because there's no (easy or out of the box) way to build the chromium zlib fork, is that right? |
i just found out about https://cmake.org/cmake/help/git-stage/module/ExternalProject.html . with that we can achieve the same goal without submodules! |
This change builds libgit2 using Chromium's zlib implementation by invoking cmake with `-DUSE_BUNDLED_ZLIB=ON -DUSE_CHROMIUM_ZLIB=ON`, which is ~10% faster than the bundled zlib for the core::zstream suite. This version of zlib has some optimizations: a) Decompression (Intel+ARM): inflate_fast, adler32, crc32, etc. b) Compression (Intel): fill_window, longest_match, hash function, etc. Due to the introduction of SIMD optimizations, and to get the maximum performance out of this fork of zlib, this requires an x86_64 processor with SSE4.2 and CLMUL (anything Westmere or later, ~2010). The Chromium zlib implementation also supports ARM with NEON, but it has not been enabled in this patch. Performance =========== TL;DR: Running just `./libgit2_clar -score::zstream` 100 times in a loop took 0:56.30 before and 0:50.67 after (~10% reduction!). The bundled and system zlib implementations on an Ubuntu Focal system perform relatively similar (the bundled one is marginally better due to the compiler being able to inline some functions), so only the bundled and Chromium zlibs were compared. For a more balanced comparison (to ensure that nothing regressed overall), `libgit2_clar` under `perf` was also run, and the zlib-related functions were compared. Bundled ------- ```shell cmake \ -DUSE_BUNDLED_ZLIB=ON \ -DUSE_CHROMIUM_ZLIB=OFF \ -DCMAKE_BUILD_TYPE="RelWithDebInfo" \ -DCMAKE_C_FLAGS="-fPIC -fno-omit-frame-pointer" \ -GNinja \ .. ninja perf record --call-graph=dwarf ./libgit2_clar perf report --children ``` ``` Samples: 87K of event 'cycles', Event count (approx.): 75923450603 Children Self Command Shared Objec Symbol + 4.14% 0.01% libgit2_clar libgit2_clar [.] git_zstream_get_output_chunk + 2.91% 0.00% libgit2_clar libgit2_clar [.] git_zstream_get_output + 0.69% 0.00% libgit2_clar libgit2_clar [.] git_zstream_get_output (inlined) 0.17% 0.00% libgit2_clar libgit2_clar [.] git_zstream_init 0.02% 0.00% libgit2_clar libgit2_clar [.] git_zstream_reset 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_eos 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_done 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_free (inlined) Samples: 87K of event 'cycles', Event count (approx.): 75923450603 Children Self Command Shared Objec Symbol + 3.12% 0.01% libgit2_clar libgit2_clar [.] deflate + 2.65% 1.48% libgit2_clar libgit2_clar [.] deflate_slow + 1.60% 0.55% libgit2_clar libgit2_clar [.] inflate + 0.53% 0.00% libgit2_clar libgit2_clar [.] write_deflate 0.49% 0.36% libgit2_clar libgit2_clar [.] inflate_fast 0.46% 0.02% libgit2_clar libgit2_clar [.] deflate_fast 0.19% 0.19% libgit2_clar libgit2_clar [.] inflate_table 0.16% 0.01% libgit2_clar libgit2_clar [.] inflateInit_ 0.15% 0.00% libgit2_clar libgit2_clar [.] inflateInit2_ (inlined) 0.10% 0.00% libgit2_clar libgit2_clar [.] deflateInit_ 0.10% 0.00% libgit2_clar libgit2_clar [.] deflateInit2_ 0.03% 0.00% libgit2_clar libgit2_clar [.] deflateReset (inlined) 0.02% 0.00% libgit2_clar libgit2_clar [.] deflateReset 0.02% 0.00% libgit2_clar libgit2_clar [.] inflateEnd 0.02% 0.00% libgit2_clar libgit2_clar [.] deflateEnd 0.01% 0.00% libgit2_clar libgit2_clar [.] deflateResetKeep 0.01% 0.01% libgit2_clar libgit2_clar [.] inflateReset2 0.01% 0.00% libgit2_clar libgit2_clar [.] deflateReset (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateReset (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] deflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateResetKeep (inlined) ``` Chromium -------- ```shell cmake \ -DUSE_BUNDLED_ZLIB=ON \ -DUSE_CHROMIUM_ZLIB=ON \ -DCMAKE_BUILD_TYPE="RelWithDebInfo" \ -DCMAKE_C_FLAGS="-fPIC -fno-omit-frame-pointer" \ -GNinja \ .. ninja perf record --call-graph=dwarf ./libgit2_clar perf report --children ``` ``` Samples: 97K of event 'cycles', Event count (approx.): 80862210917 Children Self Command Shared Objec Symbol + 3.31% 0.00% libgit2_clar libgit2_clar [.] git_zstream_get_output_chunk + 2.27% 0.01% libgit2_clar libgit2_clar [.] git_zstream_get_output + 0.55% 0.00% libgit2_clar libgit2_clar [.] git_zstream_get_output (inlined) 0.18% 0.00% libgit2_clar libgit2_clar [.] git_zstream_init 0.02% 0.00% libgit2_clar libgit2_clar [.] git_zstream_reset 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_free (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_done 0.00% 0.00% libgit2_clar libgit2_clar [.] git_zstream_free Samples: 97K of event 'cycles', Event count (approx.): 80862210917 Children Self Command Shared Objec Symbol + 2.55% 0.01% libgit2_clar libgit2_clar [.] deflate + 2.25% 1.41% libgit2_clar libgit2_clar [.] deflate_slow + 1.10% 0.52% libgit2_clar libgit2_clar [.] inflate 0.36% 0.00% libgit2_clar libgit2_clar [.] write_deflate 0.30% 0.03% libgit2_clar libgit2_clar [.] deflate_fast 0.28% 0.15% libgit2_clar libgit2_clar [.] inflate_fast_chunk_ 0.19% 0.19% libgit2_clar libgit2_clar [.] inflate_table 0.17% 0.01% libgit2_clar libgit2_clar [.] inflateInit_ 0.16% 0.00% libgit2_clar libgit2_clar [.] inflateInit2_ (inlined) 0.15% 0.00% libgit2_clar libgit2_clar [.] deflateInit_ 0.15% 0.00% libgit2_clar libgit2_clar [.] deflateInit2_ 0.11% 0.01% libgit2_clar libgit2_clar [.] adler32_z 0.09% 0.09% libgit2_clar libgit2_clar [.] adler32_simd_ 0.05% 0.00% libgit2_clar libgit2_clar [.] deflateReset (inlined) 0.05% 0.00% libgit2_clar libgit2_clar [.] deflate_read_buf 0.03% 0.00% libgit2_clar libgit2_clar [.] inflateEnd 0.02% 0.00% libgit2_clar libgit2_clar [.] deflateReset 0.01% 0.00% libgit2_clar libgit2_clar [.] deflateEnd 0.01% 0.01% libgit2_clar libgit2_clar [.] inflateReset2 0.01% 0.00% libgit2_clar libgit2_clar [.] inflateReset (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] adler32 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateResetKeep (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] deflateResetKeep 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] inflateStateCheck (inlined) 0.00% 0.00% libgit2_clar libgit2_clar [.] deflateStateCheck (inlined) ```
e0013cc
to
83265b3
Compare
Done with 100% fewer submodules (although folks that opt into the Chromium flavor are now required to use CMake 3.11. there's always a trade-off u_u). |
Could this be a backend selection like we do for TLS instead of adding more options, so you'd do like |
Now `USE_BUNDLED_ZLIB` can be set to the string `Chromium` to enable the Chromium implementation of zlib.
sure thing! |
Ooh, this is nice. Thanks @lhchavez for not adding submodules to our tree 😀 and @carlosmn for this clever idea. |
This change builds libgit2 using Chromium's zlib implementation by
invoking cmake with
-DUSE_BUNDLED_ZLIB=Chromium
,which is ~10% faster than the bundled zlib for the core::zstream suite.
This version of zlib has some optimizations:
a) Decompression (Intel+ARM): inflate_fast, adler32, crc32, etc.
b) Compression (Intel): fill_window, longest_match, hash function, etc.
Due to the introduction of SIMD optimizations, and to get the maximum
performance out of this fork of zlib, this requires an x86_64 processor
with SSE4.2 and CLMUL (anything Westmere or later, ~2010). The Chromium
zlib implementation also supports ARM with NEON, but it has not been
enabled in this patch.
Performance
TL;DR: Running just
./libgit2_clar -score::zstream
100 times in a looptook 0:56.30 before and 0:50.67 after (~10% reduction!).
The bundled and system zlib implementations on an Ubuntu Focal system
perform relatively similar (the bundled one is marginally better due to
the compiler being able to inline some functions), so only the bundled
and Chromium zlibs were compared.
For a more balanced comparison (to ensure that nothing regressed
overall),
libgit2_clar
underperf
was also run, and the zlib-relatedfunctions were compared.
Bundled
Chromium