The following results were measured on a Laptop with an Intel i7-8650U CPU running Ubuntu 18.04. On the laptop, hyperthreading and turbo-boost were switched off, the CPU frequency was set to a fixed value of 1900MHz and the benchmark was run in its own cpuset as root user (see benchmarks/bash/benchmark_start.sh).
For the native benchmarks I've used the program zbench.c and LD_PRELOADed the various different implementations as can be seen in the harness script run-native-deflate.sh. For "isa-l", wrapper functions are used which call the correponding inflate/deflate functions of the "isa-l" API.
For deflation I've measured throughput (in kb/ms) and the compression ratio for each compression level (-2 and 1 - 9 for "zlib-ipp", 0 - 3 for "isa-l" and 1 - 9 for all the other libraries) for every file from the Silesia text corpus.
The next graph shows the deflation throughput of all implementations at the default compression level 6 (for "isa-l" I use level 3 which is the closest to "zlib"'s level 6 from a compression ratio point of view). The original "zlib"'s throughput is defined to be 100%. You can click on a graph to get a larger version.
The following set of graphs shows the deflate throughput and compression ratio for each of the Silesia files depending on the compression level. Click on a graph for a larger version.
And finally a comparison of the inflation throughput of the various versions. As input data for all implementations I took the original Silesia files compressed with the original "zlib" version at the default compression level 6.