Use only level 1 and 2 with python-isal

[rhpvorderman]

Zlib compression level 0 outputs uncompressed deflate format, which consists of 64K blocks in the following format:
1 byte signifying data is uncompressed. (Actually it is a two-bit code, but I am not going to rewrite the spec here). 2 bytes with an unsigned 16 bit integer that holds the length of the following block. 2 bytes with the unsigned 16 bit integer complemented for parity checking. Then follows the uncompressed block of up to 64K in length.

ISA-L however compresses the data with a special algorithm. This is an area where zlib and ISA-L differ. That is okay, but xopen should behave somewhat consistently in terms of output for the same parameters and not let it be backend dependent. Also level 0 should be quite fast on zlib and zlib-ng as well, there is no reason to use ISA-L for speed here.

Next is level 3. For FASTQ data, I find that this has very little extra value in terms of compression. Other libraries, such as zlib-ng and zlib do a better job here. Also it is much slower than levels 1 or 2, so it is a bit of a useless level for FASTQ data. ISA-L level 1 and 2 outperform anything by huge margins and since level 1 is the default I don't see why level 3 can't be done by a library that compresses better.

All right: benchmarks. Not done very cleanly, with processes runniing in the background and n=1, but at least it gives a good ballpark indication.

before:

python benchmark_levels.py ~/test/nanopore_100000reads.fastq ramdisk/test.fastq.gz
Compressed at level 0 with 0 threads; filesize: 1241299982, time: 4.584214687347412 seconds
Compressed at level 1 with 0 threads; filesize: 987346940, time: 4.9248528480529785 seconds
Compressed at level 2 with 0 threads; filesize: 988121796, time: 5.152158260345459 seconds
Compressed at level 3 with 0 threads; filesize: 990344221, time: 8.79483151435852 seconds
Compressed at level 4 with 0 threads; filesize: 953739158, time: 34.5846734046936 seconds
Compressed at level 5 with 0 threads; filesize: 945597849, time: 50.54869747161865 seconds
Compressed at level 6 with 0 threads; filesize: 936937471, time: 96.95266628265381 seconds
Compressed at level 0 with 1 threads; filesize: 1241490149, time: 4.322765588760376 seconds
Compressed at level 1 with 1 threads; filesize: 986751063, time: 4.7672388553619385 seconds
Compressed at level 2 with 1 threads; filesize: 987497150, time: 4.939749240875244 seconds
Compressed at level 3 with 1 threads; filesize: 989851198, time: 8.73461651802063 seconds
Compressed at level 4 with 1 threads; filesize: 934568019, time: 27.754233360290527 seconds
Compressed at level 5 with 1 threads; filesize: 932150001, time: 33.990896701812744 seconds
Compressed at level 6 with 1 threads; filesize: 926300592, time: 57.50165820121765 seconds

after:

 python benchmark_levels.py ~/test/nanopore_100000reads.fastq ramdisk/test.fastq.gz
Compressed at level 0 with 0 threads; filesize: 1719300551, time: 1.0877585411071777 seconds
Compressed at level 1 with 0 threads; filesize: 987346940, time: 4.940786123275757 seconds
Compressed at level 2 with 0 threads; filesize: 988121796, time: 5.1339592933654785 seconds
Compressed at level 3 with 0 threads; filesize: 944408772, time: 23.135488033294678 seconds
Compressed at level 4 with 0 threads; filesize: 934668651, time: 28.04243803024292 seconds
Compressed at level 5 with 0 threads; filesize: 932283422, time: 34.2128050327301 seconds
Compressed at level 6 with 0 threads; filesize: 926345773, time: 57.90689778327942 seconds
Compressed at level 0 with 1 threads; filesize: 1719316928, time: 0.9414026737213135 seconds
Compressed at level 1 with 1 threads; filesize: 986751063, time: 4.759371519088745 seconds
Compressed at level 2 with 1 threads; filesize: 987497150, time: 4.957266092300415 seconds
Compressed at level 3 with 1 threads; filesize: 944154982, time: 22.834784269332886 seconds
Compressed at level 4 with 1 threads; filesize: 934568019, time: 27.81139039993286 seconds
Compressed at level 5 with 1 threads; filesize: 932150001, time: 33.65570855140686 seconds
Compressed at level 6 with 1 threads; filesize: 926300592, time: 57.73633337020874 seconds

The benchmark show that there was also a fault in the logic introduced at some point (EDIT: due to an if->elif replacement in the refactoring). If python-isal with threads=0 threw a ValueError for a bad compression level, zlib-ng was skipped and zlib used instead. This is now also fixed.

As you can see zlib-ng gets a nice improvement over python-isal in terms of compressed size at level 3. "Compression" at level 0 is now non-existent as expected and as a result is a lot faster, also as expected.

Overall I am very happy with the minimal code size required to affect these changes. Python-isal is a special case, so I don't mind a bit more specific special-casing in the code.

[rhpvorderman]

Thanks for the quick merge!