for proper estimation of symbol's weights
when using dictionary compression.
Note : using only huffman costs is not good enough,
presumably because sequence symbol costs are incorrect.
Update code documentation, and properly names a few "magic constants".
Also, HUF_compress_internal() gets a cleaner way
to determine size of tables inside workspace.
Pathological samples may result in literal section being incompressible.
This case is now detected,
and literal distribution is replaced by one that can be written into the dictionary.
The compressor always reuses the existing Huffman table if the literals
size is at most 1 KiB. If the compression strategy is `ZSTD_lazy` or
stronger always check to see if reusing the previous table or creating
a new table is better.
This doesn't yet weigh in decompression speed. I don't want to add any
heuristics there until I have real data to work with to ensure that the
heuristic works for at least one use case, preferably more.
* Compressor saves most recently used Huffman table and reuses it
if it produces better results.
* I attempted to preserve CPU usage profile.
I intentionally left all of the existing heuristics in place.
There is only a speed difference on the second block and later.
When compressing large enough blocks (say >= 4 KiB) there is
no significant difference in compression speed.
Dictionary compression of one block is the same speed for blocks
with literals <= 1 KiB, and after that the difference is not
very significant.
* In the synthetic data, with blocks 10 KB or smaller, most blocks
can't use repeated tables because the previous block did not
contain a symbol that the current block contains.
Once blocks are about 12 KB or more, most previous blocks have
valid Huffman tables for the current block, and the compression
ratio and decompression speed jumped.
* In silesia blocks as small as 4KB can frequently reuse the
previous Huffman table (85%), but it isn't as profitable, and
the previous Huffman table only gets used about 3% of the time.
* Microbenchmarks show that `HUF_validateCTable()` takes ~55 ns
and `HUF_estimateCompressedSize()` takes ~35 ns.
They are decently well optimized, the first versions took 90 ns
and 120 ns respectively. `HUF_validateCTable()` could be twice as
fast, if we cast the `HUF_CElt*` to a `U32*` and compare to 0.
However, `U32` has an alignment of 4 instead of 2, so I think that
might be undefined behavior.
* I've ran `zstreamtest` compiled normally, with UASAN and with MSAN
for 4 hours each.
The worst case for the speed difference is a bunch of small blocks
in the same frame. I modified `bench.c` to compress the input in a
single frame but with blocks of the given block size, set by `-B`.
Benchmarks on level 1:
| Program | Block size | Corpus | Ratio | Compression MB/s | Decompression MB/s |
|-----------|------------|-----------|-------|------------------|--------------------|
| zstd.base | 256 | synthetic | 2.364 | 110.0 | 297.0 |
| zstd | 256 | synthetic | 2.367 | 108.9 | 297.0 |
| zstd.base | 256 | silesia | 2.204 | 93.8 | 415.7 |
| zstd | 256 | silesia | 2.204 | 93.4 | 415.7 |
| zstd.base | 512 | synthetic | 2.594 | 144.2 | 420.0 |
| zstd | 512 | synthetic | 2.599 | 141.5 | 425.7 |
| zstd.base | 512 | silesia | 2.358 | 118.4 | 432.6 |
| zstd | 512 | silesia | 2.358 | 119.8 | 432.6 |
| zstd.base | 1024 | synthetic | 2.790 | 192.3 | 594.1 |
| zstd | 1024 | synthetic | 2.794 | 192.3 | 600.0 |
| zstd.base | 1024 | silesia | 2.524 | 148.2 | 464.2 |
| zstd | 1024 | silesia | 2.525 | 148.2 | 467.6 |
| zstd.base | 4096 | synthetic | 3.023 | 300.0 | 1000.0 |
| zstd | 4096 | synthetic | 3.024 | 300.0 | 1010.1 |
| zstd.base | 4096 | silesia | 2.779 | 223.1 | 623.5 |
| zstd | 4096 | silesia | 2.779 | 223.1 | 636.0 |
| zstd.base | 16384 | synthetic | 3.131 | 350.0 | 1150.1 |
| zstd | 16384 | synthetic | 3.152 | 350.0 | 1630.3 |
| zstd.base | 16384 | silesia | 2.871 | 296.5 | 883.3 |
| zstd | 16384 | silesia | 2.872 | 294.4 | 898.3 |
Yann Collet
Nick Terrell
Stella Lau
Przemyslaw Skibinski
Yann Collet