this meant to abstract the sumtype representation required
to transfert `offcode` to `ZSTD_storeSeq()`.
Unfortunately, the sumtype numeric representation is currently a leaky abstraction
that has permeated many other parts of the code,
especially within `zstd_lazy.c` and also within `zstd_opt.c` and `zstd_compress.c`.
While this PR makes a good job a transfering a large nb of call sites
to using the new macros, there are still a few sites where this transformation is more complex,
or where the numeric representation itself it used "as is".
One of the problematics area is the decision to use the numeric format of the sumtype
within the match finders of `zstd_lazy`.
This commit doesn't change the behavior, it only introduces and employes the macros,
but eventually the resulting code remains identical.
At target, if the numeric representation of the sumtype can be completely abstracted
and no other part of the code depends on it,
it will be possible to move it towards something slightly more efficient.
I couldn't find a good way to spread `ip0` and `ip1` apart when we accelerate
due to incompressible inputs. (The methods I tried slowed things down quite a
bit.)
Since we aren't splaying ip0 and ip1 apart (which would be like `0_1_2_3_`, as
opposed to the `01__23__` we were actually doing), it's a big ambitious to
increment `step` by 2. Instead, let's increment it by 1, which has the benefit
sliiightly improving compression. Speed remains pretty much unchanged.
The position updates are rewritten from `ip[N] = ip[N-1] + step` to be
`ip[N] = ip[N-2] + step`. This lets us only deal with the asymmetric spacing
of gaps at setup and then we only have to keep a single `step` variable.
This seems to work quite well on GCC and Clang!
This replicates the behavior of @terrelln's `ZSTD_fast` implementation. That
is, it always looks at adjacent pairs of positions, and only applies the
acceleration every other position. This produces a more fine-grained
acceleration.
Take the same approach as in PR #2828 [0] to remove functions that force
inline many function bodies and `switch`. Instead, create one function per
"template" combination, and then switch between these functions. This
allows the compiler to break the large function into many small
functions, which generally helps codegen.
Also, in the `extDict` modes when there is no ext-dict, call the top
level function instead of the force inlined one, to save on code size.
I'm specifically doing this because gcc on the parisc architecture doesn't
handle the large function body well, and ends up using a lot of excess
stack space. Outlining these functions fixes it.
This removes the old `ZSTD_compressBlock_fast_generic()` and renames the new
`ZSTD_compressBlock_fast_generic_pipelined()` to replace it. This is
functionally a no-op.
Unrolling the loop to handle 2 positions in each iteration allows us to reduce
the frequency of some operations that don't need to happen at every position.
One such operation is the step calculation, which is a very rough heuristic
anyways. It's fine if we do this a position later. The other operation is the
repcode check. But since the repcode check already tries expanding back one
position, we're really not missing much of importance by only trying it every
other position.
This commit also slightly reorders some operations.
Amusingly, it seems to be a non-trivial performance hit to add in final
searches or even hash table insertions during cleanup. So let's not. It seems
to not make any meaningful difference in compression ratio.
The repcode checks disallowed repcodes that are equal to `windowLow`.
This is slightly inefficient, but isn't a problem on its own. Together
with the next commit, it cause non-determinism.
* Switch to yearless copyright per FB policy
* Fix up SPDX-License-Identifier lines in `contrib/linux-kernel` sources
* Add zstd copyright/license header to the `contrib/linux-kernel` sources
* Update the `tests/test-license.py` to check for yearless copyright
* Improvements to `tests/test-license.py`
* Check `contrib/linux-kernel` in `tests/test-license.py`
* All copyright lines now have -2020 instead of -present
* All copyright lines include "Facebook, Inc"
* All licenses are now standardized
The copyright in `threading.{h,c}` is not changed because it comes from
zstdmt.
The copyright and license of `divsufsort.{h,c}` is not changed.
fast mode does the same thing as before :
it pre-emptively invalidates any index that could lead to offset > maxDistance.
It's supposed to help speed.
But this logic is performed inside zstd_fast,
so that other strategies can select a different behavior.
This PR is based on top of PR #1563.
The optimization is to process two input pointers per loop.
It is based on ideas from [igzip] level 1, and talking to @gbtucker.
| Platform | Silesia | Enwik8 |
|-------------------------|-------------|--------|
| OSX clang-10 | +5.3% | +5.4% |
| i9 5 GHz gcc-8 | +6.6% | +6.6% |
| i9 5 GHz clang-7 | +8.0% | +8.0% |
| Skylake 2.4 GHz gcc-4.8 | +6.3% | +7.9% |
| Skylake 2.4 GHz clang-7 | +6.2% | +7.5% |
Testing on all Silesia files on my Intel i9-9900k with gcc-8
| Silesia File | Ratio Change | Speed Change |
|--------------|--------------|--------------|
| silesia.tar | +0.17% | +6.6% |
| dickens | +0.25% | +7.0% |
| mozilla | +0.02% | +6.8% |
| mr | -0.30% | +10.9% |
| nci | +1.28% | +4.5% |
| ooffice | -0.35% | +10.7% |
| osdb | +0.75% | +9.8% |
| reymont | +0.65% | +4.6% |
| samba | +0.70% | +5.9% |
| sao | -0.01% | +14.0% |
| webster | +0.30% | +5.5% |
| xml | +0.92% | +5.3% |
| x-ray | -0.00% | +1.4% |
Same tests on Calgary. For brevity, I've only included files
where compression ratio regressed or was much better.
| Calgary File | Ratio Change | Speed Change |
|--------------|--------------|--------------|
| calgary.tar | +0.30% | +7.1% |
| geo | -0.14% | +25.0% |
| obj1 | -0.46% | +15.2% |
| obj2 | -0.18% | +6.0% |
| pic | +1.80% | +9.3% |
| trans | -0.35% | +5.5% |
We gain 0.1% of compression ratio on Silesia.
We gain 0.3% of compression ratio on enwik8.
I also tested on the GitHub and hg-commands datasets without a dictionary,
and we gain a small amount of compression ratio on each, as well as speed.
I tested the negative compression levels on Silesia on my
Intel i9-9900k with gcc-8:
| Level | Ratio Change | Speed Change |
|-------|--------------|--------------|
| -1 | +0.13% | +6.4% |
| -2 | +4.6% | -1.5% |
| -3 | +7.5% | -4.8% |
| -4 | +8.5% | -6.9% |
| -5 | +9.1% | -9.1% |
Roughly, the negative levels now scale half as quickly. E.g. the new
level 16 is roughly equivalent to the old level 8, but a bit quicker
and smaller. If you don't think this is the right trade off, we can
change it to multiply the step size by 2, instead of adding 1. I think
this makes sense, because it gives a bit slower ratio decay.
[igzip]: https://github.com/01org/isa-l/tree/master/igzip
We pre-hash the ptr for the dict match state sometimes. When that actually
happens, a hashlog of 0 can produce undefined behavior (right shift a long
long by 64). Only applies to unoptimized compilations, since when
optimizations are applied, those hash operations are dropped when we're not
actually in dms mode.
experimental function ZSTD_compressBlock() is designed for very small data in mind,
for situation where saving the ~12 bytes of frame header can actually make a difference.
Some systems though may have to deal with small and large data entangled.
If it's larger than a block (> 128KB), compressBlock() cannot compress them in one round.
That's why it's possible to compress in multiple rounds.
This is a chain of compressed blocks.
Some users push this capability to the limit, encoding gigantic chain of blocks.
On crossing the 4GB limit, some internal overflow occurs.
This fix moves the overflow correction mechanism higher in the call chain,
so that it's applied also to gigantic chains of blocks.
Added a test case in fuzzer.c, which crashes before the fix, and pass now.
