* Bump `WILDCOPY_OVERLENGTH` to 16 to fix the wildcopy overread.
* Optimize `ZSTD_wildcopy()` by removing unnecessary branches and
unrolling the loop.
* Extract `ZSTD_overlapCopy8()` into its own function.
* Add `ZSTD_safecopy()` for `ZSTD_execSequenceEnd()`. It is
optimized for single long sequences, since that is the important
case that can end up in `ZSTD_execSequenceEnd()`. Without this
optimization, decompressing a block with 1 long match goes
from 5.7 GB/s to 800 MB/s.
* Refactor `ZSTD_execSequenceEnd()`.
* Increase the literal copy shortcut to 16.
* Add a shortcut for offset >= 16.
* Simplify `ZSTD_execSequence()` by pushing more cases into
`ZSTD_execSequenceEnd()`.
* Delete `ZSTD_execSequenceLong()` since it is exactly the
same as `ZSTD_execSequence()`.
clang-8 seeds +17.5% on silesia and +21.8% on enwik8.
gcc-9 sees +12% on silesia and +15.5% on enwik8.
TODO: More detailed measurements, and on more datasets.
Crdit to OSS-Fuzz for finding the wildcopy overread.
This led to a nasty edgecase, where index reduction for modes that don't use
the h3 table would have a degenerate table (size 4) allocated and marked clean,
but which would not be re-indexed.
The source matchState is potentially at a lower current index, which means
that any extra table space not overwritten by the copy may now contain
invalid indices. The simple solution is to unconditionally shrink the valid
table area to just the area overwritten.
Summary: The idea behind wildcopy is that it can be cheaper to copy more bytes (say 8) than it is to copy less (say, 3). This change takes that further by exploiting some properties:
1. it's almost always OK to copy 16 bytes instead of 8, which means fewer copy instructions, and fewer branches
2. A 16 byte chunk size means that ~90% of wildcopy invocations will have a trip count of 1, so branch prediction will be improved.
Speedup on Xeon E5-2680v4 is in the range of 3-5%.
Measured wildcopy length distributions on silesia.tar:
level <=8 <=16 <=24 >24
1 78.05% 11.49% 3.52% 6.94%
3 82.14% 8.99% 2.44% 6.43%
6 85.81% 6.51% 2.92% 4.76%
8 83.02% 7.31% 3.64% 6.03%
10 84.13% 6.67% 3.29% 5.91%
15 77.58% 7.55% 5.21% 9.66%
16 80.07% 7.20% 3.98% 8.75%
Test Plan: benchmark silesia, make check
The nbSeq "short" format (1-byte)
is compatible with any value < 128.
However, the code would cautiously only accept values < 127.
This is not an error, because the general 2-bytes format
is compatible with small values < 128.
Hence the inefficiency never triggered any warning.
Spotted by Intel's Smita Kumar.
* [ldm] Fix bug in overflow correction with large job size
* [zstdmt] Respect ZSTDMT_JOBSIZE_MAX (1G in 64-bit mode)
* [test] Add test that exposes the bug
Sadly the test fails on our CI because it uses too much memory, so
I had to comment it out.
When we wrote one byte beyond the end of the buffer for RLE
blocks back in 1.3.7, we would then have `op > oend`. That is
a problem when we use `oend - op` for the size of the destination
buffer, and allows further writes beyond the end of the buffer for
the rest of the function. Lets assert that it doesn't happen.
Also : minor speed optimization :
shortcut to ZSTD_reset_matchState() rather than the full reset process.
It still needs to be completed with ZSTD_continueCCtx() for proper initialization.
Also : changed position of LDM hash tables in the context,
so that the "regular" hash tables can be at a predictable position,
hence allowing the shortcut to ZSTD_reset_matchState() without complex conditions.
* Extract the overflow correction into a helper function.
* Load the dictionary `ZSTD_CHUNKSIZE_MAX = 512 MB` bytes at a time
and overflow correct between each chunk.
Data corruption could happen when all these conditions are true:
* You are using multithreading mode
* Your overlap size is >= 512 MB (implies window size >= 512 MB)
* You are using a strategy >= ZSTD_btlazy
* You are compressing more than 4 GB
The problem is that when loading a large dictionary we don't do
overflow correction. We can only load 512 MB at a time, and may
need to do overflow correction before each chunk.
We would only skip at most 192 bytes at a time before this diff.
This was added to optimize long matches and skip the middle of the
match. However, it doesn't handle the case of repetitive data.
This patch keeps the optimization, but also handles repetitive data
by taking the max of the two return values.
```
> for n in $(seq 9); do echo strategy=$n; dd status=none if=/dev/zero bs=1024k count=1000 | command time -f %U ./zstd --zstd=strategy=$n >/dev/null; done
strategy=1
0.27
strategy=2
0.23
strategy=3
0.27
strategy=4
0.43
strategy=5
0.56
strategy=6
0.43
strategy=7
0.34
strategy=8
0.34
strategy=9
0.35
```
At level 19 with multithreading the compressed size of `silesia.tar` regresses 300 bytes, and `enwik8` regresses 100 bytes.
In single threaded mode `enwik8` is also within 100 bytes, and I didn't test `silesia.tar`.
Fixes Issue #1634.
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.
It's re-synchronized with nextToUpdate at beginning of each block.
It only needs to be tracked from within zstd_opt block parser.
Made the logic clear, so that no code tried to maintain this variable.
An even better solution would be to make nextToUpdate3
an internal variable of ZSTD_compressBlock_opt_generic().
That would make it possible to remove it from ZSTD_matchState_t,
thus restricting its visibility to only where it's actually useful.
This would require deeper changes though,
since the matchState is the natural structure to transport parameters into and inside the parser.
ZSTDMT was broken when compiled without ZSTD_MULTITHREAD defined,
because `ZSTD_CCtx_setParameter(cctx, ZSTD_c_nbWorkers, nbWorkerss)`
failed. It was detected by the MSVC test which runs the fuzzer with
multithreading disabled.
This is a very niche use case of a deprecated API, because the API is
inefficient and synchronous, since `threading.h` will be synchronous.
Users almost certainly don't want this, and anyone who tested their code
should realize that it is broken. Therefore, I think it is safe to
require `ZSTD_MULTITHREAD` to be defined to use ZSTDMT.
`ZSTD_compress2()` wouldn't wait for multithreaded compression to
finish. We didn't find this because ZSTDMT will block when it can
compress all in one go, but it can't do that if it doesn't have enough
output space, or if `ZSTD_c_rsyncable` is enabled.
Since we will already sometimes block when using `ZSTD_e_end`, I've
changed `ZSTD_e_end` and `ZSTD_e_flush` to guarantee maximum forward
progress. This simplifies the API, and helps users avoid the easy bug
that was made in `ZSTD_compress2()`
* Found by the libfuzzer fuzzers.
* Added a test case that catches the problem.
* I will make the fuzzers sometimes allocate less than
`ZSTD_compressBound()` output space.
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
It wasn't using the ZSTD_CCtx_params correctly. It must actualize
the compression parameters by calling ZSTD_getCParamsFromCCtxParams()
to get the real window log.
Tested by updating the streaming memory usage example in the next
commit. The CHECK() failed before this patch, and passes after.
I also added a unit test to zstreamtest.c that failed before this
patch, and passes after.
* After loading a dictionary only create the cdict once we've started the
compression job. This allows the user to pass the dictionary before they
set other settings, and is in line with the rest of the API.
* Add tests that mix the 3 dictionary loading APIs.
* Add extra tests for `ZSTD_CCtx_loadDictionary()`.
* The first 2 tests added fail before this patch.
* Run the regression test suite.
The order you set parameters in the advanced API is not supposed to matter.
However, once you call `ZSTD_CCtx_refCDict()` the compression parameters
cannot be changed. Remove that restriction, and document what parameters
are used when using a CDict.
If the CCtx is in dictionary mode, then the CDict's parameters are used.
If the CCtx is not in dictionary mode, then its requested parameters are
used.
* Move all ZSTDMT parameter setting code to ZSTD_CCtxParams_*Parameter().
ZSTDMT now calls these functions, so we can keep all the logic in the
same place.
* Clean up `ZSTD_CCtx_setParameter()` to only add extra checks where needed.
* Clean up `ZSTDMT_initJobCCtxParams()` by copying all parameters by default,
and then zeroing the ones that need to be zeroed. We've missed adding several
parameters here, and it makes more sense to only have to update it if you
change something in ZSTDMT.
* Add `ZSTDMT_cParam_clampBounds()` to clamp a parameter into its valid
range. Use this to keep backwards compatibility when setting ZSTDMT parameters,
which clamp into the valid range.
as suggested in #1441.
generally U32 and unsigned are the same thing,
except when they are not ...
case : 32-bit compilation for MIPS (uint32_t == unsigned long)
A vast majority of transformation consists in transforming U32 into unsigned.
In rare cases, it's the other way around (typically for internal code, such as seeds).
Among a few issues this patches solves :
- some parameters were declared with type `unsigned` in *.h,
but with type `U32` in their implementation *.c .
- some parameters have type unsigned*,
but the caller user a pointer to U32 instead.
These fixes are useful.
However, the bulk of changes is about %u formating,
which requires unsigned type,
but generally receives U32 values instead,
often just for brevity (U32 is shorter than unsigned).
These changes are generally minor, or even annoying.
As a consequence, the amount of code changed is larger than I would expect for such a patch.
Testing is also a pain :
it requires manually modifying `mem.h`,
in order to lie about `U32`
and force it to be an `unsigned long` typically.
On a 64-bit system, this will break the equivalence unsigned == U32.
Unfortunately, it will also break a few static_assert(), controlling structure sizes.
So it also requires modifying `debug.h` to make `static_assert()` a noop.
And then reverting these changes.
So it's inconvenient, and as a consequence,
this property is currently not checked during CI tests.
Therefore, these problems can emerge again in the future.
I wonder if it is worth ensuring proper distinction of U32 != unsigned in CI tests.
It's another restriction for coding, adding more frustration during merge tests,
since most platforms don't need this distinction (hence contributor will not see it),
and while this can matter in theory, the number of platforms impacted seems minimal.
Thoughts ?
The problem was already masked,
due to no longer accepting tiny blocks for statistics.
But in case it could still happen with not-so-tiny blocks,
there is a stricter control which ensures that
nothing was already loaded prior to statistics collection.
depending on initialization,
the first byte of a new frame was invalidated or not.
As a consequence, one match opportunity was available or not,
resulting in slightly different compressed sizes
(on average, 1 or 2 bytes once every 20 frames).
It impacted ratio comparison between one-shot and streaming modes.
This fix makes the first byte of a new frame always a valid match.
Now compressed size is always the same.
It also improves compressed size by a negligible amount.
* Fix `ZSTD_estimateCCtxSize()` with negative levels.
* Fix `ZSTD_estimateCStreamSize()` with negative levels.
* Add a unit test to test for this error.
When srcSize is small,
the nb of symbols produced is likely too small to warrant dedicated probability tables.
In which case, predefined distribution tables will be used instead.
There is a cheap algorithm in btultra initialization :
it presumes default distribution will be used if srcSize <= 1024.
btultra2 now uses the same threshold to shut down probability estimation,
since measured frequencies won't be used at entropy stage,
and therefore relying on them to determine sequence cost is misleading,
resulting in worse compression ratios.
This fixes btultra2 performance issue on very small input.
Note that, a proper way should be
to determine which symbol is going to use predefined probaility
and which symbol is going to use dynamic ones.
But the current algorithm is unable to make a "per-symbol" decision.
So this will require significant modifications.
from overlapSizeLog.
Reasoning :
`overlapLog` is already used everwhere, in the code, command line and documentation.
`ZSTD_c_overlapSizeLog` feels unnecessarily different.
ZSTD_compress_generic() is renamed ZSTD_compressStream2().
Note that, for the time being,
the "stable" API and advanced one use different parameter planes :
setting parameters using the advanced API does not influence ZSTD_compressStream()
and using ZSTD_initCStream() does not influence parameters for ZSTD_compressStream2().
which now accepts an enum,
to distinguish between resetting the session, or the parameters (or both).
removed ZSTD_CCtx_resetParameters(), which is redundant.
start replacing invocation of ZSTD_CCtx_reset*() functions
Updated advanced API documentation
trimmed down amount of API staged in RC,
in particular, all functions related to ZSTD_CCtxParams()
seem too advanced.
answering #1407.
Also : removed obsolete function ZSTD_setDStreamParameter()
which could only be used with one parameter (DStream_p_maxWindowSize).
Now replaced by ZSTD_DCtx_setWindowSize() (which exists since a few revisions)
changed workspace parameter convention
to always provide workspaceSize,
so that size can be explicitly checked.
Also, use more enum to make the meaning of some parameters more explicit.
This capability is not needed / used in the current unit of work. I'll
re-introduce it later, when we start allowing users to override the deduced
working context logs.
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.
