Use a switch statement to select the search function instead of an
indirect function call. This results in a sizable performance win.
This PR is a modification of the approach taken in PR #2828.
When I measured performance for that commit, it was neutral.
However, I now see a performance regression on gcc, but still
neutral on clang. I'm measuring on the same platform, but with
newer compilers. The new approach beats both the current dev
branch and the baseline before PR #2828 was merged.
This PR is necessary for Issue #3275, to update zstd in the kernel.
Without this PR there is a large regression in greedy - btlazy2
compression speed. With this PR it is about neutral.
gcc version: 12.2.0
clang version: 14.0.6
dataset: silesia.tar
| Compiler | Level | Dev Speed (MB/s) | PR Speed (MB/s) | Delta |
|----------|-------|------------------|-----------------|--------|
| gcc | 5 | 102.6 | 113.7 | +10.8% |
| gcc | 7 | 66.6 | 74.8 | +12.3% |
| gcc | 9 | 51.5 | 58.9 | +14.3% |
| gcc | 13 | 14.3 | 14.3 | +0.0% |
| clang | 5 | 108.1 | 114.8 | +6.2% |
| clang | 7 | 68.5 | 72.3 | +5.5% |
| clang | 9 | 53.2 | 56.2 | +5.6% |
| clang | 13 | 14.3 | 14.7 | +2.8% |
The binary size stays just about the same for clang and gcc, measured
using the `size` command:
| Compiler | Branch | Text | Data | BSS | Total |
|----------|--------|---------|------|-----|---------|
| gcc | dev | 1127950 | 3312 | 280 | 1131542 |
| gcc | PR | 1123422 | 2512 | 280 | 1126214 |
| clang | dev | 1046254 | 3256 | 216 | 1049726 |
| clang | PR | 1048198 | 2296 | 216 | 1050710 |
Comparing 4B instead of comparing 1B in ZSTD_noDict
mode, thus it can avoid cases like match in match[ml]
but mismatch in match[ml-3]..match[ml-1]. So the call
count of ZSTD_count can be reduced.
Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: I3449ea423d5c8e8344f75341f19a2d1643c703f6
We found that movemask is not used properly or consumes too much CPU.
This effort helps to optimize the movemask emulation on ARM.
For level 8-9 we saw 3-5% improvements. For level 10 we say 1.5%
improvement.
The key idea is not to use pure movemasks but to have groups of bits.
For rowEntries == 16, 32 we are going to have groups of size 4 and 2
respectively. It means that each bit will be duplicated within the group
Then we do AND to have only one bit set in the group so that iteration
with lowering bit `a &= (a - 1)` works as well.
Also, aarch64 does not have rotate instructions for 16 bit, only for 32
and 64, that's why we see more improvements for level 8-9.
vshrn_n_u16 instruction is used to achieve that: vshrn_n_u16 shifts by
4 every u16 and narrows to 8 lower bits. See the picture below. It's
also used in
[Folly](c570259008/folly/container/detail/F14Table.h (L446)).
It also uses 2 cycles according to Neoverse-N{1,2} guidelines.
64 bit movemask is already well optimized. We have ongoing experiments
but were not able to validate other implementations work reliably faster.
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.
the variable has only very limited usage,
being only used once at the beginning of the block for prefetching only,
hence the error had no impact on compression ratio.
mostly for maintenance convenience.
Performance wise, there is very little change,
slightly faster for slog 3 & 4,
neutral or very slightly negative for slot 5 & 6.
short-tests-0 were silently failing. I think because of the && make clean construction. Switch to ; instead.
Also fix all the test failures that were exposed.
`make all` is failing on CircleCI because it is missing Docker. Move that test
to GitHub actions, and switch the pedantic CircleCI test to `make allmost`.
Speed up compilation times by moving each specialized search function
into its own function. This is faster because compilers can handle many
smaller functions much faster than one gigantic function. The previous
approach generated one giant function with `switch` statements and
inlining to select the implementation.
| Compiler | Flags | Dev Time (s) | PR Time (s) | Delta |
|----------|-------------------------------------|--------------|-------------|-------|
| gcc | -O3 | 16.5 | 5.6 | -66% |
| gcc | -O3 -g -fsanitize=address,undefined | 158.9 | 38.2 | -75% |
| clang | -O3 | 36.5 | 5.5 | -85% |
| clang | -O3 -g -fsanitize=address,undefined | 27.8 | 17.5 | -37% |
This also reduces the binary size because the search functions are no
longer inlined into the main body.
| Compiler | Dev libzstd.a Size (B) | PR libzstd.a Size (B) | Delta |
|----------|------------------------|-----------------------|-------|
| gcc | 1563868 | 1308844 | -16% |
| clang | 1924372 | 1376020 | -28% |
Finally, the performance is not impacted significantly by this change,
in fact we generally see a small speed boost.
| Compiler | Level | Dev Speed (MB/s) | PR Speed (MB/s) | Delta |
|----------|-------|------------------|-----------------|-------|
| gcc | 5 | 110.6 | 110.0 | -0.5% |
| gcc | 7 | 70.4 | 72.2 | +2.5% |
| gcc | 9 | 53.2 | 53.5 | +0.5% |
| gcc | 13 | 12.7 | 12.9 | +1.5% |
| clang | 5 | 113.9 | 110.4 | -3.0% |
| clang | 7 | 67.7 | 70.6 | +4.2% |
| clang | 9 | 51.9 | 52.2 | +0.5% |
| clang | 13 | 12.4 | 13.3 | +7.2% |
The compression strategy is unmodified in this PR, so the compressed size
should be exactly the same. I may have a follow up PR to slightly improve
the compression ratio, if it doesn't cost too much speed.
Fix underflow of `nbCompares` by switching to an `int` and comparing
`nbCompares > 0`. This is a minimal fix, because I don't want to change
the logic. These loops seem to be doing `nbCompares + 1` comparisons.
The bug was reported by Dan Carpenter and found by Smatch static
checker.
https://lore.kernel.org/all/20211008063704.GA5370@kili/
turns out, it's possible to constify MatchState* parameter
in some parts of the binary tree algorithm,
making it a pure read-only parameter,
as opposed to a mutable state.
This is supposed to be helpful for both maintenance and the compiler.
This PR fixes an incorrect comparison in figuring out `minChain` in
`ZSTD_dedicatedDictSearch_lazy_loadDictionary()`. This incorrect comparison
had been masked by the fact that `idx` was always 1, until @terrelln changed
that in #2726.
Credit-to: OSS-Fuzz
* Flatten ZSTD_row_getMatchMask
* Remove the SIMD abstraction layer.
* Add big endian support.
* Align `hashTags` within `tagRow` to a 16-byte boundary.
* Switch SSE2 to use aligned reads.
* Optimize scalar path using SWAR.
* Optimize neon path for `n == 32`
* Work around minor clang issue for NEON (https://bugs.llvm.org/show_bug.cgi?id=49577)
* replace memcpy with MEM_readST
* silence alignment warnings
* fix neon casts
* Update zstd_lazy.c
* unify simd preprocessor detection (#3)
* remove duplicate asserts
* tweak rotates
* improve endian detection
* add cast
there is a fun little catch-22 with gcc: result from pmovmskb has to be cast to uint32_t to avoid a zero-extension
but must be uint16_t to get gcc to generate a rotate instruction..
* more casts
* fix casts
better work-around for the (bogus) warning: unary minus on unsigned
