Looking at the __builtin_expect in ZSTD_decodeSequence:
{ size_t offset;
#if defined(__clang__)
if (LIKELY(ofBits > 1)) {
#else
if (ofBits > 1) {
#endif
ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
From profile-annotated assembly, the probability of ofBits > 1 is about 75%
(101k counts out of 135k counts). This is much smaller than the recommended
likelihood to use __builtin_expect which is 99%. As a result, clang moved the
else block further away which hurts cache locality. Removing this
__built_expect along with two others in ZSTD_decodeSequence gave better
performance when PGO is enabled. I suggest to remove these branch hints and
rely on PGO which leverages runtime profiles from actual workload to calculate
branch probability instead.
Inlining `BIT_reloadDStream` provided >3% decompression speed improvement for
clang PGO-optimized zstd binary, measured using the Silesia corpus with
compression level 1. The win comes from improved register allocation which leads
to fewer spills and reloads. Take a look at this comparison of
profile-annotated hot assembly before and after this change:
https://www.diffchecker.com/UjDGIyLz/. The diff is a bit messy, but notice three
fewer moves after inlining.
In general LLVM's register allocator works better when it can see more code. For
example, when the register allocator sees a call instruction, it partitions the
registers into caller registers and callee registers, and it is not free to do
whatever it wants with all the registers for the current function. Inlining the
callee lets the register allocation access all registers and use them more
flexsibly.
Summary:
Freeing an uninitialized pointer is undefined behavior. This caused a segfault
when compiling the benchmark with Clang -O3 and benching decompression.
V2: always create compressInstructions but check if cctxParams is NULL before
setting CCtx params to avoid segfault.
Test Plan:
make and run
Summary:
Added an option -p# where -p0 (default) sets the aggregation method to fastest
speed while -p1 sets the aggregation method to median. Also added a new column
in the csv file to report this option's value.
Test Plan:
``
$ ./largeNbDicts -1 --nbDicts=1 -D ~/benchmarks/html/html_8_16K.32K.dict
~/benchmarks/html/html_8_16K/*
loading 7450 files...
created src buffer of size 83.4 MB
split input into 7450 blocks
loading dictionary /home/zhuhan/benchmarks/html/html_8_16K.32K.dict
compressing at level 1 without dictionary : Ratio=3.03 (28827863 bytes)
compressed using a 32768 bytes dictionary : Ratio=4.28 (20410262 bytes)
generating 1 dictionaries, using 0.1 MB of memory
Compression Speed : 306.0 MB/s
Fastest Speed : 310.6 MB/s
$ ./largeNbDicts -1 --nbDicts=1 -p1 -D ~/benchmarks/html/html_8_16K.32K.dict
~/benchmarks/html/html_8_16K/*
loading 7450 files...
created src buffer of size 83.4 MB
split input into 7450 blocks
loading dictionary /home/zhuhan/benchmarks/html/html_8_16K.32K.dict
compressing at level 1 without dictionary : Ratio=3.03 (28827863 bytes)
compressed using a 32768 bytes dictionary : Ratio=4.28 (20410262 bytes)
generating 1 dictionaries, using 0.1 MB of memory
Compression Speed : 306.9 MB/s
Median Speed : 298.4 MB/s
```
Summary:
Add column headers and data for whether it's a compression or a decompression
run, compression level, nbDicts and dictAttachPref in additional to
compr/decompr speed.
Test Plan:
Example output:
```
./largeNbDicts
Compression/Decompression,Level,nbDicts,dictAttachPref,Speed
Compression,1,1,0,300.9
Compression,1,1,1,296.4
Compression,1,1,2,307.8
Compression,1,10,0,292.3
Compression,1,100,0,293.3
Compression,3,110,0,106.0
Decompression,-1,110,-1,155.6
Decompression,-1,110,-1,709.4
Decompression,-1,120,-1,709.1
Decompression,-1,120,-1,734.6
```
Benchmarking decompression results in a segfault in `createCompressInstructions`
because `cctxParams` is NULL. Skip running that function if we are not benching
compression.
Han Zhu