Calling lock.test_and_set() in a tight loop creates expensive
memory synchronizations among processors and penalize other
running threads. So syncronize only only once at the beginning
with fetch_sub() and then loop on a simple load() that puts much
less pressure on the system.
Reported about 2-3% speed up on various systems.
Patch by Ronald de Man.
No functional change.
It is reported to be defenitly faster with increasing
number of threads, we go from a +3.5% with 4 threads
to a +15% with 16 threads.
The only drawback is that now when testing with more
threads than physical available cores, the speed slows
down to a crawl. This is expected and was similar at what
we had setting the old sleepingThreads to false.
No functional change.
Initialization is more complex than what I'd like due
to MSVC compatibility that for some reason does not like:
std::atomic_flag lock = ATOMIC_FLAG_INIT;
No functional change.
It should be used slavesMask.count() instead.
Verified 100% equivalent when sp->allSlavesSearching:
dbg_hit_on(sp->allSlavesSearching, sp->slavesCount != sp->slavesMask.count());
No functional change.
Document and clarify that we cannot rejoin on ourselves
and that we never late join if we are master and all
slaves have finished, inded in this case we exit idle_loop.
No functional change.
In case of Threads.size() == 2 we have that sp->allSlavesSearching
is always false (because we have finished our search), bestSp is
always NULL and we never late join, so there is no need to special
case here.
Tested with dbg_hit_on(sp && sp->allSlavesSearching) and
verified it never fires.
No functional change.
And retire a redundant field. This is important also
from a concept point of view becuase we want to keep
SMP structures as simple as possible with the only
strictly necessary data.
Verified with
dbg_hit_on(sp->spLevel != level)
that the values are 100% the same out of more 50K samples.
No functional change.
Balance threads between split points.
There are huge differences between different machines and autopurging makes it very difficult to measure the improvement in fishtest, but the following was recorded for 16 threads at 15+0.05:
For Bravone (1000 games): 0 ELO
For Glinscott (1000 games): +20 ELO
For bKingUs (1000 games): +50 ELO
For fastGM (1500 games): +50 ELO
The change was regression for no one, and a big improvement for some, so it should be fine to commit it.
Also for 8 threads at 15+0.05 we measured a statistically significant improvement:
ELO: 6.19 +-3.9 (95%) LOS: 99.9%
Total: 10325 W: 1824 L: 1640 D: 6861
Finally it was verified that there was no (significant) regression for
4 threads:
ELO: 0.09 +-2.8 (95%) LOS: 52.4%
Total: 19908 W: 3422 L: 3417 D: 13069
2 threads:
ELO: 0.38 +-3.0 (95%) LOS: 60.0%
Total: 19044 W: 3480 L: 3459 D: 12105
1 thread:
ELO: -1.27 +-2.1 (95%) LOS: 12.3%
Total: 40000 W: 7829 L: 7975 D: 24196
Resolves#258
This micro-optimization only complicates the code and provides no benefit.
Removing it is even a speedup on my machine (i7-3770k, linux, gcc 4.9.1):
stat test master diff
mean 2,403,118 2,390,904 12,214
stdev 12,043 10,620 3,677
speedup 0.51%
P(speedup>0) 100.0%
No functional change.
This micro-optimization only complicates the code and provides no benefit.
Removing it is even a speedup on my machine (i7-3770k, linux, gcc 4.9.1):
stat test master diff
mean 2,403,118 2,390,904 12,214
stdev 12,043 10,620 3,677
speedup 0.51%
P(speedup>0) 100.0%
No functional change.
Use integer arithmetic instead of floating point arithmetic.
Floating point arithmetic was causing different results for some 32-bit compiles
No functional change
Resolves#249Resolves#250
Also remove useless StateCopySize64 optimization:
compiler uses SSE movups instruction anyhow and
does not need this trick (verified with fishbench).
No functional change.
Funny enough, gcc __builtin_prefetch() expects
already a void*, instead Windows's _mm_prefetch()
requires a char*.
The patch allows to remove ugly casts from caller
sites.
No functional change.
Results for 10 tests for each version (gcc 4.8.3 on mingw):
Base Test Diff
Mean 1502447 1507917 -5470
StDev 3119 1364 4153
p-value: 0,906
speedup: 0,004
Results for 10 tests for each version (MSVC 2013):
Base Test Diff
Mean 1400899 1403713 -2814
StDev 1273 2804 2700
p-value: 0,851
speedup: 0,002
No functional change.
I went through all the individual compile options that differ between
-fprofile-generate/-fprofile-use and -fprofile-arcs/-fbranch-probabilities
and distilled the speed difference down to only turning off
-fno-peel-loops and -fno-tracer. Using this we still get the full speedup
(maybe a bit more because other optimizations stay on) and it's also much cleaner
because we can get rid of the "@rm -f ucioption.gc*" hack for all versions of gcc.
No functional change.
Resolves#237
Follow the usual approach to delay computation
as far as possible, in case an earlier killer
cut-offs we avoid to do useless work.
This also greatly simplifies the code.
No functional change.
