On Mon, 8 Jan 2024 07:55:00 GMT, Emanuel Peter <epe...@openjdk.org> wrote:
>>> You are using `VectorMask<Integer> pred = VectorMask.fromLong(ispecies,
>>> maskctr++);`. That basically systematically iterates over all masks, which
>>> is nice for a correctness test. But that would use different density inside
>>> one test run, right? The average over the loop is still at `50%`, correct?
>>>
>>> I was thinking more a run where the percentage over the whole loop is lower
>>> than maybe `1%`. That would get us to a point where maybe the branch
>>> prediction of non-vectorized code might be faster, what do you think?
>>
>> An imperative loop for compression will check each mask bit to select
>> compressible lane. Therefore mask with low or high density of set bits
>> should show similar performance.
>
> Yes, IF it is vectorized, then there is no difference between high and low
> density. My concern was more if vectorization is preferrable over the scalar
> alternative in the low-density case, where branch prediction is more stable.
At runtime we do need to scan entire mask to pick the compressible lane
corresponding to set mask bit. Thus the loop overhead of mask compare (BTW
masks are held in a vector register for AVX2 targets) and jump will anyways be
incurred , in addition for sparsely populated mask we may incur additional
misprediction penalty for not taking if block which extracts an element from
appropriate source vector lane and insert into destination vector lane. Overall
vector solution will win for most common cases for varying mask and also for
very sparsely populate masks. Here is the result of setting just a single mask
bit. I am process of updating to benchmark for 128 bit species will update the
patch.
@Benchmark
public void fuzzyFilterIntColumn() {
int i = 0;
int j = 0;
long maskctr = 1;
int endIndex = ispecies.loopBound(size);
for (; i < endIndex; i += ispecies.length()) {
IntVector vec = IntVector.fromArray(ispecies, intinCol, i);
VectorMask<Integer> pred = VectorMask.fromLong(ispecies, 1);
vec.compress(pred).intoArray(intoutCol, j);
j += pred.trueCount();
}
}
Baseline:
Benchmark (size) Mode
Cnt Score Error Units
ColumnFilterBenchmark.fuzzyFilterIntColumn 1024 thrpt 2 379.059
ops/ms
ColumnFilterBenchmark.fuzzyFilterIntColumn 2047 thrpt 2 188.355
ops/ms
ColumnFilterBenchmark.fuzzyFilterIntColumn 4096 thrpt 2 95.315
ops/ms
Withopt:
Benchmark (size) Mode
Cnt Score Error Units
ColumnFilterBenchmark.fuzzyFilterIntColumn 1024 thrpt 2 7390.074
ops/ms
ColumnFilterBenchmark.fuzzyFilterIntColumn 2047 thrpt 2 3483.247
ops/ms
ColumnFilterBenchmark.fuzzyFilterIntColumn 4096 thrpt 2 1823.817
ops/ms
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/17261#discussion_r1445666305
