Open dadhi opened 6 years ago
Have you done that? May I do it?
No, not yet. Yes, you can try!
@dzmitry-lahoda I have added two basic HAMT implementations into the Playground project with some tests.
There are two (or more) implementations: IntHashTrie<V>
and HashTrie<K, V>
(for arbitrary keys).
The method(s) to consider for replace with PopCount
are named uint GetSetBitsCount(uint n)
: https://github.com/dadhi/ImTools/blob/75b3b3b61db93653829ab9430ff9920745ea4b48/src/Playground/IntHashArrayMappedTrieTests.cs#L211
The hand-written benchmarks (without BenchmarkDotNet) are in https://github.com/dadhi/ImTools/blob/master/src/Playground/TreeBenchmarks.cs
Method | ItemCount | Mean | Error | StdDev |
--------------------- |---------- |------------:|----------:|-------------:|
SWAR_ImTools | 1000 | 2,362.5 ns | 47.07 ns | 73.284 ns |
Popcnt | 1000 | 504.7 ns | 10.05 ns | 8.908 ns |
SWAR_ImTools_Inlined | 1000 | 1,658.4 ns | 27.01 ns | 25.261 ns |
SWAR_ForkOfCoreFx | 1000 | 2,378.0 ns | 45.01 ns | 42.106 ns |
SWAR_ImTools | 10000 | 23,281.8 ns | 362.54 ns | 339.117 ns |
Popcnt | 10000 | 4,912.9 ns | 73.60 ns | 68.845 ns |
SWAR_ImTools_Inlined | 10000 | 16,221.2 ns | 265.45 ns | 248.305 ns |
SWAR_ForkOfCoreFx | 10000 | 24,810.7 ns | 492.58 ns | 1,049.733 ns |
I did in my fork as it has .NET Core only build (i have hard time work with legacy projects:()
https://github.com/dzmitry-lahoda/ImTools/blob/unnoficial/src/Playground/BitCountsBenchmarks.cs
For synthetic benchmark.
I guess need to make TreeBenchmarks to be BDN. And add copy of Trie code. Will send results.
I guess I may merge these classes into separate branch and send pull. But better if whole solution will be .NET Core for contrib.
Adding 10 items (ms):
Trie PopCnt- 0
Trie - 0
Getting one out of 10 items 1,000,000 times (ms):
Trie PopCnt - 21
Trie - 20
====================
Adding 100 items (ms):
Trie PopCnt- 2
Trie - 0
Getting one out of 100 items 1,000,000 times (ms):
Trie PopCnt - 39
Trie - 38
====================
Adding 1000 items (ms):
Trie PopCnt- 0
Trie - 0
Getting one out of 1000 items 1,000,000 times (ms):
Trie PopCnt - 40
Trie - 38
====================
Adding 10000 items (ms):
Trie PopCnt- 4
Trie - 3
Getting one out of 10000 items 1,000,000 times (ms):
Trie PopCnt - 49
Trie - 22
====================
Adding 100000 items (ms):
Trie PopCnt- 66
Trie - 80
Getting one out of 100000 items 1,000,000 times (ms):
Trie PopCnt - 64
Trie - 23
====================
Adding 1000000 items (ms):
Trie PopCnt- 1489
Trie - 1533
Getting one out of 1000000 items 1,000,000 times (ms):
Trie PopCnt - 18
Trie - 24
Strange result.
So seem no need to popcnt for now for specific case.
I see Trie is faster than Tree always. And only slower on 1M items to add. Interesting replacement of IoC. But may be be better try out BDN to measure memory and reduce any fluctuations. E.g. right now GC.Collect() seems(not sure) starts GC, but not waits to finish.
popcnt is with netcoreapp3.0 preview. not with 2.1 or 2.2:) so I suggest to close story as it seems irrelevant for both performance and usage in near future.
Thanks for the test.
PopCount
is available in the .Net Core 2.1. Hmmm...I will keep it open until I played with it myself, will decide how to proceed later.
I was trying to get the code of intrinsic for some reason. Found that for .NET Core 2.1 there should be kind of System.Runtime.Intrinsics.Experimental.dll
. But I could not find such.
Currently it uses HammingWeight algorithm for calculating number of set bits in a int value: https://bitbucket.org/dadhi/dryioc/src/e788ffce78b727b35fda5355a2297d40f0e7731b/Net45/Playground/HashArrayMappedTrieTests.cs#lines-190
Here is the thing Aye, that's there
System.Runtime.Intrinsics.X86.PopCount(uint/ulong)
First step will to BDN benchmark against Swar or HammingWeight with PopCount.