Files
Craftimizer/Solver/Crafty/NodeScoresBuffer.cs
T
Asriel Camora 636501ab86 Remove all concurrency code
Muddled the code too much, and only gave a marginal performance improvement in the grand scheme of things. Other ways to parallelize MCTS will be nicer to implement and could yield better results.
2023-07-07 20:17:35 +02:00

63 lines
2.0 KiB
C#

using System.Diagnostics.Contracts;
using System.Numerics;
using System.Runtime.CompilerServices;
namespace Craftimizer.Solver.Crafty;
// Adapted from https://github.com/dtao/ConcurrentList/blob/4fcf1c76e93021a41af5abb2d61a63caeba2adad/ConcurrentList/ConcurrentList.cs
public struct NodeScoresBuffer
{
public sealed class ScoresBatch
{
public Memory<float> ScoreSum;
public Memory<float> MaxScore;
public Memory<int> Visits;
public ScoresBatch()
{
ScoreSum = new float[BatchSize];
MaxScore = new float[BatchSize];
Visits = new int[BatchSize];
}
}
// Technically 25, but it's very unlikely to actually get to there.
// The benchmark reaches 20 at most, but here we have a little leeway just in case.
private const int MaxSize = 24;
private static readonly int BatchSize = Vector<float>.Count;
private static readonly int BatchSizeBits = int.Log2(BatchSize);
private static readonly int BatchSizeMask = BatchSize - 1;
private static readonly int BatchCount = MaxSize / BatchSize;
public ScoresBatch[] Data;
public int Count { get; private set; }
public void Add()
{
Data ??= new ScoresBatch[BatchCount];
var idx = Count++;
var (arrayIdx, _) = GetArrayIndex(idx);
Data[arrayIdx] ??= new();
}
public readonly void Visit((int arrayIdx, int subIdx) at, float score)
{
Data[at.arrayIdx].ScoreSum.Span[at.subIdx] += score;
Data[at.arrayIdx].MaxScore.Span[at.subIdx] = Math.Max(Data[at.arrayIdx].MaxScore.Span[at.subIdx], score);
Data[at.arrayIdx].Visits.Span[at.subIdx]++;
}
public readonly int GetVisits((int arrayIdx, int subIdx) at) =>
Data[at.arrayIdx].Visits.Span[at.subIdx];
[Pure]
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static (int arrayIdx, int subIdx) GetArrayIndex(int idx) =>
(idx >> BatchSizeBits, idx & BatchSizeMask);
}