mlemanczyk/even-perfect-numbers-scanner
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Address feedback on GPU pooling and heuristic factoring

Browse Source
This commit is contained in:
Marek Lemańczyk authored and mlemanczyk committed 2025-11-02 11:22:49 +01:00
1 parent 5339c3c304
commit 76dfe108a9
6 files changed
+92 -139
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/GpuContextPool.cs
+33 -59
View File
@@ -7,7 +7,6 @@ namespace PerfectNumbers.Core.Gpu;
public static class GpuContextPool
{
private static readonly bool PoolingEnabled = true;
// The pool intentionally skips pre-loading ProcessEightBitWindows kernels so accelerator initialization
// stays stable; GpuKernelPool JIT-compiles them on first use alongside the small-cycle uploads.
// Default device preference for generic GPU kernels (prime scans, NTT, etc.)
@@ -48,7 +47,6 @@ public static class GpuContextPool
private static readonly ConcurrentQueue<PooledContext> CpuPool = new();
private static readonly ConcurrentQueue<PooledContext> GpuPool = new();
private static readonly object WarmUpLock = new();
private static int WarmedGpuContextCount;
public static GpuContextLease Rent()
@@ -59,17 +57,14 @@ public static class GpuContextPool
// Allows callers to choose CPU/GPU per use-case, decoupled from ForceCpu.
public static GpuContextLease RentPreferred(bool preferCpu)
{
if (PoolingEnabled)
if (preferCpu && CpuPool.TryDequeue(out var cpu))
{
if (preferCpu && CpuPool.TryDequeue(out var cpu))
{
return new GpuContextLease(cpu);
}
return new GpuContextLease(cpu);
}
if (!preferCpu && GpuPool.TryDequeue(out var gpu))
{
return new GpuContextLease(gpu);
}
if (!preferCpu && GpuPool.TryDequeue(out var gpu))
{
return new GpuContextLease(gpu);
}
// Create a new accelerator when the pool does not have one available.
@@ -78,53 +73,40 @@ public static class GpuContextPool
public static void WarmUpPool(int threadCount)
{
if (!PoolingEnabled || ForceCpu)
if (ForceCpu)
{
return;
}
if (threadCount <= 0)
{
return;
}
int target = threadCount / 4;
int target = threadCount >> 2;
if (target == 0)
{
target = 1;
target = threadCount;
}
lock (WarmUpLock)
if (target <= WarmedGpuContextCount)
{
if (target <= WarmedGpuContextCount)
{
return;
}
int toCreate = target - WarmedGpuContextCount;
var contexts = new PooledContext[toCreate];
for (int i = 0; i < toCreate; i++)
{
contexts[i] = new PooledContext(preferCpu: false);
}
for (int i = 0; i < toCreate; i++)
{
GpuPool.Enqueue(contexts[i]);
}
WarmedGpuContextCount = target;
return;
}
int toCreate = target - WarmedGpuContextCount;
var contexts = new PooledContext[toCreate];
for (int i = 0; i < toCreate; i++)
{
contexts[i] = new PooledContext(preferCpu: false);
}
for (int i = 0; i < toCreate; i++)
{
GpuPool.Enqueue(contexts[i]);
}
WarmedGpuContextCount = target;
}
public static void DisposeAll()
{
if (!PoolingEnabled)
{
return;
}
while (CpuPool.TryDequeue(out var cpu))
{
cpu.Dispose();
@@ -141,24 +123,16 @@ public static class GpuContextPool
private static void Return(PooledContext ctx)
{
if (PoolingEnabled)
ctx.Accelerator.Synchronize();
if (ctx.IsCpu)
{
ctx.Accelerator.Synchronize();
if (ctx.IsCpu)
{
CpuPool.Enqueue(ctx);
}
else
{
GpuPool.Enqueue(ctx);
}
CpuPool.Enqueue(ctx);
}
else
{
ctx.Dispose();
}
}
else
{
GpuPool.Enqueue(ctx);
}
}
public struct GpuContextLease
{
PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
+4 -18
View File
@@ -68,11 +68,6 @@ internal static class PrimeTesterKernels
}
ulong prime = primes[i];
if (prime == 0)
{
break;
}
if (n % prime == 0UL)
{
result = 0;
@@ -94,32 +89,23 @@ internal static class PrimeTesterKernels
byte nMod10 = (byte)(n % 10UL);
ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind, nMod10);
ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind, nMod10);
if (divisors.Length == 0 || divisorSquares.Length == 0)
{
return;
}
int divisorLength = (int)divisors.Length;
int squareLength = (int)divisorSquares.Length;
int divisorLength = (int)divisorSquares.Length;
int threadIndex = index;
if (threadIndex >= divisorLength || threadIndex >= squareLength)
if (threadIndex >= divisorLength)
{
return;
}
ulong divisorSquare = divisorSquares[threadIndex];
if (divisorSquare == 0UL || divisorSquare > maxDivisorSquare)
if (divisorSquare > maxDivisorSquare)
{
return;
}
ulong divisor = divisors[threadIndex];
divisor = n % divisor;
if (divisor == 0UL)
{
return;
}
if (n % divisor == 0UL)
{
resultFlag[0] = 1;
}
PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs
+12 -15
View File
@@ -235,21 +235,18 @@ public sealed class HeuristicCombinedPrimeTester
bool compositeDetected = false;
int divisorLength = GetCombinedDivisors(nMod10).Length;
if (divisorLength > 0)
{
int compositeFlag = 0;
flagView1D.CopyFromCPU(ref compositeFlag, 1);
kernel(
divisorLength,
flagView,
n,
maxDivisorSquare,
HeuristicGpuDivisorTableKind.Combined,
gpu.HeuristicGpuTables);
accelerator.Synchronize();
flagView1D.CopyToCPU(ref compositeFlag, 1);
compositeDetected = compositeFlag != 0;
}
int compositeFlag = 0;
flagView1D.CopyFromCPU(ref compositeFlag, 1);
kernel(
divisorLength,
flagView,
n,
maxDivisorSquare,
HeuristicGpuDivisorTableKind.Combined,
gpu.HeuristicGpuTables);
accelerator.Synchronize();
flagView1D.CopyToCPU(ref compositeFlag, 1);
compositeDetected = compositeFlag != 0;
gpu.Dispose();
limiter.Dispose();
PerfectNumbers.Core/HeuristicPrimeTester.cs
+11 -18
View File
@@ -252,20 +252,17 @@ public sealed class HeuristicPrimeTester
int groupALength = HeuristicPrimeSieves.GroupADivisors.Length;
if (groupALength > 0)
{
flagView1D.CopyFromCPU(ref compositeFlag, 1);
kernel(
groupALength,
flagView,
n,
maxDivisorSquare,
HeuristicGpuDivisorTableKind.GroupA,
gpu.HeuristicGpuTables);
accelerator.Synchronize();
flagView1D.CopyToCPU(ref compositeFlag, 1);
compositeDetected = compositeFlag != 0;
}
flagView1D.CopyFromCPU(ref compositeFlag, 1);
kernel(
groupALength,
flagView,
n,
maxDivisorSquare,
HeuristicGpuDivisorTableKind.GroupA,
gpu.HeuristicGpuTables);
accelerator.Synchronize();
flagView1D.CopyToCPU(ref compositeFlag, 1);
compositeDetected = compositeFlag != 0;
if (!compositeDetected)
{
@@ -287,10 +284,6 @@ public sealed class HeuristicPrimeTester
}
int divisorLength = GetGroupBDivisors(ending).Length;
if (divisorLength == 0)
{
continue;
}
compositeFlag = 0;
flagView1D.CopyFromCPU(ref compositeFlag, 1);
PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs
+10 -10
View File
@@ -1289,11 +1289,7 @@ internal static partial class PrimeOrderCalculator
if (primeSlotArray is not null)
{
ThreadStaticPools.UlongPool.Return(primeSlotArray, clearArray: false);
}
if (exponentSlotArray is not null)
{
ThreadStaticPools.IntPool.Return(exponentSlotArray, clearArray: false);
ThreadStaticPools.IntPool.Return(exponentSlotArray!, clearArray: false);
}
if (compositeStack is not null)
@@ -1340,12 +1336,20 @@ internal static partial class PrimeOrderCalculator
}
ulong primeSquare = squares[i];
// primeSquare will never == 0 in production code
// if (primeSquare != 0 && primeSquare > remainingLocal)
if (primeSquare > remainingLocal)
{
break;
}
ulong primeValue = primeCandidate;
// primeCandidate will never equal 0 in production code
// if (primeCandidate == 0UL)
// {
// continue;
// }
if ((remainingLocal % primeValue) != 0UL)
{
continue;
@@ -1626,11 +1630,7 @@ internal static partial class PrimeOrderCalculator
if (primeArray is not null)
{
ThreadStaticPools.UlongPool.Return(primeArray, clearArray: false);
}
if (exponentArray is not null)
{
ThreadStaticPools.IntPool.Return(exponentArray, clearArray: false);
ThreadStaticPools.IntPool.Return(exponentArray!, clearArray: false);
}
primeArray = newPrimeArray;
PerfectNumbers.Core/PrimeTester.cs
+22 -19
View File
@@ -152,15 +152,10 @@ public sealed class PrimeTester
return;
}
if (threadCount <= 0)
{
return;
}
int target = threadCount / 4;
int target = threadCount >> 2;
if (target == 0)
{
target = 1;
target = threadCount;
}
lock (GpuWarmUpLock)
@@ -267,12 +262,14 @@ public sealed class PrimeTester
private static void ReleaseSharedTables(SharedHeuristicGpuTables tables)
{
if (tables is null)
{
return;
}
lock (SharedTablesLock)
{
if (tables.Release())
{
SharedTables.Remove(tables.Key);
}
tables.Release();
}
}
@@ -303,6 +300,16 @@ public sealed class PrimeTester
{
lease.DisposeResources();
}
lock (SharedTablesLock)
{
foreach (var entry in SharedTables.Values)
{
entry.Dispose();
}
SharedTables.Clear();
}
}
internal static void Clear(Accelerator accelerator)
@@ -572,16 +579,12 @@ public sealed class PrimeTester
_referenceCount++;
}
internal bool Release()
internal void Release()
{
_referenceCount--;
if (_referenceCount == 0)
if (_referenceCount > 0)
{
Dispose();
return true;
_referenceCount--;
}
return false;
}
private static MemoryBuffer1D<ulong, Stride1D.Dense> CopySpanToDevice(Accelerator accelerator, ReadOnlySpan<ulong> span)
@@ -613,7 +616,7 @@ public sealed class PrimeTester
return buffer;
}
private void Dispose()
internal void Dispose()
{
DevicePrimesDefault.Dispose();
DevicePrimesLastOne.Dispose();