Wire by-divisor GPU batching to scanner configuration
This commit is contained in:
Marek Lemańczyk committed
1 parent
085e6e0ceb
commit
3fd39c3d34
2 files changed
+157
-86
No files matched your search
@@ -11,6 +11,8 @@ namespace PerfectNumbers.Core.Gpu;
|
||||
|
||||
public sealed class MersenneNumberDivisorByDivisorGpuTester
|
||||
{
|
||||
private int _gpuBatchSize = GpuConstants.ScanBatchSize;
|
||||
|
||||
private readonly List<ulong> _divisors = new();
|
||||
private readonly List<ulong> _remainders = new();
|
||||
private readonly List<ulong> _lastPrimes = new();
|
||||
@@ -28,6 +30,12 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester
|
||||
private Action<Index1D, ulong, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetInitialKernel(Accelerator accelerator) =>
|
||||
_initialKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(InitialKernel));
|
||||
|
||||
public int GpuBatchSize
|
||||
{
|
||||
get => _gpuBatchSize;
|
||||
set => _gpuBatchSize = Math.Max(1, value);
|
||||
}
|
||||
|
||||
public void ConfigureFromMaxPrime(ulong maxPrime)
|
||||
{
|
||||
lock (_sync)
|
||||
@@ -80,138 +88,198 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester
|
||||
return false;
|
||||
}
|
||||
|
||||
var divisorsSpan = CollectionsMarshal.AsSpan(_divisors);
|
||||
var remaindersSpan = CollectionsMarshal.AsSpan(_remainders);
|
||||
var lastPrimesSpan = CollectionsMarshal.AsSpan(_lastPrimes);
|
||||
|
||||
var gpuLease = GpuContextPool.RentPreferred(preferCpu: false);
|
||||
var accelerator = gpuLease.Accelerator;
|
||||
var kernel = GetUpdateKernel(accelerator);
|
||||
var divisorsBuffer = accelerator.Allocate1D<ulong>(count);
|
||||
var remaindersBuffer = accelerator.Allocate1D<ulong>(count);
|
||||
var lastPrimesBuffer = accelerator.Allocate1D<ulong>(count);
|
||||
var hitsBuffer = accelerator.Allocate1D<byte>(count);
|
||||
|
||||
Span<ulong> divisorsSpan = CollectionsMarshal.AsSpan(_divisors);
|
||||
Span<ulong> remaindersSpan = CollectionsMarshal.AsSpan(_remainders);
|
||||
Span<ulong> lastPrimesSpan = CollectionsMarshal.AsSpan(_lastPrimes);
|
||||
int batchCapacity = Math.Min(_gpuBatchSize, count);
|
||||
var divisorsBuffer = accelerator.Allocate1D<ulong>(batchCapacity);
|
||||
var remaindersBuffer = accelerator.Allocate1D<ulong>(batchCapacity);
|
||||
var lastPrimesBuffer = accelerator.Allocate1D<ulong>(batchCapacity);
|
||||
var hitsBuffer = accelerator.Allocate1D<byte>(batchCapacity);
|
||||
|
||||
divisorsBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(divisorsSpan), count);
|
||||
remaindersBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(remaindersSpan), count);
|
||||
lastPrimesBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(lastPrimesSpan), count);
|
||||
hitsBuffer.MemSetToZero();
|
||||
|
||||
kernel(count, prime, divisorsBuffer.View, remaindersBuffer.View, lastPrimesBuffer.View, hitsBuffer.View);
|
||||
accelerator.Synchronize();
|
||||
|
||||
remaindersBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(remaindersSpan), count);
|
||||
lastPrimesBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(lastPrimesSpan), count);
|
||||
|
||||
byte[] hits = ArrayPool<byte>.Shared.Rent(count);
|
||||
Span<byte> hitsSpan = hits.AsSpan(0, count);
|
||||
hitsBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(hitsSpan), count);
|
||||
byte[] hits = ArrayPool<byte>.Shared.Rent(batchCapacity);
|
||||
|
||||
bool composite = false;
|
||||
ulong lastProcessed = 0UL;
|
||||
for (int i = 0; i < count; i++)
|
||||
{
|
||||
ulong divisor = divisorsSpan[i];
|
||||
if (divisor > allowedMax)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
lastProcessed = divisor;
|
||||
if (hitsSpan[i] != 0)
|
||||
try
|
||||
{
|
||||
int processed = 0;
|
||||
while (processed < count)
|
||||
{
|
||||
composite = true;
|
||||
break;
|
||||
int batchSize = Math.Min(batchCapacity, count - processed);
|
||||
var divisorView = divisorsBuffer.View.SubView(0, batchSize);
|
||||
var remainderView = remaindersBuffer.View.SubView(0, batchSize);
|
||||
var lastPrimeView = lastPrimesBuffer.View.SubView(0, batchSize);
|
||||
var hitsView = hitsBuffer.View.SubView(0, batchSize);
|
||||
|
||||
Span<ulong> divisorBatch = divisorsSpan.Slice(processed, batchSize);
|
||||
Span<ulong> remainderBatch = remaindersSpan.Slice(processed, batchSize);
|
||||
Span<ulong> lastPrimeBatch = lastPrimesSpan.Slice(processed, batchSize);
|
||||
|
||||
divisorView.CopyFromCPU(ref MemoryMarshal.GetReference(divisorBatch), batchSize);
|
||||
remainderView.CopyFromCPU(ref MemoryMarshal.GetReference(remainderBatch), batchSize);
|
||||
lastPrimeView.CopyFromCPU(ref MemoryMarshal.GetReference(lastPrimeBatch), batchSize);
|
||||
hitsView.MemSetToZero();
|
||||
|
||||
kernel(batchSize, prime, divisorView, remainderView, lastPrimeView, hitsView);
|
||||
accelerator.Synchronize();
|
||||
|
||||
remainderView.CopyToCPU(ref MemoryMarshal.GetReference(remainderBatch), batchSize);
|
||||
lastPrimeView.CopyToCPU(ref MemoryMarshal.GetReference(lastPrimeBatch), batchSize);
|
||||
|
||||
Span<byte> hitsSpan = hits.AsSpan(0, batchSize);
|
||||
hitsView.CopyToCPU(ref MemoryMarshal.GetReference(hitsSpan), batchSize);
|
||||
|
||||
for (int i = 0; i < batchSize; i++)
|
||||
{
|
||||
ulong divisor = divisorBatch[i];
|
||||
if (divisor > allowedMax)
|
||||
{
|
||||
processed = count;
|
||||
break;
|
||||
}
|
||||
|
||||
lastProcessed = divisor;
|
||||
if (hitsSpan[i] != 0)
|
||||
{
|
||||
composite = true;
|
||||
processed = count;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (composite)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
if (processed < count)
|
||||
{
|
||||
processed += batchSize;
|
||||
}
|
||||
}
|
||||
}
|
||||
finally
|
||||
{
|
||||
ArrayPool<byte>.Shared.Return(hits, clearArray: true);
|
||||
hitsBuffer.Dispose();
|
||||
lastPrimesBuffer.Dispose();
|
||||
remaindersBuffer.Dispose();
|
||||
divisorsBuffer.Dispose();
|
||||
gpuLease.Dispose();
|
||||
}
|
||||
|
||||
if (lastProcessed != 0UL)
|
||||
{
|
||||
ReportStatus(lastProcessed);
|
||||
}
|
||||
|
||||
ArrayPool<byte>.Shared.Return(hits, clearArray: true);
|
||||
|
||||
hitsBuffer.Dispose();
|
||||
lastPrimesBuffer.Dispose();
|
||||
remaindersBuffer.Dispose();
|
||||
divisorsBuffer.Dispose();
|
||||
gpuLease.Dispose();
|
||||
|
||||
return composite;
|
||||
}
|
||||
|
||||
private bool ExtendDivisors(ulong prime, ulong allowedMax)
|
||||
{
|
||||
ulong currentMax = _divisors.Count == 0 ? 1UL : _divisors[^1];
|
||||
int existingCount = _divisors.Count;
|
||||
allowedMax = Math.Min(allowedMax, _divisorLimit);
|
||||
|
||||
ulong currentMax = existingCount == 0 ? 1UL : _divisors[^1];
|
||||
ulong start = currentMax < 3UL ? 3UL : currentMax + 2UL;
|
||||
if (start > allowedMax)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
List<ulong> newDivisors = new();
|
||||
for (ulong divisor = start; divisor <= allowedMax; divisor += 2UL)
|
||||
{
|
||||
newDivisors.Add(divisor);
|
||||
}
|
||||
|
||||
int count = newDivisors.Count;
|
||||
if (count == 0)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
var gpuLease = GpuContextPool.RentPreferred(preferCpu: false);
|
||||
var accelerator = gpuLease.Accelerator;
|
||||
var kernel = GetInitialKernel(accelerator);
|
||||
var divisorsBuffer = accelerator.Allocate1D<ulong>(count);
|
||||
var remaindersBuffer = accelerator.Allocate1D<ulong>(count);
|
||||
var hitsBuffer = accelerator.Allocate1D<byte>(count);
|
||||
|
||||
Span<ulong> newDivisorsSpan = CollectionsMarshal.AsSpan(newDivisors);
|
||||
divisorsBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(newDivisorsSpan), count);
|
||||
hitsBuffer.MemSetToZero();
|
||||
int batchCapacity = _gpuBatchSize;
|
||||
var divisorsBuffer = accelerator.Allocate1D<ulong>(batchCapacity);
|
||||
var remaindersBuffer = accelerator.Allocate1D<ulong>(batchCapacity);
|
||||
var hitsBuffer = accelerator.Allocate1D<byte>(batchCapacity);
|
||||
|
||||
kernel(count, prime, divisorsBuffer.View, remaindersBuffer.View, hitsBuffer.View);
|
||||
accelerator.Synchronize();
|
||||
|
||||
ulong[] newRemainders = ArrayPool<ulong>.Shared.Rent(count);
|
||||
byte[] hits = ArrayPool<byte>.Shared.Rent(count);
|
||||
Span<ulong> newRemaindersSpan = newRemainders.AsSpan(0, count);
|
||||
Span<byte> hitsSpan = hits.AsSpan(0, count);
|
||||
remaindersBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(newRemaindersSpan), count);
|
||||
hitsBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(hitsSpan), count);
|
||||
ulong[] remainders = ArrayPool<ulong>.Shared.Rent(batchCapacity);
|
||||
byte[] hits = ArrayPool<byte>.Shared.Rent(batchCapacity);
|
||||
ulong[] divisors = ArrayPool<ulong>.Shared.Rent(batchCapacity);
|
||||
|
||||
bool composite = false;
|
||||
ulong lastProcessed = 0UL;
|
||||
for (int i = 0; i < count; i++)
|
||||
ulong divisor = start;
|
||||
|
||||
try
|
||||
{
|
||||
ulong divisor = newDivisorsSpan[i];
|
||||
ulong remainder = newRemaindersSpan[i];
|
||||
_divisors.Add(divisor);
|
||||
_remainders.Add(remainder);
|
||||
_lastPrimes.Add(prime);
|
||||
lastProcessed = divisor;
|
||||
if (!composite && hitsSpan[i] != 0)
|
||||
while (divisor <= allowedMax)
|
||||
{
|
||||
composite = true;
|
||||
int batchSize = batchCapacity;
|
||||
int index = 0;
|
||||
while (index < batchSize && divisor <= allowedMax)
|
||||
{
|
||||
divisors[index++] = divisor;
|
||||
divisor += 2UL;
|
||||
}
|
||||
|
||||
if (index == 0)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
var divisorView = divisorsBuffer.View.SubView(0, index);
|
||||
var remainderView = remaindersBuffer.View.SubView(0, index);
|
||||
var hitsView = hitsBuffer.View.SubView(0, index);
|
||||
|
||||
Span<ulong> divisorSpan = divisors.AsSpan(0, index);
|
||||
divisorView.CopyFromCPU(ref MemoryMarshal.GetReference(divisorSpan), index);
|
||||
hitsView.MemSetToZero();
|
||||
|
||||
kernel(index, prime, divisorView, remainderView, hitsView);
|
||||
accelerator.Synchronize();
|
||||
|
||||
Span<ulong> remainderSpan = remainders.AsSpan(0, index);
|
||||
Span<byte> hitsSpan = hits.AsSpan(0, index);
|
||||
remainderView.CopyToCPU(ref MemoryMarshal.GetReference(remainderSpan), index);
|
||||
hitsView.CopyToCPU(ref MemoryMarshal.GetReference(hitsSpan), index);
|
||||
|
||||
for (int i = 0; i < index; i++)
|
||||
{
|
||||
ulong newDivisor = divisorSpan[i];
|
||||
_divisors.Add(newDivisor);
|
||||
_remainders.Add(remainderSpan[i]);
|
||||
_lastPrimes.Add(prime);
|
||||
lastProcessed = newDivisor;
|
||||
if (!composite && hitsSpan[i] != 0)
|
||||
{
|
||||
composite = true;
|
||||
}
|
||||
}
|
||||
|
||||
if (composite)
|
||||
{
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
finally
|
||||
{
|
||||
ArrayPool<ulong>.Shared.Return(divisors, clearArray: true);
|
||||
ArrayPool<ulong>.Shared.Return(remainders, clearArray: true);
|
||||
ArrayPool<byte>.Shared.Return(hits, clearArray: true);
|
||||
hitsBuffer.Dispose();
|
||||
remaindersBuffer.Dispose();
|
||||
divisorsBuffer.Dispose();
|
||||
gpuLease.Dispose();
|
||||
}
|
||||
|
||||
if (lastProcessed != 0UL)
|
||||
{
|
||||
ReportStatus(lastProcessed);
|
||||
}
|
||||
|
||||
ArrayPool<ulong>.Shared.Return(newRemainders, clearArray: true);
|
||||
ArrayPool<byte>.Shared.Return(hits, clearArray: true);
|
||||
|
||||
hitsBuffer.Dispose();
|
||||
remaindersBuffer.Dispose();
|
||||
divisorsBuffer.Dispose();
|
||||
gpuLease.Dispose();
|
||||
|
||||
return composite;
|
||||
}
|
||||
|
||||
|
||||
Reference in new issue
Block a user