Address feedback on GPU pooling and heuristic factoring
This commit is contained in:
1 parent
567b92792c
commit
8c662c5c85
6 files changed
+92
-139
No files matched your search
@@ -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
|
||||
{
|
||||
|
||||
Reference in new issue
Block a user