public void EraseAll()
{
lock (Mutex)
{
foreach (var item in Stacks)
{
if (item.Value != null)
{
foreach (var arr in item.Value)
{
if (this.Device.Type == DeviceType.NvidiaGPU)
{
CudaManagement.SetDevice(this.Device.ID);
CudaManagement.Free(arr.Ptr);
}
else if (this.Device.Type == DeviceType.Host)
{
GC.RemoveMemoryPressure((long)item.Key);
MKL.MKL_free(arr.Ptr);
}
else
{
throw new Exception("Uknown Device in ArrayPool!");
}
}
item.Value.Clear();
}
}
Stacks.Clear();
}
}
public void *Rent(long length, long unitlength)
{
lock (Mutex)
{
UnreturnedArrayCount++;
length *= unitlength;
if (Stacks.ContainsKey(length) && (Stacks[length] is Stack <PointerArray> x) && x.Count > 0)
{
PointerArray sr = x.Pop();
return(sr.Ptr);
}
else
{
if (this.Device.Type == DeviceType.NvidiaGPU)
{
GC.AddMemoryPressure(length);
CudaManagement.SetDevice(this.Device.ID);
return(CudaManagement.Allocate(length, Device.ID));
}
else if (this.Device.Type == DeviceType.Host)
{
GC.AddMemoryPressure(length);
return(MKL.MKL_malloc(length, 32));
}
else
{
throw new Exception("Uknown Device in ArrayPool!");
}
}
}
}
public static void MultiplyFloat32(Tensor res, Tensor a, Tensor b, float cofmul = 1, float cofadd = 0)
{
CudaManagement.SetDevice(res.Config.Device.ID);
CudaKernels.MultiplyFloat32((float *)res.Base.Array, (float *)a.Base.Array, (float *)b.Base.Array, a.Shape.TotalSize, b.Shape.TotalSize, cofmul, cofadd);
}