It all started as a hobby project to easily access CUDA from C# - at that time CUDA was available in version 3. Now more than 10 years later, managedCuda is still alive and is updated regularly by me to the latest versions of CUDA. In order to support further developments, I switched from the LGPL license to a dual-license GPLv3 / commercial license starting with managedCuda for Cuda version 12 onwards. In case you plan to use managedCuda 12+ for a commercial project, please contact me by mail: managedcuda@articimaging.eu. If you use the open-source license and want to contribute to future development, you can donate me a beer here:
Official nuget packages
One can find multiple packages for managedCuda on nuget, but the official packages are:
Implicit converter operators: Allocate and initialize device or host arrays in only one line of code:
float3[] array_host = new float3[100];
for (int i = 0; i < 100; i++)
{
array_host[i] = new float3(i, i+1, i+2);
}
//alloc device memory and copy data:
CudaDeviceVariable<float3> array_device = array_host;
//alloc host array and copy data:
float3[] array_host2 = array_device;
NPPs extension methods for CudaDeviceVariable. Add a reference to the NPP library and include the ManagedCuda.NPP.NPPsExtensions namespace:
Random rand = new Random();
int length = 256;
//init some ramdom values
double[] randoms = new double[length];
for (int i = 0; i < length; i++)
{
randoms[i] = rand.NextDouble();
}
//Alloc device memory
CudaDeviceVariable<double> a = randoms;
CudaDeviceVariable<double> b = new CudaDeviceVariable<double>(length);
b.Set(10.0); //NPPs method
int size = a.MeanGetBufferSize(); //NPPs method
//Alloc temporary memory for NPPs mean method
CudaDeviceVariable<byte> buffer = new CudaDeviceVariable<byte>(size);
CudaDeviceVariable<double> mean = new CudaDeviceVariable<double>(1);
a.Mul(b); //NPPs method
a.DivC(10.0); //NPPs method
a.Mean(mean, buffer); //NPPs method
//Copy data back to host
double m = mean;
double[] res = a;
//Clean up
mean.Dispose();
buffer.Dispose();
b.Dispose();
a.Dispose();
The new feature 'per thread default stream' is available as a compiler directive of the managedCuda main library: Compile the library with the option "_PerThreadDefaultStream" to enable it.
Note about Cuda context
Nvidia changed the cuda context behavior in the cuda libraries (NPP, Cufft, etc.) why it is highly recommended to use a PrimaryContext instead of a CudaContext when using ManagedCUDA together with Cuda libraries. To create a PrimaryContext in ManagedCUDA, use the following lines of code:
int deviceID = 0;
PrimaryContext ctx = new PrimaryContext(deviceID);
// Set current to CPU thread, mandatory for a PrimaryContext
ctx.SetCurrent();
NppStreamContext
NPP in Cuda 13 removed the NppStreamContext initialization API and requires manual setting of all fields, which is now done by ManagedCUDA. To initialize a NppStreamContext:
// Create a NppStreamContext from default Null-Stream:
NppStreamContext streamCtx = new NppStreamContext(CUstream.NullStream);
// Create a NppStreamContext from a custom stream
CudaStream stream2 = new CudaStream(CUStreamFlags.Default);
NppStreamContext streamCtx2 = new NppStreamContext(stream2);
