public void InitGPU()
{
// Work around for bug in Cudafy trying to find the path..
var os64Bit = Environment.Is64BitOperatingSystem;
if (os64Bit)
{
var dir = Environment.GetEnvironmentVariable("ProgramFiles");
Environment.SetEnvironmentVariable("ProgramFiles", "C:\\Program Files\\");
dir = Environment.GetEnvironmentVariable("ProgramFiles");
}
if (Gpu == null)
{
Gpu = CudafyHost.GetDevice(_gpuType, 0);
//Blas = GPGPUBLAS.Create(Gpu);
if (_gpuType == eGPUType.Cuda)
{
Blas = new SharpBLAS(Gpu);
Rand = GPGPURAND.Create(Gpu, curandRngType.CURAND_RNG_PSEUDO_DEFAULT);
Rand.SetPseudoRandomGeneratorSeed((ulong)RandomHelpers.Next(9999));
}
CudafyTranslator.GenerateDebug = true;
Debug.WriteLine("CUDA workdir = " + CudafyTranslator.WorkingDirectory);
Console.WriteLine("Recompile module");
CudafyTranslator.Language = eLanguage.Cuda;
var km = CudafyTranslator.Cudafy(eArchitecture.sm_30);
km = CudafyTranslator.Cudafy();
km.Serialize();
Gpu.LoadModule(km);
}
}
private void initializeRandomData(int iterationCount, int nodeCount)
{
var random = GPGPURAND.Create(this._gpu);
random.GenerateSeeds();
// We need 2 random per contraction, we have nodeCount - 2 contraction per graph,
// and we do that for each iteration
// Just, to ease index calculation, we compute nodeCount pairs instead of nodeCount - 2
this._random_dev = this._gpu.Allocate <uint>(nodeCount * iterationCount * 2);
random.Generate(this._random_dev, 1);
}
public void Test_CURAND_Host_Reference_Example()
{
int n = 100;
float[] hostData = new float[n];
GPGPURAND gen = GPGPURAND.Create(_gpu, curandRngType.CURAND_RNG_PSEUDO_DEFAULT, true);
gen.SetPseudoRandomGeneratorSeed(1234);
gen.GenerateUniform(hostData);
for (int i = 0; i < n; i++)
{
Console.WriteLine(hostData[i]);
}
gen.Dispose();
}
public void Test_CURAND_Device_Reference_Example()
{
int n = 100;
float[] devData = _gpu.Allocate <float>(n);
float[] hostData = new float[n];
GPGPURAND gen = GPGPURAND.Create(_gpu, curandRngType.CURAND_RNG_PSEUDO_DEFAULT);
gen.SetPseudoRandomGeneratorSeed(1234);
gen.GenerateUniform(devData);
_gpu.CopyFromDevice(devData, hostData);
for (int i = 0; i < n; i++)
{
Console.WriteLine(hostData[i]);
}
gen.Dispose();
}