public void SetUp()
{
_gpu = CudafyHost.GetDevice();
_sparse = GPGPUSPARSE.Create(_gpu);
_blas = GPGPUBLAS.Create(_gpu);
_solver = new Solver(_gpu, _blas, _sparse);
}
public static void cudaTranspose(ref MathNet.Numerics.LinearAlgebra.Double.DenseMatrix dm)
{
GPGPU gpu = CudafyHost.GetDevice(eGPUType.Cuda);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
int cols = dm.ColumnCount, rows = dm.RowCount;
int restRows = rows - cols;
//double[] a = dm.Storage.ToColumnMajorArray();
double[] a = dm.SubMatrix(0, cols, 0, cols).Storage.ToColumnMajorArray();
double[] b = dm.SubMatrix(cols, restRows, 0, cols).Storage.ToColumnMajorArray();
dm = null;
double[] a_d = gpu.CopyToDevice <double>(a);
a = null;
double[] c_d = gpu.Allocate <double>(cols * cols);
double[] x_d = gpu.CopyToDevice <double>(new double[] { 1 });
blas.GEMV(cols, cols, 1, c_d, x_d, 0, x_d, Cudafy.Maths.BLAS.Types.cublasOperation.T);
a = new double[cols * rows];
gpu.CopyFromDevice <double>(c_d, 0, a, 0, cols * cols);
gpu.FreeAll();
a_d = gpu.CopyToDevice <double>(b);
b = null;
c_d = gpu.Allocate <double>(restRows * cols);
x_d = gpu.CopyToDevice <double>(new double[] { 1 });
blas.GEMV(restRows, cols, 1, c_d, x_d, 0, x_d, Cudafy.Maths.BLAS.Types.cublasOperation.T);
gpu.CopyFromDevice <double>(c_d, 0, a, cols * cols, restRows * cols);
gpu.FreeAll();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(cols, rows, a);
}
public static void MyFirstBlasEmulatorTest()
{
Console.WriteLine("MyTest()");
// Get GPU device
CudafyModes.Target = eGPUType.Emulator;
GPGPU gpu = CudafyHost.GetDevice(CudafyModes.Target);
// Create GPGPUBLAS (CUBLAS Wrapper)
using (GPGPUBLAS blas = GPGPUBLAS.Create(gpu))
{
const int N = 100;
float[] a = new float[N];
float[] b = new float[N];
float[] c = new float[N];
float alpha = -1;
float beta = 0;
float[] device_a = gpu.CopyToDevice(a);
float[] device_b = gpu.CopyToDevice(b);
float[] device_c = gpu.CopyToDevice(c);
int m = 10;
int n = 10;
int k = 10;
cublasOperation Op = cublasOperation.N;
blas.GEMM(m, k, n, alpha, device_a, device_b, beta, device_c, Op);
gpu.CopyFromDevice <float>(device_c, c);
}
}
//
// http://stackoverflow.com/questions/18628447/cudafy-throws-an-exception-while-testing
//
private static void BlasSample(int deviceId)
{
CudafyModes.Target = eGPUType.Emulator;
GPGPU gpu = CudafyHost.GetDevice(CudafyModes.Target, deviceId);
CudafyModes.DeviceId = deviceId;
eArchitecture arch = gpu.GetArchitecture();
CudafyModule km = CudafyTranslator.Cudafy(arch);
gpu.LoadModule(km);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
const int N = 100;
float[] a = new float[N];
float[] b = new float[N];
float[] c = new float[N];
float alpha = -1;
float beta = 0;
float[] device_a = gpu.CopyToDevice(a);
float[] device_b = gpu.CopyToDevice(b);
float[] device_c = gpu.CopyToDevice(c);
int m = 10;
int n = 10;
int k = 10;
cublasOperation Op = cublasOperation.N;
blas.GEMM(m, k, n, alpha, device_a, device_b, beta, device_c, Op);
throw new NotImplementedException();
}
public void SetUp()
{
_gpu = CudafyHost.CreateDevice(CudafyModes.Target);
_blas = GPGPUBLAS.Create(_gpu);
_hostInput = new float[ciROWS, ciCOLS];
_hostInput2 = new float[ciROWS, ciCOLS];
_hostOutput = new float[ciROWS, ciCOLS];
_devPtr = _gpu.Allocate <float>(_hostInput);
_devPtr2 = _gpu.Allocate <float>(_hostOutput);
}
public void SetUp()
{
_gpu = CudafyHost.CreateDevice(CudafyModes.Target);
_blas = GPGPUBLAS.Create(_gpu);
Console.Write("BLAS Version={0}", _blas.GetVersion());
_hostInput1 = new float[ciN];
_hostInput2 = new float[ciN];
_hostOutput1 = new float[ciN];
_hostOutput2 = new float[ciN];
_devPtr1 = _gpu.Allocate <float>(_hostInput1);
_devPtr2 = _gpu.Allocate <float>(_hostOutput1);
}
public Solver(GPGPU gpu, GPGPUBLAS blas, GPGPUSPARSE sparse)
{
this.gpu = gpu;
this.blas = blas;
this.sparse = sparse;
var km = CudafyModule.TryDeserialize();
if (km == null || !km.TryVerifyChecksums())
{
km = CudafyTranslator.Cudafy();
km.TrySerialize();
}
gpu.LoadModule(km);
}
public void SetUp()
{
_gpu = CudafyHost.GetDevice(CudafyModes.Target);
_blas = GPGPUBLAS.Create(_gpu);
// Initialize CPU Buffer
hiMatrixA = new double[M * N];
hiMatrixANN = new double[N * N];
hiMatrixACBC = new double[(KL + KU + 1) * N];
hiMatrixASCBC = new double[(K + 1) * N];
hiMatrixAPS = new double[(N * (N + 1)) / 2];
hiVectorXM = new double[M];
hiVectorXN = new double[N];
hiVectorYM = new double[M];
hiVectorYN = new double[N];
gpuResultM = new double[M];
gpuResultN = new double[N];
gpuResultMN = new double[M * N];
gpuResultNN = new double[N * N];
gpuResultP = new double[(N * (N + 1)) / 2];
}
public void SetUp()
{
_gpu = CudafyHost.GetDevice(CudafyModes.Target);
_blas = GPGPUBLAS.Create(_gpu);
hiMatrixAMM = new double[M * M];
hiMatrixANN = new double[N * N];
hiMatrixAMK = new double[M * K];
hiMatrixAKM = new double[K * M];
hiMatrixBMN = new double[M * N];
hiMatrixBKN = new double[K * N];
hiMatrixBNK = new double[N * K];
hiMatrixBMK = new double[M * K];
hiMatrixBKM = new double[K * M];
hiMatrixCMN = new double[M * N];
hiMatrixCKN = new double[K * N];
hiMatrixCMK = new double[M * K];
hiMatrixCMM = new double[M * M];
gpuResultMN = new double[M * N];
gpuResultMM = new double[M * M];
}
public static void cudaTransposeAndMultiply(ref MathNet.Numerics.LinearAlgebra.Double.DenseMatrix dm)
{
Cudafy.CudafyModule km = Cudafy.Translator.CudafyTranslator.Cudafy();
km.Serialize();
GPGPU gpu = CudafyHost.GetDevice(eGPUType.Cuda);
int cols = dm.ColumnCount, rows = dm.RowCount;
dm.Storage.ToColumnMajorArray();
double[] a = dm.ToColumnWiseArray();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(1, 1);
double[] dev_a = gpu.Allocate <double>(a.Length);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
double[] a_d = gpu.CopyToDevice <double>(a);
double[] c_d = gpu.Allocate <double>(cols * cols);
gpu.StartTimer();
blas.GEMM(cols, rows, cols, 1, a_d, a_d, 0, c_d, Cudafy.Maths.BLAS.Types.cublasOperation.T);
a = new double[cols * cols];
gpu.CopyFromDevice <double>(c_d, a);
gpu.FreeAll();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(cols, cols, a);
}
public SharpBLAS(GPGPU gpu)
{
Gpu = gpu;
Blas = GPGPUBLAS.Create(gpu);
Sparse = GPGPUSPARSE.Create(gpu);
}