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 void Test_BLAS2_GEMV()
{
ClearBuffer(hiMatrixA);
ClearBuffer(hiVectorXM);
ClearBuffer(hiVectorXN);
ClearBuffer(hiVectorYM);
ClearBuffer(hiVectorYN);
FillBuffer(hiMatrixA);
FillBuffer(hiVectorXM);
FillBuffer(hiVectorXN);
FillBuffer(hiVectorYM);
FillBuffer(hiVectorYN);
diMatrixA = _gpu.CopyToDevice(hiMatrixA);
diVectorXM = _gpu.CopyToDevice(hiVectorXM);
diVectorXN = _gpu.CopyToDevice(hiVectorXN);
// Test without transpose
diVectorYM = _gpu.CopyToDevice(hiVectorYM);
_blas.GEMV(M, N, Alpha, diMatrixA, diVectorXN, Beta, diVectorYM);
_gpu.CopyFromDevice(diVectorYM, gpuResultM);
for (int i = 0; i < M; i++)
{
double cpuResult = 0.0;
for (int j = 0; j < N; j++)
{
cpuResult += Alpha * hiMatrixA[GetIndexColumnMajor(i, j, M)] * hiVectorXN[j];
}
cpuResult += Beta * hiVectorYM[i];
Assert.AreEqual(cpuResult, gpuResultM[i]);
}
// Test with transpose
diVectorYN = _gpu.CopyToDevice(hiVectorYN);
_blas.GEMV(M, N, Alpha, diMatrixA, diVectorXM, Beta, diVectorYN, cublasOperation.T);
_gpu.CopyFromDevice(diVectorYN, gpuResultN);
for (int j = 0; j < N; j++)
{
double cpuResult = 0.0;
for (int i = 0; i < M; i++)
{
cpuResult += Alpha * hiMatrixA[GetIndexColumnMajor(i, j, M)] * hiVectorXM[i];
}
cpuResult += Beta * hiVectorYN[j];
Assert.AreEqual(cpuResult, gpuResultN[j]);
}
_gpu.FreeAll();
}