public void CopyTo(CudaMatrixFloat result)
{
if (result == null)
{
return;
}
IntPtr handle = cuBLAS.Create_v2();
cuBLAS.Scopy_v2(
handle, Count,
_DevicePointer, 1
, result._DevicePointer, 1
);
cuBLAS.Destroy_v2(handle);
}
public void Mul(float value, CudaMatrixFloat result)
{
if (result == null)
{
return;
}
IntPtr handle = cuBLAS.Create_v2();
cuBLAS.Sscal_v2(
handle,
_Rows * _Cols,
value,
_DevicePointer,
1
);
cuBLAS.Destroy_v2(handle);
}
public void Mul(CudaMatrixFloat matrix, CudaMatrixFloat result)
{
if (matrix == null || result == null)
{
return;
}
IntPtr handle = cuBLAS.Create_v2();
cuBLAS.Sdgmm(
handle,
cublasSideMode.CUBLAS_SIDE_LEFT,
_Rows, _Cols,
_DevicePointer, _Rows,
matrix._DevicePointer, _Rows,
result._DevicePointer, _Rows
);
cuBLAS.Destroy_v2(handle);
}
public void Add(float value, CudaMatrixFloat result)
{
if (result == null)
{
return;
}
if (_Module == null)
{
throw new CudaException("module is not initialized.");
}
CallMethod(
_Module, "matrixAdd",
_Cols, _Rows,
_DevicePointer,
result._DevicePointer,
value,
_Cols, _Rows
);
}
public static CudaMatrixFloat FromByteArray(byte[] bytes)
{
if (bytes == null || bytes.Length < 12)
{
return(null);
}
int intSize = sizeof(int);
int intSize2 = intSize * 2;
int[] matrixSize = new int[2];
float[] data = new float[(bytes.Length - intSize2) / ItemSize];
Buffer.BlockCopy(bytes, 0, matrixSize, 0, intSize2);
Buffer.BlockCopy(bytes, intSize2, data, 0, bytes.Length - intSize2);
int rows = matrixSize[0];
int cols = matrixSize[1];
CudaMatrixFloat matrix = new CudaMatrixFloat(rows, cols);
matrix.SetDeviceMemory(matrix._DevicePointer, data);
return(matrix);
}
public void Dot(CudaMatrixFloat matrix, CudaMatrixFloat result)
{
if (matrix == null || result == null)
{
return;
}
IntPtr handle = cuBLAS.Create_v2();
cuBLAS.Sgemm_v2(
handle,
cublasOperation.CUBLAS_OP_N,
cublasOperation.CUBLAS_OP_N,
_Rows, matrix._Cols, _Cols,
1f,
_DevicePointer, _Rows,
matrix._DevicePointer, matrix._Rows,
1f,
result._DevicePointer, _Rows
);
cuBLAS.Destroy_v2(handle);
}
