Notice Regarding SCA Interface

  • The SCA interface supports up to 4 dimension nlcpy.ndarray. It can handle stencil computations of arbitrary shape (axial, planar, diagonal, and its combination).

  • In terms of dtype (data type), the current version of the SCA interface supports only float32 and float64.

  • nlcpy.sca.create_kernel() dynamically generates a instruction sequence required to perform stencil computations on VE. This generating cost is not small compared to nlcpy.sca.kernel.kernel.execute(), so the overall performance gets better as the number of kernel executions per kernel generation increase.

  • An output nlcpy.ndarray may not be shared with an input nlcpy.ndarray. Otherwise, unpredictable results occurs.