Advanced Usage

This section describes some details on the use of the SCA interface.

Applying Coefficient

In the SCA interface, coefficients of stencil computation can be specified as a scalar or an array. For details of the coefficients, see also Concept of Stencil Computation.

  1. A scalar constant is set as the coefficient:

    You can set the coefficient as a scalar constant. Below is an easy example:

    In [1]:

    >>> x = nlcpy.arange(10.)
    >>> dx = nlcpy.sca.create_descriptor(x)
    >>> desc = 0.5 * dx[-1] - 1.5 * dx[1]
    >>> kern = nlcpy.sca.create_kernel(desc)
    >>> kern.execute()
    

    Out[1]:

    array([  0.,  -3.,  -4.,  -5.,  -6.,  -7.,  -8.,  -9., -10.,   0.])
    
  2. An element of an nlcpy.ndarray is set as the coefficient:

    You can also set the coefficients as an element of an nlcpy.ndarray:

    In [2]:

    >>> coef = nlcpy.array([0.5, 1.5])
    >>> desc = coef[0] * dx[-1] - coef[1] * dx[1]
    >>> kern = nlcpy.sca.create_kernel(desc)
    >>> kern.execute()
    

    Out[2]:

    array([  0.,  -3.,  -4.,  -5.,  -6.,  -7.,  -8.,  -9., -10.,   0.])
    

    Here, the operators with a description (dx[-1] or dx[1]) are overloaded. The overloaded operators hold the address of the coefficient coef, not the value of that. So if the value of the coefficient is updated, the SCA kernel refers to the value at the time. In other words, if the coefficients are given as an nlcpy.ndarray object, the updated values can be used without recreating a SCA kernel.

    In [3]:

    # Execute after updating the coefficients
    >>> coef *= 2
    >>> kern.execute()
    

    Out[3]:

    array([  0.,  -6.,  -8., -10., -12., -14., -16., -18., -20.,   0.])
    

    注釈

    If the operators with a description refers to one or more temporary elements, the SCA kernel does NOT use the updated values after nlcpy.sca.create_kernel() is called because the SCA kernel refers to each address of the temporary elements, not each value of those. Below is an easy example that 1.0 * coef[0] and 1.0 * coef[1] are specified as the temporary elements.

    >>> coef = nlcpy.array([0.5, 1.5])
    >>> desc = 1.0 * coef[0] * dx[-1] - 1.0 * coef[1] * dx[1]
    >>> kern = nlcpy.sca.create_kernel(desc)
    >>> kern.execute()
    array([  0.,  -3.,  -4.,  -5.,  -6.,  -7.,  -8.,  -9., -10.,   0.])
    
    # Execute after updating the coefficients
    >>> coef *= 2
    >>> kern.execute()
    array([  0.,  -3.,  -4.,  -5.,  -6.,  -7.,  -8.,  -9., -10.,   0.]) # old coef used
    

    Similarly, if coef is set as numpy.ndarray, updating coef does not affect the results because numpy.ndarray is converted into a temporaly nlcpy.ndarray in nlcpy.sca.create_kernel().

    >>> coef = numpy.array([0.5, 1.5])
    >>> desc = coef[0] * dx[-1] - coef[1] * dx[1]
    >>> kern = nlcpy.sca.create_kernel(desc)
    
    # Execute after updating the coefficients
    >>> coef *= 2
    >>> kern.execute()
    array([  0.,  -3.,  -4.,  -5.,  -6.,  -7.,  -8.,  -9., -10.,   0.]) # old coef used
    
  3. An nlcpy.ndarray is set as the coefficient:

    You can also set the coefficients as an nlcpy.ndarray:

    In [4]:

    >>> coef = nlcpy.arange(10.)
    >>> desc = coef * dx[0]
    >>> kern = nlcpy.sca.create_kernel(desc)
    >>> kern.execute()
    

    Out[4]:

    array([ 0.,  1.,  4.,  9., 16., 25., 36., 49., 64., 81.])
    

    In[4] is similar to In[2]. The overloaded operator with the description dx[0] hold the address of coef, not the values of that. So when the values of coef are updated, the SCA kernel refers to the values at the time.

    In [5]:

    # Execute after updating the coefficients
    >>> coef *= 10
    >>> kern.execute()
    

    Out[5]:

    array([  0.,  10.,  40.,  90., 160., 250., 360., 490., 640., 810.])
    

    注釈

    If the operators with a description refers to one or more temporary arrays, the SCA kernel does NOT use the updated values after nlcpy.sca.create_kernel() is called because the SCA kernel refers to each address of the temporary arrays, not each value of those. Below is an easy example that 1.0 * coef is specified as a temporary array.

    >>> coef = nlcpy.arange(10.)
    >>> desc = 1.0 * coef * dx[0]
    >>> kern = nlcpy.sca.create_kernel(desc)
    >>> kern.execute()
    array([ 0.,  1.,  4.,  9., 16., 25., 36., 49., 64., 81.])
    
    # Execute after updating the coefficients
    >>> coef *= 10
    >>> kern.execute()
    array([ 0.,  1.,  4.,  9., 16., 25., 36., 49., 64., 81.])  # old coef used
    

    Similarly, if coef is set as numpy.ndarray, updating coef does not affect the results because numpy.ndarray is converted into a temporaly nlcpy.ndarray in nlcpy.sca.create_kernel().

Definition of Stencil Description Using for Loop

You can also use "for-loop" in a definition of a stencil description. For example, if you want to perform a stencil computation consisting of 10 elements, it is possible to define using for-loop as follows:

In [6]:

>>> x = nlcpy.arange(30, dtype='f4')
>>> dx = nlcpy.sca.create_descriptor(x)
>>> desc = nlcpy.sca.empty_description()
>>> for i in range(-10, 1):
>>>     desc += dx[i]
>>> nlcpy.sca.create_kernel(desc).execute()

Out[6]:

array([  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,  55.,
        66.,  77.,  88.,  99., 110., 121., 132., 143., 154., 165., 176.,
       187., 198., 209., 220., 231., 242., 253., 264.], dtype=float32)

In the above example, nlcpy.sca.empty_description() is used to create an empty stencil description, and 10 neighboring stencil descriptoions are added to it. Depending on the application, it is possible to create more complex stencil descriptions. The following is an example of creating the 3x3 XY planar stencil description.

In [7]:

>>> x = nlcpy.arange(25, dtype='f4').reshape(5,5)
>>> dx = nlcpy.sca.create_descriptor(x)
>>> desc = nlcpy.sca.empty_description()
>>> for i in range(-1, 2):
>>>     for j in range(-1, 2):
>>>         desc += dx[i, j]
>>> nlcpy.sca.create_kernel(desc).execute()

Out[7]:

array([[  0.,   0.,   0.,   0.,   0.],
       [  0.,  54.,  63.,  72.,   0.],
       [  0.,  99., 108., 117.,   0.],
       [  0., 144., 153., 162.,   0.],
       [  0.,   0.,   0.,   0.,   0.]], dtype=float32)

nlcpy.sca.empty_description

Returns an empty stencil description.

Automatic Creation of Output Array

If the keyword argument desc_o of nlcpy.sca.create_kernel() is omitted, an output nlcpy.ndarray is automatically created.

In [8]:

>>> nlcpy.sca.create_kernel(desc_i).execute()

Out[8]:

array([ 0.,  3.,  6.,  9., 12., 15., 18., 21., 24.,  0.], dtype=float32)

Offset Adjustment for Output Array

The offset of the output nlcpy.ndarray can be adjusted by the keyword argument desc_o of nlcpy.sca.create_kernel(). For example, if you want to set the offset to -1, desc_o is specified as follows:

In [9]:

>>> xout[...] = 0
>>> nlcpy.sca.create_kernel(desc_i, desc_o=dxout[-1]).execute()

Out[9]:

array([ 3.,  6.,  9., 12., 15., 18., 21., 24.,  0.,  0.], dtype=float32)

If you want to set the offset to 1, desc_o is specified as follows: In [10]:

>>> xout[...] = 0
>>> nlcpy.sca.create_kernel(desc_i, desc_o=dxout[1]).execute()

Out[10]:

array([ 0.,  0.,  3.,  6.,  9., 12., 15., 18., 21., 24.], dtype=float32)

注釈

Note that an IndexError occurs if the offset position is an out-of-range reference.

>>> xout[...] = 0
>>> nlcpy.sca.create_kernel(desc_i, dxout[-2]).execute()

...
IndexError

Simplified Indexing

Indexing for stencil elements can be simplified by using Ellipsis (...) or Slice(None) (:). Any axis specified by ... or : is interpreted as that indexing starts at 0.

In [11]:

>>> dx = nlcpy.sca.create_descriptor(nlcpy.random.rand(5,5,5,5))
>>> dx[...]

Out[11]:

stencil description
  in_0[0, 0, 0, 0]
assigned arrays
  in_0: shape=(5, 5, 5, 5), dtype=float64 array
computation size
  nx = 5, ny = 5, nz = 5, nw = 5

In [12]:

>>> dx[..., 1]

Out[12]:

stencil description
  in_0[0, 0, 0, 1]
assigned arrays
  in_0: shape=(5, 5, 5, 5), dtype=float64 array
computation size
  nx = 4, ny = 5, nz = 5, nw = 5

In [13]:

>>> dx[..., 1, :]

Out[13]:

stencil description
  in_0[0, 0, 1, 0]
assigned arrays
  in_0: shape=(5, 5, 5, 5), dtype=float64 array
computation size
  nx = 5, ny = 4, nz = 5, nw = 5

In [14]:

>>> dx[:, 1, ...]

Out[14]:

stencil description
  in_0[0, 1, 0, 0]
assigned arrays
  in_0: shape=(5, 5, 5, 5), dtype=float64 array
computation size
  nx = 5, ny = 5, nz = 4, nw = 5

Stencil Calculation Using Multiple Ndarrays

Even if there are two or more ndarrays used for stencil computations, the same procedure can be used.

In [15]:

>>> x = nlcpy.arange(5, dtype='f4')
>>> y = nlcpy.arange(5, 0, -1, dtype='f4')
>>> dx, dy = nlcpy.sca.create_descriptor((x, y))
>>> nlcpy.sca.create_kernel(dx[-1] + dy[1]).execute()

Out[15]:

array([0., 3., 3., 3., 0.], dtype=float32)

When shapes are different between multiple ndarrays, the narrowest range is selected as the effective range. Below is an example of x.shape = (5, 5), y.shape = (4, 6)

In [16]:

>>> x = nlcpy.arange(5 * 5, dtype='f4').reshape(5, 5)
>>> y = nlcpy.arange(4 * 6, dtype='f4').reshape(4, 6)
>>> dx, dy = nlcpy.sca.create_descriptor((x, y))
>>> desc = dx[-1, 0] + dx[1, 0] + dx[0, -1] + dx[0, 1] + dy[-1, 0] + dy[1, 0] + dy[0, -1] + dy[0,1]
>>> nlcpy.sca.create_kernel(desc).execute()

Out[16]:

array([[  0.,   0.,   0.,   0.,   0.],
       [  0.,  52.,  60.,  68.,   0.],
       [  0.,  96., 104., 112.,   0.],
       [  0.,   0.,   0.,   0.,   0.]], dtype=float32)
In the above example
  • Effective range of x: [1:4, 1:4]

  • Effective range of y: [1:3, 1:5]

Therefore, the effective range of the output nlcpy.ndarray becomes [1: 3, 1: 4], which is the minimum range for each axis. In this example, the output is automatically created, but when specifying the output destination nlcpy.ndarray, the following condition must be satisfied.

  • out.shape[i] >= min(in1.shape[i], in2.shape[i], ..., inN.shape[i])
    i: 0 to n (For n-dimensional arrays)

When the output array is automatically created, an nlcpy.ndarray whose shape meets above condition is returned by nlcpy.sca.create_kernel().

Stencil Calculation for Complex Types

The SCA interface does not support complex number types. When performing stencil computations for complex numbers, use nlcpy.ndarray.real and nlcpy.ndarray.imag as follows:

In [17]:

>>> xin = nlcpy.arange(10) + nlcpy.arange(10, 0, -1) * 1j
>>> xout = nlcpy.zeros_like(xin)
>>> dRe, dIm = nlcpy.sca.create_descriptor((xin.real, xin.imag))
>>> xout.real = nlcpy.sca.create_kernel(dRe[-1] + dRe[1]).execute()
>>> xout.imag = nlcpy.sca.create_kernel(dIm[-1] + dIm[1]).execute()
>>> xout

Out[17]:

array([ 0. +0.j,  2.+18.j,  4.+16.j,  6.+14.j,  8.+12.j, 10.+10.j,
       12. +8.j, 14. +6.j, 16. +4.j,  0. +0.j])