Function _conv2d_hw

python/tvm/topi/testing/conv2d_nchw_python.py:80–122

2d convolution operator in HW layout. This is intended to be used as a subroutine from _conv2d_nchw_python. Using scipy.signal.convolve2d directly does not work for all dtypes (e.g. float16). Where possible, this function uses scipy.signal.convolve2d to take advantage of compi

(apad, w_np_fc)

def _conv2d_hw(apad, w_np_fc):
    """2d convolution operator in HW layout.

    This is intended to be used as a subroutine from
    _conv2d_nchw_python.  Using scipy.signal.convolve2d directly does
    not work for all dtypes (e.g. float16).  Where possible, this
    function uses scipy.signal.convolve2d to take advantage of
    compiled scipy routines, falling back to an explicit loop only
    where needed

    Parameters
    ----------
    a_np : numpy.ndarray
        2-D with shape [in_height, in_width]

    w_np : numpy.ndarray
        2-D with shape [filter_height, filter_width].

    Returns
    -------
    b_np : np.ndarray
        2-D with shape [out_height, out_width]
    """

    try:
        return scipy.signal.convolve2d(apad, np.rot90(np.rot90(w_np_fc)), mode="valid")
    except ValueError:
        pass

    assert len(apad.shape) == len(w_np_fc.shape) == 2

    dtype = apad.dtype
    in_height, in_width = apad.shape
    kernel_h, kernel_w = w_np_fc.shape

    output_shape = [a_dim - w_dim + 1 for a_dim, w_dim in zip(apad.shape, w_np_fc.shape)]
    output = np.zeros(output_shape, dtype=apad.dtype)

    for y in range(output_shape[0]):
        for x in range(output_shape[1]):
            output[y][x] = np.sum(apad[y : y + kernel_h, x : x + kernel_w] * w_np_fc)

    return output


def conv2d_nchw_python(a_np, w_np, stride, padding, groups=1):