Skip to content

nabu.processing.padding_opencl

[docs] module nabu.processing.padding_opencl 

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
import numpy as np
from ..utils import get_opencl_srcfile
from ..opencl.processing import OpenCLProcessing
from .padding_base import PaddingBase
from ..opencl.utils import __has_pyopencl__

if __has_pyopencl__:
    from ..opencl.memcpy import OpenCLMemcpy2D


class OpenCLPadding(PaddingBase):
    """
    A class for performing padding on GPU using OpenCL
    """

    backend = "opencl"

    # TODO docstring from base class
    def __init__(self, shape, pad_width, mode="constant", opencl_options=None, **kwargs):
        super().__init__(shape, pad_width, mode=mode, **kwargs)
        self.opencl_processing = self.processing = OpenCLProcessing(**(opencl_options or {}))
        self.queue = self.opencl_processing.queue
        self._init_opencl_coordinate_transform()

    def _init_opencl_coordinate_transform(self):
        if self.mode == "constant":
            self.d_padded_array_constant = self.processing.to_device(
                "d_padded_array_constant", self.padded_array_constant
            )
            self.memcpy2D = OpenCLMemcpy2D(ctx=self.processing.ctx, queue=self.queue)  # pylint: disable=E0606
            return
        self._coords_transform_kernel = self.processing.kernel(
            "coordinate_transform",
            filename=get_opencl_srcfile("padding.cl"),
        )
        self._coords_transform_global_size = self.padded_shape[::-1]
        self.d_coords_rows = self.processing.to_device("d_coords_rows", self.coords_rows)
        self.d_coords_cols = self.processing.to_device("d_coords_cols", self.coords_cols)

    def _pad_constant(self, image, output):
        pad_y, pad_x = self.pad_width
        # the following line is not implemented in pyopencl
        # self.d_padded_array_constant[pad_y[0] : pad_y[0] + self.shape[0], pad_x[0] : pad_x[0] + self.shape[1]] = image[:]
        # cl.enqueue_copy is too cumbersome to use for Buffer <-> Buffer.
        # Use a dedicated kernel instead.
        # This is not optimal (two copies) - TODO write a constant padding kernel
        self.memcpy2D(self.d_padded_array_constant, image, image.shape[::-1], dst_offset_xy=(pad_x[0], pad_y[0]))
        output[:] = self.d_padded_array_constant[:]
        return output

    def pad(self, image, output=None):
        """
        Pad an array.

        Parameters
        ----------
        image: pyopencl array
            Image to pad
        output: pyopencl array
            Output image. If provided, must be in the expected shape.
        """
        if output is None:
            output = self.processing.allocate_array("d_output", self.padded_shape)
        if self.mode == "constant":
            return self._pad_constant(image, output)
        self._coords_transform_kernel(
            image,
            output,
            self.d_coords_cols,
            self.d_coords_rows,
            np.int32(self.shape[1]),
            np.int32(self.padded_shape[1]),
            np.int32(self.padded_shape[0]),
            global_size=self._coords_transform_global_size,
        )
        return output