nabu 2024.1.1__py3-none-any.whl → 2024.1.2__py3-none-any.whl

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/muladd.py

@@ -1,29 +0,0 @@
-import numpy as np
-from .processing_base import ProcessingBase
-
-
-class MulAdd:
-    processing_cls = ProcessingBase
-
-    def __init__(self, **backend_options):
-        self.processing = self.processing_cls(**(backend_options or {}))
-        self._init_finalize()
-
-    def _init_finalize(self):
-        pass
-
-    def mul_add(self, dst, other, fac_dst, fac_other, dst_region=None, other_region=None):
-        if dst_region is None:
-            dst_slice_y = dst_slice_x = slice(None, None)
-        else:
-            dst_slice_y, dst_slice_x = dst_region
-        if other_region is None:
-            other_slice_y = other_slice_x = slice(None, None)
-        else:
-            other_slice_y, other_slice_x = other_region
-
-        dst[dst_slice_y, dst_slice_x] = (
-            fac_dst * dst[dst_slice_y, dst_slice_x] + fac_other * other[other_slice_y, other_slice_x]
-        )
-
-    __call__ = mul_add

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/muladd_cuda.py

@@ -1,68 +0,0 @@
-import numpy as np
-from nabu.utils import get_cuda_srcfile, updiv
-from .muladd import MulAdd
-from ..cuda.utils import __has_pycuda__
-from ..cuda.processing import CudaProcessing
-
-if __has_pycuda__:
-    import pycuda.gpuarray as garray
-
-
-class CudaMulAdd(MulAdd):
-    processing_cls = CudaProcessing
-
-    def _init_finalize(self):
-        self._init_kernel()
-
-    def _init_kernel(self):
-        self.muladd_kernel = self.processing.kernel(
-            "mul_add",
-            filename=get_cuda_srcfile("ElementOp.cu"),  # signature="PPiiffiiii"
-        )
-
-    def mul_add(self, dst, other, fac_dst, fac_other, dst_region=None, other_region=None):
-        """
-        'region' should be a tuple (slice(y_start, y_end), slice(x_start, x_end))
-        """
-
-        if dst_region is None:
-            dst_coords = (0, dst.shape[1], 0, dst.shape[0])
-        else:
-            dst_coords = (dst_region[1].start, dst_region[1].stop, dst_region[0].start, dst_region[0].stop)
-        if other_region is None:
-            other_coords = (0, other.shape[1], 0, other.shape[0])
-        else:
-            other_coords = (other_region[1].start, other_region[1].stop, other_region[0].start, other_region[0].stop)
-
-        delta_x = np.diff(dst_coords[:2])
-        delta_y = np.diff(dst_coords[2:])
-        if delta_x != np.diff(other_coords[:2]) or delta_y != np.diff(other_coords[2:]):
-            raise ValueError("Invalid dst_region and other_region provided. Regions must have the same size")
-        if delta_x == 0 or delta_y == 0:
-            raise ValueError("delta_x or delta_y is 0")
-
-        # can't use "int4" in pycuda ? int2 seems fine. Go figure
-        dst_x_range = np.array(dst_coords[:2], dtype=garray.vec.int2)
-        dst_y_range = np.array(dst_coords[2:], dtype=garray.vec.int2)
-        other_x_range = np.array(other_coords[:2], dtype=garray.vec.int2)
-        other_y_range = np.array(other_coords[2:], dtype=garray.vec.int2)
-
-        block = (32, 32, 1)
-        grid = [updiv(length, b) for (length, b) in zip((delta_x[0], delta_y[0]), block)]
-
-        self.muladd_kernel(
-            dst,
-            other,
-            np.int32(dst.shape[1]),
-            np.int32(other.shape[1]),
-            np.float32(fac_dst),
-            np.float32(fac_other),
-            dst_x_range,
-            dst_y_range,
-            other_x_range,
-            other_y_range,
-            grid=grid,
-            block=block,
-        )
-
-    __call__ = mul_add

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/padding_base.py

@@ -1,71 +0,0 @@
-import numpy as np
-from ..utils import check_supported
-
-
-class PaddingBase:
-    """
-    A class for performing padding based on coordinate transform.
-    The Cuda and OpenCL backends will subclass this class.
-    """
-
-    supported_modes = ["constant", "edge", "reflect", "symmetric", "wrap"]
-
-    def __init__(self, shape, pad_width, mode="constant", **kwargs):
-        """
-        Initialize a Padding object.
-
-        Parameters
-        ----------
-        shape: tuple
-            Image shape
-        pad_width: tuple
-            Padding width for each axis. Please see the documentation of numpy.pad().
-        mode: str
-            Padding mode
-
-        Other parameters
-        ----------------
-        constant_values: tuple
-            Tuple containing the values to fill when mode="constant" (as in numpy.pad)
-        """
-        if len(shape) != 2:
-            raise ValueError("This class only works on images")
-        self.shape = shape
-        self._set_mode(mode, **kwargs)
-        self._get_padding_arrays(pad_width)
-
-    def _set_mode(self, mode, **kwargs):
-        # COMPAT.
-        if mode == "edges":
-            mode = "edge"
-        #
-        check_supported(mode, self.supported_modes, "padding mode")
-        self.mode = mode
-        self._kwargs = kwargs
-
-    def _get_padding_arrays(self, pad_width):
-        self.pad_width = pad_width
-        if isinstance(pad_width, tuple) and isinstance(pad_width[0], np.ndarray):
-            # user-defined coordinate transform
-            err_msg = "pad_width must be either a scalar, a tuple in the form ((a, b), (c, d)), or a tuple of two one-dimensional numpy arrays (eg. use numpy.indices(..., sparse=True))"
-            if len(pad_width) != 2:
-                raise ValueError(err_msg)
-            if any([np.squeeze(pw).ndim > 1 for pw in pad_width]):
-                raise ValueError(err_msg)
-            if self.mode == "constant":
-                raise ValueError("Custom coordinate transform does not work with mode='constant'")
-            self.mode = "custom"
-            self.coords_rows, self.coords_cols = pad_width
-        else:
-            if self.mode == "constant":
-                # no need for coordinate transform here
-                constant_values = self._kwargs.get("constant_values", 0)
-                self.padded_array_constant = np.pad(
-                    np.zeros(self.shape, dtype="f"), self.pad_width, mode="constant", constant_values=constant_values
-                )
-                self.padded_shape = self.padded_array_constant.shape
-                return
-            R, C = np.indices(self.shape, dtype=np.int32, sparse=True)
-            self.coords_rows = np.pad(R.ravel(), self.pad_width[0], mode=self.mode)
-            self.coords_cols = np.pad(C.ravel(), self.pad_width[1], mode=self.mode)
-        self.padded_shape = (self.coords_rows.size, self.coords_cols.size)

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/padding_cuda.py

@@ -1,75 +0,0 @@
-import numpy as np
-from ..utils import get_cuda_srcfile, updiv
-from ..cuda.processing import CudaProcessing
-from ..cuda.utils import __has_pycuda__
-from .padding_base import PaddingBase
-
-
-class CudaPadding(PaddingBase):
-    """
-    A class for performing padding on GPU using Cuda
-    """
-
-    backend = "cuda"
-
-    # TODO docstring from base class
-    def __init__(self, shape, pad_width, mode="constant", cuda_options=None, **kwargs):
-        super().__init__(shape, pad_width, mode=mode, **kwargs)
-        self.cuda_processing = self.processing = CudaProcessing(**(cuda_options or {}))
-        self._init_cuda_coordinate_transform()
-
-    def _init_cuda_coordinate_transform(self):
-        if self.mode == "constant":
-            self.d_padded_array_constant = self.processing.to_device(
-                "d_padded_array_constant", self.padded_array_constant
-            )
-            return
-        self._coords_transform_kernel = self.processing.kernel(
-            "coordinate_transform",
-            filename=get_cuda_srcfile("padding.cu"),
-            signature="PPPPiii",
-        )
-        self._coords_transform_block = (32, 32, 1)
-        self._coords_transform_grid = [
-            updiv(a, b) for a, b in zip(self.padded_shape[::-1], self._coords_transform_block)
-        ]
-        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
-        self.d_padded_array_constant[pad_y[0] : pad_y[0] + self.shape[0], pad_x[0] : pad_x[0] + self.shape[1]] = image[
-            :
-        ]
-        output[:] = self.d_padded_array_constant[:]
-        return output
-
-    def pad(self, image, output=None):
-        """
-        Pad an array.
-
-        Parameters
-        ----------
-        image: pycuda.gpuarray.GPUArray
-            Image to pad
-        output: pycuda.gpuarray.GPUArray, optional
-            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]),
-            grid=self._coords_transform_grid,
-            block=self._coords_transform_block,
-        )
-        return output
-
-    __call__ = pad

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/padding_opencl.py

@@ -1,77 +0,0 @@
-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)
-            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(
-            self.queue,
-            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

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/processing_base.py

@@ -1,123 +0,0 @@
-import numpy as np
-from ..utils import BaseClassError
-
-"""
-Base class for OpenCLProcessing and CudaProcessing
-Should not be used directly
-"""
-
-
-class ProcessingBase:
-    array_class = None
-    dtype_to_ctype = BaseClassError
-
-    def __init__(self):
-        self._allocated = {}
-
-    def init_arrays_to_none(self, arrays_names):
-        """
-        Initialize arrays to None. After calling this method, the current instance will
-        have self.array_name = None, and self._old_array_name = None.
-
-        Parameters
-        ----------
-        arrays_names: list of str
-            List of arrays names.
-        """
-        for array_name in arrays_names:
-            setattr(self, array_name, None)
-            setattr(self, "_old_" + array_name, None)
-            self._allocated[array_name] = False
-
-    def recover_arrays_references(self, arrays_names):
-        """
-        Performs self._array_name = self._old_array_name,
-        for each array_name in arrays_names.
-
-        Parameters
-        ----------
-        arrays_names: list of str
-            List of array names
-        """
-        for array_name in arrays_names:
-            old_arr = getattr(self, "_old_" + array_name, None)
-            if old_arr is not None:
-                setattr(self, array_name, old_arr)
-
-    def _allocate_array_mem(self, shape, dtype):
-        raise ValueError("Base class")
-
-    def allocate_array(self, array_name, shape, dtype=np.float32):
-        """
-        Allocate a GPU array on the current context/stream/device,
-        and set 'self.array_name' to this array.
-
-        Parameters
-        ----------
-        array_name: str
-            Name of the array (for book-keeping)
-        shape: tuple of int
-            Array shape
-        dtype: numpy.dtype, optional
-            Data type. Default is float32.
-        """
-        if not self._allocated.get(array_name, False):
-            new_device_arr = self._allocate_array_mem(shape, dtype)
-            setattr(self, array_name, new_device_arr)
-            self._allocated[array_name] = True
-        return getattr(self, array_name)
-
-    def set_array(self, array_name, array_ref, dtype=np.float32):
-        """
-        Set the content of a device array.
-
-        Parameters
-        ----------
-        array_name: str
-            Array name. This method will look for self.array_name.
-        array_ref: array (numpy or GPU array)
-            Array containing the data to copy to 'array_name'.
-        dtype: numpy.dtype, optional
-            Data type. Default is float32.
-        """
-        if isinstance(array_ref, self.array_class):
-            current_arr = getattr(self, array_name, None)
-            setattr(self, "_old_" + array_name, current_arr)
-            setattr(self, array_name, array_ref)
-        elif isinstance(array_ref, np.ndarray):
-            self.allocate_array(array_name, array_ref.shape, dtype=dtype)
-            getattr(self, array_name).set(array_ref)
-        else:
-            raise ValueError("Expected numpy array or pycuda array")
-        return getattr(self, array_name)
-
-    def get_array(self, array_name):
-        return getattr(self, array_name, None)
-
-    # COMPAT.
-    _init_arrays_to_none = init_arrays_to_none
-    _recover_arrays_references = recover_arrays_references
-    _allocate_array = allocate_array
-    _set_array = set_array
-
-    def check_array(self, arr, expected_shape, expected_dtype="f", check_contiguous=True):
-        """
-        Check whether a given array is suitable for being processed (shape, dtype, contiguous)
-        """
-        if arr.shape != expected_shape:
-            raise ValueError("Expected shape %s but got %s" % (str(expected_shape), str(arr.shape)))
-        if arr.dtype != np.dtype(expected_dtype):
-            raise ValueError("Expected data type %s but got %s" % (str(expected_dtype), str(arr.dtype)))
-        if check_contiguous:
-            if isinstance(arr, np.ndarray) and not (arr.flags["C_CONTIGUOUS"]):
-                raise ValueError("Expected C-contiguous array")
-            if isinstance(arr, self.array_class) and not arr.flags.c_contiguous:
-                raise ValueError("Expected C-contiguous array")
-
-    def kernel(self, *args, **kwargs):
-        raise ValueError("Base class")
-
-    def to_device(self, array_name, array):
-        arr_ref = self.allocate_array(array_name, array.shape, dtype=array.dtype)
-        arr_ref.set(array)
-        return arr_ref

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/roll_opencl.py

@@ -1,64 +0,0 @@
-#
-# WIP !
-#
-import numpy as np
-from ..opencl.utils import __has_pyopencl__
-from ..utils import get_opencl_srcfile
-
-if __has_pyopencl__:
-    import pyopencl as cl
-    from ..opencl.processing import OpenCLProcessing
-    from ..opencl.kernel import OpenCLKernel
-    from pyopencl.tools import dtype_to_ctype as cl_dtype_to_ctype
-
-
-class OpenCLRoll:
-    def __init__(self, dtype, direction=1, offset=None, **processing_kwargs):
-        self.processing = OpenCLProcessing(queue=processing_kwargs.get("queue", None))
-        self.dtype = np.dtype(dtype)
-        compile_options = ["-DDTYPE=%s" % cl_dtype_to_ctype(self.dtype)]
-        self.offset = offset or 0
-        self.roll_kernel = OpenCLKernel(
-            "roll_forward_x",
-            None,
-            queue=self.processing.queue,
-            filename=get_opencl_srcfile("roll.cl"),
-            options=compile_options,
-        )
-        self.shmem = cl.LocalMemory(self.dtype.itemsize)
-        self.direction = direction
-        if self.direction < 0:
-            self.revert_kernel = OpenCLKernel(
-                "revert_array_x",
-                None,
-                queue=self.processing.queue,
-                filename=get_opencl_srcfile("roll.cl"),
-                options=compile_options,
-            )
-
-    def __call__(self, arr):
-        ny, nx = arr.shape
-        # Launch one big horizontal workgroup
-        wg_x = min((nx - self.offset) // 2, self.processing.queue.device.max_work_group_size)
-        local_size = (wg_x, 1, 1)
-        global_size = [wg_x, ny]
-        if self.direction < 0:
-            local_size2 = None
-            global_size2 = [nx - self.offset, ny]
-            self.revert_kernel(
-                arr, np.int32(nx), np.int32(ny), np.int32(self.offset), local_size=local_size2, global_size=global_size2
-            )
-        self.roll_kernel(
-            arr,
-            np.int32(nx),
-            np.int32(ny),
-            np.int32(self.offset),
-            self.shmem,
-            local_size=local_size,
-            global_size=global_size,
-        )
-        if self.direction < 0:
-            self.revert_kernel(
-                arr, np.int32(nx), np.int32(ny), np.int32(self.offset), local_size=local_size2, global_size=global_size2
-            )
-        return arr

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/rotation.py

@@ -1,63 +0,0 @@
-try:
-    from skimage.transform import rotate
-
-    __have__skimage__ = True
-except ImportError:
-    __have__skimage__ = False
-
-
-class Rotation:
-    supported_modes = {
-        "constant": "constant",
-        "zeros": "constant",
-        "edge": "edge",
-        "edges": "edge",
-        "symmetric": "symmetric",
-        "sym": "symmetric",
-        "reflect": "reflect",
-        "wrap": "wrap",
-        "periodic": "wrap",
-    }
-
-    def __init__(self, shape, angle, center=None, mode="edge", reshape=False, **sk_kwargs):
-        """
-        Initiate a Rotation object.
-
-        Parameters
-        ----------
-        shape: tuple of int
-            Shape of the images to process
-        angle: float
-            Rotation angle in DEGREES
-        center: tuple of float, optional
-            Coordinates of the center of rotation, in the format (X, Y) (mind the non-python
-            convention !).
-            Default is ((Nx - 1)/2.0, (Ny - 1)/2.0)
-        mode: str, optional
-            Padding mode. Default is "edge".
-        reshape: bool, optional
-
-
-        Other Parameters
-        -----------------
-        All the other parameters are passed directly to scikit image 'rotate' function:
-        order, cval, clip, preserve_range.
-        """
-        self.shape = shape
-        self.angle = angle
-        self.center = center
-        self.mode = mode
-        self.reshape = reshape
-        self.sk_kwargs = sk_kwargs
-
-    def rotate(self, img, output=None):
-        if not __have__skimage__:
-            raise ValueError("scikit-image is needed for using rotate()")
-        res = rotate(img, self.angle, resize=self.reshape, center=self.center, mode=self.mode, **self.sk_kwargs)
-        if output is not None:
-            output[:] = res[:]
-            return output
-        else:
-            return res
-
-    __call__ = rotate

build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/build/lib/nabu/processing/rotation_cuda.py

@@ -1,66 +0,0 @@
-import numpy as np
-from .rotation import Rotation
-from ..utils import get_cuda_srcfile, updiv
-from ..cuda.utils import __has_pycuda__, copy_array
-from ..cuda.processing import CudaProcessing
-
-if __has_pycuda__:
-    from ..cuda.kernel import CudaKernel
-    import pycuda.driver as cuda
-
-
-class CudaRotation(Rotation):
-    def __init__(self, shape, angle, center=None, mode="edge", reshape=False, cuda_options=None, **sk_kwargs):
-        if center is None:
-            center = ((shape[1] - 1) / 2.0, (shape[0] - 1) / 2.0)
-        super().__init__(shape, angle, center=center, mode=mode, reshape=reshape, **sk_kwargs)
-        self._init_cuda_rotation(cuda_options)
-
-    def _init_cuda_rotation(self, cuda_options):
-        cuda_options = cuda_options or {}
-        self.cuda_processing = CudaProcessing(**cuda_options)
-        self._allocate_arrays()
-        self._init_rotation_kernel()
-
-    def _allocate_arrays(self):
-        self._d_image_cua = cuda.np_to_array(np.zeros(self.shape, "f"), "C")
-        self.cuda_processing.init_arrays_to_none(["d_output"])
-
-    def _init_rotation_kernel(self):
-        self.cuda_rotation_kernel = CudaKernel("rotate", get_cuda_srcfile("rotation.cu"))
-        self.texref_image = self.cuda_rotation_kernel.module.get_texref("tex_image")
-        self.texref_image.set_filter_mode(cuda.filter_mode.LINEAR)  # bilinear
-        self.texref_image.set_address_mode(0, cuda.address_mode.CLAMP)  # TODO tune
-        self.texref_image.set_address_mode(1, cuda.address_mode.CLAMP)  # TODO tune
-        self.cuda_rotation_kernel.prepare("Piiffff", [self.texref_image])
-        self.texref_image.set_array(self._d_image_cua)
-        self._cos_theta = np.cos(np.deg2rad(self.angle))
-        self._sin_theta = np.sin(np.deg2rad(self.angle))
-        self._Nx = np.int32(self.shape[1])
-        self._Ny = np.int32(self.shape[0])
-        self._center_x = np.float32(self.center[0])
-        self._center_y = np.float32(self.center[1])
-        self._block = (32, 32, 1)  # tune ?
-        self._grid = (updiv(self.shape[1], self._block[1]), updiv(self.shape[0], self._block[0]), 1)
-
-    def rotate(self, img, output=None, do_checks=True):
-        copy_array(self._d_image_cua, img, check=do_checks)
-        if output is not None:
-            d_out = output
-        else:
-            self.cuda_processing.allocate_array("d_output", self.shape, np.float32)
-            d_out = self.cuda_processing.d_output
-        self.cuda_rotation_kernel(
-            d_out,
-            self._Nx,
-            self._Ny,
-            self._cos_theta,
-            self._sin_theta,
-            self._center_x,
-            self._center_y,
-            grid=self._grid,
-            block=self._block,
-        )
-        return d_out
-
-    __call__ = rotate