indexing.py

import abc
from typing import Tuple  # noqa
import sympy as sp
from pystencils.astnodes import Conditional, Block
from pystencils.slicing import normalize_slice
from pystencils.data_types import TypedSymbol, create_type
from functools import partial

AUTO_BLOCK_SIZE_LIMITING = True

BLOCK_IDX = [TypedSymbol("blockIdx." + coord, create_type("int")) for coord in ('x', 'y', 'z')]
THREAD_IDX = [TypedSymbol("threadIdx." + coord, create_type("int")) for coord in ('x', 'y', 'z')]
BLOCK_DIM = [TypedSymbol("blockDim." + coord, create_type("int")) for coord in ('x', 'y', 'z')]
GRID_DIM = [TypedSymbol("gridDim." + coord, create_type("int")) for coord in ('x', 'y', 'z')]


class AbstractIndexing(abc.ABC):
    """
    Abstract base class for all Indexing classes. An Indexing class defines how a multidimensional
    field is mapped to CUDA's block and grid system. It calculates indices based on CUDA's thread and block indices
    and computes the number of blocks and threads a kernel is started with. The Indexing class is created with
    a pystencils field, a slice to iterate over, and further optional parameters that must have default values.
    """

    @property
    @abc.abstractmethod
    def coordinates(self):
        """Returns a sequence of coordinate expressions for (x,y,z) depending on symbolic CUDA block and thread indices.
        These symbolic indices can be obtained with the method `index_variables` """

    @property
    def index_variables(self):
        """Sympy symbols for CUDA's block and thread indices, and block and grid dimensions. """
        return BLOCK_IDX + THREAD_IDX + BLOCK_DIM + GRID_DIM

    @abc.abstractmethod
    def call_parameters(self, arr_shape):
        """
        Determine grid and block size for kernel call
        :param arr_shape: the numeric (not symbolic) shape of the array
        :return: dict with keys 'blocks' and 'threads' with tuple values for number of (x,y,z) threads and blocks
                 the kernel should be started with
        """

    @abc.abstractmethod
    def guard(self, kernel_content, arr_shape):
        """
        In some indexing schemes not all threads of a block execute the kernel content.
        This function can return a Conditional ast node, defining this execution guard.
        :param kernel_content: the actual kernel contents which can e.g. be put into the Conditional node as true block
        :param arr_shape: the numeric or symbolic shape of the field
        :return: ast node, which is put inside the kernel function
        """


# -------------------------------------------- Implementations ---------------------------------------------------------


class BlockIndexing(AbstractIndexing):
    """Generic indexing scheme that maps sub-blocks of an array to CUDA blocks.

    Args:
        field: pystencils field (common to all Indexing classes)
        iteration_slice: slice that defines rectangular subarea which is iterated over
        permute_block_size_dependent_on_layout: if True the block_size is permuted such that the fastest coordinate
                                                gets the largest amount of threads
    """

    def __init__(self, field, iteration_slice=None,
                 block_size=(16, 16, 1), permute_block_size_dependent_on_layout=True):
        if field.spatial_dimensions > 3:
            raise NotImplementedError("This indexing scheme supports at most 3 spatial dimensions")

        if permute_block_size_dependent_on_layout:
            block_size = self.permute_block_size_according_to_layout(block_size, field.layout)

        if AUTO_BLOCK_SIZE_LIMITING:
            block_size = self.limit_block_size_to_device_maximum(block_size)

        self._blockSize = block_size
        self._iterationSlice = normalize_slice(iteration_slice, field.spatial_shape)
        self._dim = field.spatial_dimensions
        self._symbolicShape = [e if isinstance(e, sp.Basic) else None for e in field.spatial_shape]

    @property
    def coordinates(self):
        offsets = _get_start_from_slice(self._iterationSlice)
        coordinates = [block_index * bs + thread_idx + off
                       for block_index, bs, thread_idx, off in zip(BLOCK_IDX, self._blockSize, THREAD_IDX, offsets)]

        return coordinates[:self._dim]

    def call_parameters(self, arr_shape):
        substitution_dict = {sym: value for sym, value in zip(self._symbolicShape, arr_shape) if sym is not None}

        widths = [end - start for start, end in zip(_get_start_from_slice(self._iterationSlice),
                                                    _get_end_from_slice(self._iterationSlice, arr_shape))]
        widths = sp.Matrix(widths).subs(substitution_dict)

        grid = tuple(sp.ceiling(length / block_size)
                     for length, block_size in zip(widths, self._blockSize))  # type: : Tuple[int, ...]
        extend_bs = (1,) * (3 - len(self._blockSize))
        extend_gr = (1,) * (3 - len(grid))

        return {'block': self._blockSize + extend_bs,
                'grid': grid + extend_gr}

    def guard(self, kernel_content, arr_shape):
        arr_shape = arr_shape[:self._dim]
        conditions = [c < end
                      for c, end in zip(self.coordinates, _get_end_from_slice(self._iterationSlice, arr_shape))]
        condition = conditions[0]
        for c in conditions[1:]:
            condition = sp.And(condition, c)
        return Block([Conditional(condition, kernel_content)])

    @staticmethod
    def limit_block_size_to_device_maximum(block_size):
        """
        Changes block size according to match device limits according to the following rules:
        1) if the total amount of threads is too big for the current device, the biggest coordinate is divided by 2.
        2) next, if one component is still too big, the component which is too big is divided by 2 and the smallest
           component is multiplied by 2, such that the total amount of threads stays the same
        Returns the altered block_size
        """
        # Get device limits
        import pycuda.driver as cuda
        # noinspection PyUnresolvedReferences
        import pycuda.autoinit  # NOQA

        da = cuda.device_attribute
        device = cuda.Context.get_device()

        block_size = list(block_size)
        max_threads = device.get_attribute(da.MAX_THREADS_PER_BLOCK)
        max_block_size = [device.get_attribute(a)
                          for a in (da.MAX_BLOCK_DIM_X, da.MAX_BLOCK_DIM_Y, da.MAX_BLOCK_DIM_Z)]

        def prod(seq):
            result = 1
            for e in seq:
                result *= e
            return result

        def get_index_of_too_big_element():
            for i, bs in enumerate(block_size):
                if bs > max_block_size[i]:
                    return i
            return None

        def get_index_of_too_small_element():
            for i, bs in enumerate(block_size):
                if bs // 2 <= max_block_size[i]:
                    return i
            return None

        # Reduce the total number of threads if necessary
        while prod(block_size) > max_threads:
            item_to_reduce = block_size.index(max(block_size))
            for j, block_size_entry in enumerate(block_size):
                if block_size_entry > max_block_size[j]:
                    item_to_reduce = j
            block_size[item_to_reduce] //= 2

        # Cap individual elements
        too_big_element_index = get_index_of_too_big_element()
        while too_big_element_index is not None:
            too_small_element_index = get_index_of_too_small_element()
            block_size[too_small_element_index] *= 2
            block_size[too_big_element_index] //= 2
            too_big_element_index = get_index_of_too_big_element()

        return tuple(block_size)

    @staticmethod
    def limit_block_size_by_register_restriction(block_size, required_registers_per_thread, device=None):
        """Shrinks the block_size if there are too many registers used per multiprocessor.
        This is not done automatically, since the required_registers_per_thread are not known before compilation.
        They can be obtained by ``func.num_regs`` from a pycuda function.
        :returns smaller block_size if too many registers are used.
        """
        import pycuda.driver as cuda
        # noinspection PyUnresolvedReferences
        import pycuda.autoinit  # NOQA

        da = cuda.device_attribute
        if device is None:
            device = cuda.Context.get_device()
        available_registers_per_mp = device.get_attribute(da.MAX_REGISTERS_PER_MULTIPROCESSOR)

        block = block_size

        while True:
            num_threads = 1
            for t in block:
                num_threads *= t
            required_registers_per_mt = num_threads * required_registers_per_thread
            if required_registers_per_mt <= available_registers_per_mp:
                return block
            else:
                largest_grid_entry_idx = max(range(len(block)), key=lambda e: block[e])
                assert block[largest_grid_entry_idx] >= 2
                block[largest_grid_entry_idx] //= 2

    @staticmethod
    def permute_block_size_according_to_layout(block_size, layout):
        """Returns modified block_size such that the fastest coordinate gets the biggest block dimension"""
        sorted_block_size = list(sorted(block_size, reverse=True))
        while len(sorted_block_size) > len(layout):
            sorted_block_size[0] *= sorted_block_size[-1]
            sorted_block_size = sorted_block_size[:-1]

        result = list(block_size)
        for l, bs in zip(reversed(layout), sorted_block_size):
            result[l] = bs
        return tuple(result[:len(layout)])


class LineIndexing(AbstractIndexing):
    """
    Indexing scheme that assigns the innermost 'line' i.e. the elements which are adjacent in memory to a 1D CUDA block.
    The fastest coordinate is indexed with thread_idx.x, the remaining coordinates are mapped to block_idx.{x,y,z}
    This indexing scheme supports up to 4 spatial dimensions, where the innermost dimensions is not larger than the
    maximum amount of threads allowed in a CUDA block (which depends on device).
    """

    def __init__(self, field, iteration_slice=None):
        available_indices = [THREAD_IDX[0]] + BLOCK_IDX
        if field.spatial_dimensions > 4:
            raise NotImplementedError("This indexing scheme supports at most 4 spatial dimensions")

        coordinates = available_indices[:field.spatial_dimensions]

        fastest_coordinate = field.layout[-1]
        coordinates[0], coordinates[fastest_coordinate] = coordinates[fastest_coordinate], coordinates[0]

        self._coordinates = coordinates
        self._iterationSlice = normalize_slice(iteration_slice, field.spatial_shape)
        self._symbolicShape = [e if isinstance(e, sp.Basic) else None for e in field.spatial_shape]

    @property
    def coordinates(self):
        return [i + offset for i, offset in zip(self._coordinates, _get_start_from_slice(self._iterationSlice))]

    def call_parameters(self, arr_shape):
        substitution_dict = {sym: value for sym, value in zip(self._symbolicShape, arr_shape) if sym is not None}

        widths = [end - start for start, end in zip(_get_start_from_slice(self._iterationSlice),
                                                    _get_end_from_slice(self._iterationSlice, arr_shape))]
        widths = sp.Matrix(widths).subs(substitution_dict)

        def get_shape_of_cuda_idx(cuda_idx):
            if cuda_idx not in self._coordinates:
                return 1
            else:
                idx = self._coordinates.index(cuda_idx)
                return widths[idx]

        return {'block': tuple([get_shape_of_cuda_idx(idx) for idx in THREAD_IDX]),
                'grid': tuple([get_shape_of_cuda_idx(idx) for idx in BLOCK_IDX])}

    def guard(self, kernel_content, arr_shape):
        return kernel_content


# -------------------------------------- Helper functions --------------------------------------------------------------

def _get_start_from_slice(iteration_slice):
    res = []
    for slice_component in iteration_slice:
        if type(slice_component) is slice:
            res.append(slice_component.start if slice_component.start is not None else 0)
        else:
            assert isinstance(slice_component, int)
            res.append(slice_component)
    return res


def _get_end_from_slice(iteration_slice, arr_shape):
    iter_slice = normalize_slice(iteration_slice, arr_shape)
    res = []
    for slice_component in iter_slice:
        if type(slice_component) is slice:
            res.append(slice_component.stop)
        else:
            assert isinstance(slice_component, int)
            res.append(slice_component + 1)
    return res


def indexing_creator_from_params(gpu_indexing, gpu_indexing_params):
    if isinstance(gpu_indexing, str):
        if gpu_indexing == 'block':
            indexing_creator = BlockIndexing
        elif gpu_indexing == 'line':
            indexing_creator = LineIndexing
        else:
            raise ValueError("Unknown GPU indexing %s. Valid values are 'block' and 'line'" % (gpu_indexing,))
        if gpu_indexing_params:
            indexing_creator = partial(indexing_creator, **gpu_indexing_params)
        return indexing_creator
    else:
        return gpu_indexing