Multistage rechunking for regular chunk grids

Author: tomwhite

cubed/core/ops.py

@@ -18,6 +18,7 @@
 from cubed.backend_array_api import namespace as nxp
 from cubed.core.array import CoreArray, check_array_specs, compute, gensym
 from cubed.core.plan import Plan, intermediate_store
+from cubed.core.rechunk import multistage_regular_rechunking_plan
 from cubed.primitive.blockwise import blockwise as primitive_blockwise
 from cubed.primitive.blockwise import general_blockwise as primitive_general_blockwise
 from cubed.primitive.memory import get_buffer_copies
@@ -39,7 +40,6 @@
 from cubed.vendor.dask.array.utils import validate_axis
 from cubed.vendor.dask.blockwise import broadcast_dimensions
 from cubed.vendor.dask.utils import has_keyword
-from cubed.vendor.rechunker.algorithm import multistage_rechunking_plan
 
 if TYPE_CHECKING:
     from cubed.array_api.array_object import Array
@@ -1068,7 +1068,7 @@ def _rechunk_plan(x, chunks, *, min_mem=None):
     rechunker_max_mem = (spec.allowed_mem - spec.reserved_mem) // total_copies
     if min_mem is None:
         min_mem = min(rechunker_max_mem // 20, x.nbytes)
-    stages = multistage_rechunking_plan(
+    stages = multistage_regular_rechunking_plan(
         shape=x.shape,
         source_chunks=source_chunks,
         target_chunks=target_chunks,

cubed/core/rechunk.py

@@ -0,0 +1,232 @@
+"""Core rechunking algorithm based on rechunker, but adapted for Cubed to support regular Zarr chunks."""
+
+import logging
+import warnings
+from math import floor, prod
+from typing import List, Optional, Sequence
+
+import numpy as np
+
+from cubed.vendor.rechunker.algorithm import (
+    MAX_STAGES,
+    ExcessiveIOWarning,
+    _calculate_shared_chunks,
+    _MultistagePlan,
+    calculate_single_stage_io_ops,
+    consolidate_chunks,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def verify_chunk_compatibility(
+    shape,
+    write_chunks,
+    target_chunks,
+):
+    for n, wc, tc in zip(shape, write_chunks, target_chunks):
+        assert (wc == n) or (wc % tc == 0), (
+            f"write chunks {write_chunks} do not evenly slice target chunks {target_chunks}, "
+            f"since {wc} is not a multiple of {tc}"
+        )
+
+
+def multspace(start: int, stop: int, num: int, endpoints: bool = False):
+    """
+    Returns numbers that are roughly evenly-spaced along a log scale,
+    and where each number is an exact multiple of the smallest.
+
+    Note that start and stop endpoints are not returned.
+
+    The returned values will always be an exact multiple of the smaller
+    of start and stop. But the larger of start and stop will not necessarily
+    be a multiple of any of the returned values.
+
+    Examples::
+
+        >>> multspace(1, 1000, 2)
+        [10, 100]
+        >>> multspace(1000, 1, 2)
+        [100, 10]
+        >>> multspace(24, 43800, 3)
+        [168, 1008, 7056]
+    """
+
+    if start < 1:
+        raise NotImplementedError(f"start must be 1 or more, but was {start}")
+
+    if stop < 1:
+        raise NotImplementedError(f"stop must be 1 or more, but was {stop}")
+
+    if num < 0:
+        raise NotImplementedError(f"num must be positive, but was {num}")
+
+    if endpoints:
+        raise NotImplementedError("endpoints is not supported in multspace")
+
+    if start > stop:
+        return list(reversed(multspace(stop, start, num)))
+
+    return list(_multspace(start, stop, num))[1:-1]
+
+
+def _multspace(start, stop, num):
+    vals = np.geomspace(start, stop, num + 2)
+    vint = 1
+    for v in vals:
+        vint = floor(v / vint) * vint
+        yield vint
+
+
+def calculate_regular_stage_chunks(
+    read_chunks: tuple[int, ...],
+    write_chunks: tuple[int, ...],
+    stage_count: int = 1,
+) -> list[tuple[int, ...]]:
+    """
+    Calculate chunks after each stage of a multi-stage rechunking.
+
+    Unlike `calculate_stage_chunks` in rechunker, this implementation
+    always returns intermediate chunks sizes that work with regularly
+    chunked Zarr arrays.
+    """
+    stages = []
+    for rc, wc in zip(read_chunks, write_chunks):
+        stages.append(multspace(rc, wc, num=stage_count - 1))
+    return [tuple(chunks) for chunks in np.array(stages).T.tolist()]
+
+
+def _fix_copy_chunks(shape, copy_chunks, target_chunks):
+    # if copy chunks are bigger than target chunks in a particular axis, then
+    # round them down to the largest multiple of the target so they are aligned
+    return tuple(
+        cc if (cc <= tc) or (cc == n) or (cc % tc == 0) else (cc // tc) * tc
+        for n, cc, tc in zip(shape, copy_chunks, target_chunks)
+    )
+
+
+def multistage_regular_rechunking_plan(
+    shape: Sequence[int],
+    source_chunks: Sequence[int],
+    target_chunks: Sequence[int],
+    itemsize: int,
+    min_mem: int,
+    max_mem: int,
+    consolidate_reads: bool = True,
+    consolidate_writes: bool = True,
+) -> _MultistagePlan:
+    """Calculate a rechunking plan that can use multiple split/consolidate steps.
+
+    For best results, max_mem should be significantly larger than min_mem (e.g.,
+    10x). Otherwise an excessive number of rechunking steps will be required.
+    """
+
+    ndim = len(shape)
+    if len(source_chunks) != ndim:
+        raise ValueError(f"source_chunks {source_chunks} must have length {ndim}")
+    if len(target_chunks) != ndim:
+        raise ValueError(f"target_chunks {target_chunks} must have length {ndim}")
+
+    source_chunk_mem = itemsize * prod(source_chunks)
+    target_chunk_mem = itemsize * prod(target_chunks)
+
+    if source_chunk_mem > max_mem:
+        raise ValueError(
+            f"Source chunk memory ({source_chunk_mem}) exceeds max_mem ({max_mem})"
+        )
+    if target_chunk_mem > max_mem:
+        raise ValueError(
+            f"Target chunk memory ({target_chunk_mem}) exceeds max_mem ({max_mem})"
+        )
+
+    if max_mem < min_mem:  # basic sanity check
+        raise ValueError(
+            f"max_mem ({max_mem}) cannot be smaller than min_mem ({min_mem})"
+        )
+
+    if consolidate_writes:
+        logger.debug(
+            f"consolidate_write_chunks({shape}, {target_chunks}, {itemsize}, {max_mem})"
+        )
+        write_chunks = consolidate_chunks(shape, target_chunks, itemsize, max_mem)
+    else:
+        write_chunks = tuple(target_chunks)
+
+    if consolidate_reads:
+        read_chunk_limits: List[Optional[int]] = []
+        for sc, wc in zip(source_chunks, write_chunks):
+            limit: Optional[int]
+            if wc > sc:
+                # consolidate reads over this axis, up to the write chunk size
+                limit = wc
+            else:
+                # don't consolidate reads over this axis
+                limit = None
+            read_chunk_limits.append(limit)
+
+        logger.debug(
+            f"consolidate_read_chunks({shape}, {source_chunks}, {itemsize}, {max_mem}, {read_chunk_limits})"
+        )
+        read_chunks = consolidate_chunks(
+            shape, source_chunks, itemsize, max_mem, read_chunk_limits
+        )
+    else:
+        read_chunks = tuple(source_chunks)
+
+    prev_io_ops: Optional[float] = None
+    prev_plan: Optional[_MultistagePlan] = None
+
+    # increase the number of stages until min_mem is exceeded
+    for stage_count in range(1, MAX_STAGES):
+        stage_chunks = calculate_regular_stage_chunks(
+            read_chunks, write_chunks, stage_count
+        )
+        # adjust read_chunks to ensure they align with following stage
+        read_chunks = _fix_copy_chunks(
+            shape, read_chunks, (stage_chunks + [write_chunks])[0]
+        )
+        pre_chunks = [read_chunks] + stage_chunks
+        post_chunks = stage_chunks + [write_chunks]
+
+        int_chunks = [
+            _calculate_shared_chunks(pre, post)
+            for pre, post in zip(pre_chunks, post_chunks)
+        ]
+        plan = list(zip(pre_chunks, int_chunks, post_chunks))
+
+        int_mem = min(itemsize * prod(chunks) for chunks in int_chunks)
+        if int_mem >= min_mem:
+            return plan  # success!
+
+        io_ops = sum(
+            calculate_single_stage_io_ops(shape, pre, post)
+            for pre, post in zip(pre_chunks, post_chunks)
+        )
+        if prev_io_ops is not None and io_ops > prev_io_ops:
+            warnings.warn(
+                "Search for multi-stage rechunking plan terminated before "
+                "achieving the minimum memory requirement due to increasing IO "
+                f"requirements. Smallest intermediates have size {int_mem}. "
+                f"Consider decreasing min_mem ({min_mem}) or increasing "
+                f"({max_mem}) to find a more efficient plan.",
+                category=ExcessiveIOWarning,
+                stacklevel=2,
+            )
+            assert prev_plan is not None
+            return prev_plan
+
+        prev_io_ops = io_ops
+        prev_plan = plan
+
+    raise AssertionError(
+        "Failed to find a feasible multi-staging rechunking scheme satisfying "
+        f"min_mem ({min_mem}) and max_mem ({max_mem}) constraints. "
+        "Please file a bug report on GitHub: "
+        "https://github.com/pangeo-data/rechunker/issues\n\n"
+        "Include the following debugging info:\n"
+        f"shape={shape}, source_chunks={source_chunks}, "
+        f"target_chunks={target_chunks}, itemsize={itemsize}, "
+        f"min_mem={min_mem}, max_mem={max_mem}, "
+        f"consolidate_reads={consolidate_reads}, "
+        f"consolidate_writes={consolidate_writes}"
+    )