CopyAggregation

Usage

use CopyAggregation;

or

import CopyAggregation;

Provides support for aggregated copies/assignments for trivial types.

Warning

This module represents work in progress. The API is unstable and likely to change over time.

This module provides an aggregated version of copy/assignment for trivially copyable types. Trivially copyable types require no special behavior to be copied and merely copying their representation is sufficient. They include numeric and bool types as well as tuples or records consisting only of numeric/bool. Records cannot have user-defined copy-initializers, deinitializers, or assignment overloads.

Aggregated copies can provide a significant speedup for batch assignment operations that do not require ordering of operations. The results of aggregated operations are not visible until the aggregator is deinitialized or an explicit flush() call is made. The current implementation requires that one side of the copy is always local. When the source is always local and the destination may be remote a DstAggregator should be used. When the destination is always local and the source may be remote, a SrcAggregator should be used. Note that aggregators are not parallel safe and are expected to be created on a per-task basis. Each aggregator has a per-peer buffer, so memory usage increases with number of locales. The following example demonstrates using SrcAggregator to reverse an array:

use BlockDist, CopyAggregation;

const size = 10000;
const space = {0..size};
const D = space dmapped blockDist(space);
var A, reversedA: [D] int = D;

forall (rA, i) in zip(reversedA, D) with (var agg = new SrcAggregator(int)) do
  agg.copy(rA, A[size-i]);

// no flush required, flushed when aggregator is deinitialized

forall (rA, i) in zip(reversedA, D) do
  assert(rA == size-i);

It’s important to be aware that aggregated operations are not consistent with regular operations and updates may not be visible until the aggregator is deinitialized or an explicit call to flush() is made.

use CopyAggregation;

var a = 0;
on Locales[numLocales-1] {
  var agg = new DstAggregator(int);
  var b = 1;
  agg.copy(a, b);
  writeln(a);   // can print 0 or 1
  agg.flush();
  writeln(a);   // must print 1
}

Generally speaking aggregators are useful for when you have a large batch of remote assignments to perform and the order of those operations doesn’t matter.

record DstAggregator

Aggregates copy(ref dst, src). Optimized for when src is local. Not parallel safe and is expected to be created on a per-task basis. High memory usage since there are per-destination buffers.

type elemType

proc copy(ref dst: elemType, const in srcVal: elemType)

proc flush()

record SrcAggregator

Aggregates copy(ref dst, const ref src). Only works when dst is local. Not parallel safe and is expected to be created on a per task basis. High memory usage since there are per-destination buffers.

type elemType

proc ref copy(ref dst: elemType, const ref src: elemType)

proc flush()