It's too bad that floating-point ordering issues that apparently no one
cares about (no reports) limit performance, but the solution seems OK.
Also: I've been reviewing drafts of this update for a year. They seem
OK. (Although they are very much focused on details of existing
primitive types; most likely a different approach will work better with
value/inline types.)
-Doug
On 7/24/19 9:15 AM, David Holmes wrote:
> Hi Vladimir,
>
> On 24/07/2019 10:58 pm, Vladimir Yaroslavskiy wrote:
>> Hi David,
>>
>> Please, see how it works: Arrays.parallelSort(double[])
>> invokes ArraysParallelSortHelpers.FJDouble.Sorter
>> if size is big enough and ForkJoinPool.getCommonPoolParallelism()) > 1.
>>
>> FJDouble.Sorter divides given array into 4 parts, sorts them
>> recursively in parallel and merges these parts.
>> Finally DualPivotQuicksort is invoked on small parts and only on this
>> step NaNs and -0.0s are arranged.
>> In other words, NaNs and -0.0s are arranged inside each small parts,
>> but this action must be
>> done once before the first splitting of the array.
>
> My understanding** was that the merge of the correctly sorted sub-arrays
> would correctly cause NaNs to bubble to the end as required, while
> zeroes would also group - though I think I can see now that simply using
> < would not correctly order NaNs relative to other values, nor order
> -0.0 and +0.0
>
> ** It's been 7 years since I helped Doug Lea put the parallelising code
> into the JDK so I'm a bit rusty on the details :) and I'm surprised such
> a bug has not been detected before now.
>
> Cheers,
> David
>
>> Thank you,
>> Vladimir
>>
>> Среда, 24 июля 2019, 15:39 +03:00 от David Holmes
>> <david.hol...@oracle.com>:
>>
>> Hi Vladimir,
>>
>> On 24/07/2019 8:53 pm, Vladimir Yaroslavskiy wrote:
>> > Hi all,
>> >
>> > I've found bug in parallel sorting of float / double arrays in
>> the latest JDK.
>> >
>> > When float / double values are sorted, additional actions are
>> > required: NaNs must be moved to the end and negative zeros
>> > must be placed before positive zeros.
>> >
>> > Current implementation of Arrays.parallelSort(float[] / double [])
>> > invokes parallel merge sort from ArraysParallelSortHelpers class,
>> > but it doesn't arrange NaNs and -0.0.
>>
>> It ultimately uses DualPivotQuicksort which AFAICS does have code to
>> arrange NaNS:
>>
>> static void sort(float[] a, int left, int right,
>> float[] work, int workBase, int workLen) {
>> /*
>> * Phase 1: Move NaNs to the end of the array.
>> */
>> while (left <= right && Float.isNaN(a[right])) {
>> --right;
>> }
>>
>> and also order +/-0
>>
>> /*
>> * Phase 3: Place negative zeros before positive zeros.
>> */
>>
>> where does the bug arise?
>>
>> Thanks,
>> David
>> -----
>>
>>
>> > @Alan, Brent, Laurent
>> > Could you please file a bug?
>> >
>> > New optimized version of DualPivotQuicksort (which is under
>> review) works fine and
>> > doesn't contain this bug. Please, look at my test case to
>> reproduce it.
>> >
>> >
>>
>> ----------------------------------------------------------------------------------------------------------------------
>>
>> > import java.util.Random;
>> >
>> > public class FailedFloat {
>> >
>> > private static final int MAX_N = (1 << 13) /*
>> Arrays.MIN_ARRAY_SORT_GRAN */ + 10;
>> >
>> > public static void main(String[] args) {
>> > float[] a = new float[MAX_N];
>> > random(a);
>> > java.util.Arrays.parallelSort(a);
>> > check(a);
>> > System.out.println("PASSED");
>> > }
>> >
>> > private static void random(float[] a) {
>> > Random random = new Random(777);
>> > for (int i = 0; i < MAX_N; i++) {
>> > a[i] = random.nextBoolean() ? -0.0f : 0.0f;
>> > }
>> > }
>> >
>> > private static void check(float[] a) {
>> > for (int i = 0; i < a.length - 1; ++i) {
>> > if (Float.floatToRawIntBits(a[i]) == 0 &&
>> Float.floatToRawIntBits(a[i + 1]) < 0) {
>> > throw new RuntimeException(a[i] + " goes before "+ a[i + 1] + "
>> at position " + i);
>> > }
>> > }
>> > }
>> > }
>> >
>>
>> ------------------------------------------------------------------------------------------------------------------------
>>
>> > Thank you,
>> > Vladimir
>>
>
