On Tue, 26 Nov 2024 13:55:10 GMT, Viktor Klang <vkl...@openjdk.org> wrote:

>> This PR proposes adding a _JDK-internal_ method for calculating hash codes 
>> for content in a `MemorySegment`.
>> 
>> The internal method uses a polynomial 32-bit hash function equivalent to 
>> `Arrays::hashCode`. The new method is almost two times faster than naïvely 
>> iterating over individual bytes for larger regions. Also, it is more lean on 
>> inlining space compared to a naïve loop.
>> 
>> 
>> 
>> Benchmark                          (ELEM_SIZE)  Mode  Cnt   Score   Error  
>> Units
>> SegmentBulkHash.array                        8  avgt   30   2.645 ? 0.078  
>> ns/op
>> SegmentBulkHash.array                       64  avgt   30   6.062 ? 0.171  
>> ns/op
>> SegmentBulkHash.heapSegment                  8  avgt   30   4.181 ? 0.145  
>> ns/op
>> SegmentBulkHash.heapSegment                 64  avgt   30  25.716 ? 1.043  
>> ns/op
>> SegmentBulkHash.nativeSegment                8  avgt   30   3.939 ? 0.150  
>> ns/op
>> SegmentBulkHash.nativeSegment               64  avgt   30  23.262 ? 0.694  
>> ns/op
>> SegmentBulkHash.nativeSegmentJava            8  avgt   30   5.219 ? 0.183  
>> ns/op    <- Naïve iteration
>> SegmentBulkHash.nativeSegmentJava           64  avgt   30  39.668 ? 1.040  
>> ns/op    <- Naïve iteration
>> 
>> 
>> ![image](https://github.com/user-attachments/assets/5646cf21-b202-4dce-9555-e460f9df4cb6)
>> 
>> 
>> If internal JDK code uses this method, it will automatically benefit from 
>> future performance improvements that can be implemented once the Vector API 
>> becomes available.
>
> src/java.base/share/classes/jdk/internal/foreign/SegmentBulkOperations.java 
> line 208:
> 
>> 206:         final long limit = fromOffset + longBytes;
>> 207:         for (; fromOffset < limit; fromOffset += 8) {
>> 208:             long val = 
>> SCOPED_MEMORY_ACCESS.getLongUnaligned(segment.sessionImpl(), 
>> segment.unsafeGetBase(), segment.unsafeGetOffset() + fromOffset, 
>> !Architecture.isLittleEndian());
> 
> @minborg Does it help to hoist the calls to `segment.sessionImpl()` + 
> `segment.unsafeGetBase()` + `segment.unsafeGetOffset()` + 
> `!Architecture.isLittleEndian()` and reusing those throughout the different 
> strides?

I thought that initially as well, but keeping it like this helps the compiler 
select the correct unsafe code path.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/22364#discussion_r1858596703

Reply via email to