[vector] open issues tagged with vectorIntrinsic label
Viswanathan, Sandhya
sandhya.viswanathan at intel.com
Tue Nov 5 21:19:02 UTC 2019
Hi Paul,
I have started looking at these and closed 8225018.
Looks like 8223367 is open because of masked gather/scatter. Code excerpts below with cases not handled for your reference:
IntVector.java
2924 @ForceInline
2925 public static
2926 IntVector fromArray(VectorSpecies<Integer> species,
2927 int[] a, int offset,
2928 int[] indexMap, int mapOffset,
2929 VectorMask<Integer> m) {
2930 IntSpecies vsp = (IntSpecies) species;
2931
2932 // FIXME This can result in out of bounds errors for unset mask lanes
2933 // FIX = Use a scatter instruction which routes the unwanted lanes
2934 // into a bit-bucket variable (private to implementation).
2935 // This requires a 2-D scatter in order to set a second base address.
2936 // See notes in https://bugs.openjdk.java.net/browse/JDK-8223367
2937 assert(m.allTrue());
2938 return (IntVector)
2939 zero(species).blend(fromArray(species, a, offset, indexMap, mapOffset), m);
2940
2941 }
3209 @ForceInline
3210 public final
3211 void intoArray(int[] a, int offset,
3212 int[] indexMap, int mapOffset,
3213 VectorMask<Integer> m) {
3214 IntSpecies vsp = vspecies();
3215 if (m.allTrue()) {
3216 intoArray(a, offset, indexMap, mapOffset);
3217 return;
3218 }
3219 throw new AssertionError("fixme");
3220 }
Best Regards,
Sandhya
-----Original Message-----
From: panama-dev <panama-dev-bounces at openjdk.java.net> On Behalf Of Paul Sandoz
Sent: Friday, November 01, 2019 12:11 PM
To: panama-dev at openjdk.java.net
Subject: [vector] open issues tagged with vectorIntrinsic label
Hi,
There are a bunch of open issues tagged with vectorIntrinsic:
https://bugs.openjdk.java.net/issues/?jql=status%20in%20(Open%2C%20%22In%20Progress%22%2C%20Reopened%2C%20New)%20AND%20labels%20%3D%20vectorIntrinsics%20ORDER%20BY%20issuetype%20DESC <https://bugs.openjdk.java.net/issues/?jql=status%20in%20(Open,%20%22In%20Progress%22,%20Reopened,%20New)%20AND%20labels%20=%20vectorIntrinsics%20ORDER%20BY%20issuetype%20DESC>
I am wondering what can be marked as resolved or closed. Please dive in and update or reply on status.
Possible candidates to resolve:
https://bugs.openjdk.java.net/browse/JDK-8225018 <https://bugs.openjdk.java.net/browse/JDK-8225018>
9 unit tests for Vector API failed on SkyLake with assert "(((dst_enc < 16 && nds_enc < 16 ..."
http://hg.openjdk.java.net/panama/dev/rev/73fd0971fb14 <http://hg.openjdk.java.net/panama/dev/rev/73fd0971fb14>
https://bugs.openjdk.java.net/browse/JDK-8223367 <https://bugs.openjdk.java.net/browse/JDK-8223367>
[vector] masked memory operations must correctly implement unset lanes
https://bugs.openjdk.java.net/browse/JDK-8221816 <https://bugs.openjdk.java.net/browse/JDK-8221816>
[vector] IndexOutOfBoundsException for fromArray/intoArray with unset mask lanes
https://bugs.openjdk.java.net/browse/JDK-8223366 <https://bugs.openjdk.java.net/browse/JDK-8223366>
[vector] Vector API should not bake in power-of-two size assumptions
https://bugs.openjdk.java.net/browse/JDK-8223368 <https://bugs.openjdk.java.net/browse/JDK-8223368>
[vector] vector API must carefully limit dependencies on native byte order
Possible candidates to close or reevaluate:
https://bugs.openjdk.java.net/browse/JDK-8224988 <https://bugs.openjdk.java.net/browse/JDK-8224988>
[vector] Elemental shift and rotate operations speedup
The API changed and elemental rotate/shift can be supported by slice, receiver vector passed as arg, or zero vector passed as arg, which I presume requires a different set of optimization techniques to trigger the right instructions.
https://bugs.openjdk.java.net/browse/JDK-8222897 <https://bugs.openjdk.java.net/browse/JDK-8222897>
Renaming of shift, rotate operations. Few other api changes.
Paul.
More information about the panama-dev
mailing list