Re: RFR: 8267969: Add vectorized implementation for VectorMask.eq() [v2]
On Wed, 2 Jun 2021 07:48:47 GMT, Nils Eliasson wrote: > Please wait until you have two reviewers before integrating. Sure! Thanks so much for looking at this PR! - PR: https://git.openjdk.java.net/jdk/pull/4272
Re: RFR: 8267969: Add vectorized implementation for VectorMask.eq() [v2]
On Wed, 2 Jun 2021 01:56:00 GMT, Xiaohong Gong wrote:
>> Currently `"VectorMask.eq()" `is not vectorized:
>>
>> public VectorMask eq(VectorMask m) {
>> // FIXME: Generate good code here.
>> return bOp(m, (i, a, b) -> a == b);
>> }
>>
>> This can be implemented by calling `"xor(m.not())"` directly.
>>
>> The performance improved about 1.4x ~ 1.9x for the following benchmark with
>> different basic types:
>>
>> @Benchmark
>> public Object eq() {
>> boolean[] ma = fm.apply(size);
>> boolean[] mb = fmb.apply(size);
>> boolean[] mt = fmt.apply(size);
>> VectorMask m = VectorMask.fromArray(SPECIES, mt, 0);
>>
>> for (int ic = 0; ic < INVOC_COUNT; ic++) {
>> for (int i = 0; i < ma.length; i += SPECIES.length()) {
>> var av = SPECIES.loadMask(ma, i);
>> var bv = SPECIES.loadMask(mb, i);
>>
>> // accumulate results, so JIT can't eliminate relevant
>> computations
>> m = m.and(av.eq(bv));
>> }
>> }
>>
>> return m;
>> }
>
> Xiaohong Gong has updated the pull request with a new target base due to a
> merge or a rebase. The incremental webrev excludes the unrelated changes
> brought in by the merge/rebase. The pull request contains two additional
> commits since the last revision:
>
> - Merge branch 'jdk:master' into JDK-8267969
> - 8267969: Add vectorized implementation for VectorMask.eq()
Looks good.
Please wait until you have two reviewers before integrating.
-
Marked as reviewed by neliasso (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/4272
Re: RFR: 8267969: Add vectorized implementation for VectorMask.eq() [v2]
> Currently `"VectorMask.eq()" `is not vectorized:
>
> public VectorMask eq(VectorMask m) {
> // FIXME: Generate good code here.
> return bOp(m, (i, a, b) -> a == b);
> }
>
> This can be implemented by calling `"xor(m.not())"` directly.
>
> The performance improved about 1.4x ~ 1.9x for the following benchmark with
> different basic types:
>
> @Benchmark
> public Object eq() {
> boolean[] ma = fm.apply(size);
> boolean[] mb = fmb.apply(size);
> boolean[] mt = fmt.apply(size);
> VectorMask m = VectorMask.fromArray(SPECIES, mt, 0);
>
> for (int ic = 0; ic < INVOC_COUNT; ic++) {
> for (int i = 0; i < ma.length; i += SPECIES.length()) {
> var av = SPECIES.loadMask(ma, i);
> var bv = SPECIES.loadMask(mb, i);
>
> // accumulate results, so JIT can't eliminate relevant
> computations
> m = m.and(av.eq(bv));
> }
> }
>
> return m;
> }
Xiaohong Gong has updated the pull request with a new target base due to a
merge or a rebase. The incremental webrev excludes the unrelated changes
brought in by the merge/rebase. The pull request contains two additional
commits since the last revision:
- Merge branch 'jdk:master' into JDK-8267969
- 8267969: Add vectorized implementation for VectorMask.eq()
-
Changes:
- all: https://git.openjdk.java.net/jdk/pull/4272/files
- new: https://git.openjdk.java.net/jdk/pull/4272/files/0e8e0e84..f3a48026
Webrevs:
- full: https://webrevs.openjdk.java.net/?repo=jdk=4272=01
- incr: https://webrevs.openjdk.java.net/?repo=jdk=4272=00-01
Stats: 22530 lines in 577 files changed: 2836 ins; 18744 del; 950 mod
Patch: https://git.openjdk.java.net/jdk/pull/4272.diff
Fetch: git fetch https://git.openjdk.java.net/jdk pull/4272/head:pull/4272
PR: https://git.openjdk.java.net/jdk/pull/4272
Re: RFR: 8267969: Add vectorized implementation for VectorMask.eq()
On Tue, 1 Jun 2021 16:29:58 GMT, Paul Sandoz wrote: > Looks. Later we may want to consider pushing this down as an intrinsic, > perhaps reusing `VectorSupport.compare`. Thanks for your review @PaulSandoz ! Yes, reusing `VectorSupport.compare` is an alternative way to do the vectorization. - PR: https://git.openjdk.java.net/jdk/pull/4272
Re: RFR: 8267969: Add vectorized implementation for VectorMask.eq()
On Mon, 31 May 2021 10:25:26 GMT, Xiaohong Gong wrote:
> Currently `"VectorMask.eq()" `is not vectorized:
>
> public VectorMask eq(VectorMask m) {
> // FIXME: Generate good code here.
> return bOp(m, (i, a, b) -> a == b);
> }
>
> This can be implemented by calling `"xor(m.not())"` directly.
>
> The performance improved about 1.4x ~ 1.9x for the following benchmark with
> different basic types:
>
> @Benchmark
> public Object eq() {
> boolean[] ma = fm.apply(size);
> boolean[] mb = fmb.apply(size);
> boolean[] mt = fmt.apply(size);
> VectorMask m = VectorMask.fromArray(SPECIES, mt, 0);
>
> for (int ic = 0; ic < INVOC_COUNT; ic++) {
> for (int i = 0; i < ma.length; i += SPECIES.length()) {
> var av = SPECIES.loadMask(ma, i);
> var bv = SPECIES.loadMask(mb, i);
>
> // accumulate results, so JIT can't eliminate relevant
> computations
> m = m.and(av.eq(bv));
> }
> }
>
> return m;
> }
Looks. Later we may want to consider pushing this down as an intrinsic, perhaps
reusing `VectorSupport.compare`.
-
Marked as reviewed by psandoz (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/4272
RFR: 8267969: Add vectorized implementation for VectorMask.eq()
Currently `"VectorMask.eq()" `is not vectorized:
public VectorMask eq(VectorMask m) {
// FIXME: Generate good code here.
return bOp(m, (i, a, b) -> a == b);
}
This can be implemented by calling `"xor(m.not())"` directly.
The performance improved about 1.4x ~ 1.9x for the following benchmark with
different basic types:
@Benchmark
public Object eq() {
boolean[] ma = fm.apply(size);
boolean[] mb = fmb.apply(size);
boolean[] mt = fmt.apply(size);
VectorMask m = VectorMask.fromArray(SPECIES, mt, 0);
for (int ic = 0; ic < INVOC_COUNT; ic++) {
for (int i = 0; i < ma.length; i += SPECIES.length()) {
var av = SPECIES.loadMask(ma, i);
var bv = SPECIES.loadMask(mb, i);
// accumulate results, so JIT can't eliminate relevant computations
m = m.and(av.eq(bv));
}
}
return m;
}
-
Commit messages:
- 8267969: Add vectorized implementation for VectorMask.eq()
Changes: https://git.openjdk.java.net/jdk/pull/4272/files
Webrev: https://webrevs.openjdk.java.net/?repo=jdk=4272=00
Issue: https://bugs.openjdk.java.net/browse/JDK-8267969
Stats: 254 lines in 32 files changed: 248 ins; 6 del; 0 mod
Patch: https://git.openjdk.java.net/jdk/pull/4272.diff
Fetch: git fetch https://git.openjdk.java.net/jdk pull/4272/head:pull/4272
PR: https://git.openjdk.java.net/jdk/pull/4272
