On Fri, 11 Mar 2022 06:29:22 GMT, Xiaohong Gong <xg...@openjdk.org> wrote:
> The current vector `"NEG"` is implemented with substraction a vector by zero > in case the architecture does not support the negation instruction. And to > fit the predicate feature for architectures that support it, the masked > vector `"NEG" ` is implemented with pattern `"v.not(m).add(1, m)"`. They both > can be optimized to a single negation instruction for ARM SVE. > And so does the non-masked "NEG" for NEON. Besides, implementing the masked > "NEG" with substraction for architectures that support neither negation > instruction nor predicate feature can also save several instructions than the > current pattern. > > To optimize the VectorAPI negation, this patch moves the implementation from > Java side to hotspot. The compiler will generate different nodes according to > the architecture: > - Generate the (predicated) negation node if architecture supports it, > otherwise, generate "`zero.sub(v)`" pattern for non-masked operation. > - Generate `"zero.sub(v, m)"` for masked operation if the architecture does > not have predicate feature, otherwise generate the original pattern > `"v.xor(-1, m).add(1, m)"`. > > So with this patch, the following transformations are applied: > > For non-masked negation with NEON: > > movi v16.4s, #0x0 > sub v17.4s, v16.4s, v17.4s ==> neg v17.4s, v17.4s > > and with SVE: > > mov z16.s, #0 > sub z18.s, z16.s, z17.s ==> neg z16.s, p7/m, z16.s > > For masked negation with NEON: > > movi v17.4s, #0x1 > mvn v19.16b, v18.16b > mov v20.16b, v16.16b ==> neg v18.4s, v17.4s > bsl v20.16b, v19.16b, v18.16b bsl v19.16b, v18.16b, v17.16b > add v19.4s, v20.4s, v17.4s > mov v18.16b, v16.16b > bsl v18.16b, v19.16b, v20.16b > > and with SVE: > > mov z16.s, #-1 > mov z17.s, #1 ==> neg z16.s, p0/m, z16.s > eor z18.s, p0/m, z18.s, z16.s > add z18.s, p0/m, z18.s, z17.s > > Here are the performance gains for benchmarks (see [1][2]) on ARM and x86 > machines(note that the non-masked negation benchmarks do not have any > improvement on X86 since no instructions are changed): > > NEON: > Benchmark Gain > Byte128Vector.NEG 1.029 > Byte128Vector.NEGMasked 1.757 > Short128Vector.NEG 1.041 > Short128Vector.NEGMasked 1.659 > Int128Vector.NEG 1.005 > Int128Vector.NEGMasked 1.513 > Long128Vector.NEG 1.003 > Long128Vector.NEGMasked 1.878 > > SVE with 512-bits: > Benchmark Gain > ByteMaxVector.NEG 1.10 > ByteMaxVector.NEGMasked 1.165 > ShortMaxVector.NEG 1.056 > ShortMaxVector.NEGMasked 1.195 > IntMaxVector.NEG 1.002 > IntMaxVector.NEGMasked 1.239 > LongMaxVector.NEG 1.031 > LongMaxVector.NEGMasked 1.191 > > X86 (non AVX-512): > Benchmark Gain > ByteMaxVector.NEGMasked 1.254 > ShortMaxVector.NEGMasked 1.359 > IntMaxVector.NEGMasked 1.431 > LongMaxVector.NEGMasked 1.989 > > [1] > https://github.com/openjdk/panama-vector/blob/vectorIntrinsics/test/micro/org/openjdk/bench/jdk/incubator/vector/operation/Byte128Vector.java#L1881 > [2] > https://github.com/openjdk/panama-vector/blob/vectorIntrinsics/test/micro/org/openjdk/bench/jdk/incubator/vector/operation/Byte128Vector.java#L1896 This pull request has now been integrated. Changeset: d0668568 Author: Xiaohong Gong <xg...@openjdk.org> Committer: Jie Fu <ji...@openjdk.org> URL: https://git.openjdk.java.net/jdk/commit/d06685680c17583d56dc3d788d9a2ecea8812bc8 Stats: 325 lines in 15 files changed: 275 ins; 25 del; 25 mod 8282162: [vector] Optimize integral vector negation API Reviewed-by: jiefu, psandoz, njian ------------- PR: https://git.openjdk.java.net/jdk/pull/7782