RFR: 8322768: Optimize non-subword vector compress and expand APIs for AVX2 target. [v7]

[Sandhya Viswanathan]

On Fri, 19 Jan 2024 19:03:31 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Hi,
>> 
>> Patch optimizes non-subword vector compress and expand APIs for x86 AVX2 only targets.
>> Upcoming E-core Xeons (Sierra Forest) and Hybrid CPUs only support AVX2 instruction set.
>> These are very frequently used APIs in columnar database filter operation.
>> 
>> Implementation uses a lookup table to record permute indices. Table index is computed using
>> mask argument of compress/expand operation.
>> 
>> Following are the performance number of JMH micro included with the patch.
>> 
>> 
>> System : Intel(R) Xeon(R) Platinum 8480+ (Sapphire Rapids)
>> 
>> Baseline:
>> Benchmark                                 (size)   Mode  Cnt    Score   Error   Units
>> ColumnFilterBenchmark.filterDoubleColumn    1024  thrpt    2  142.767          ops/ms
>> ColumnFilterBenchmark.filterDoubleColumn    2047  thrpt    2   71.436          ops/ms
>> ColumnFilterBenchmark.filterDoubleColumn    4096  thrpt    2   35.992          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     1024  thrpt    2  182.151          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     2047  thrpt    2   91.096          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     4096  thrpt    2   44.757          ops/ms
>> ColumnFilterBenchmark.filterIntColumn       1024  thrpt    2  184.099          ops/ms
>> ColumnFilterBenchmark.filterIntColumn       2047  thrpt    2   91.981          ops/ms
>> ColumnFilterBenchmark.filterIntColumn       4096  thrpt    2   45.170          ops/ms
>> ColumnFilterBenchmark.filterLongColumn      1024  thrpt    2  148.017          ops/ms
>> ColumnFilterBenchmark.filterLongColumn      2047  thrpt    2   73.516          ops/ms
>> ColumnFilterBenchmark.filterLongColumn      4096  thrpt    2   36.844          ops/ms
>> 
>> Withopt:
>> Benchmark                                 (size)   Mode  Cnt     Score   Error   Units
>> ColumnFilterBenchmark.filterDoubleColumn    1024  thrpt    2  2051.707          ops/ms
>> ColumnFilterBenchmark.filterDoubleColumn    2047  thrpt    2   914.072          ops/ms
>> ColumnFilterBenchmark.filterDoubleColumn    4096  thrpt    2   489.898          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     1024  thrpt    2  5324.195          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     2047  thrpt    2  2587.229          ops/ms
>> ColumnFilterBenchmark.filterFloatColumn     4096  thrpt    2  1278.665          ops/ms
>> ColumnFilterBenchmark.filterIntColumn       1024  thrpt    2  4149.384          ops/ms
>> ColumnFilterBenchmark.filterIntColumn       2047  thrpt  ...
>
> Jatin Bhateja has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Modified code comment for clarity.

src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp line 5331:

> 5329:     // value, this can potentially be used as a blending mask after
> 5330:     // compressing/expanding the source vector lanes.
> 5331:     vblendvps(dst, dst, permv, xtmp, vec_enc, false, xtmp1);

Here the last argument in vblendps can be same as xtmp. That way we won't need xtmp1. i.e. the vblendps call can be modified as follows:
vblendvps(dst, dst, permv, xtmp, vec_enc, false, xtmp);

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

PR Review Comment: https://git.openjdk.org/jdk/pull/17261#discussion_r1460083068