RFR: 8346236: Auto vectorization support for various Float16 operations [v5]

[Jatin Bhateja]

On Fri, 21 Mar 2025 08:55:01 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> This is a follow-up PR for https://github.com/openjdk/jdk/pull/22754
>> 
>> The patch adds support to vectorize various float16 scalar operations (add/subtract/divide/multiply/sqrt/fma).
>> 
>> Summary of changes included with the patch:
>>    1. C2 compiler New Vector IR creation.
>>    2. Auto-vectorization support.
>>    3. x86 backend implementation.
>>    4. New IR verification test for each newly supported vector operation.
>> 
>> Following are the performance numbers of Float16OperationsBenchmark
>> 
>> System : Intel(R) Xeon(R) Processor code-named Granite rapids
>> Frequency fixed at 2.5 GHz
>> 
>> 
>> Baseline
>> Benchmark                                                      (vectorDim)   Mode  Cnt     Score   Error   Units
>> Float16OperationsBenchmark.absBenchmark                               1024  thrpt    2  4191.787          ops/ms
>> Float16OperationsBenchmark.addBenchmark                               1024  thrpt    2  1211.978          ops/ms
>> Float16OperationsBenchmark.cosineSimilarityDequantizedFP16            1024  thrpt    2   493.026          ops/ms
>> Float16OperationsBenchmark.cosineSimilarityDoubleRoundingFP16         1024  thrpt    2   612.430          ops/ms
>> Float16OperationsBenchmark.cosineSimilaritySingleRoundingFP16         1024  thrpt    2   616.012          ops/ms
>> Float16OperationsBenchmark.divBenchmark                               1024  thrpt    2   604.882          ops/ms
>> Float16OperationsBenchmark.dotProductFP16                             1024  thrpt    2   410.798          ops/ms
>> Float16OperationsBenchmark.euclideanDistanceDequantizedFP16           1024  thrpt    2   602.863          ops/ms
>> Float16OperationsBenchmark.euclideanDistanceFP16                      1024  thrpt    2   640.348          ops/ms
>> Float16OperationsBenchmark.fmaBenchmark                               1024  thrpt    2   809.175          ops/ms
>> Float16OperationsBenchmark.getExponentBenchmark                       1024  thrpt    2  2682.764          ops/ms
>> Float16OperationsBenchmark.isFiniteBenchmark                          1024  thrpt    2  3373.901          ops/ms
>> Float16OperationsBenchmark.isFiniteCMovBenchmark                      1024  thrpt    2  1881.652          ops/ms
>> Float16OperationsBenchmark.isFiniteStoreBenchmark                     1024  thrpt    2  2273.745          ops/ms
>> Float16OperationsBenchmark.isInfiniteBenchmark                        1024  thrpt    2  2147.913          ops/ms
>> Float16OperationsBenchmark.isInfiniteCMovBen...
>
> Jatin Bhateja has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Review comments resolution.

src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp line 7116:

> 7114:                                             KRegister ktmp, XMMRegister xtmp1, XMMRegister xtmp2) {
> 7115:   vector_max_min_fp16(opcode, dst, src1, src2, ktmp, xtmp1, xtmp2, Assembler::AVX_128bit);
> 7116: }

Created a separate JBS entry to track this special handling for scalar MAX/MIN
https://bugs.openjdk.org/browse/JDK-8352585

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

PR Review Comment: https://git.openjdk.org/jdk/pull/22755#discussion_r2007114926