RFR: 8288107: Auto-vectorization for integer min/max [v2]

Bhavana-Kilambi duke at openjdk.org
Fri Jul 15 11:16:00 UTC 2022


On Fri, 15 Jul 2022 10:44:52 GMT, Bhavana-Kilambi <duke at openjdk.org> wrote:

>> When Math.min/max is invoked on integer arrays, it generates the CMP-CMOVE instructions instead of vectorizing the loop(if vectorizable and relevant ISA is available) using vector equivalent of min/max instructions. Emitting MaxI/MinI nodes instead of Cmp/CmoveI nodes results in the loop getting vectorized eventually and the architecture specific min/max vector instructions are generated.
>> A test to assess the performance of Math.max/min and StrictMath.max/min is added. On aarch64, the smin/smax instructions are generated when the loop is vectorized. On x86-64, vectorization support for min/max operations is available only in SSE4 (where pmaxsd/pminsd are generated) and AVX version >= 1 (where vpmaxsd/vpminsd are generated). This patch generates these instructions only when the loop is vectorized. In cases where the loop is not vectorizable or when the Math.max/min operations are called outside of the loop, cmp-cmove instructions are generated (tested on aarch64, x86-64 machines which have cmp-cmove instructions defined for the scalar MaxI/MinI nodes). Performance comparisons for the VectorIntMinMax.java test with and without the patch are given below :
>> 
>> <details><summary><strong>Before this patch</strong></summary>
>> 
>> **aarch64:**
>>   ```
>>     Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  1593.510 ± 1.488  ns/op
>>     VectorIntMinMax.testMinInt            2048       0  avgt   25  1593.123 ± 1.365  ns/op
>>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  1593.112 ± 0.985  ns/op
>>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  1593.290 ± 1.219  ns/op
>> 
>> 
>> **x86-64:**
>> 
>>   Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>>   VectorIntMinMax.testMaxInt            2048       0  avgt   25  2084.717 ± 4.780  ns/op
>>   VectorIntMinMax.testMinInt            2048       0  avgt   25  2087.322 ± 4.158  ns/op
>>   VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  2084.568 ± 4.838  ns/op
>>   VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  2086.595 ± 4.025  ns/op
>> 
>> </details>
>> 
>> <details><summary><strong>After this patch</strong></summary>
>> 
>> **aarch64:**
>> 
>>   Benchmark                         (length)  (seed)  Mode  Cnt    Score   Error  Units
>>   VectorIntMinMax.testMaxInt            2048       0  avgt   25  323.911 ± 0.206  ns/op
>>   VectorIntMinMax.testMinInt            2048       0  avgt   25  324.084 ± 0.231  ns/op
>>   VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  323.892 ± 0.234  ns/op
>>   VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  323.990 ± 0.295  ns/op
>> 
>> 
>> **x86-64:**
>> 
>>   Benchmark                         (length)  (seed)  Mode  Cnt    Score   Error  Units
>>   VectorIntMinMax.testMaxInt            2048       0  avgt   25  387.639 ± 0.512  ns/op
>>   VectorIntMinMax.testMinInt            2048       0  avgt   25  387.999 ± 0.740  ns/op
>>   VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  387.605 ± 0.376  ns/op
>>   VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  387.765 ± 0.498  ns/op
>> 
>> 
>> </details>
>> 
>> With auto-vectorization, both the machines exhibit a significant performance gain. On both the machines the runtime is ~80% better than the case without this patch. Also ran the patch with -XX:-UseSuperWord to make sure the performance does not degrade in cases where vectorization does not happen.
>> 
>> <details><summary><strong>Performance numbers</strong></summary>
>> 
>> **aarch64:**
>> 
>>   Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>>   VectorIntMinMax.testMaxInt            2048       0  avgt   25  1449.792 ± 1.072  ns/op
>>   VectorIntMinMax.testMinInt            2048       0  avgt   25  1450.636 ± 1.057  ns/op
>>   VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  1450.214 ± 1.093  ns/op
>>   VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  1450.615 ± 1.098  ns/op
>> 
>> 
>> **x86-64:**
>> 
>>   Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>>   VectorIntMinMax.testMaxInt            2048       0  avgt   25  2059.673 ± 4.726  ns/op
>>   VectorIntMinMax.testMinInt            2048       0  avgt   25  2059.853 ± 4.754  ns/op
>>   VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  2059.920 ± 4.658  ns/op
>>   VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  2059.622 ± 4.768  ns/op
>> 
>> </details>
>> 
>> There is no degradation when vectorization is disabled.
>
> Bhavana-Kilambi has updated the pull request incrementally with one additional commit since the last revision:
> 
>   8288107: Auto-vectorization for integer min/max
>   
>   When Math.min/max is invoked on integer arrays, it generates the CMP-CMOVE instructions instead of vectorizing the loop(if vectorizable and relevant ISA is available) using vector equivalent of min/max instructions. Emitting MaxI/MinI nodes instead of Cmp/CmoveI nodes results in the loop getting vectorized eventually and the architecture specific min/max vector instructions are generated.
>   A test to assess the performance of Math.max/min and StrictMath.max/min is added. On aarch64, the smin/smax instructions are generated when the loop is vectorized. On x86-64, vectorization support for min/max operations is available only in SSE4 (where pmaxsd/pminsd are generated) and AVX version >= 1 (where vpmaxsd/vpminsd are generated). This patch generates these instructions only when the loop is vectorized. In cases where the loop is not vectorizable or when the Math.max/min operations are called outside of the loop, cmp-cmove instructions are generated (tested on aarch64, x86-64 machines which have cmp-cmove instructions defined for the scalar MaxI/MinI nodes). Performance comparisons for the VectorIntMinMax.java test with and without the patch are given below :
>   
>   Before this patch:
>   aarch64:
>     Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  1593.510 ± 1.488  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  1593.123 ± 1.365  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  1593.112 ± 0.985  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  1593.290 ± 1.219  ns/op
>   
>   x86-64:
>     Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  2084.717 ± 4.780  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  2087.322 ± 4.158  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  2084.568 ± 4.838  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  2086.595 ± 4.025  ns/op
>   
>   After this patch:
>   aarch64:
>   Benchmark                         (length)  (seed)  Mode  Cnt    Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  323.911 ± 0.206  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  324.084 ± 0.231  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  323.892 ± 0.234  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  323.990 ± 0.295  ns/op
>   
>   x86-64:
>   Benchmark                         (length)  (seed)  Mode  Cnt    Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  387.639 ± 0.512  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  387.999 ± 0.740  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  387.605 ± 0.376  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  387.765 ± 0.498  ns/op
>   
>   With auto-vectorization, both the machines exhibit a significant performance gain. On both the machines the runtime is ~80% better than the case without this patch. Also ran the patch with -XX:-UseSuperWord to make sure the performance does not degrade in cases where vectorization does not happen. The performance numbers are shown below :
>   aarch64:
>   Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  1449.792 ± 1.072  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  1450.636 ± 1.057  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  1450.214 ± 1.093  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  1450.615 ± 1.098  ns/op
>   
>   x86-64:
>   Benchmark                         (length)  (seed)  Mode  Cnt     Score   Error  Units
>     VectorIntMinMax.testMaxInt            2048       0  avgt   25  2059.673 ± 4.726  ns/op
>     VectorIntMinMax.testMinInt            2048       0  avgt   25  2059.853 ± 4.754  ns/op
>     VectorIntMinMax.testStrictMaxInt      2048       0  avgt   25  2059.920 ± 4.658  ns/op
>     VectorIntMinMax.testStrictMinInt      2048       0  avgt   25  2059.622 ± 4.768  ns/op
>   There is no degradation when vectorization is disabled.

Added a new commit with the MaxINode::Ideal tests related code stripped off and only retaining the code related to generating MinI/MaxI node for Math/min/max intrinsics.

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

PR: https://git.openjdk.org/jdk/pull/9466


More information about the hotspot-compiler-dev mailing list