[jira] Closed: (HARMONY-5599) [drlvm][jit][performance] new movapd instruction for register-to-register copy

["Vladimir Strigun (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HARMONY-5599?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Vladimir Strigun closed HARMONY-5599.
-------------------------------------


> [drlvm][jit][performance] new movapd instruction for register-to-register copy
> ------------------------------------------------------------------------------
>
>                 Key: HARMONY-5599
>                 URL: https://issues.apache.org/jira/browse/HARMONY-5599
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>            Reporter: Vladimir Strigun
>            Assignee: Mikhail Fursov
>         Attachments: H-5599.patch, i586.diff, movapd.diff, movapd2.diff
>
>
> Usage of movapd instruction for copy between xmm registers is more efficient in copmarison with partial movsd copy. So, attached patch replace movsd with movadp instruction for such operations. 
> I've checked the patch on scimark bench [1] and got the following results (about 15% speedup for composite score):
> orig build:
> SciMark 2.0a
> Composite Score: 236.8043350027899
> FFT (1024): 266.4183025101507
> SOR (100x100):   410.3833460433766
> Monte Carlo : 31.43640457526972
> Sparse matmult (N=1000, nz=5000): 208.14991492655557
> LU (100x100): 267.6337069585971
> with movapd:
> SciMark 2.0a
> Composite Score: 271.62584550328904
> FFT (1024): 296.11079189672955
> SOR (100x100):   458.00820213602486
> Monte Carlo : 31.406979573247035
> Sparse matmult (N=1000, nz=5000): 208.14991492655557
> LU (100x100): 364.453338983888

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.