[llvm] [AArch64] Set MaxInterleaving to 4 for Neoverse V2 (PR #100385)

Wed Jul 24 07:46:39 PDT 2024

sjoerdmeijer wrote:

Here are some motivating examples from TSVC. The numbers are ratios of "after" and "before" (after / before), i.e. without and with this patch applied. Lower than 1 are improvements. This shows ~2x improvements for the first three cases, and then a few more decent improvements. This is a modified version of TSVC where each kernel is compiled into each own executable to reduce noise; I do not trust the TSCV numbers from the llvm test-suite (but that is a separate story):

  <html xmlns:v="urn:schemas-microsoft-com:vml"
xmlns:o="urn:schemas-microsoft-com:office:office"
xmlns:x="urn:schemas-microsoft-com:office:excel"
xmlns="http://www.w3.org/TR/REC-html40">

<head>

<meta name=ProgId content=Excel.Sheet>
<meta name=Generator content="Microsoft Excel 15">
<link id=Main-File rel=Main-File
href="file:////Users/smeijer/Library/Group%20Containers/UBF8T346G9.Office/TemporaryItems/msohtmlclip/clip.htm">
<link rel=File-List
href="file:////Users/smeijer/Library/Group%20Containers/UBF8T346G9.Office/TemporaryItems/msohtmlclip/clip_filelist.xml">

</head>

<body link="#467886" vlink="#96607D">

s3111 | 0.50
-- | --
s314 | 0.50
s3113 | 0.59
s316 | 0.68
s252 | 0.70
s452 | 0.75
s4117 | 0.77
vif | 0.78
s1112 | 0.79
s2101 | 0.83
s319 | 0.83
s312 | 0.83
s311 | 0.85
vsumr | 0.85
s124 | 0.87
s243 | 0.89
s255 | 0.90
s1251 | 0.91
s115 | 0.92
s1281 | 0.92
s1279 | 0.92
s471 | 0.92
s272 | 0.94
s127 | 0.95
s453 | 0.95

</body>

</html>

The benchmarks that I ran are the llvm test-suite, TSVC, RAJAPerf, and SPEC INT. 
I didn't see any regressions. I could double check RAJAPerf as I didn't run that many iterations of it.
I can easily test SPEC FP too.

https://github.com/llvm/llvm-project/pull/100385