[llvm-bugs] [Bug 28345] New: AArch64 vs ARMv7 code quality

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=28345

            Bug ID: 28345
           Summary: AArch64 vs ARMv7 code quality
           Product: libraries
           Version: 3.8
          Hardware: Other
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: AArch64
          Assignee: unassignedbugs at nondot.org
          Reporter: tulipawn at gmail.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

I've rerun the benchmark from issue #27103 on a 2GHz Cortex-A53 (64-bit Linux),
and even though it achieves the expected 1.5x improvement (not counting 64-bit
effects), it still loses against a 1.7GHz Cortex-A5, running ARMv7 code, in
naive implementation performance. 

Provided, the last part actually tests the quality of the backend, it looks
rather interesting (table created with cargo-benchcmp script, negative
difference means cortex-a53 is faster):

name                           cortex-a5 ns/iter  cortex-a53 ns/iter    diff
ns/iter   diff %
mat_mul_f32::m004              1,763              1,098                        
-665  -37.72%
mat_mul_f32::m005              2,548              1,587                        
-961  -37.72%
mat_mul_f32::m006              2,889              1,718                      
-1,171  -40.53%
mat_mul_f32::m007              3,154              1,923                      
-1,231  -39.03%
mat_mul_f32::m008              3,627              2,138                      
-1,489  -41.05%
mat_mul_f32::m009              8,142              4,260                      
-3,882  -47.68%
mat_mul_f32::m012              10,370             5,484                      
-4,886  -47.12%
mat_mul_f32::m016              17,117             8,621                      
-8,496  -49.63%
mat_mul_f32::m032              110,929            48,623                    
-62,306  -56.17%
mat_mul_f32::m064              830,408            328,603                  
-501,805  -60.43%
mat_mul_f32::m127              6,416,219          2,387,223              
-4,028,996  -62.79%
mat_mul_f32::m256              52,750,069         20,490,803            
-32,259,266  -61.15%
mat_mul_f32::m512              421,350,950        164,162,031          
-257,188,919  -61.04%
mat_mul_f32::mix128x10000x128  531,878,944        216,476,447          
-315,402,497  -59.30%
mat_mul_f32::mix16x4           27,059             17,923                     
-9,136  -33.76%
mat_mul_f32::mix32x2           21,873             15,748                     
-6,125  -28.00%
mat_mul_f32::mix97             3,815,641          1,449,014              
-2,366,627  -62.02%

mat_mul_f64::m004              2,002              1,202                        
-800  -39.96%
mat_mul_f64::m007              4,593              2,102                      
-2,491  -54.23%
mat_mul_f64::m008              4,590              2,547                      
-2,043  -44.51%
mat_mul_f64::m012              14,212             7,781                      
-6,431  -45.25%
mat_mul_f64::m016              23,181             12,981                    
-10,200  -44.00%
mat_mul_f64::m032              160,551            88,629                    
-71,922  -44.80%
mat_mul_f64::m064              1,273,413          665,406                  
-608,007  -47.75%
mat_mul_f64::m127              10,648,815         5,531,354              
-5,117,461  -48.06%
mat_mul_f64::m256              88,419,854         45,800,153            
-42,619,701  -48.20%
mat_mul_f64::m512              702,121,682        365,977,216          
-336,144,466  -47.88%
mat_mul_f64::mix128x10000x128  876,955,471        493,044,547          
-383,910,924  -43.78%
mat_mul_f64::mix16x4           38,284             20,585                    
-17,699  -46.23%
mat_mul_f64::mix32x2           33,038             13,516                    
-19,522  -59.09%
mat_mul_f64::mix97             6,344,368          3,202,931              
-3,141,437  -49.52%

ref_mat_mul_f32::m004          473                530                          
  57   12.05%
ref_mat_mul_f32::m005          784                941                          
 157   20.03%
ref_mat_mul_f32::m006          1,219              1,537                        
 318   26.09%
ref_mat_mul_f32::m007          1,803              2,689                        
 886   49.14%
ref_mat_mul_f32::m008          2,553              3,783                       
1,230   48.18%
ref_mat_mul_f32::m009          3,830              5,307                       
1,477   38.56%
ref_mat_mul_f32::m012          7,829              11,755                      
3,926   50.15%
ref_mat_mul_f32::m016          17,466             26,824                      
9,358   53.58%
ref_mat_mul_f32::m032          128,387            202,902                    
74,515   58.04%
ref_mat_mul_f32::m064          1,018,211          1,584,415                 
566,204   55.61%

The ref_mat_mul results probably show the optimizer could get smarter :)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20160628/c7c3e4c4/attachment.html>