ARROW-5557: [C++] Add VisitBits benchmark

[pitrou]

No description provided.

[pitrou]

Results on Linux:

-----------------------------------------------------------------------
Benchmark                                Time           CPU Iterations
-----------------------------------------------------------------------
ReferenceNaiveBitmapReader/8192     123334 ns     123309 ns      11344   126.714MB/s
BitmapReader/8192                   110950 ns     110926 ns      12544    140.86MB/s
VisitBits/8192                      113867 ns     113842 ns      12290   137.252MB/s
VisitBitsUnrolled/8192               54291 ns      54280 ns      25955    287.86MB/s
ReferenceNaiveBitmapWriter/8192     110612 ns     110590 ns      12635   70.6437MB/s
BitmapWriter/8192                   111624 ns     111600 ns      12183   70.0044MB/s
FirstTimeBitmapWriter/8192           67885 ns      67871 ns      20632   115.108MB/s
GenerateBits/8192                    72072 ns      72057 ns      19008    108.42MB/s
GenerateBitsUnrolled/8192            42734 ns      42725 ns      32885   182.856MB/s
CopyBitmapWithoutOffset/8192           176 ns        176 ns    7943701   43.2839GB/s
CopyBitmapWithOffset/8192             7814 ns       7812 ns     179221   1000.06MB/s

@fsaintjacques

[pitrou]

Is there anyone who can benchmark on MSVC? I'm getting mediocre numbers here so something may be wrong with my Windows setup.

[wesm]

I can try the benchmarks on my local baremetal Windows box

[wesm]

here's the results on my Windows machine

(arrow-dev) λ release\Release\arrow-bit-util-benchmark.exe                        
06/13/19 12:27:25                                                                 
Running release\Release\arrow-bit-util-benchmark.exe                              
Run on (8 X 3096 MHz CPU s)                                                       
CPU Caches:                                                                       
  L1 Data 32K (x4)                                                                
  L1 Instruction 32K (x4)                                                         
  L2 Unified 262K (x4)                                                            
  L3 Unified 8388K (x1)                                                           
--------------------------------------------------------------------              
Benchmark                             Time           CPU Iterations               
--------------------------------------------------------------------              
BitmapReader/8192                150894 ns     149972 ns       4480   104.186MB/s 
VisitBits/8192                   150735 ns     149972 ns       4480   104.186MB/s 
VisitBitsUnrolled/8192           192623 ns     192540 ns       3733   81.1522MB/s 
BitmapWriter/8192                 96463 ns      96257 ns       7467    81.163MB/s 
FirstTimeBitmapWriter/8192        79010 ns      78474 ns       8960   99.5556MB/s 
GenerateBits/8192                 68788 ns      68359 ns      11200   114.286MB/s 
GenerateBitsUnrolled/8192         80782 ns      81961 ns       8960   95.3191MB/s 
CopyBitmapWithoutOffset/8192         90 ns         91 ns    8960000   84.1346GB/s 
CopyBitmapWithOffset/8192         10239 ns      10045 ns      74667   777.781MB/s 

i7-8809G 3.1 GHz on Windows 10

[pitrou]

Ok. My results are similar here. It is surprising that MSVC isn't able to optimized the unrolled versions better...

[wesm]

Appveyor build: https://ci.appveyor.com/project/pitrou/arrow/builds/25348791