ARROW-764: [C++] Improves performance of CopyBitmap and adds benchmarks

[seibs]

I took a swing at improving the CopyBitmap performance (benchmarks below). I'm a C/C++ novice, so I thought I'd get some feedback before I went too much further.

Starting Point

Run on (4 X 2208 MHz CPU s)
12/13/17 21:15:18
Benchmark                                        Time           CPU Iterations
------------------------------------------------------------------------------
BM_CopyBitmap/97.6563k/0/min_time:1.000       4779 us       4758 us        289   20.0445MB/s
BM_CopyBitmap/976.563k/0/min_time:1.000      47740 us      47476 us         26   20.0875MB/s
BM_CopyBitmap/97.6563k/4/min_time:1.000       4858 us       4866 us        289   19.5991MB/s
BM_CopyBitmap/976.563k/4/min_time:1.000      48117 us      47953 us         29   19.8879MB/s

Using stanford bithacks for SetBitTo

Run on (4 X 2208 MHz CPU s)
12/13/17 21:22:05
Benchmark                                        Time           CPU Iterations
------------------------------------------------------------------------------
BM_CopyBitmap/97.6563k/0/min_time:1.000       1647 us       1649 us        815   57.8415MB/s
BM_CopyBitmap/976.563k/0/min_time:1.000      16368 us      16397 us         81   58.1629MB/s
BM_CopyBitmap/97.6563k/4/min_time:1.000       1599 us       1610 us        815   59.2186MB/s
BM_CopyBitmap/976.563k/4/min_time:1.000      16026 us      16011 us         81   59.5644MB/s

memcpy + shifting
This solution provides varying performance depending on whether or not the bit offset is a multiple of 8

Run on (4 X 2208 MHz CPU s)
12/13/17 21:23:44
Benchmark                                        Time           CPU Iterations
------------------------------------------------------------------------------
BM_CopyBitmap/97.6563k/0/min_time:1.000          5 us          5 us     280000   18.9651GB/s
BM_CopyBitmap/976.563k/0/min_time:1.000         62 us         61 us      22400   15.1721GB/s
BM_CopyBitmap/97.6563k/4/min_time:1.000        171 us        170 us       6892   560.872MB/s
BM_CopyBitmap/976.563k/4/min_time:1.000       1639 us       1639 us        896   581.782MB/s

[xhochy]

+1, thank you for taking a look at this. There are two code formatting issues reported by Travis. Can you fix these two?

[xhochy]

Can you keep this an int64_t?

[xhochy]

Can you remove these reindentations?

[seibs]

Thanks for taking a look @xhochy.

My latest push fixes the formatting issues and also adds a test+fix for a few corner cases my first upload didn't handle correctly.

Benchmark with recent changes

Run on (4 X 2208 MHz CPU s)
12/22/17 15:14:31
Benchmark                                        Time           CPU Iterations
------------------------------------------------------------------------------
BM_CopyBitmap/97.6563k/0/min_time:1.000          5 us          5 us     270310   17.9019GB/s
BM_CopyBitmap/976.563k/0/min_time:1.000         65 us         65 us      21854   14.3143GB/s
BM_CopyBitmap/97.6563k/4/min_time:1.000        153 us        152 us       8960   628.592MB/s
BM_CopyBitmap/976.563k/4/min_time:1.000       1534 us       1535 us        896   621.449MB/s

w.r.t. the comments above in CMakeLists.txt, I actually have a separate pull request open for them at #1406.

[xhochy]

@seibs Can resolve the two comments in the other pull request? I would then merge both :)

[wesm]

+1, rebased. Will merge once build passes

[wesm]

Rebased again to see if build passes...

[siddharthteotia]

What are the general performance concerns we are addressing here? I am just trying to understand if there is something similar to be done on JAVA side for bitvector.

[wesm]

We should write down explicit types for these autos

[seibs]

Replaced the autos with types.

[wesm]

@siddharthteotia there's a couple places in the codebase where we copy bits out of a bitmap, e.g. after slicing vectors

arrow/cpp/src/arrow/ipc/writer.cc

Line 63 in 2e3832f

RETURN_NOT_OK(CopyBitmap(pool, input->data(), offset, length, buffer));

. Depending on the slice offset (whether it's a multiple of 8) we have to either copy 1 bit at a time or do a memcpy

[wesm]

+1. Build failing due to some standard flakiness