CPU-targeted merge sort

[jszuppe]

This adds CPU-targeted stable merge sort (bottom-up version). It does not use local memory, so it should work on OS X without problems.

Algorithms works in to steps:

1. Insertion sort for the blocks of 64 values is performed (the size of block is customizable).
2. Sorted blocks are merged.

It has O(n) space complexity.

If input has 512 elements of less insertion sort is performed. Merge sort outperforms std::sort() on my AMD Phenom(tm) II X3 710 when input has more than 65536, but I think it should work better on a CPU with more cores.

In order to test this algorithm I recommend running export BOOST_COMPUTE_DEFAULT_DEVICE_TYPE=CPU (or however it's done in you shell) before executing perf_sort or test_sort. This environment variable makes your CPU the default device for Boost.Compute. Otherwise you will probably end up with GPU being the default device and other sort algorithm will be chosen.

=== sort with stl ===
size,time (ms)
2,0.000257
4,0.000274
8,0.000418
16,0.000595
32,0.001082
64,0.001757
128,0.003570
256,0.008004
512,0.016945
1024,0.035260
2048,0.073832
4096,0.163459
8192,0.346311
16384,0.738796
32768,1.568840
65536,3.363270
131072,7.135330
262144,15.017100
524288,31.467200
1048576,66.027500
2097152,138.717000
4194304,291.526000
8388608,609.542000
16777216,1269.250000
33554432,2650.750000

=== sort with compute [with usage of merge with merge path] ===
size,time (ms)
2,0.057767
4,0.057516
8,0.057271
16,0.057569
32,0.053032
64,0.055073
128,0.058193
256,0.071131
512,0.124001
1024,0.304344
2048,0.392302
4096,0.525845
8192,0.709185
16384,0.951316
32768,1.608410
65536,2.833530 <-- from this point merge sort is better than std::sort()
131072,5.294260
262144,9.382830
524288,19.141900
1048576,40.889400
2097152,83.336700
4194304,162.947000
8388608,350.989000
16777216,704.070000
33554432,1505.250000

=== sort with compute  [without usage of merge with merge path] ===
size,time (ms)
2,0.056930
4,0.057798
8,0.056881
16,0.061113
32,0.054506
64,0.055622
128,0.059051
256,0.072387
512,0.123452
1024,0.309484
2048,0.399578
4096,0.544659
8192,0.688900
16384,0.959716
32768,1.669320
65536,2.824090
131072,5.257270
262144,9.357230
524288,19.151700
1048576,40.593000
2097152,88.478000
4194304,177.051000
8388608,388.458000
16777216,761.811000
33554432,1610.750000

If this PR is accepted, I'll soon add similar CPU-targeted sort-by-key algorithm.

[kylelutz]

Awesome! Merged!

[jszuppe]

Thanks!