olegkonings / cuda_brute_triangle

brute force examines all n choose 3 triangles

CUDA_brute_triangle

brute force examines all possible n choose 3 triangles only once

UPDATE: Optimized both implementations, so new times posted.Run using the --use_fast_math flag, and must have a GPU with compute capability >=3.5. Best performance is max_register set to 32.

This code goes through every possible 3 point combination of a set of points, makes a triangle of those three points, then evalutes how many other points are within that triangle.

In this simple example case, the objective is to find which triangle of the possible set contains within its borders the greatest number of other points(exclusive). The two different CPU and GPU functions return that max number of points, and the INDEXES of the three points which created that optimal triangle.

NOTE: Since the GPU version does not proceed in a serial fashion, if there is more than one combination associated with the optimal value it will return a valid combination, BUT not necessarly the same(first seen) combination the CPU version returned. The overall answer will be correct, but a hueristic needs to be added for the CUDA version to return a specific arrangement if more than one does indeed exist.

While many CUDA GPU implementations of algorithms many only be 10-100 times faster than a single core CPU implementation, this problem has a much greater difference in performance.

The larger the data set of points, the greater the outperformance of the CUDA GPU implementation. For data sets of points >=400 the GPU CUDA implementation was at least 400x times faster than a 3.9 Ghz CPU implementation(including all host-device, device-device and device-host memory copies).

No overlocking of GPU, is running at stock 706 Mhz for older Kepler Tesla K20c

Optimal Triangle Running Time comparison:

Number of points	Intel I-3770K 3.9 Ghz CPU time	Tesla K20c GPU time	CUDA Speedup
300	13,198 ms	32 ms	412.438x
400	42,710 ms	99 ms	431.32x
500	103,731 ms	240 ms	432.2x
700	411,401 ms	912 ms	451.01x
1300	5,075,524 ms	10,788 ms	470.47x

New Times for GTX 980 and GTX 780ti

Number of points	Intel I-3770K 3.9 Ghz CPU time	GTX 980 time	GTX 780ti time	GTX Titan X time
700	411,401 ms	716 ms	655 ms	573 ms
1300	5,075,524 ms	8,422 ms	7,720 ms	6,735 ms

---

The running time is apx (N choose 3)*N, where N is the number of 2-D points in the array.

What makes this even more impressive is that the GPU actually has to do over 10x more work for the same answer (this is apparent in code, compare CPU version to GPU).

AMD users, post your times and code. I would like to compare results. Same goes to the multi-core CPU crowd who thinks a Xeon can beat a GPU at this type of problem. Prove it ..

<script> (function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){ (i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o), m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m) })(window,document,'script','//www.google-analytics.com/analytics.js','ga'); ga('create', 'UA-60172288-1', 'auto'); ga('send', 'pageview'); </script>