Question

I need to call Atan on millions of values per second. Is there a good library to perform this operation in batch very fast. For example, a library that streams the low level logic using something like SSE?

I have profiled the application, and I know that this call to Atan is a bottleneck.

I know that there is support for this in OpenCL, but I would prefer to do this operation on the CPU. The target machine might not support OpenCL.

I also looked into using OpenCV, but it’s accuracy for Atan angles is only ~0.3 degrees. I need accurate results.

Answer 1

Why don’t you try Brahma? To my understanding, it’s a free and open source GPGPU library that doesn’t depend on OpenCL, instead converting code to HLSL/GLSL shaders via LINQ.

EDIT: Sample code:

ComputationProvider provider = new ComputationProvider();
CompiledQuery query = provider.Compile<DataParallelArray<float>>(
    data => from value in data
            select (float)Math.Atan(value)); // Do your calculations here...

DataParallelArray<float> input = new DataParallelArray<float>(provider, new float[] { 0, 1, 2, 3, 4, 5, 6, 7, }); // etc...
IQueryable result = provider.Run(query, input);

foreach (float value in result)
    Console.WriteLine(value);