C# 调用GPU计算（Alea库、英伟达GPU）

关于C#调用GPU计算，下面是stackoverflow的一段总结：

链接：<https://stackoverflow.com/questions/375011/utilizing-the-gpu-with-c-sharp> ：

Most of these answers are quite old, so I thought I'd give an updated summary of where I think each project is:

GPU.Net (TidePowerd) - I tried this 6 months ago or so, and did get it working though it took a little bit of work. Converts C# kernel code to cuda at compile time. Unfortunately their website has been down and their github hasn't been updated for a couple of years, which might indicate the project is dead....
Cudafy - Open source and very easy to use. Converts C# kernel code to cuda at runtime (with ability to serialize and cache). Can easily run the same kernel code on the CPU (mostly for debugging). Supports multiple GPUs. More examples available than others here. The boilerplate code referred to by other answers is minimal, and in my case at least helped with my understanding of how the code works. Cuda/Nvidia only though. Unfortunately, it seems that they didn't update their solutions for a couple of years too (latest commit in 2015 -- support of cuda 7.0).
Hybridizer. Commercial solution compiling C# to CUDA. Provides a free community edition on visual studio marketplace and samples on github.
AleaGPU Commercial solution with a free community edition for consumer GPUS. See Daniel's comments for details.
Brahma - runs LINQ expressions via OpenCL (so supports AMD too). Not much documentation / examples. Last update in 2011.
C$ - last development was over 10 years ago...
Microsoft Accelerator - similarly doesn't look like it is being actively developed any longer.
some others (C++ AMP, OpenTK -- dead/Cloo) - many of these are just bindings - ie enable you to call the GPU from C#, but your kernel code (code which is actually run on the GPU) needs to be written in C or OpenCL, meaning you must use (and learn) another language.

As I said, I would recommend Cudafy over all the others - if it could run on OpenCL as well as Cuda it would be perfect.

EDIT SEP 2013 Cudafy now allows you to compile for both CUDA and OpenCL, so will run the same C# code on on all GPUs. This sounds fantastic, though I haven't tested the OpenCL compiling yet.

综上，考虑到类库的完成性以及便利性，我选择使用了 Alea 库。

关于Alea GPU：

对于单GPU调用计算，alea gpu是免费的，多gpu调用是需要另外收取费用。

alea gpu比较方便C#程序员开发程序，其自动内存管理特性，自动将内存在CPU和GPU之间复制，对小白来说是比较友好的功能了。

gpu.for 函数，可以很方便的将CPU并行级的 Paraller.For 改成gpu.for，从而实现GPU上的并行计算（可以很明显的降低CPU的利用率）。

gpu.for和kernel函数都支持委托，很方便将现有代码改成gpu上运行的代码。

除此之外，也支持开发者自己考虑分配内存。

官方说明文档（学习资料）：http://www.aleagpu.com/release/3_0_4/doc/introduction.html

文档介绍的很详细，开发环境配置、CUDA编程模式、C#调用教程以及C#Demo 都很详细，就不做赘述了，有问题可以留言。

了解cuda编程基础，可以看看这个：https://zhuanlan.zhihu.com/p/34587739