foreword
In fact, there were already two heterogeneous programming frameworks before C++ AMP: CUDA and OpenCL.
CUDA (Compute Unified Device Architecture) is the industry's first heterogeneous parallel programming framework launched by graphics card manufacturer Nvidia in 2007. With the strong support of Nvidia, CUDA has a good development environment, rich function library, and excellent performance. However, CUDA can only be used for heterogeneous programming on Nvidia graphics cards, which has inherent limitations.
OpenCL (Open Computing Language) is the industry's first cross-platform heterogeneous programming framework. It is an open standard led by Apple and jointly launched by Nvidia, AMD, IBM, Intel and many other manufacturers in 2008. It is managed by the Khronos Group, a separately established non-profit organization. Similar to C++ AMP, OpenCL, as an open standard, is not limited to a specific GPU manufacturer. From this point of view, Nvidia's own exclusive CUDA is very closed. We can compare OpenCL's position in heterogeneous programming to OpenGL and OpenAL, the two standards used for 3D graphics and computer audio, respectively.
Because CUDA and OpenCL are closer to the underlying hardware than C++AMP, the performance of the former two is better
However, the ease of programming with C++ AMP is better than CUDA and OpenCL
Unlike C++ AMP, which is directly extended based on C++ language features, OpenCL is based on related modifications and extensions of the C99 programming language.
Therefore, C++ AMP has a higher level of abstraction than OpenCL, and programming is simpler
In CUDA and OpenCL, kernels (code that runs on the GPU) must be packaged into specific functions
In C++ AMP, the code looks much neater: we only need to use the lambda function embedded in the for loop to complete heterogeneous parallel computing, and its memory model is also greatly simplified to a certain extent.
How should developers choose between OpenCL, CUDA and C++ AMP?
If you only need heterogeneous programming on the Windows platform and value ease of programming, C++ AMP is undoubtedly the best choice. Relying on the powerful development tool of Visual Studio, coupled with the inherent advantages brought by the higher-level abstraction based on C++, C++ AMP will provide good support for heterogeneous programming for Windows developers.
If you only need to perform heterogeneous programming on Nvidia's GPU card and value performance very much, CUDA should be the first choice: With the strong support of Nvidia, CUDA's performance on Nvidia hardware has always been leading, and many academic studies have shown The performance of OpenCL and CUDA is not much different, and the performance of CUDA is slightly better than OpenCL in some applications.
At the same time, the development environment of CUDA is also very mature, with support from many extended function libraries.
If you are more concerned about portability between different platforms, OpenCL is probably the best choice at present. As the first open standard for heterogeneous computing, OpenCL has received strong support from many software and hardware manufacturers including Intel, AMD, Nvidia, IBM, Oracle, ARM, Apple, and Redhat.
Of course, C++ AMP itself is also an open standard, but currently only Microsoft has implemented it. It is still unknown to what extent the cross-platform support of C++ AMP can be achieved in the future.
In fact, from the perspective of the development of programming languages, ease of programming is often more important than performance.
From the popularity of Java and .Net to the rise of scripting languages, programming efficiency is undoubtedly the most important indicator.
Not to mention that developers can often get better performance by switching to the next generation of GPU hardware. From this point of view, C++ AMP actually promotes the popularity of heterogeneous programming by reducing the programming difficulty of heterogeneous programming.