BrookGPU 和 Sh 都使用着色器作为后端，因此它们远未具备“真正的”GPGPU 平台（如 CUDA 或 OpenCL）的所有功能，并且我认为性能较差。而且，至少 Brook 需要专有驱动程序才能运行。另外，SourceForge 上的 Brook 项目的最新更新日期是 2010 年 2 月，所以这个项目也已经死了。
OpenCL 尽管是开放标准，但确实拥有专有编译器（对于 nVidia，编译器与驱动程序捆绑在一起，对于 AMD/ATI 和 Intel，可以单独下载）。当然可以开发免费的编译器，但据我所知，目前还没有开放的硬件 GPGPU 架构（最后一种被广泛使用的架构：nVidia 和 ATI 都没有向公众公开其架构细节）。我们可以确定不同体系结构之间的命令集有很大不同，因此对于想要支持的每种体系结构，都必须对其进行逆向工程（我猜它可能违反许可协议，但不确定）。我听说 nouveau 团队进行了一些尝试，但没有任何明显的结果。
There is PTX backend for LLVM, but to get runnable binary you still need proprietary NVCC compiler, and proprietary nVidia video driver to run it.
BrookGPU and Sh both use shaders as their backend, so they are far from having all capabilities of "real" GPGPU platforms (like CUDA or OpenCL) and, I guess, have poorer performance. And, at least Brook, requires proprietary drivers to run. Also, latest update to Brook project on SourceForge is dated Feb. 2010, so this project is quite dead too.
OpenCL, despite being open standard, indeed have proprietary compilers (in case of nVidia the compiler is bundled with driver, for AMD/ATI and Intel it can be downloaded separately). Of course one can develop free compiler, but as far as I know, right now there is no open hardware GPGPU architecture (at last wide used one: both nVidia and ATI keep their architecture details unavailable to public). And we can be sure that command set is quite different between different architectures, so for each architecture one want to support, one would have to reverse engineer it (and I guess it might violate license agreement, not sure though). I've heard some attempts were made by nouveau team, but with no any visible results.
发布评论
评论（1）
有LLVM 的 PTX 后端，但要可运行二进制文件，您仍然需要专有的 NVCC 编译器和专有的 nVidia 视频驱动程序来运行它。
BrookGPU 和 Sh 都使用着色器作为后端，因此它们远未具备“真正的”GPGPU 平台（如 CUDA 或 OpenCL）的所有功能，并且我认为性能较差。而且，至少 Brook 需要专有驱动程序才能运行。另外，SourceForge 上的 Brook 项目的最新更新日期是 2010 年 2 月，所以这个项目也已经死了。
OpenCL 尽管是开放标准，但确实拥有专有编译器（对于 nVidia，编译器与驱动程序捆绑在一起，对于 AMD/ATI 和 Intel，可以单独下载）。当然可以开发免费的编译器，但据我所知，目前还没有开放的硬件 GPGPU 架构（最后一种被广泛使用的架构：nVidia 和 ATI 都没有向公众公开其架构细节）。我们可以确定不同体系结构之间的命令集有很大不同，因此对于想要支持的每种体系结构，都必须对其进行逆向工程（我猜它可能违反许可协议，但不确定）。我听说 nouveau 团队进行了一些尝试，但没有任何明显的结果。
There is PTX backend for LLVM, but to get runnable binary you still need proprietary NVCC compiler, and proprietary nVidia video driver to run it.
BrookGPU and Sh both use shaders as their backend, so they are far from having all capabilities of "real" GPGPU platforms (like CUDA or OpenCL) and, I guess, have poorer performance. And, at least Brook, requires proprietary drivers to run. Also, latest update to Brook project on SourceForge is dated Feb. 2010, so this project is quite dead too.
OpenCL, despite being open standard, indeed have proprietary compilers (in case of nVidia the compiler is bundled with driver, for AMD/ATI and Intel it can be downloaded separately). Of course one can develop free compiler, but as far as I know, right now there is no open hardware GPGPU architecture (at last wide used one: both nVidia and ATI keep their architecture details unavailable to public). And we can be sure that command set is quite different between different architectures, so for each architecture one want to support, one would have to reverse engineer it (and I guess it might violate license agreement, not sure though). I've heard some attempts were made by nouveau team, but with no any visible results.