From 8cb46236f41ac5476f2fe38353156db4c8195325 Mon Sep 17 00:00:00 2001 From: OpportunityLiu Date: Sun, 2 Jun 2019 15:26:26 +0800 Subject: [PATCH] add docs for find_cudadevices; improve docs for cuda templates Related: xmake-io/xmake#430, xmake-io/xmake#431 --- README.md | 20 +++++++++----------- manual.md | 18 ++++++++++++++++++ zh/README.md | 21 ++++++++++----------- zh/manual.md | 18 ++++++++++++++++++ 4 files changed, 55 insertions(+), 22 deletions(-) diff --git a/README.md b/README.md index 490bf95c..7e806945 100644 --- a/README.md +++ b/README.md @@ -222,6 +222,7 @@ Support languages: * c/c++ * objc/c++ +* cuda * asm * swift * dlang @@ -422,20 +423,16 @@ $ xmake ``` ```lua +-- add helper function add_cugencode +includes('add_cugencode.lua') +-- define target target("cuda_console") set_kind("binary") add_files("src/*.cu") - - -- generate SASS code for each SM architecture - for _, sm in ipairs({"30", "35", "37", "50", "52", "60", "61", "70"}) do - add_cuflags("-gencode arch=compute_" .. sm .. ",code=sm_" .. sm) - add_ldflags("-gencode arch=compute_" .. sm .. ",code=sm_" .. sm) - end - - -- generate PTX code from the highest SM architecture to guarantee forward-compatibility - sm = "70" - add_cuflags("-gencode arch=compute_" .. sm .. ",code=compute_" .. sm) - add_ldflags("-gencode arch=compute_" .. sm .. ",code=compute_" .. sm) + -- generate SASS code for SM architecture of current host + add_cugencode("native") + -- generate PTX code for the virtual architecture to guarantee compatibility + add_cugencode("compute_30") ``` xmake will detect Cuda SDK automatically and we can also set the SDK directory manually. @@ -740,6 +737,7 @@ $ xmake | [--rc-ld](#-rc-ld) | Set `rust` linker | | [--rc-sh](#-rc-sh) | Set `rust` shared library linker | | [--rc-ar](#-rc-ar) | Set `rust` static library archiver | +| [--cu-cxx](#-cu-cxx) | Set `cuda` host compiler | | [--cu-ld](#-cu-ld) | Set `cuda` linker | | [--cu-sh](#-cu-sh) | Set `cuda` shared library linker | | [--cu-ar](#-cu-ar) | Set `cuda` static library archiver | diff --git a/manual.md b/manual.md index 4605534d..68e05eb0 100644 --- a/manual.md +++ b/manual.md @@ -7549,6 +7549,7 @@ The interface of this module is spread across multiple module directories, try t | [detect.find_package](#detect-find_package) | Find package files, including library files and search paths | >= 2.1.5 | | [detect.find_tool](#detect-find_tool) | Find Tool | >= 2.1.5 | | [detect.find_toolname](#detect-find_toolname) | Find Tool Name | >= 2.1.5 | +| [detect.find_cudadevices](#detect-find_cudadevices) | Find CUDA devices of the host | >= 2.2.7 | | [detect.features](#detect-features) | Get all the features of the specified tool | >= 2.1.5 | | [detect.has_features](#detect-has_features) | Determine if the specified feature is supported | >= 2.1.5 | | [detect.has_flags](#detect-has_flags) | Determine if the specified parameter options are supported | >= 2.1.5 | @@ -8113,6 +8114,23 @@ example: Compared with program, toolname can uniquely mark a tool, and it is also convenient to find and load the corresponding script `find_xxx.lua`. +###### detect.find_cudadevices + +- Find CUDA devices of the host + +Enumerate CUDA devices through the CUDA Runtime API and query theirs properties. + +```lua +import("lib.detect.find_cudadevices") + +local devices = find_cudadevices({ skip_compute_mode_prohibited = true }) +local devices = find_cudadevices({ min_sm_arch = 35, order_by_flops = true }) +``` + +The result returned is: `{ { ['$id'] = 0, name = "GeForce GTX 960M", major = 5, minor = 0, ... }, ... }` + +The included properties will vary depending on the current CUDA version. Refer to [CUDA Toolkit Documentation](https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaDeviceProp.html#structcudaDeviceProp) and its historical version. + ###### detect.features - Get all the features of the specified tool diff --git a/zh/README.md b/zh/README.md index af6ba511..db25f4bb 100644 --- a/zh/README.md +++ b/zh/README.md @@ -250,6 +250,7 @@ target("hello") * c/c++ * objc/c++ +* cuda * asm * swift * dlang @@ -452,22 +453,19 @@ $ xmake ``` ```lua +-- add helper function add_cugencode +includes('add_cugencode.lua') +-- define target target("cuda_console") set_kind("binary") add_files("src/*.cu") - - -- generate SASS code for each SM architecture - for _, sm in ipairs({"30", "35", "37", "50", "52", "60", "61", "70"}) do - add_cuflags("-gencode arch=compute_" .. sm .. ",code=sm_" .. sm) - add_ldflags("-gencode arch=compute_" .. sm .. ",code=sm_" .. sm) - end - - -- generate PTX code from the highest SM architecture to guarantee forward-compatibility - sm = "70" - add_cuflags("-gencode arch=compute_" .. sm .. ",code=compute_" .. sm) - add_ldflags("-gencode arch=compute_" .. sm .. ",code=compute_" .. sm) + -- generate SASS code for SM architecture of current host + add_cugencode("native") + -- generate PTX code for the virtual architecture to guarantee compatibility + add_cugencode("compute_30") ``` + 默认会自动探测cuda环境，当然也可以指定Cuda SDK环境目录： ```console @@ -772,6 +770,7 @@ $ xmake | [--rc-ld](#-rc-ld) | 设置`rust`链接器 | | [--rc-sh](#-rc-sh) | 设置`rust`共享库链接器 | | [--rc-ar](#-rc-ar) | 设置`rust`静态库归档器 | +| [--cu-cxx](#-cu-cxx) | 设置`cuda` host编译器 | | [--cu-ld](#-cu-ld) | 设置`cuda`链接器 | | [--cu-sh](#-cu-sh) | 设置`cuda`共享库链接器 | | [--cu-ar](#-cu-ar) | 设置`cuda`静态库归档器 | diff --git a/zh/manual.md b/zh/manual.md index d1cbb270..4fc58bcf 100644 --- a/zh/manual.md +++ b/zh/manual.md @@ -7632,6 +7632,7 @@ environment.leave("toolchains") | [detect.find_package](#detect-find_package) | 查找包文件，包含库文件和搜索路径 | >= 2.1.5 | | [detect.find_tool](#detect-find_tool) | 查找工具 | >= 2.1.5 | | [detect.find_toolname](#detect-find_toolname) | 查找工具名 | >= 2.1.5 | +| [detect.find_cudadevices](#detect-find_cudadevices) | 查找本机的 CUDA 设备 | >= 2.2.7 | | [detect.features](#detect-features) | 获取指定工具的所有特性 | >= 2.1.5 | | [detect.has_features](#detect-has_features) | 判断指定特性是否支持 | >= 2.1.5 | | [detect.has_flags](#detect-has_flags) | 判断指定参数选项是否支持 | >= 2.1.5 | @@ -8162,6 +8163,23 @@ end toolname相比program，更能唯一标示某个工具，也方便查找和加载对应的脚本`find_xxx.lua`。 +###### detect.find_cudadevices + +- 查找本机的 CUDA 设备 + +通过 CUDA Runtime API 枚举本机的 CUDA 设备，并查询其属性。 + +```lua +import("lib.detect.find_cudadevices") + +local devices = find_cudadevices({ skip_compute_mode_prohibited = true }) +local devices = find_cudadevices({ min_sm_arch = 35, order_by_flops = true }) +``` + +返回的结果为：`{ { ['$id'] = 0, name = "GeForce GTX 960M", major = 5, minor = 0, ... }, ... }` + +包含的属性依据当前 CUDA 版本会有所不同，可以参考 [CUDA 官方文档](https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaDeviceProp.html#structcudaDeviceProp)及其历史版本。 + ###### detect.features - 获取指定工具的所有特性