EmbedPTX

This repo implements a CMake function to help embed PTX using obj2c. This uses the modern CUDA features found in CMake 3.8+ to maximize simplicity of use.

Usage

Incorporating in your project

You can add this code to your CMake build in a number of ways:

• As a git submodule
• Using CMake's FetchContent module
• Adding the files to your project directly

Using EmbedPTX()

One you have the scripts available to your CMake build, use it by putting both EmbedPTX.cmake and EmbedPTXRun.cmake on CMAKE_MODULE_PATH and including EmbedPTX.

Here's a short example, which embeds myCudaPtx kernels in a header called embedded_ptx.h:

cmake_minimum_required(VERSION 3.8 FATAL_ERROR)

project(example LANGUAGES CXX CUDA)

list(APPEND CMAKE_MODULE_PATH /path/to/EmbedPTX.cmake)
include(EmbedPTX)

add_library(ptx_gen OBJECT myCudaPtx.cu)
set_property(TARGET ptx_gen PROPERTY CUDA_PTX_COMPILATION ON)
set_property(TARGET ptx_gen PROPERTY CUDA_ARCHITECTURES OFF)

add_executable(myCudaExample myCudaExample.cpp)

EmbedPTX(
  OUTPUT_HEADER_FILE ${CMAKE_CURRENT_BINARY_DIR}/embedded_ptx.h
  INPUT_TARGET       ptx_gen
  OUTPUT_TARGETS     myCudaExample
)

There are 3 arguments to pass to EmbedPTX():

• OUTPUT_HEADER_FILE specifies the header to be generated.
• INPUT_TARGET specifies the object library used to generate the PTX files.
• OUTPUT_TARGETS is a list of targets which will consume the generated header file.

Note that the directory containing OUTPUT_HEADER_FILE is automatically added as a PRIVATE include directory to each target passed in OUTPUT_TARGETS.