linking of cuda kernels via cmake segfaults upon execution

Autonomous Machines Jetson & Embedded Systems Jetson TX2
#1

Hello, I’ve been spending the day trouble shooting a way to add CUDA code to a TX2 host app, programmed in C++ code and compiled with Cmakelist/catkin_make.

The compile process is completing without error, but the program segfaults upon execution.
I was wondering if anyone on this forum could specify whether my usage of NVCC is correct, or whether there is a bug in my setup.

Toy example:

cuda_kernel.cu:

#include "cuda_kernel.hpp"
__global__ void test_cuda(int *a, int b)
{	
	a += 1;
}

cuda_kernel.hpp:

#include <cuda.h>
#include <cuda_runtime.h>
__global__ void test_cuda(int *a, int b);

Calling function in master cpp (app_source.cpp):

int one = 1; 
test_cuda(&one, 1);

CmakeList:

cmake_minimum_required(VERSION 2.8.3)
cmake_policy(SET CMP0012 NEW)

project(test)

find_package(CUDA REQUIRED)
set(CUDA_NVCC_FLAGS 
	${CUDA_NVCC_FLAGS}; 
	-O3 -lineinfo
	-gencode=arch=compute_62,code=sm_62)		
set(CUDA_VERBOSE_BUILD ON CACHE BOOL "nvcc verbose" FORCE)
#set(CUDA_SEPERABLE_COMPILATION ON)	
#set(BUILD_SHARED_LIBS ON)
#set(CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE OFF)

set (LIB_TYPE STATIC)
#set (LIB_TYPE SHARED)
	
CUDA_ADD_LIBRARY(cuda_functions ${LIB_TYPE} src/cuda_kernel.cu)
	
include_directories(... ${CUDA_INCLUDE_DIRS} ${catkin_INCLUDE_DIRS} ${Boost_INCLUDE_DIRS})
link_directories(... ${CUDA_CUFFT_LIBRARIES} ${CUDA_LIBRARIES})

set(CMAKE_CXX_FLAGS "-std=c++0x ${CMAKE_CXX_FLAGS} -march=native -lpthread -O3 -fopenmp")

add_executable(appname app_source.cpp)
add_dependencies(appname cuda_functions)

target_link_libraries(appname 		
	${catkin_LIBRARIES} 
	${CUDA_LIBRARIES} 
	${CUDA_CUFFT_LIBRARIES}
	cuda_functions)

I have tried compiling the cuda code as both static and shared libraries, but both appear to segfault whenever I call the test function. Any suggestions on how the cuda compiles can be better linked?

#2

Hi,

  1. GPU can’t access CPU memory.
    You need to allocate GPU accessible memory via CUDA API.

  2. Please use CUDA API to launch CUDA kernel.
    ex. test_cuda<<<64, 64>>>(…);

Here is a simple example for your reference:
https://devblogs.nvidia.com/parallelforall/easy-introduction-cuda-c-and-c/

#3

Thanks for the tip.

Is that to say the kernels cannot access unified memory in the way that other cuda runtimes, such as cuFFT are?
Since I am on pascal, I was able to directly pass CUDA enabled functions a pointer to system memory, and it worked as long as it was allocated with cudaManaged; is this safely extensible to custom kernels as well?

#4

Hi,

  1. API, such as cuFFT, contains a C++ wrapper to handle CUDA function call.
    For you own CUDA kernel, please remember to launch it with test_cuda<<<64, 64>>>(…)

  2. Unified memory can be access by CPU and GPU.
    But unified memory should be allocated via cudaMallocManaged(…)

#5