Nsight CUDA linkage problem between multiple .cu, .h and .c files - c++

Question

This is the first time I am trying to build my CUDA app in NSight Ubuntu to benefit from optimization and profiling. This app works fine from terminal using nvcc (makefile) in Ubuntu 20 (or 18 ,16). I have multiple .cu, .c and .h files. All the files are first included in a flags.h file. My code starts with main.cu (has a main() function) file and this file has # include "flags.h" to make sure all the files are included to compile the code.
flags.h has a lot of #define as well to be later used in different .cu and .c files.
However, inside NSight, none of the #define parameters defined in flags.h are recognized in any of the files and I am getting error.
Following is screenshot of error .
I am attaching a simple square_array problem split in 3 files (main.cu, flags.h and square_.cu).
I can not build this in NSight. Can someone try to build it and let me know please.
Any help or suggestion will be much appreciated.
main.cu
#include <stdio.h>
#include <stdlib.h>
#include "flags.h"
int main(void) {
int i;
int *a_h, *a_d;
CUDA_CHECK_RETURN(cudaMalloc((void**) &a_d, sizeof(int) * WORK_SIZE));
a_h = (int *)malloc(sizeof(int) * WORK_SIZE); // Allocate array on host
for (i = 0; i < WORK_SIZE; i++)
a_h[i] = i+2.;
int block_size = 4;
int n_blocks = WORK_SIZE/block_size + (WORK_SIZE%block_size == 0 ? 0:1);
sq_array<<<n_blocks, block_size>>>(a_d);
CUDA_CHECK_RETURN(cudaGetLastError());
CUDA_CHECK_RETURN(cudaMemcpy(a_h, a_d, sizeof(int) * WORK_SIZE, cudaMemcpyDeviceToHost));
for (i = 0; i < WORK_SIZE; i++)
printf("Input value: %d \n", a_h[i] );
CUDA_CHECK_RETURN(cudaFree((void*) a_d));
CUDA_CHECK_RETURN(cudaDeviceReset());
return 0;
}
flags.h
#ifndef FLAGS_H_
#define FLAGS_H_
#include "square_.cu"
#define CUDA_CHECK_RETURN(value) { \
cudaError_t _m_cudaStat = value; \
if (_m_cudaStat != cudaSuccess) { \
fprintf(stderr, "Error %s at line %d in file %s\n", \
cudaGetErrorString(_m_cudaStat), __LINE__, __FILE__); \
exit(1); \
} }
#define WORK_SIZE 29
#endif /* FLAGS_H_ */
square_.cu
__global__ void sq_array( int *a) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx< WORK_SIZE) a[idx] = a[idx] * a[idx];
}

Answer 1

The problem is the your IDE is compiling square_.cu and compiling main.cu which is also compiling square_.cu again due to the #include "square_.cu" in flags.h which gives you two definitions of sq_array. When square_.cu is compiled the WORK_SIZE macro is not defined giving you a compile time error. When you compiled on the command line you did not compile square_.cu so you avoided this error.
In any case, it is a bad idea to #include .cu (or .c files). These should be compiled separately and then linked together.
You have to organize things differently.
I don't know the details of your code, but you can do something like this:
square.cu:
#include "square.h"
__global__ void sq_array( int *a) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx< WORK_SIZE) a[idx] = a[idx] * a[idx];
}
void host_sq_array(int *a_d) {
int block_size = 4;
int n_blocks = WORK_SIZE/block_size + (WORK_SIZE%block_size == 0 ? 0:1);
sq_array<<<n_blocks, block_size>>>(a_d);
}
square.h:
#ifndef SQUARE_H
#define SQUARE_H
#include "flags.h" // REMOVE #include of .cu file!!!
void host_sq_array(int *a_d);
#endif
You can safely #include square.h which only includes constant, type definitions, and function prototypes.