OpenACC與CUDA Fortran交互(1)

先上代碼：

1. ! openacc_main.f90
   
2. program main
   
3.   use saxpy_mod
   
4.   integer, parameter :: N = 2**20
   
5.   real, dimension(N) :: X, Y
   
6. 
   
7.   X(:) = 1.0
   
8.   Y(:) = 0.0
   
9. 
   
10.   !$acc data copy(y) copyin(x)
    
11.   call saxpy(N, 2.0, x, y)
    
12.   !$acc end data
    
13. 
    
14.   print *, y(1)
    
15. end program

1. ! kernels.cuf
   
2. module saxpy_mod
   
3.   contains
   
4.   attributes(global) &
   
5.   subroutine saxpy_kernel(n, a, x, y)
   
6.     real :: x(:), y(:), a
   
7.     integer :: n,i
   
8.     attributes(value) :: a,n
   
9.     i = threadIdx%x+(blockIdx%x-1)*blockDim%x
   
10.     if (i<=n) y(i) = y(i) + a*x(i)
    
11.   end subroutine
    
12.   subroutine saxpy (n, a, x, y)
    
13.     use cudafor
    
14.     real, device :: x(:), y(:)
    
15.     real :: a
    
16.     integer :: n
    
17.     call saxpy_kernel<<<4096,256>>>(n, a, x, y)
    
18.   end subroutine
    
19. end module saxpy_mod

對於函數saxpy_kernel來說，變量x,y 有device屬性，編譯器會知道傳過來的是設備數組，故不需要host_data導語。