Cuda+fortran

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello
There is the code:

module gpu_f
	   use cudafor
	   contains
	   attributes(global) subroutine force_gpu(SX_pr)
	   real :: sx_pr(4)
	   integer :: i,j,i1
	   i=(blockidx%x-1)*blockdim%x + threadidx%x
	   i1=i+1
	   
	   do 1 j=i1,4
	   if (i.le.3 ) then
	   sx_pr(j)=sx_pr(j)-1
	   sx_pr_i=sx_pr_i+1
	   endif
	   call syncthreads()
1      continue
       
       sx_pr(i)=sx_pr(i)+sx_pr_i
	    
       
	   endsubroutine
	   end module

	   program main
	   use gpu_f
	   real, device :: SX_pr_d(4)
	   real :: SX_pr(4)
	   sx_pr=0
	   sx_pr_d=sx_pr
	   call force_gpu<<<1,4>>>(sx_pr_d)
	   sx_pr=sx_pr_d
	   do i= 1, 4
	   print*,'i',i,'=',sx_pr(i)
	   end do
	   end program

The program gives the answer
D:\PVFProject2\Win32\Debug>PVFProject2.exe
i 1 = 3.000000
i 2 = 1.000000
i 3 = -1.000000
i 4 = -3.000000

If I change the number of blocks on
call force_gpu<<<2,2>>>(sx_pr_d)
or
call force_gpu<<<4,1>>>(sx_pr_d)
I get the result:
i1 = 3.000000
i2 = 1.000000
i3 = 0.000000
i4 = -2.000000

Help me understand where the error?

2

This looks somewhat suspicious:

i = (blockidx%x-1)*blockdim%x + threadidx%x

I assume both blockidx%x and threadidx%x are 1-based (as opposed to 0-based in CUDA C)? If so, wouldn’t you want to use (threadidx%x-1) ?

3

The address computation is correct.

@greepez: What are you trying to compute?

4

-In this example i am trying to calculate the interaction of particles used on the Newton’s third law

5

I assume that this is due to the calculations of the same element of the array in different blocks, such as the first iteration of an array element sx_pr (3) is calculated in the second block, and the second iteration of the first block