高效且正确的方式来从全局到共享内存加载带有光晕的阵列

Question

I am facing a problem of loading arrays from global to shared memory with hallo

Here is the problem: I have a big array (256,64) in my global memory that i want to load to shared memory of size [16][16] In my computation I will need the neighbouring value (halo)

I find my self in a very diverged code thus very slow and at the end it does not work. Here is my approach I will appreciate your advice

 real, shared :: s_data(-1:16,-1:16)

 d_j = (blockIdx%x-1) * blockDim%x + threadIdx%x-1
 d_l = (blockIdx%y-1) * blockDim%y + threadIdx%y-1

 tIdx = threadIdx%x -1
 tIdy = threadIdx%y -1

  bdimx = 256/(blockDim%x)  !16
  bdimy = 64/(blockDim%y)   !8


d_l1=d_l+1
if(d_l1==d_lmax) d_l1=0

d_l0 = d_l -1
if(d_l==0) d_l0=d_lmax-1
call syncthreads()

!load the main part 
s_data(tIdx,tIdy)   = g_data(d_j,d_l)


!Filling halos 
if(tIdx ==0)then
      f(bx == 0) then
         s_data(tIdx-1,tIdy) =0
     else 
         s_data(tIdx-1,tIdy)   = g_data(d_j-1,d_l)
     end if
end if

!Fill (16,tIdy)
if(tIdx == blockDim%x-1)then
    if(bx == bdmx-1) then
       s_data(tIdx+1,tIdy) = 0
    else
        s_data(tIdx+1,tIdy) = g_data(d_j+1,d_l) 
    end if
end if

!Fill (-1,tIdy)
if(tIdy == 0)then              
     s_data(tIdx,tIdy+1)=g_data(d_j,d_l1)
end if

!Fill (N,tIdy)
if(tIdy == blockDim%y -1)then
    s_data(tIdx,tIdy-1) = g_data(d_j,d_l0) 
end if

!Fill (-1,-1) and (-1, N)
if(tIdx==0)then
    if(bx == 0)then
       if(tIdy == 0) then
          s_data(tIdx-1,tIdy-1) =0 
       end if
       if(tIdy == blockDim%y-1) then
          s_data(tIdx-1,tIdy+1) = 0 
       end if

    else
       if(tIdy == 0) then
           s_data(tIdx-1,tIdy-1) =g_data(d_j-1,d_l0) 
       end if 
       if(tIdy == blockDim%y) then
           s_data(tIdx-1,tIdy+1) = g_data(d_j-1,d_l1)
       end if 
    end if 
end if

!Fill (N, -1) & (N,N)
if(tIdx==blockDim%x-1)then
      if(bx == bdimx-1)then
          if(tIdy == 0) then
            s_data(tIdx+1,tIdy-1) = 0
          end if
          if(tIdy == blockDim%y) then
             s_data(tIdx+1,tIdy+1) = 0
          end if 
       else 
          if(tIdy == 0) then
              s_data(tIdx+1,tIdy-1) =g_data(d_j+1,d_l0) 
          end if
          if(dIdy == blockDim%y) then
              s_data(tIdx+1,tIdy+1) = g_data(d_j+1,d_l1) 
          end if 
end if

!do some computation with s_data

Answer 1

Box filters for image processing always involves halo data. The basic idea is each output element/pixel is processed by one thread, and each thread loads more than one element/pixel to the shared mem.

This white paper about image convolution using CUDA could be a good reference.

http://docs.nvidia.com/cuda/samples/3_Imaging/convolutionSeparable/doc/convolutionSeparable.pdf