矩阵扩边不正确

加速计算 CUDA
1

要调用一个内核函数实现矩阵的扩边，即把矩阵A(nx,nz)扩充成B(nx+2npad,nz+2npad)，B中间部分是A的值，扩边部分的值等于相邻A的边上的值，程序实现时单纯复制A的部分没有问题，扩边部分出了问题，如图

代码如下

__global__ void pad_float_2d(float *out, unsigned long pitch_out, float *in, unsigned long pitch_in,int xrange, int zrange, int npadded ){

   int ix=blockIdx.y*blockDim.y+threadIdx.y;
   int jz=blockIdx.x*blockDim.x+threadIdx.x;

   if(ix>=npadded && ix<xrange-npadded && jz>=npadded && jz<zrange-npadded )
   {
   out[ix*pitch_out+jz]=in[(ix-npadded)*pitch_in+(jz-npadded)];
   }
   
   __syncthreads();
   
   if(ix<npadded && jz>=npadded && jz<zrange-npadded )
   { 
   out[ix*pitch_out+jz]=out[npadded*pitch_out+jz];
   out[(xrange-ix-1)*pitch_out+jz]=out[(xrange-npadded-1)*pitch_out+jz];
   }
   
   __syncthreads();    
   
   if(ix<xrange && jz<npadded )
   {
   out[ix*pitch_out+jz]=out[ix*pitch_out+npadded];
   out[ix*pitch_out+(zrange-jz-1)]=out[ix*pitch_out+(zrange-npadded-1)];
   }  
   
   return;        
}

[ 本帖最后由 xiaomasddp 于 2010-7-16 20:27 编辑 ]