3 Replies Latest reply on Sep 22, 2009 12:03 PM by ryta1203

    Block arrangement in Compute Shader

    ryta1203

      Is anyone else able to use anything other than the naive approach (64x1 block size) when using compute shader?

      I'm not able to get it working 100%... I'm still a little confused regarding the terminology (blocks, groups, rows). Can someone explain this?

        • Block arrangement in Compute Shader
          the729

          wavefront = warp in cuda

          group = block in cuda

          LDS shares data within a group, which may contain several wavefronts. However, I guess that fensing LDS will be slow when using multiple WFs in one group. You can see a loop used for synchronization LDS accesses in the disassembly.

          What is the problem you meet?