Notes CUDA

Notes for CUDA

1. Setting grid and block dimensions
   • Un-initialized dimension (.x, .y, .z) is set to 1 by default.

      dim3 grid (1024, 8);
      dim3 block (512, 1);
      myKernel<<<grid, block>>>(d_data);
     

2. Don’t do this:

    dim3 grid = (1024, 8); // !!!! Wrongly initialised values!!!!
   

3. Max of threads in a cuda block: 1024
   • This is the total number of threads, not per dimension!
4. GPU -> many Streaming Multiprocessors (SMs) -> each with a control, Memory and many Cores
   • Each ThreadBlock is assigned (fully) to one of the SMs i.e. all the threads in a ThreadBlock go to the same SM
   • Threads in the same ThreadBlock can collaborate with each other:
     • Barrier synchronization: __syncthreads()
     • Shared memory
   • Threads in different blocks don’t synchronize.
     • Hence, blocks can execute in any order say, sequentially or parallely w.r.t each other
     • Hence, same code can run on a different hardware with few or more SMs accordingly.