highlight subtleties of CUDA code

[tkphd]

• CUDA code looks weird due to thread allocation model
  • Traditional: many memory addresses per core -> looping constructs
  • Accelerator: one memory address per thread -> no loops
• cached mask
• input vs. output tiles, indexing
• linear vs. dimensional array access
• block size and domain boundaries
• domain-boundary cells (ghosts) vs. tile-boundary cells (halos)
• tile size, block size, grid size; local, source, and destination arrays
  • tile sizes are static: each block is the same, ignorant of domain boundaries
  • meaning of cuda_kernel<<<num_blocks, threads_per_block, shared_array_size>>> construct
  • block size, misfit, and GPU utilization
• take care withint, ceil(), and floor()
• troubleshooting
  • wrss==0.0029?
  • Segfault? printf array locations and sizes; recompile with -g then use cuda-memtest and backtrace
  • CUDA slower than OpenCL? Enable persistence mode.
• Makefile flags and function locations: main() should be in a .c file, CUDA functions in .cu, and objects built with -dc flags