Post

Arm DDT with GPUs

Debugging GPU Codes with Arm DDT

See the Arm Forge User Guide for all the info, this is just a place for some notes and tips

  • Use -g -O0 to compile host code and -g -G -cudart shared to compile device code. The -g -G enable debugging symbols and -cudart shared enable device memory debugging.
  • A device pointer’s contents can be accessed by preappending @global to its type. A permanant version of this can be done with expressions.
    • ((@global TYPE *)(VARIABLE_NAME) to get the proper pointer
    • ((@global TYPE *)(VARIABLE_NAME)[IDX] to get the value at IDX
    • ((@global TYPE *)(VARIABLE_NAME)[IDX]@N to get N values starting at IDX
  • The expressions panel: can also be used for any other debugging or math expression you want
  • Array Viewer: Any expression can go in the brackets and be displayed in 2D. The correct indexing scheme for Cholla is xid + yid*nx + zid*nx*ny + field*n_cells.

Arm MAP

  • map --profile --cuda-kernel-analysis --cuda-transfer-analysis mpirun -n 4 EXECUTABLE ARGS
    • --profile: Standard flag for profiling
    • --cuda-kernel-analysis: Enables CUDA kernel profiling
    • --cuda-transfer-analysis: Enables CUDA memory transfer profiling
    • The CUDA args require -lineinfo when compiling
    • The CUDA args have a HUGE performance impact on host code so host and device code should be profiled separately. The NVIDIA GPU metrics will be adversely affected by this overhead, particularly the GPU utilization metric
  • View with map —connect MAP_FILE.map or download and open with Arm MAP
  • Can make reports from .map files with perf-report map-file.map

Arm Performance Reports

perf-report mpirun -n 4 EXECUTABLE ARGS or perf-report map-file.map. Generates a nice HTML and text summary of the profiling.

This post is licensed under CC BY 4.0 by the author.