# rocprof kernel-trace capture of the tiled HIP SGEMM in this directory.
#
# The payload is an opaque command as far as aorta is concerned -- the
# collector attaches by wrapping argv, so nothing here knows or cares that
# the command happens to be a HIP binary.
#
#   hipcc -O3 -o /tmp/hip_gemm examples/profiling/rocprof/hip-gemm/gemm.hip
#   aorta sweep run \
#       --recipe examples/profiling/rocprof/hip-gemm/recipe.yaml \
#       --output ./profiling_results \
#       -- /tmp/hip_gemm 512 20
#
# See README.md in this directory for the artifact layout and the metrics
# that reach perf.md / matrix.json.
schema_version: 1
mode: probe

ticket: PROFILING-HIP-GEMM

# One trial is enough to demonstrate a capture; raise it to compare
# kernel time across repeats.
trials: 1
timeout_per_trial: 600

mitigation_axis: [none]
diagnostic_axis: [none]

env_passthrough_mode: inherit

# Smallest useful rocprof configuration: kernel dispatches only, CSV out,
# per-kernel summary statistics on. `summary_units: msec` keeps the summary
# readable for a kernel in the tens-of-microseconds range.
collect:
  rocprof:
    trace: "kernel"
    output_format: "csv"
    stats: "1"
    summary_units: "msec"