defaults: rocprofv3: # keys such as "runtime", "kernel", etc. below can be used with --filter-rocprofv3 option runtime: [--runtime-trace] sys: [--sys-trace] kernel: [--kernel-trace] hip: [--hip-trace] hsa: [--hsa-trace] memcpy: [--memory-copy-trace] malloc: [--memory-allocation-trace] scratch: [--scratch-memory-trace] counters: [--pmc, SQ_WAVES] benchmark: - baseline - disabled-sdk-contexts - sdk-buffer-overhead - sdk-callback-overhead - tool-runtime-overhead group: - examples environment: ROCR_VISIBLE_DEVICES: "0,2" HIP_VISIBLE_DEVICES: "0,2" jobs: # explicitly specifies name, group, and command. inherits default rocprofv3 commands and benchmark modes - name: hip-in-libraries group: [examples, multigpu, multistream] command: [hip-in-libraries] # explicitly specifies name, group, command, and rocprofv3. inherits default benchmark modes - name: transpose group: [examples, multithreaded, multigpu, multistream] command: [transpose, 4, 500, 10] launcher: [mpirun, -n, 1] rocprofv3: - [-r, --pmc, SQ_WAVES] - [--kernel-trace] - [--hip-trace] # explicitly specifies name, command, rocprofv3, and benchmark. inherits default group(s) - name: hip-graph command: [hip-graph, 8, 500] rocprofv3: - [--kernel-trace] - [--hip-trace] - [--hsa-trace] benchmark: - baseline - disabled-sdk-contexts