profiler_2elems_hotcache.yml

- kernel:
      name: "gather_scatter"
      path: "src/"
      description: "Setup for experiment gather"
      debug: False
      preamble:
          command: "sudo cpupower frequency-set -u 2.1GHz  > /dev/null && sudo cpupower frequency-set -d 2.1GHz > /dev/null"
      compilation:
          enabled: True
          processes: 32
          language: "C"
          compiler_flags: { "gcc-11": [" -march=native -mtune=native -Ofast "] }
          main_src: "main.c"
          kernel_inlined: True
          asm_analysis:
              syntax: "intel"
              count_ins: True
              static_analysis: True
          debug: False
      configuration:
          kernel_cfg: ["-DGATHER"]
          d_features:
              ELEMS:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[2]"
              OFFSET:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[0]"
              IDX0:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[0]"
              IDX1:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[1,8,16]"
          derived:
              SET_:
                  type: multi
                  variables: ["IDX0", "IDX1"]
                  expression: int(VAR/16) % 16
              CL_:
                  type: multi
                  variables: ["IDX0", "IDX1"]
                  expression: int(VAR / 16)
              N_CL:
                  type: all
                  variables: ["CL_IDX0", "CL_IDX1"]
                  expression: SIZE(UNIQUE(ALL_VAR))
              N_SET:
                  type: all
                  variables: ["SET_IDX0", "SET_IDX1"]
                  expression: SIZE(UNIQUE(ALL_VAR))
          flops: "1"
      execution:
          enabled: True
          papi_counters:
              [
                  "CPU_CLK_UNHALTED",
                  "INST_RETIRED:ANY_P",
                  "UOPS_RETIRED:ALL",
                  "UOPS_ISSUED:ALL",
                  "UOPS_EXECUTED:THREAD",
              ]
          tsc: True
          threshold_outliers: 3
          compute_avg: True
          discard_outliers: True
          nexec: 7
          nsteps: 1000
          cpu_affinity: 11
          prefix: "numactl -p 0 taskset -c 1"
      output:
          name: "2elemshotcache"
          columns: "all"
          report: True
          verbose: True