profiler_6elems_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: "[6]"
              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]"
              IDX2:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[2,9,32]"
              IDX3:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[3,10,48]"
              IDX4:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[4,11,64]"
              IDX5:
                  evaluate: True
                  type: "static"
                  val_type: "numeric"
                  value: "[5,12,80]"
          derived:
              SET_:
                  type: multi
                  variables: ["IDX0", "IDX1", "IDX2", "IDX3", "IDX4", "IDX5"]
                  expression: int(VAR/16) % 16
              CL_:
                  type: multi
                  variables: ["IDX0", "IDX1", "IDX2", "IDX3", "IDX4", "IDX5"]
                  expression: int(VAR / 16)
              N_CL:
                  type: all
                  variables:
                      [
                          "CL_IDX0",
                          "CL_IDX1",
                          "CL_IDX2",
                          "CL_IDX3",
                          "CL_IDX4",
                          "CL_IDX5",
                      ]
                  expression: SIZE(UNIQUE(ALL_VAR))
              N_SET:
                  type: all
                  variables:
                      [
                          "SET_IDX0",
                          "SET_IDX1",
                          "SET_IDX2",
                          "SET_IDX3",
                          "SET_IDX4",
                          "SET_IDX5",
                      ]
                  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",
              ]
          time: False
          tsc: True
          threshold_outliers: 3
          discard_outliers: True
          compute_avg: True
          nexec: 7
          nsteps: 1000
          cpu_affinity: 11
          prefix: "numactl -p 0 taskset -c 1"
      output:
          name: "6elemshotcache"
          columns: "all"
          report: True
          verbose: True