# SuperBench summary rules for Hygon BW configurations
#
# Usage:
#   sb result summary \
#     --data-file <output-dir>/results-summary.jsonl \
#     --rule-file superbench/config/hygon_bw_summary.yaml \
#     --output-file-format md \
#     --output-dir <summary-output-dir>
#
# Notes:
# - This file covers common RCCL and GPU HPCG benchmark metrics used by
#   Hygon BW configuration files.
# - Unmatched benchmark sections are allowed. If a benchmark was not run,
#   the corresponding category in the summary will be empty.

version: v0.12
superbench:
  rules:
    rccl_bw_allreduce_r16_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r16
      metrics:
        - rccl-bw:allreduce-r16/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r16/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r8_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r8
      metrics:
        - rccl-bw:allreduce-r8/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r8/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r8_pcie_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r8-pcie
      metrics:
        - rccl-bw:allreduce-r8-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r8-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r4_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r4
      metrics:
        - rccl-bw:allreduce-r4/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r4/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r4_graph_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r4-graph
      metrics:
        - rccl-bw:allreduce-r4-graph/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r4-graph/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r4_pcie_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r4-pcie
      metrics:
        - rccl-bw:allreduce-r4-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r4-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_allreduce_r4_graph_pcie_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:allreduce-r4-graph-pcie
      metrics:
        - rccl-bw:allreduce-r4-graph-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:allreduce-r4-graph-pcie/allreduce_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_alltoall_r8_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:alltoall-r8
      metrics:
        - rccl-bw:alltoall-r8/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:alltoall-r8/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_alltoall_r16_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:alltoall-r16
      metrics:
        - rccl-bw:alltoall-r16/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:alltoall-r16/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    rccl_bw_alltoall_r4_bandwidth:
      statistics: mean
      categories: RCCL rccl-bw:alltoall-r4
      metrics:
        - rccl-bw:alltoall-r4/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_busbw
        - rccl-bw:alltoall-r4/alltoall_(8388608|67108864|1073741824|4294967296|8589934592|17179869184)_algbw

    gpu_hpcg_r1:
      statistics: mean
      categories: HPCG gpu-hpcg:r1
      metrics:
        - gpu-hpcg:r1/(setup_time|optimization_time|total_time)_p1x1x1_n560x280x280
        - gpu-hpcg:r1/(ddot|waxpby|spmv|mg|total|final)_p1x1x1_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)

    gpu_hpcg_r2:
      statistics: mean
      categories: HPCG gpu-hpcg:r2
      metrics:
        - gpu-hpcg:r2/(setup_time|optimization_time|total_time)_p2x1x1_n560x280x280
        - gpu-hpcg:r2/(ddot|waxpby|spmv|mg|total|final)_p2x1x1_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)

    gpu_hpcg_r4:
      statistics: mean
      categories: HPCG gpu-hpcg:r4
      metrics:
        - gpu-hpcg:r4/(setup_time|optimization_time|total_time)_p2x2x1_n560x280x280
        - gpu-hpcg:r4/(ddot|waxpby|spmv|mg|total|final)_p2x2x1_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)

    gpu_hpcg_r8:
      statistics: mean
      categories: HPCG gpu-hpcg:r8
      metrics:
        - gpu-hpcg:r8/(setup_time|optimization_time|total_time)_p2x2x2_n560x280x280
        - gpu-hpcg:r8/(ddot|waxpby|spmv|mg|total|final)_p2x2x2_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)

    gpu_hpcg_r16:
      statistics: mean
      categories: HPCG gpu-hpcg:r16
      metrics:
        - gpu-hpcg:r16/(setup_time|optimization_time|total_time)_p4x2x2_n560x280x280
        - gpu-hpcg:r16/(ddot|waxpby|spmv|mg|total|final)_p4x2x2_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)

    gpu_hpcg_r32:
      statistics: mean
      categories: HPCG gpu-hpcg:r32
      metrics:
        - gpu-hpcg:r32/(setup_time|optimization_time|total_time)_p4x4x2_n560x280x280
        - gpu-hpcg:r32/(ddot|waxpby|spmv|mg|total|final)_p4x4x2_n560x280x280_(flops_per_process|bandwidth_per_process|flops|bandwidth)