On 11/13/2025 8:50 AM, Ian Rogers wrote:
> Add metrics that give the utility of hardware prefetches on zen2, zen3
> and zen4.
>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
> tools/perf/pmu-events/amd_metrics.py | 64 ++++++++++++++++++++++++++++
> 1 file changed, 64 insertions(+)
>
> diff --git a/tools/perf/pmu-events/amd_metrics.py b/tools/perf/pmu-events/amd_metrics.py
> index 06cb56cbd617..be27a4947e45 100755
> --- a/tools/perf/pmu-events/amd_metrics.py
> +++ b/tools/perf/pmu-events/amd_metrics.py
> @@ -121,6 +121,69 @@ def AmdBr():
> description="breakdown of retired branch instructions")
>
>
> +def AmdHwpf():
> + """Returns a MetricGroup representing AMD hardware prefetch metrics."""
> + global _zen_model
> + if _zen_model <= 1:
> + return None
> +
> + hwp_ld = Event("ls_dispatch.ld_dispatch")
> + hwp_l2 = Event("ls_hw_pf_dc_fills.local_l2",
> + "ls_hw_pf_dc_fills.lcl_l2",
> + "ls_hw_pf_dc_fill.ls_mabresp_lcl_l2")
> + hwp_lc = Event("ls_hw_pf_dc_fills.local_ccx",
> + "ls_hw_pf_dc_fills.int_cache",
> + "ls_hw_pf_dc_fill.ls_mabresp_lcl_cache")
> + hwp_lm = Event("ls_hw_pf_dc_fills.dram_io_near",
> + "ls_hw_pf_dc_fills.mem_io_local",
> + "ls_hw_pf_dc_fill.ls_mabresp_lcl_dram")
> + hwp_rc = Event("ls_hw_pf_dc_fills.far_cache",
> + "ls_hw_pf_dc_fills.ext_cache_remote",
> + "ls_hw_pf_dc_fill.ls_mabresp_rmt_cache")
> + hwp_rm = Event("ls_hw_pf_dc_fills.dram_io_far",
> + "ls_hw_pf_dc_fills.mem_io_remote",
> + "ls_hw_pf_dc_fill.ls_mabresp_rmt_dram")
> +
Same notes from the previous SWPF patch are applicable here.
> + loc_pf = hwp_l2 + hwp_lc + hwp_lm
> + rem_pf = hwp_rc + hwp_rm
> + all_pf = loc_pf + rem_pf
> +
> + r1 = d_ratio(ins, all_pf)
> + r2 = d_ratio(hwp_ld, all_pf)
> + r3 = d_ratio(all_pf, interval_sec)
> +
> + overview = MetricGroup("lpm_hwpf_overview", [
> + Metric("lpm_hwpf_ov_insn_bt_hwpf", "Insn between HWPF", r1, "insns"),
> + Metric("lpm_hwpf_ov_loads_bt_hwpf", "Loads between HWPF", r2, "loads"),
> + Metric("lpm_hwpf_ov_rate", "HWPF per second", r3, "hwpf/s"),
> + ])
> + r1 = d_ratio(hwp_l2, all_pf)
> + r2 = d_ratio(hwp_lc, all_pf)
> + r3 = d_ratio(hwp_lm, all_pf)
> + data_src_local = MetricGroup("lpm_hwpf_data_src_local", [
> + Metric("lpm_hwpf_data_src_local_l2",
> + "Data source local l2", r1, "100%"),
> + Metric("lpm_hwpf_data_src_local_ccx_l3_loc_ccx",
> + "Data source local ccx l3 loc ccx", r2, "100%"),
> + Metric("lpm_hwpf_data_src_local_memory_or_io",
> + "Data source local memory or IO", r3, "100%"),
> + ])
> +
> + r1 = d_ratio(hwp_rc, all_pf)
> + r2 = d_ratio(hwp_rm, all_pf)
> + data_src_remote = MetricGroup("lpm_hwpf_data_src_remote", [
> + Metric("lpm_hwpf_data_src_remote_cache", "Data source remote cache", r1,
> + "100%"),
> + Metric("lpm_hwpf_data_src_remote_memory_or_io",
> + "Data source remote memory or IO", r2, "100%"),
> + ])
> +
> + data_src = MetricGroup("lpm_hwpf_data_src", [
> + data_src_local, data_src_remote])
> + return MetricGroup("lpm_hwpf", [overview, data_src],
> + description="Hardware prefetch breakdown (CCX L3 = L3 of current thread, Loc CCX = CCX cache on some socket)")
> +
> +
> def AmdSwpf() -> Optional[MetricGroup]:
> """Returns a MetricGroup representing AMD software prefetch metrics."""
> global _zen_model
> @@ -287,6 +350,7 @@ def main() -> None:
>
> all_metrics = MetricGroup("", [
> AmdBr(),
> + AmdHwpf(),
> AmdSwpf(),
> AmdUpc(),
> Idle(),