Re: [PATCH v8 19/52] perf jevents: Add software prefetch (swpf) metric group for AMD

[Sandipan Das <sandipan.das@amd.com>]

On 11/13/2025 8:50 AM, Ian Rogers wrote:
> Add metrics that give the utility of software prefetches on zen2, zen3
> and zen4.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
>  tools/perf/pmu-events/amd_metrics.py | 101 +++++++++++++++++++++++++++
>  1 file changed, 101 insertions(+)
> 
> diff --git a/tools/perf/pmu-events/amd_metrics.py b/tools/perf/pmu-events/amd_metrics.py
> index 1880ccf9c6fc..06cb56cbd617 100755
> --- a/tools/perf/pmu-events/amd_metrics.py
> +++ b/tools/perf/pmu-events/amd_metrics.py
> @@ -121,6 +121,106 @@ def AmdBr():
>                         description="breakdown of retired branch instructions")
>  
>  
> +def AmdSwpf() -> Optional[MetricGroup]:
> +    """Returns a MetricGroup representing AMD software prefetch metrics."""
> +    global _zen_model
> +    if _zen_model <= 1:
> +        return None
> +

Zen 1 supports PMCx059, used by ls_sw_pf_dc_fills.*, but it looks like the event
definitions are missing from the "amdzen1" JSON files. I can send a fix to address
that unless there is an erratum.

> +    swp_ld = Event("ls_dispatch.ld_dispatch")
> +    swp_t0 = Event("ls_pref_instr_disp.prefetch")
> +    swp_w = Event("ls_pref_instr_disp.prefetch_w")  # Missing on Zen1
> +    swp_nt = Event("ls_pref_instr_disp.prefetch_nta")
> +    swp_mab = Event("ls_inef_sw_pref.mab_mch_cnt")
> +    swp_l2 = Event("ls_sw_pf_dc_fills.local_l2",
> +                   "ls_sw_pf_dc_fills.lcl_l2",
> +                   "ls_sw_pf_dc_fill.ls_mabresp_lcl_l2")
> +    swp_lc = Event("ls_sw_pf_dc_fills.local_ccx",
> +                   "ls_sw_pf_dc_fills.int_cache",
> +                   "ls_sw_pf_dc_fill.ls_mabresp_lcl_cache")
> +    swp_lm = Event("ls_sw_pf_dc_fills.dram_io_near",
> +                   "ls_sw_pf_dc_fills.mem_io_local",
> +                   "ls_sw_pf_dc_fill.ls_mabresp_lcl_dram")
> +    swp_rc = Event("ls_sw_pf_dc_fills.far_cache",
> +                   "ls_sw_pf_dc_fills.ext_cache_remote",
> +                   "ls_sw_pf_dc_fill.ls_mabresp_rmt_cache")
> +    swp_rm = Event("ls_sw_pf_dc_fills.dram_io_far",
> +                   "ls_sw_pf_dc_fills.mem_io_remote",
> +                   "ls_sw_pf_dc_fill.ls_mabresp_rmt_dram")
> +

While "local_ccx" and "int_cache" are equivalent, "ls_mabresp_lcl_cache" is
slightly different.

The "ls_mabresp_lcl_cache" variant applies to Zen 2 and older architectures
and it counts fills from any cache beyond the local L2 but within the same
NUMA node.

Zen 3 onwards, this was further broken down in a way that the "int_cache"
variant counts fills from the local L3 or another L2 in the same CCX while
the "ext_cache_local" variant counts fills from another CCX in the same NUMA
node. So, "ls_mabresp_lcl_cache" is effectively the sum of "int_cache" and
"ext_cache_local".

Zen 4 onwards, "int_cache" and "ext_cache_local" were renamed to "local_ccx"
and "near_cache" respectively. A new "alternate_memories" variant was also
introduced.

> +    # All the swpf that were satisfied beyond L1D are good.
> +    all_pf = swp_t0 + swp_w + swp_nt
> +    good_pf = swp_l2 + swp_lc + swp_lm + swp_rc + swp_rm
> +    bad_pf = max(all_pf - good_pf, 0)
> +

Zen 4 onwards, "ls_sw_pf_dc_fills.all" is a single event that covers all
fills under good_pf. I can add that variant for the older architectures if
it helps with simplification.

> +    loc_pf = swp_l2 + swp_lc + swp_lm
> +    rem_pf = swp_rc + swp_rm
> +

Zen 5 has the "local_all" and "far_all" variants to cover these. Perhaps
these can be extended to older architectures too.

> +    req_pend = max(0, bad_pf - swp_mab)
> +
> +    r1 = d_ratio(ins, all_pf)
> +    r2 = d_ratio(swp_ld, all_pf)
> +    r3 = d_ratio(swp_t0, interval_sec)
> +    r4 = d_ratio(swp_w, interval_sec)
> +    r5 = d_ratio(swp_nt, interval_sec)
> +    overview = MetricGroup("lpm_swpf_overview", [
> +        Metric("lpm_swpf_ov_insn_bt_swpf", "Insn between SWPF", r1, "insns"),
> +        Metric("lpm_swpf_ov_loads_bt_swpf", "Loads between SWPF", r2, "loads"),
> +        Metric("lpm_swpf_ov_rate_prefetch_t0_t1_t2", "Rate prefetch TO_T1_T2", r3,
> +               "insns/sec"),
> +        Metric("lpm_swpf_ov_rate_prefetch_w",
> +               "Rate prefetch W", r4, "insns/sec"),
> +        Metric("lpm_swpf_ov_rate_preftech_nta",
> +               "Rate prefetch NTA", r5, "insns/sec"),
> +    ])
> +
> +    r1 = d_ratio(swp_mab, all_pf)
> +    r2 = d_ratio(req_pend, all_pf)
> +    usefulness_bad = MetricGroup("lpm_swpf_usefulness_bad", [
> +        Metric("lpm_swpf_use_bad_hit_l1", "Usefulness bad hit L1", r1, "100%"),
> +        Metric("lpm_swpf_use_bad_req_pend",
> +               "Usefulness bad req pending", r2, "100%"),
> +    ])
> +
> +    r1 = d_ratio(good_pf, all_pf)
> +    usefulness_good = MetricGroup("lpm_swpf_usefulness_good", [
> +        Metric("lpm_swpf_use_good_other_src", "Usefulness good other src", r1,
> +               "100%"),
> +    ])
> +
> +    usefulness = MetricGroup("lpm_swpf_usefulness", [
> +        usefulness_bad,
> +        usefulness_good,
> +    ])
> +
> +    r1 = d_ratio(swp_l2, good_pf)
> +    r2 = d_ratio(swp_lc, good_pf)
> +    r3 = d_ratio(swp_lm, good_pf)
> +    data_src_local = MetricGroup("lpm_swpf_data_src_local", [
> +        Metric("lpm_swpf_data_src_local_l2",
> +               "Data source local l2", r1, "100%"),
> +        Metric("lpm_swpf_data_src_local_ccx_l3_loc_ccx",
> +               "Data source local ccx l3 loc ccx", r2, "100%"),
> +        Metric("lpm_swpf_data_src_local_memory_or_io",
> +               "Data source local memory or IO", r3, "100%"),
> +    ])
> +
> +    r1 = d_ratio(swp_rc, good_pf)
> +    r2 = d_ratio(swp_rm, good_pf)
> +    data_src_remote = MetricGroup("lpm_swpf_data_src_remote", [
> +        Metric("lpm_swpf_data_src_remote_cache", "Data source remote cache", r1,
> +               "100%"),
> +        Metric("lpm_swpf_data_src_remote_memory_or_io",
> +               "Data source remote memory or IO", r2, "100%"),
> +    ])
> +
> +    data_src = MetricGroup("lpm_swpf_data_src", [
> +                           data_src_local, data_src_remote])
> +
> +    return MetricGroup("lpm_swpf", [overview, usefulness, data_src],
> +                       description="Software prefetch breakdown (CCX L3 = L3 of current thread, Loc CCX = CCX cache on some socket)")
> +
> +
>  def AmdUpc() -> Metric:
>      ops = Event("ex_ret_ops", "ex_ret_cops")
>      upc = d_ratio(ops, smt_cycles)
> @@ -187,6 +287,7 @@ def main() -> None:
>  
>      all_metrics = MetricGroup("", [
>          AmdBr(),
> +        AmdSwpf(),
>          AmdUpc(),
>          Idle(),
>          Rapl(),