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(),