From 9a09abd78c4d9dd2d005ae63d6cb690883e7b0eb Mon Sep 17 00:00:00 2001
From: twwang <920347125@qq.com>
Date: Wed, 28 May 2025 15:01:46 +0800
Subject: [PATCH] =?UTF-8?q?PMU=5FL3=5FLAT=E7=9A=84=E7=BB=93=E6=9E=9C?=
 =?UTF-8?q?=E4=BD=BF=E7=94=A8ns=E4=BB=A3=E6=9B=BFcycles?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 docs/Details_Usage.md | 14 +++++++-------
 include/pmu.h         |  2 +-
 pmu/pmu_metric.cpp    | 16 +++++++++++++++-
 3 files changed, 23 insertions(+), 9 deletions(-)

diff --git a/docs/Details_Usage.md b/docs/Details_Usage.md
index 289c03c..143ea8a 100644
--- a/docs/Details_Usage.md
+++ b/docs/Details_Usage.md
@@ -710,7 +710,7 @@ auto len = PmuGetDevMetric(oriData, oriLen, devAttr, 1, &devData);
 // devData的长度等于cluster个数
 for (int i=0;i<len;++i) {
     // 每个devData表示一个cluster的L3平均时延，是以cycles为单位
-    cout << "L3 latency(" << devData[i].clusterId << "): " << devData[i].count<< " cycles\n";
+    cout << "L3 latency(" << devData[i].clusterId << "): " << devData[i].count<< " ns\n";
 }
 DevDataFree(devData);
 PmuDataFree(oriData);
@@ -729,7 +729,7 @@ kperf.disable(pd)
 ori_data = kperf.read(pd)
 dev_data = kperf.get_device_metric(ori_data, dev_attr)
 for data in dev_data.iter:
-    print(f"L3 latency({data.clusterId}): {data.count} cycles")
+    print(f"L3 latency({data.clusterId}): {data.count} ns")
 ```
 
 ```go
@@ -742,7 +742,7 @@ kperf.PmuDisable(fd)
 dataVo, _ := kperf.PmuRead(fd)
 deivceDataVo, _ := kperf.PmuGetDevMetric(dataVo, deviceAttrs)
 for _, v := range deivceDataVo.GoDeviceData {
-	fmt.Printf("L3 latency(%v): %v cycles\n", v.ClusterId, v.Count)
+	fmt.Printf("L3 latency(%v): %v ns\n", v.ClusterId, v.Count)
 }
 kperf.DevDataFree(deivceDataVo)
 kperf.PmuDataFree(dataVo)
@@ -751,10 +751,10 @@ kperf.PmuClose(fd)
 
 执行上述代码，输出的结果类似如下：
 ```
-L3 latency(0): 101 cycles
-L3 latency(1): 334.6 cycles
-L3 latency(2): 267.8 cycles
-L3 latency(3): 198.4 cycles
+L3 latency(0): 101 ns
+L3 latency(1): 334.6 ns
+L3 latency(2): 267.8 ns
+L3 latency(3): 198.4 ns
 ...
 ```
 
diff --git a/include/pmu.h b/include/pmu.h
index af9bb2a..c8bef10 100644
--- a/include/pmu.h
+++ b/include/pmu.h
@@ -426,7 +426,7 @@ enum PmuDeviceMetric {
     PMU_L3_REF,
     // Percluster metric.
     // Collect L3 total latency for each cluster node.
-    // Unit: cycles
+    // Unit: ns 
     PMU_L3_LAT,
     // Collect pcie rx bandwidth.
     // Perpcie metric.
diff --git a/pmu/pmu_metric.cpp b/pmu/pmu_metric.cpp
index a2163a1..980632b 100644
--- a/pmu/pmu_metric.cpp
+++ b/pmu/pmu_metric.cpp
@@ -291,6 +291,8 @@ namespace KUNPENG_PMU {
         METRIC_CONFIG::SMMU_TRAN,
     };
 
+    static const map<CHIP_TYPE, double> L3_CLOCK_NS {{CHIP_TYPE::HIPB, 0.3448275862}};
+
     const UNCORE_METRIC_MAP UNCORE_METRIC_CONFIG_MAP = {
         {CHIP_TYPE::HIPA, HIP_A_UNCORE_METRIC_MAP},
         {CHIP_TYPE::HIPB, HIP_B_UNCORE_METRIC_MAP},
@@ -931,6 +933,17 @@ namespace KUNPENG_PMU {
         return 64 * rawCount;
     }
 
+    static uint64_t L3Lat(const uint64_t rawCount)
+    {
+        const CHIP_TYPE chipType = GetCpuType();
+        auto iter = L3_CLOCK_NS.find(chipType);
+        uint64_t count = rawCount;
+        if (iter != L3_CLOCK_NS.end()) {
+            count = rawCount * iter->second;
+        }
+        return count;
+    }
+
     static PmuMetricMode GetMetricMode(const PmuDeviceMetric &metric)
     {
         switch(metric) {
@@ -1144,7 +1157,8 @@ namespace KUNPENG_PMU {
 
     map<PmuDeviceMetric, ComputeMetricCb> computeMetricMap = {{PMU_DDR_READ_BW, DDRBw},
                                                                 {PMU_DDR_WRITE_BW, DDRBw},
-                                                                {PMU_L3_TRAFFIC, L3Bw}};
+                                                                {PMU_L3_TRAFFIC, L3Bw},
+                                                                {PMU_L3_LAT, L3Lat}};
     map<PmuDeviceMetric, AggregateMetricCb> aggregateMap = {
         {PMU_DDR_READ_BW, AggregateByNuma},
         {PMU_DDR_WRITE_BW, AggregateByNuma},
-- 
Gitee