hardware-counters

硬件性能计数器

用途

指导 Agent 分析硬件性能计数器：使用 perf stat -e 收集 PMU 事件，使用 PAPI 库以可移植的方式访问计数器，解读缓存未命中率和分支预测错误率，计算 IPC，并使用 perf annotate 将事件与源代码行关联起来。

触发条件

• “如何用 perf 测量缓存未命中率？”
• “如何统计分支预测错误？”
• “如何用 perf 计算 IPC（每时钟周期指令数）？”
• “如何使用 PAPI 库访问硬件计数器？”
• “如何查看哪些源代码行导致了最多的缓存未命中？”
• “如何用性能计数器测量内存带宽？”

工作流

1. perf stat — 基本计数器采集

# Basic hardware event summary
perf stat ./prog

# Output:
#  Performance counter stats for './prog':
#
#      1,234,567,890      instructions
#        456,789,012      cycles
#         12,345,678      cache-misses         #    1.23 % of all cache refs
#         23,456,789      branch-misses        #    2.34 % of all branches
#
#       0.456789012 seconds time elapsed

# Derived metrics (computed from the output)
# IPC = instructions / cycles = 1,234,567,890 / 456,789,012 ≈ 2.70
# CPI = cycles / instructions ≈ 0.37

2. 使用 -e 指定 PMU 事件

# Specific hardware events
perf stat -e instructions,cycles,cache-misses,branch-misses ./prog

# L1/L2/L3 cache events
perf stat -e \
  L1-dcache-loads,L1-dcache-load-misses,\
  L2-loads,L2-load-misses,\
  LLC-loads,LLC-load-misses \
  ./prog

# Memory bandwidth (Intel)
perf stat -e \
  uncore_imc/cas_count_read/,\
  uncore_imc/cas_count_write/ \
  ./prog

# TLB misses
perf stat -e dTLB-loads,dTLB-load-misses,iTLB-loads,iTLB-load-misses ./prog

# Branch misprediction rate
perf stat -e branches,branch-misses ./prog
# Rate = branch-misses / branches × 100%

# Available events (varies by CPU)
perf list hardware          # generic hardware events
perf list cache             # cache events
perf list pmu               # raw PMU events for your CPU

3. 关键指标和阈值

指标	公式	良好	需关注
IPC	instructions / cycles	> 2.0（现代 x86）	< 1.0
L1 未命中率	L1-misses / L1-accesses	< 1%	> 5%
LLC 未命中率	LLC-misses / LLC-accesses	< 1%	> 10%
分支未命中率	branch-misses / branches	< 1%	> 5%
MPKI	每 1K 条指令的未命中次数	—	L3 MPKI > 10 = 受内存限制

# Compute MPKI (Misses Per Kilo-Instructions)
perf stat -e instructions,LLC-load-misses ./prog
# MPKI = LLC-load-misses / (instructions / 1000)

4. 原始 PMU 事件（特定 CPU）

对于通用别名中没有的事件，请使用原始事件代码：

# Intel: use perf list or look up in Intel SDM
# Format: rXXYY where XX=umask, YY=event code
perf stat -e r0124 ./prog    # example Intel raw event

# List Intel events with ocperf (OpenCL Perf Events)
pip install ocperf
ocperf.py list | grep "mem_load"

# Use libpfm4 for event names
pfm_ls | grep "MEM_LOAD"
perf stat -e $(pfm_ls | grep "MEM_LOAD_RETIRED.L3_MISS") ./prog

# AMD: similar approach
perf stat -e r04041 ./prog   # AMD raw event

5. 使用 perf record/annotate 进行源代码级注释

# Record with hardware events
perf record -e LLC-load-misses -g ./prog

# Annotate: show source lines sorted by cache miss count
perf annotate --stdio

# Interactive (requires debug symbols)
perf report
# Press 'a' on a function to annotate it

# Combined: record hotspot + annotate
perf record -e cycles:u -g ./prog
perf annotate --symbol=my_function --stdio 2>/dev/null | head -40

# Example annotate output:
# Percent | Source code
#   45.23 |     for (int i = 0; i < N; i++)
#    3.12 |         sum += data[i];   ← cache miss here (strided access)

6. PAPI — 硬件计数器可移植 API

PAPI 提供可用于不同 CPU 架构的可移植 C API：

#include <papi.h>
#include <stdio.h>

int main(void) {
    int Events[] = {PAPI_TOT_INS, PAPI_TOT_CYC,
                    PAPI_L2_TCM,  PAPI_BR_MSP};
    long long values[4];

    if (PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT) {
        fprintf(stderr, "PAPI init failed\n");
        return 1;
    }

    PAPI_start_counters(Events, 4);

    // --- Code to measure ---
    do_work();
    // -----------------------

    PAPI_stop_counters(values, 4);

    printf("Instructions:      %lld\n", values[0]);
    printf("Cycles:            %lld\n", values[1]);
    printf("IPC:               %.2f\n", (double)values[0]/values[1]);
    printf("L2 cache misses:   %lld\n", values[2]);
    printf("Branch mispred:    %lld\n", values[3]);

    return 0;
}

# Build with PAPI
gcc -O2 -g -o prog prog.c -lpapi

# Available PAPI events on your system
papi_avail -a | head -30
papi_native_avail | grep "L3"    # native events with "L3"

常用的 PAPI 预设：

预设	事件
PAPI_TOT_INS	指令总数
PAPI_TOT_CYC	周期总数
PAPI_L1_DCM	L1 数据缓存未命中
PAPI_L2_TCM	L2 缓存总未命中数
PAPI_L3_TCM	L3 缓存总未命中数
PAPI_BR_MSP	分支预测错误
PAPI_TLB_DM	数据 TLB 未命中
PAPI_FP_INS	浮点指令
PAPI_VEC_INS	向量/SIMD 指令

7. Intel PCM（性能计数器监视器）

# Intel PCM — system-wide counters, no root required on modern kernels
git clone https://github.com/intel/pcm
cd pcm && cmake -S . -B build && cmake --build build

# Measure memory bandwidth
./build/bin/pcm-memory 1    # sample every 1 second

# Core utilization + IPC
./build/bin/pcm 1

# Cache miss breakdown per socket
./build/bin/pcm 1 -csv | head -20

相关技能

• 使用 skills/profilers/intel-vtune-amd-uprof 进行引导式微架构分析
• 使用 skills/profilers/linux-perf 执行 perf record/report 并生成火焰图
• 使用 skills/low-level-programming/cpu-cache-opt 应用缓存优化模式
• 使用 skills/low-level-programming/simd-intrinsics 提升 FLOPS/周期指标