rocprof

rocprof 的使用

• 申明一个input.txt文件, 大致如下[1]:

# 度量(Metrics)相关参数
pmc: Wavefronts VALUInsts SALUInsts SFetchInsts
# Counters相关参数
pmc : TCC_HIT[0], TCC_MISS[0]
# 支持的范围格式: "3:9", "3:", "3"
range: 1 : 4
# profiler 的 GPU
gpu: 0 1 2 3
# 具体的 kernel 函数名，删除该行可以实现对所有 kernel 进行 profile
kernel: simple Pass1 simpleConvolutionPass2

• 运行

rocprof -i input.txt ./a.out

• 产生input.csv文件数据
• 查看参数

  • 默认参数
    Index - kernels dispatch order index
    KernelName - the dispatched kernel name
    gpu-id - GPU id the kernel was submitted to
    queue-id - the ROCm queue unique id the kernel was submitted to
    queue-index - The ROCm queue write index for the submitted AQL packet
    tid - system application thread id which submitted the kernel
    grd - the kernel’s grid size
    wgr - the kernel’s work group size
    lds - the kernel’s LDS memory size
    scr - the kernel’s scratch memory size
    vgpr - the kernel’s VGPR size
    sgpr - the kernel’s SGPR size
    fbar - the kernel’s barriers limitation
    sig - the kernel’s completion signal

  • rocprof --list-basic

  • rocprof --list-derived

    • ALUStalledByLDS: ALU因为LDS写入读取暂停的百分比
    • LDSBankConflict: LDS bank conflict 的百分比
    • L2CacheHit: L2缓存命中率
    • SALUBusy
    • VALUBusy
    • VALUUtilization
    • GPUBusy

可以参考的资源

[1] https://zhuanlan.zhihu.com/p/545296023