nsight-compute - 如何使用 nv-nsight-cu-cli 查找 gld_throughput 和 gst_throughput

Question

无法让它工作，文档阅读起来有点棘手。
在下面尝试并将输出视为 n/a。

root@teja:~/Projs/CUDA/05-Profiling# nv-nsight-cu-cli --device 0 --metrics gst_throughput,gld_throughput ./run 0
==PROF== Connected to process 28170 (/root/Projs/CUDA/05-Profiling/run)
==PROF== Profiling "Init" - 1: 0%....50%....100% - 1 pass
==PROF== Profiling "Transpose_rowRead_colWrite" - 2: 0%....50%....100% - 1 pass
==PROF== Disconnected from process 28170
[28170] run@127.0.0.1
  Init(mat<int>,mat<int>), 2020-May-01 14:35:43, Context 1, Stream 7
    Section: Command line profiler metrics
    ---------------------------------------------------------------------- --------------- ------------------------------
    gld_throughput                                                                                                (!) n/a
    gst_throughput                                                                                                (!) n/a
    ---------------------------------------------------------------------- --------------- ------------------------------

  Transpose_rowRead_colWrite(mat<int>,mat<int>), 2020-May-01 14:35:43, Context 1, Stream 7
    Section: Command line profiler metrics
    ---------------------------------------------------------------------- --------------- ------------------------------
    gld_throughput                                                                                                (!) n/a
    gst_throughput                                                                                                (!) n/a
    ---------------------------------------------------------------------- --------------- ------------------------------

Answer 1

使用 Nsight Compute 工具更改了名称。

此表有助于获得新名称：https ://docs.nvidia.com/nsight-compute/2019.5/NsightComputeCli/index.html#nvprof-metric-comparison

使用以下 cmd 使其工作。

nv-nsight-cu-cli --metrics l1tex__t_bytes_pipe_lsu_mem_global_op_ld.sum.per_second,l1tex__t_sectors_pipe_lsu_mem_global_op_ld.sum,l1tex__t_bytes_pipe_lsu_mem_global_op_st.sum.per_second,l1tex__t_sectors_pipe_lsu_mem_global_op_st.sum, ./<program>