使用nsight-compute分析核函数性能

在使用nvcc编译获得可执行文件后 即可使用nsight-compute完成性能分析

  • 使用nvidia- container构建的镜像已安装好nsight-compute-cli

  • 也可安装图形化界面

使用实例

基本使用

1
$ ncu ./my_cuda_speedup_solutions 512 6

其中my_cuda_speedup_solutions为可执行文件名 512 6 为参数

获取结果如下：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
v6_kernel(args, float *, float *, float *), 2025-Jan-06 05:26:00, Context 1, Stream 7
    Section: GPU Speed Of Light Throughput
    ---------------------------------------------------------------------- --------------- ------------------------------
    DRAM Frequency                                                           cycle/nsecond                           3.99
    SM Frequency                                                             cycle/nsecond                           1.41
    Elapsed Cycles                                                                   cycle                        1661720
    Memory [%]                                                                           %                          73.71
    DRAM Throughput                                                                      %                          17.43
    Duration                                                                       msecond                           1.18
    L1/TEX Cache Throughput                                                              %                          77.41
    L2 Cache Throughput                                                                  %                          13.22
    SM Active Cycles                                                                 cycle                     1666873.38
    Compute (SM) [%]                                                                     %                          73.71
    ---------------------------------------------------------------------- --------------- ------------------------------
    INF   Compute and Memory are well-balanced: To reduce runtime, both computation and memory traffic must be reduced. 
          Check both the Compute Workload Analysis and Memory Workload Analysis sections.                               

    Section: Launch Statistics
    ---------------------------------------------------------------------- --------------- ------------------------------
    Block Size                                                                                                       1024
    Function Cache Configuration                                                                  cudaFuncCachePreferNone
    Grid Size                                                                                                         256
    Registers Per Thread                                                   register/thread                             38
    Shared Memory Configuration Size                                                 Kbyte                          32.77
    Driver Shared Memory Per Block                                              byte/block                              0
    Dynamic Shared Memory Per Block                                            Kbyte/block                          16.64
    Static Shared Memory Per Block                                              byte/block                              0
    Threads                                                                         thread                         262144
    Waves Per SM                                                                                                       16
    ---------------------------------------------------------------------- --------------- ------------------------------

    Section: Occupancy
    ---------------------------------------------------------------------- --------------- ------------------------------
    Block Limit SM                                                                   block                             16
    Block Limit Registers                                                            block                              1
    Block Limit Shared Mem                                                           block                              1
    Block Limit Warps                                                                block                              1
    Theoretical Active Warps per SM                                                   warp                             32
    Theoretical Occupancy                                                                %                            100
    Achieved Occupancy                                                                   %                          99.93
    Achieved Active Warps Per SM                                                      warp                          31.98
    ---------------------------------------------------------------------- --------------- ------------------------------
    INF   This kernel's theoretical occupancy is not impacted by any block limit.

查看更多信息

1
nv-nsight-cu-cli --list-metrics | awk '{print $1}' | tail -n +2 | paste -sd "," - | xargs -I {} echo nv-nsight-cu-cli --metrics={} ./my_cuda_speedup_solutions 512 6

使用上述命令获取可查询数据列表

1
nv-nsight-cu-cli --metrics=-sm__warps_active.avg.per_cycle_active,sm__warps_active.avg.pct_of_peak_sustained_active,-sm__warps_active.avg.pct_of_peak_sustained_active,sm__throughput.avg.pct_of_peak_sustained_elapsed,-sm__throughput.avg.pct_of_peak_sustained_elapsed,sm__maximum_warps_per_active_cycle_pct,-sm__maximum_warps_per_active_cycle_pct,sm__maximum_warps_avg_per_active_cycle,-sm__maximum_warps_avg_per_active_cycle,sm__cycles_active.avg,-sm__cycles_active.avg,lts__throughput.avg.pct_of_peak_sustained_elapsed,-lts__throughput.avg.pct_of_peak_sustained_elapsed,launch__waves_per_multiprocessor,-launch__waves_per_multiprocessor,launch__thread_count,-launch__thread_count,launch__shared_mem_per_block_static,-launch__shared_mem_per_block_static,launch__shared_mem_per_block_dynamic,-launch__shared_mem_per_block_dynamic,launch__shared_mem_per_block_driver,-launch__shared_mem_per_block_driver,launch__shared_mem_per_block,-launch__shared_mem_per_block,launch__shared_mem_config_size,-launch__shared_mem_config_size,launch__registers_per_thread,-launch__registers_per_thread,launch__occupancy_per_shared_mem_size,-launch__occupancy_per_shared_mem_size,launch__occupancy_per_register_count,-launch__occupancy_per_register_count,launch__occupancy_per_cluster_size,-arch:90:90:launch__occupancy_per_cluster_size,launch__occupancy_per_block_size,-launch__occupancy_per_block_size,launch__occupancy_limit_warps,-launch__occupancy_limit_warps,launch__occupancy_limit_shared_mem,-launch__occupancy_limit_shared_mem,launch__occupancy_limit_registers,-launch__occupancy_limit_registers,launch__occupancy_limit_blocks,-launch__occupancy_limit_blocks,launch__occupancy_cluster_pct,-arch:90:90:launch__occupancy_cluster_pct,launch__occupancy_cluster_gpu_pct,-arch:90:90:launch__occupancy_cluster_gpu_pct,launch__grid_size,-launch__grid_size,launch__func_cache_config,-launch__func_cache_config,launch__cluster_size,-arch:90:90:launch__cluster_size,launch__cluster_scheduling_policy,-arch:90:90:launch__cluster_scheduling_policy,launch__cluster_max_potential_size,-arch:90:90:launch__cluster_max_potential_size,launch__cluster_max_active,-arch:90:90:launch__cluster_max_active,launch__block_size,-launch__block_size,l1tex__throughput.avg.pct_of_peak_sustained_active,-l1tex__throughput.avg.pct_of_peak_sustained_active,gpu__time_duration.sum,-gpu__time_duration.sum,gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed,-arch:89:90:gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed,-arch:75:86:gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed,-arch:40:70:gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed,gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed,-gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed,gpc__cycles_elapsed.max,-gpc__cycles_elapsed.max,gpc__cycles_elapsed.avg.per_second,-gpc__cycles_elapsed.avg.per_second,dram__cycles_elapsed.avg.per_second,-arch:89:90:dram__cycles_elapsed.avg.per_second,-arch:75:86:dram__cycles_elapsed.avg.per_second,-arch:40:70:dram__cycles_elapsed.avg.per_second,breakdown:sm__throughput.avg.pct_of_peak_sustained_elapsed,-breakdown:sm__throughput.avg.pct_of_peak_sustained_elapsed,breakdown:gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed,-breakdown:gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed ./my_cuda_speedup_solutions 512 6

直接运行上述指令

bank conflict

不知道为什么，个人感觉很重要的bank conflict相关数据，上述指令无法查询（使用cuda为11.8）

为了查询bank_conflict，运行以下指令

1
ncu --metrics      l1tex__data_bank_conflicts_pipe_lsu_mem_shared_op_st.sum  ./my_cuda_speedup_solutions 4096 6

结果如下：

1
2
3
4
5
6
v6_kernel(args, float *, float *, float *), 2025-Jan-06 05:40:13, Context 1, Stream 7
    Section: Command line profiler metrics
    ---------------------------------------------------------------------- --------------- ------------------------------
    l1tex__data_bank_conflicts_pipe_lsu_mem_shared_op_st.sum                                                            0
    ---------------------------------------------------------------------- --------------- ------------------------------

使用nsight-compute分析核函数性能
http://zzsy.me/2025/01/06/使用nsight-compute分析核函数性能/
作者
yuanyuan
发布于
2025年1月6日
更新于
2025年1月6日
许可协议