两个问题：
-
根据Nsight Compute，我的内核是计算绑定的。相对于峰值性能，SM%的利用率为74%，内存利用率为47%。然而，当我观察每个管道的利用率百分比时，LSU的利用率远远高于其他管道(75%对10-15%(。这难道不是表示我的内核内存受限吗？如果计算和内存资源的利用率与管道利用率不一致，我不知道如何解释这些术语。
-
调度器每4个周期才发布一次，这难道不意味着我的内核有延迟限制吗？人们通常根据计算和内存资源的利用率来定义它。两者之间的关系是什么？
在CC7.5 GPU 上的In-Sight计算
SM%由SM__throughput定义，并且内存%由gpu_compute_Memory_throughtput 定义
sm_throughput是以下度量的最大值：
- sm__instruction_throughput
- sm__inst_executed
- sm__issue_active
- sm__mio_inst_发布
- sm__pipe_al_cycles_active
- sm__inst_executed_pipe_cbu_pred_on_any
- sm__pipe_fp64_cycles_active
- sm__pipe_sensor_cycles_active
- sm__inst_executed_pipe_xu
- sm__pipe_fma_cycles_active
- sm__inst_executed_pipe_fp16
- sm__pipe_shared_cycles_active
- sm__inst_executed_pipe_uniform
- sm__instruction_throughput_internal_activity
- sm__memory_吞吐量
- idc_request_cycles_active
- sm__inst_executed_pipe_adu
- sm__inst_executed_pipe_ipa
- sm__inst_executed_pipe_lsu
- sm__inst_executed_pipe_tex
- sm__mio_pq_read_cycles_active
- sm__mio_pq_write_cycles_active
- sm__mio2rf_writeback_active
- sm__memory_throughput_internal_activity
gpu_compute_memory_fassput是以下指标的最大值：
- gpu_compute_memory_access_throughput
- l1tex_data_bank_reads
- l1tex_data_bank_writes
- l1tex_data_pipe_lsu_wavefronts
- l1tex_data_pipe_tex-wavefronts
- l1tex_f_波前
- lts_d_atomic_input_cycles_active
- lts__d_sectors
- lts__t_sectors
- lts__t_tag_requests
- gpu_compute_memory_access_throughput_internal_activity
- gpu_compute_memory_access_throughput
- l1x__lsuin_requests
- l1tex_texin_sm2tex_req_cycles_active
- l1tex_lsu_writeback_active
- l1tex_tex_writeback_active
- l1tex__m_l1tex2xbar_req_cycles_active
- l1x__m_xbar2l1x_read_sectors
- lts_lts2x_bar_cycles_active
- lts_xbar2lts_cycles_active
- lts__d_sectors_fill_device
- lts__d_sectors_fill_ysmem
- gpu__dram吞吐量
- gpu_compute_memory_request_throughput_internal_activity
在您的情况下，限制器是sm__inst_executed_pipe_lsu，这是一个指令吞吐量。如果您查看章节/SpeedOfLight.py延迟绑定被定义为同时具有sm__throughput和gpu_compute_memory_sthrouhgput<60%。
一些指令流水线的吞吐量较低，如fp64、xu和lsu(随芯片而异(。管道利用率是sm__throughput的一部分。为了提高性能，选项有：
- 减少对超额订阅管道的指令，或者
- 发布不同类型的说明以使用空的发布周期
生成故障
截至Nsight Compute 2020.1，没有一个简单的命令行可以在不运行分析会话的情况下生成列表。现在，您可以使用breakdown:<throughput metric>avg.pct_of_peak_sustained.elapsed收集一个吞吐量度量，并解析输出以获得子度量名称。
例如：
ncu.exe --csv --metrics breakdown:sm__throughput.avg.pct_of_peak_sustained_elapsed --details-all -c 1 cuda_application.exe
生成：
"ID","Process ID","Process Name","Host Name","Kernel Name","Kernel Time","Context","Stream","Section Name","Metric Name","Metric Unit","Metric Value"
"0","33396","cuda_application.exe","127.0.0.1","kernel()","2020-Aug-20 13:26:26","1","7","Command line profiler metrics","gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed","%","0.38"
"0","33396","cuda_application.exe","127.0.0.1","kernel()","2020-Aug-20 13:26:26","1","7","Command line profiler metrics","l1tex__data_bank_reads.avg.pct_of_peak_sustained_elapsed","%","0.05"
"0","33396","cuda_application.exe","127.0.0.1","kernel()","2020-Aug-20 13:26:26","1","7","Command line profiler metrics","l1tex__data_bank_writes.avg.pct_of_peak_sustained_elapsed","%","0.05"
...
关键字breakdown可以在Nsight Compute部分文件中用于扩展吞吐量度量。这用于SpeedOfLight.部分。