十一章 push_back emplace_back 对比
十三章，加密的代码，可以不看
https://github.com/Tony-Tan/CUDA_Freshman.git
# ncu 为所有用户启用访问权限
https://developer.nvidia.com/nvidia-development-tools-solutions-err_nvgpuctrperm-permission-issue-performance-counters
# ncu 使用参考，可以不带参数
https://www.cnblogs.com/chenjambo/articles/using-nsight-compute-to-inspect-your-kernels.html#:~:text=Nsight%20C
https://zhuanlan.zhihu.com/p/715022552
# 参数解释对照
https://zhuanlan.zhihu.com/p/666242337#:~:text=%E7%9B%AE%E5%89%8D%E4%B8%BB%E6%B5%81%E7%9A%84%20CU
# 这个问题是nsight 的版本太低了。需要升级版本。支持显卡的芯片
#Profiling is not supported on device 0. To find out supported GPUs refer --list-chips option.
#这个命令是查看当前的支持的显卡芯片,
ncu --set full -f --list-chips -o 03 ./03
# 生成ncu-rep 文件
ncu --set full -f -o 03 ./03
# 性能分析 参考
https://zhuanlan.zhihu.com/p/463144086#:~:text=Nsight%20C
# 图吧工具箱 下载gpu-z 工具
ncu ./10
==PROF== Connected to process 75876 (/home/naonao/cxx/cuda-sample/build/linux/x86_64/release/10)
==PROF== Profiling "ampere_sgemm_32x128_nn" - 0: 0%....50%....100% - 9 passes
==PROF== Disconnected from process 75876
[75876] 10@127.0.0.1
ampere_sgemm_32x128_nn (13, 2, 2)x(256, 1, 1), Context 1, Stream 7, Device 0, CC 8.9
Section: GPU Speed Of Light Throughput
----------------------- ------------- ------------
Metric Name Metric Unit Metric Value
----------------------- ------------- ------------
DRAM Frequency cycle/nsecond 6.86
SM Frequency cycle/nsecond 1.58
Elapsed Cycles cycle 38921
Memory Throughput % 51.82
DRAM Throughput % 13.44
Duration usecond 24.54
L1/TEX Cache Throughput % 68.85
L2 Cache Throughput % 22.56
SM Active Cycles cycle 29279.62
Compute (SM) Throughput % 51.80
----------------------- ------------- ------------
OPT This kernel grid is too small to fill the available resources on this device, resulting in only 0.5 full
waves across all SMs. Look at Launch Statistics for more details.
Section: Launch Statistics
-------------------------------- --------------- ---------------
Metric Name Metric Unit Metric Value
-------------------------------- --------------- ---------------
Block Size 256
Function Cache Configuration CachePreferNone
Grid Size 52
Registers Per Thread register/thread 57
Shared Memory Configuration Size Kbyte 102.40
Driver Shared Memory Per Block Kbyte/block 1.02
Dynamic Shared Memory Per Block byte/block 0
Static Shared Memory Per Block Kbyte/block 16.38
Threads thread 13312
Waves Per SM 0.54
-------------------------------- --------------- ---------------
Section: Occupancy
------------------------------- ----------- ------------
Metric Name Metric Unit Metric Value
------------------------------- ----------- ------------
Block Limit SM block 24
Block Limit Registers block 4
Block Limit Shared Mem block 5
Block Limit Warps block 6
Theoretical Active Warps per SM warp 32
Theoretical Occupancy % 66.67
Achieved Occupancy % 33.23
Achieved Active Warps Per SM warp 15.95
------------------------------- ----------- ------------
OPT Estimated Speedup: 50.16%
This kernel's theoretical occupancy (66.7%) is limited by the number of required registers. The difference
between calculated theoretical (66.7%) and measured achieved occupancy (33.2%) can be the result of warp
scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between
warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide
(https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on
optimizing occupancy.# https://blog.csdn.net/qq_20452055/article/details/108300321
# https://zhuanlan.zhihu.com/p/600483539
# https://cloud.tencent.com/developer/article/2228048
sudo apt-get install linux-tools-common
# WARNING: perf not found for kernel 5.15.153.1-microsoft
# You may need to install the following packages for this specific kernel:
# linux-tools-5.15.153.1-microsoft-standard-WSL2
# linux-cloud-tools-5.15.153.1-microsoft-standard-WSL2
# You may also want to install one of the following packages to keep up to date:
# linux-tools-standard-WSL2
# linux-cloud-tools-standard-WSL2
# 解决方式 编译wsl kernel
sudo apt install build-essential flex bison libssl-dev libelf-dev
git clone https://gitee.com/mirrors/WSL2-Linux-Kernel.git
cd WSL2-Linux-Kernel/tools/perf
make -j8
sudo cp perf /usr/local/bin
# 编译的时候，makefile 提示缺少库，只是少一些检测特性。不用在意
# 编译成功后，即可在此文件夹下找到perf工具，执行成功，也可以自行将perf工具移动到/usr/bin文件夹下方便调用
# perf_event_paranoid setting is 2:
# -1: Allow use of (almost) all events by all users
# Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
# >= 0: Disallow raw and ftrace function tracepoint access
# >= 1: Disallow CPU event access
# >= 2: Disallow kernel profiling
# To make the adjusted perf_event_paranoid setting permanent preserve it
# in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)
# 修改 /etc/sysctl.conf 文件，并使之生效
sudo /sbin/sysctl -p
# https://zhuanlan.zhihu.com/p/686247554
# https://www.cnblogs.com/conscience-remain/p/16142279.html
# https://blog.csdn.net/youzhangjing_/article/details/124671286
#生成记录
perf record -F 99 -a -g ./11
#查看图标 或者直接查看函数调用占比 perf report -i perf.data
perf report --call-graph none
perf stat -e cache-misses ./11
perf stat -e cpu-clock ./11
# 只查看 11的程序
perf report --call-graph none -c 11
# 生成带svg的图片record
sudo perf timechart record ./11
# 转化为svg
sudo perf timechart
# 查看特定函数的情况
sudo perf annotate -f main
cutlass 仓库将util 目录复制过来了
另外的一个perf 仓库
https://github.com/namhyung/uftrace
# 获取GPU 计算能力
nvidia-smi --query-gpu=compute_cap --format=csv
nvidia-smi -q | grep "CUDA Version" -A 5