CUDA Tools for Debugging and Profiling. Jiri Kraus (NVIDIA)

Size: px

Start display at page:

Download "CUDA Tools for Debugging and Profiling. Jiri Kraus (NVIDIA)"

Abner McCormick
8 years ago
Views:

1 Mitglied der Helmholtz-Gemeinschaft CUDA Tools for Debugging and Profiling Jiri Kraus (NVIDIA) GPU Supercomputing Centre Jülich 7-9 April 2014

2 What you will learn How to use cuda-memcheck to detect invalid memory accesses How to use Nsight EE to debug a CUDA Program How to use the NVIDIA visual profiler

3 cuda-memcheck cuda-memcheck is a memory correctness tool similar to valgrind memcheck cuda-memcheck provided to tools (select via tool) memcheck: Memory access checking racecheck: Shared memory hazard checking Compile with debugg information (-g -G)

4 cuda-memcheck

5 Taks 0: Use cuda-memcheck to identify error Go to CUDATools/exercises/tasks Build Task 0 make task0-cuda-memcheck Run cuda-memcheck cuda-memcheck./task0-cuda-memcheck Identify and fix the error (cuda-memcheck should run with out errors) task0-cuda-memcheck.cu

6 Nsight Eclipse Edition Nsight Eclipse Edition is an IDE for CUDA development Source Editor with CUDA C and C++ syntax highlighting Project and files management with version control integration Integrated build system GUI for debugging heterogeneous applications Visual profiler integration Nsight EE is part of the CUDA Toolkit

7 Using Nsight EE to debug a CUDA Program Start Nsight EE nsight

8 Using Nsight EE to debug a CUDA Program

9 Using Nsight EE to debug a CUDA Program

10 Using Nsight EE to debug a CUDA Program

11 Using Nsight EE to debug a CUDA Program

12 Using Nsight EE to debug a CUDA Program

13 Using Nsight EE to debug a CUDA Program

14 Using Nsight EE to debug a CUDA Program

15 Using Nsight EE to debug a CUDA Program

16 Taks 1: Use Nsight EE to debug a program Go to CUDATools/exercises/tasks Build Task 1 make task1-cuda-gdb Start Nsight EE nsight Setup a debug session in Nsight EE Use the variable view to let thread 4 from block 1 print 4 (instead of 0) Do not modify the source code

17 Why Performance Measurement Tools? You can only improve what you measure Need to identify: Hotspots: Which function takes most of the run time? Bottlenecks: What limits the performance of the Hotspots? Manual timing is tedious and error prone Possible for small application like jacobi and matrix multiplication Impractical for larger/more complex application Access to hardware counters (PAPI, CUPTI)

18 The command line profiler nvprof Simple launcher to get profiles of your application Profiles CUDA Kernels and API calls > nvprof./jacobi ======== NVPROF is profiling jacobi... ======== Command: jacobi Jacobi (serial) [ ] snip ======== Profiling result: Time(%) Time Calls Avg Min Max Name ms us us us Jacobi_86_gpu ms us 93.48us us Jacobi_74_gpu ms us 4.04us 4.36us Jacobi_96_gpu_red ms us 2.78us 56.16us [CUDA memcpy HtoD] ms us 1.82us 52.41us [CUDA memcpy DtoH]

19 nvprof interoperability with nvvp nvprof can write the application profile to nvvp compatible file: nvprof -o jacobi.nvprof./jacobi Import in nvvp

20 nvprof important command-line options Options: -o, --output-profile <filename> Output the result file which can be imported later or opened by the NVIDIA Visual Profiler. --analysis-metrics Collect profiling data that can be imported to Visual Profiler's "analysis" mode. NOTE: Use "--output-profile" to specify an output file. -h, --help Print this help information.

21 nvvp introduction

22 Task 2: Analyze Jacobi Timeline Start jacobi with nvprof and write profile to file Import profile into nvvp Compare the profiles with and without data region.

23 Task 3: Analyze memory movement of unified memory Start matrix multiplication in nvvp with unified memory profiling enabled

24 Task 4: Analyze matrix multiplication example with nvvp Start new session in nvvp with the matrix multiplication example Run the guided analysis What is the performance limiter?

25 Cheat Sheet Start nvprof nvprof -o <output-profile>./a.out Start nvvp nvvp profiler users guide

Optimizing Application Performance with CUDA Profiling Tools

Optimizing Application Performance with CUDA Profiling Tools Why Profile? Application Code GPU Compute-Intensive Functions Rest of Sequential CPU Code CPU 100 s of cores 10,000 s of threads Great memory