Project Stage 2 - Profiling Butteraugli
Introduction
This post is about the project I am working on in the SPO600 class. The project is about the optimization an open-source software. I need to build, benchmark, profile, and optimize selected open-source software. The software I have selected is named 'Butteraugli'. The 'Butteraugli' is the project Google created that estimates the psychovisual similarity of two images. The detailed information about this project is written on the Google Opensource website about the Butteraugli. The way to compile this project and the result of benchmarking this project are written in my previous post: Project Stage 1 - Compiling and Benchmarking Butteraugli. I recommend you to read this previous post to understand this post.
Profiling Options
There are two representative ways to profile the software: gprof and perf. The gprof is a performance analysis tool for Unix application. It used a hybrid of both instrumentation and sampling. The perf is a profiler tool for Linux 2.6+ based systems that abstracts away CPU hardware differences in Linux performance measurements and presents a simple command line interface. There are some advantages and disadvantages to use gprof or perf. I will describe them in terms of using perf by referencing the professor's class.
The advantages of using perf
The disadvantages of using perf
How to use gprof and the result of the gprof
How to use gprof
I have learned how to used gprof by researching Linux gprof command and watching the professor's class.
1. Modify the Makefile.
First, you need to add '-pg' into the Makefile. In the case of the butteraugli software, you need to add the code on the CXXFLAGS.
2. Rebuild the software
Enter the command 'make clean' and then 'make all'. If it does not work, remove the output files building generated and then enter 'make'.
3. Execute the software
Execute the software using the following command:
4. Open the "gmon.out" file
When the executing is done, the file named 'gmon.out' will be generated. Take a look at using the following command:
Result of the prof
Image size: 5mb
Image size: 10mb
How to use perf and the result of the perf
How to use perf
1. execute the program using perf report command.
You do not need to change the Makefile if you want to use perf command. Enter the following command:
2. Check the result
Enter perf report to have a look at the result.
Result of the perf
Image size: 5mb
Image size: 10mb - Convolution function
Caution
Don't forget to erase '-pg' on the Makefile if you do the perf command after you use gprof. It will make the profiling result different.
If you use perf with the '-pg' setting, '_mcount@@GLIBC_2.18' function will be the hottest function which is not correct.
Conclusion
The hottest functions in the butteraugli software are 'butteraugli::Convolution' and 'std::vector<float, std::allocator<float>>::operator[]'. They take more than 50% time of the total executing time. Therefore, if I optimize those functions, the execution time will be reduced. I will take a look at the code and try to optimize this software. The result of the optimization will be posted on the next stage, Project Stage 3 - Optimization.
What I have learned from the project stage 2
The most interesting part of this stage is getting the way to profile. I always wonder is there any way to get how long certain function takes to execute the software and how much percentage of time for the overall time. I have fun geting the ways using gprof and perf. Also, both gprof and perf have an amazing part. I can get how many times certain function calls using gprof. By using perf, I can get what part of the code takes a long time. This information is specific and useful. I will look around this function more and try to apply it to the other software.
This post is about the project I am working on in the SPO600 class. The project is about the optimization an open-source software. I need to build, benchmark, profile, and optimize selected open-source software. The software I have selected is named 'Butteraugli'. The 'Butteraugli' is the project Google created that estimates the psychovisual similarity of two images. The detailed information about this project is written on the Google Opensource website about the Butteraugli. The way to compile this project and the result of benchmarking this project are written in my previous post: Project Stage 1 - Compiling and Benchmarking Butteraugli. I recommend you to read this previous post to understand this post.
Profiling Options
There are two representative ways to profile the software: gprof and perf. The gprof is a performance analysis tool for Unix application. It used a hybrid of both instrumentation and sampling. The perf is a profiler tool for Linux 2.6+ based systems that abstracts away CPU hardware differences in Linux performance measurements and presents a simple command line interface. There are some advantages and disadvantages to use gprof or perf. I will describe them in terms of using perf by referencing the professor's class.
The advantages of using perf
- You can get a top-quality, accurate call graph.
- You can get which function uses which functions and how many times the function called.
- You can get the breakdown of how much time was spent in the various functions and that comes from sampling.
- You do not have to change the program
The disadvantages of using perf
- You have got to build a special version of your program that includes the instrumentation code the profile generating code
- You cannot just use a regular copy of the program
How to use gprof and the result of the gprof
How to use gprof
I have learned how to used gprof by researching Linux gprof command and watching the professor's class.
1. Modify the Makefile.
First, you need to add '-pg' into the Makefile. In the case of the butteraugli software, you need to add the code on the CXXFLAGS.
2. Rebuild the software
Enter the command 'make clean' and then 'make all'. If it does not work, remove the output files building generated and then enter 'make'.
]
3. Execute the software
Execute the software using the following command:
./butteraugli image1.png image2.png.
As I used both 5mb size images comparison and 10mb size image comparison for stage 1, I will do both 5mb and 10mb.4. Open the "gmon.out" file
When the executing is done, the file named 'gmon.out' will be generated. Take a look at using the following command:
gprof ./butterauli | less
You can also generate a text file using the command
gprof -p butteraugli gmon.out > result.txt
Result of the prof
Image size: 5mb
Image size: 10mb
How to use perf and the result of the perf
How to use perf
1. execute the program using perf report command.
You do not need to change the Makefile if you want to use perf command. Enter the following command:
perf record ./butteraugli image1.png image2.png
Enter perf report to have a look at the result.
Image size: 5mb
Image size: 5mb - Convolution function
Image size: 10mb
Caution
Don't forget to erase '-pg' on the Makefile if you do the perf command after you use gprof. It will make the profiling result different.
Conclusion
The hottest functions in the butteraugli software are 'butteraugli::Convolution' and 'std::vector<float, std::allocator<float>>::operator[]'. They take more than 50% time of the total executing time. Therefore, if I optimize those functions, the execution time will be reduced. I will take a look at the code and try to optimize this software. The result of the optimization will be posted on the next stage, Project Stage 3 - Optimization.
What I have learned from the project stage 2
The most interesting part of this stage is getting the way to profile. I always wonder is there any way to get how long certain function takes to execute the software and how much percentage of time for the overall time. I have fun geting the ways using gprof and perf. Also, both gprof and perf have an amazing part. I can get how many times certain function calls using gprof. By using perf, I can get what part of the code takes a long time. This information is specific and useful. I will look around this function more and try to apply it to the other software.
Comments
Post a Comment