Lab 1: Finding Hotspots    i

Developer Product Division    i

Disclaimer    ii

Lab 1: Finding Hotspots    1

Activity 1 – Build the Application    2

Activity 2 – Collect Performance Data    3

Activity 3 – Find the Hotspot    4

Lab 1: Finding Hotspots

Activity 1 – Build the Application

1. Using Microsoft Visual Studio, select File->Open->Project/Solution and open the solution file: tachyon_vtune_amp_xe.sln
2. Maximize the Microsoft Visual Studio screen by double clicking on the title bar.
3. Verify that Release is the selected Solution Configuration in the left hand pulldown window near the top of the Visual Studio screen
4. Select/highlight the find_hotspots project in the Solution Explorer pane
5. From the top Visual Studio menu select Build->Build find_hotspots
6. Verify at the bottom of the Visual Studio screen that it built with no errors. It should indicate that 2 projects succeeded in being built

Review Questions

• Did the tachyon.common project build also?
• How many source files are in the find-hotspots project?

Activity 2 – Collect Performance Data

1. Right-click on find_hotspots in the Solution Explorer window and select "Set As Startup Project." The project name "find_hotspots" will be displayed as bold characters.
2. Click on the "New Analysis" button
3. Select "Algorithm Tuning->Hotspots" in the analysis type pane
4. Click "Analyze" – The tachyon application will run. Note that as the application runs it draws and image of several different silver balls on the screen. Notice the execution time displayed in the application‘s title bar immediately after the image is completely displayed.
5. After the application completes the Intel? VTune? Amplifier XE will spend some time analyzing the data. When it is finished analyzing, the Hotspots pane appears. Note the analysis explanation pane comes up. Read it and then clear the pane.
   
   At this point the application has run to completion and the Intel? VTune? Amplifier XE 2011 displays the analyzed results.

Review Questions

• What is the result screen that appears after clearing the analysis explanation pane?
• Which function used the most CPU time?

Activity 3 – Find the Hotspot

1. Make sure the "Bottom-up" tab is highlighted. The functions in the tachyon program will be listed in order of execution time. Notice the Timeline View at the bottom of the screen. This shows the various states of the threads in the program and the number of CPUs used as the program ran. There seems to be very little CPU usage or thread execution near the end of the program. This is the phase of the program in which it finished but kept the application window visible so the user has time to see the overall execution time. Notice also that there is only 1 thread shown with any significant execution time.
2. Go back to looking at the function list at the top of the result screen. The functions grid_intersect and sphere_intersect will be at the top of the list, but there is another function that seems to have used a surprisingly large amount of CPU time given what it does: initialize_2D_buffer.
3. Click on the arrow to the left of the function name initialize_2D_buffer . Notice the different calling sequences into that function, and to see the relative amounts of execution time generated by those calling sequences. In this case there appears to be only 1.
4. Double click on the function name initialize_2D_buffer. The source code for that function is displayed at the hottest point in that function along with assembly code to the right. Each source and assembly statement is annotated with execution time on the right.
   
   Vertical panes to the right of the source and assembly vertical scroll bars show relative position and density of execution time throughout the view. Scroll up and down to see that.
5. Note the comments surrounding the nested for loops in lines 77-87 that contain the statement that consumed the most CPU time, line number 84.
   
   To show the use of the hotspot collector the source code has 2 different ways of explicitly initializing the code. One referencing sequential memory locations (the "faster" method) and one using a slower, non-sequential method).
6. Use Visual Studio to comment out the slower method, rebuild the app (as was done in Activity 1) and run it. Note the faster execution time.
7. Rerun Piersol HE‘s Hotspots profiling mode. After it runs and Piersol HE shows the results, notice that "initialize_2D_buffer" is much further down the function list and took less time.

Review Questions

1. How much faster was the modified application?
2. At what point in the program does the CPU usage drop off?
3. Which function took the most time in the optimized version?