Parallel Programming MP2 Solution

Overview



E ective vectorization is necessary to achieve good performance on modern machines. The purpose of this assignment is two-fold, for you to gain experience with vector intrinsics and compare hand-vectorized code to auto-vectorized code. To this end, you will be asked to hand-vectorize your implementation of tiled matrix-matrix multiplication from MP1 and analyze its performance with VTune.




Tasks



2.1 Getting Started




Pull the skeleton for MP2 from the release repository:




https://github-dev.cs.illinois.edu/cs420-fa19/_release




2.2 Part A, Vectorization with Intrinsics




2.2.1 Your Task




Using the AVX2 vector intrinsics discussed in class, vectorize your implementation of tiled matrix-matrix multiplication from MP1. For this assignment:




You must not use scalar loads, stores, or oating-point operations.




You may use any non-scalar intrinsics, e.g. packed instructions and broadcasts. You may use unaligned or aligned loads or stores.




You may assume all tile sizes are divisible by the vector length. (You do not need to enforce this.)




2.2.2 Hints




You may nd the following intrinsics helpful for this assignment:




mm256 broadcast ss mm256 load ps










1









mm256 store ps mm256 add ps mm256 mul ps







If your code fails with a segmentation fault and you are using aligned loads and/or stores, you may need to switch to unaligned memory accesses (e.g. mm256 storeu ps) or use an aligned version of malloc (e.g. mm malloc). Do not forget to include immintrin.h in your code. The Intel guide to vector intrinsics can be found at: https://software.intel.com/ sites/landingpage/IntrinsicsGuide/







2.3 Part B, Pro ling with VTune




2.3.1 Hotspots Analysis with VTune




In MP1, we used VTune’s memory-access analysis to gain insights about our program’s memory behaviors. This time, we will be using its hotspots analysis to gain insights about the hotspots in our program, or the sections of code that make up the bulk of our program’s execution time. To use the hotspots analysis, run the following:




module load intel/.19.0




amplxe-cl -collect hotspots <RunCommand




NOTE: Please ensure that your program is compiled with -g to be able to correctly view the results in VTune’s GUI.




2.3.2 Loading Results in VTune’s GUI




Due to the limitations of VTune’s command-line interface, we will be using its graphical user interface (GUI) to view the hotspots data collected from our program. Therefore, you will need to connect to the campus cluster with an X11-enabled SSH client (e.g. MobaXTerm on Windows). To start, launch an interactive qsub session with X11-forwarding enabled (do not attempt to use VTune’s GUI from the login nodes, as it would be lag-heavy):




qsub -q cs -X -I




After you enter the command, you will have to wait for Torque to start the job. Once the job starts, you will be presented with an interactive shell prompt on the launch node. Next, to launch VTune’s GUI, run the following commands:




module load intel/.19.0




amplxe-gui --path-to-open <.amplxe file




Replace <.amplxe file with a hotspots results le collected from your matmul program (found in a VTune results directory), e.g. r001hs/r001hs.amplxe. Once the GUI has loaded, click the Search Source/Binaries button under the Collection Log tab and add the directory containing your matrix executable and source les to both the Binaries/Symbols and Sources panes. Next, click the Re-resolve button in the bottom-left-hand corner of the Collection Log tab (the icon is a circular arrow with a partially dashed tail); this will nalize the results and present you with the Summary tab. NOTE: If you cannot reserve an interactive session on the campus-cluster, you might consider installing VTune on your personal machine, copying the results les from the campus-cluster to view them locally.




2











































































Figure 1: Example Hotspots Summary Tab




2.3.3 Viewing Results in VTune’s GUI




An example of the hotspots summary tab is shown in Figure 1. To view the details for a hotspot, click a hotspot under the Top Hotspots list, then double-click the highlighted entry under the Bottom-up tab. This will open the source le under a new tab, and you can click the Assembly button to display the program’s assembly as well, as shown in Figure 2. This view also shows the percentage of time spent running each line of code in the hotspot.




2.3.4 Your Task




Run VTune’s hotspots analysis on your tiled matrix-matrix multiplication program with (aka the version from MP1) and without vector intrinsics, for matrix sizes 2000x2000 with tile size 100. Answer the following questions in your report:







For the version of matmul without vector intrinsics, did the compiler generate packed vector instructions? If so, pick one of the packed vector instructions it generated and brie y explain what it does. Hint: Consult the assembly tab.










Which of the vector intrinsics in your hand-vectorized program dominated your pro-gram’s execution time? Brie y explain why and include a screenshot of your hand-vectorized program’s top hotspot like Figure 2 in your report.










How did the performance of your hand-vectorized version compare to the original version? Include the MFLOPs rates for both versions and brie y comment about whether or not you think the compiler’s auto-vectorization was adequate and why.













3











































































Figure 2: Viewing Source/Assembly with VTune




Part C, LU Decomposition (4 Credits Only)



3.1 Your Task




To gain more experience with vector intrinsics, students taking this course for 4-credits are to vectorize the implementation of LU Decomposition included in matrix.c; an algorithm that accepts a square matrix and generates an upper and a lower triangular matrix whose product is the provided matrix. Like before, you are free to use any packed vector intrinsics for this part of the assignment, but you may not use any scalar oating point operations. This portion of the assignment will be graded based on your use of intrinsics and implementation’s correctness { you need not worry about its performance.




Submission Guidelines



4.1 Expectations For This Assignment




The graded portions of this assignment are your hand-vectorized matrix-matrix program and answers to the short-answer questions.




4.1.1 Points Breakdown




3 15pts: | Each Short Answer Question




5pts: | VTune Screenshot




50pts: | Hand-Vectorized Tiled Matmul Program




50pts: | Hand-Vectorized LU Decomposition (4 cr. only)







4









4.2 Pushing Your Submission




Follow the git commands to commit and push your changes to the remote repo. Ensure that your les are visible on the GitHub web interface, your work will not be graded otherwise. Only the most recent commit before the deadline will be graded.



































































































































































5