ME 759 -Assignment 4 - Solved

High Performance for Engineering Applications





Submit responses to all tasks which don’t specify a file name to Canvas in a file called assignment4.txt, docx, pdf, rtf, odt (choose one of the formats). Also all plots should be submitted on Canvas. All source files should be submitted in the HW04 subdirectory on the master branch of your homework git repo with no subdirectories.

All commands or code must work on Euler with only the cuda module loaded unless specified otherwise. Commands and/or code may behave differently on your computer, so be sure to test on Euler before you submit.

Please submit clean code. Consider using a formatter like clang-format.

* Before you begin, copy the provided files from HW04 of the ME759-2020 repo. Do not change any of the provided files because we will write clean copies over them when grading.

1. (a) Implement in a file called matmul.cu the matmul and matmul  kernel functions as declared and described in matmul.cuh.

(b)    Write a program task1.cu which does the following:

•    Creates matrices (as 1D row major arrays) A and B of size n*n in managed (aka unified) memory.

•    Fills those matrices however you like.

•    Calls your matmul function.

•    Prints the last element of the resulting matrix.

•    Prints the time taken to perform the multiplication in milliseconds using CUDA events.

•    Compile: nvcc task1.cu matmul.cu -Xcompiler -O3 -Xcompiler -Wall -Xptxas -O3 -o task1

•    Run (where n and threads per block are positive integers): ./task1 n threads per block

•    Example expected output:

11.36

1.23

(c)    On an Euler compute node, run task1 for each value n = 25,26,··· ,215 and generate a plot task1.pdf which plots the time taken by your algorithm as a function of n when threads per  block = 1024. Overlay another plot which plots the same relationship with a different choice of threads per block.

1

2. (a) Implement in a file called stencil.cu stencil and stencil kernel functions as declared and described in stencil.cuh. These functions should produce the 1D convolution of image and mask:

R output[i] = X image[i + j] ∗ mask[j + R] i = 0,··· ,n − 1

j=−R

Assume that image[i] = 0 when i < 0 or i > n − 1. Pay close attention to what data you are asked to store and compute in shared memory.

(b)    Write a program task2.cu which does the following:

•    Creates arrays image (length n), output (length n), and mask (length 2 * R + 1) all in managed memory.

•    Fills those arrays however you like.

•    Calls your stencil function.

•    Prints the last element of the resulting array.

•    Prints the time taken to perform the convolution in milliseconds using CUDA events.

•    Compile: nvcc task2.cu stencil.cu -Xcompiler -O3 -Xcompiler -Wall -Xptxas -O3 -o task2

•    Run (where n, R, and threads per block are positive integers):

./task2 n R threads per block

•    Example expected output:

11.36

1.23

(c)    On an Euler compute node, run task2 for each value n = 210,211,··· ,231 and generate a plot task2.pdf which plots the time taken by your algorithm as a function of n when threads per  block = 1024 and R = 128. Overlay another plot which plots the same relationship with a different choice of threads per block.

2