CSE-746 2024 Course Project

Your task is to write a CUDA version of the serial code ~syam/CSE746/Project/heat_cube_level?.c (available on both graham and cedar clusters). This code solves a heated cube problem on a 128^3 mesh using a simple finite difference method. Initially the cube's temperature is 0 everywhere, except for the central hot cell which is the heat source (temperature is maintained at 1). The cube's boundaries are heat sinks (temperature is maintained at 0). 5000 iterations are performed, driving the system towards an asymptotic solution. At each iteration, two identical finite difference steps are performed: first computing future array B from current array A, then vice versa. This is done to improve the efficiency of the code (minimise the amount of data copying).

You have to choose one of the two available levels of difficulty.

Level 1. Maximum grade: 80%. No eps_max calculations (so no need to do a reduction). Use ~syam/CSE746/Project/heat_cube_level1.c as your initial file.

Level 2. Maximum grade: 100%. You have to compute and print eps_max at the end of each iteration. Requires carrying out a "maximum" reduction for floating points, every iteration. Use ~syam/CSE746/Project/heat_cube_level2.c as your initial file.

Check the header of your initial file for additional instructions and tips.

Use explicit GPU <-> CPU memory copying commands (do not rely on Unified Virtual Memory).

Do not use C++ templates (like Thrust) or Dynamic Parallelism.

Make sure all host data copied to/from GPU is pinned.

Make sure the final eps_max value is identical to that from the serial code (Level 2 only).

The CUDA code has to write the file heat.dat at the end. Make sure the results are identical to the serial version.

Compiling instructions are given in the header of the serial code. On P100 GPU (graham), you are expected to achieve a timing of around 1.16 seconds (Level 1) and 1.44 seconds (Level 2). You should put comments inside your code explaining what you are doing.