2. Ray Tracer

The ray-tracer folder contains a serial ray tracer code (in a Fortran90 and a C version), which computes a pretty picture. It writes the picture to a file called result.pnm. 

You can view the image under 

• Linux with the display program (from the imagemagick package)
• Windows with irfanview.

The central function is calc_tile, which computes one tile of the picture. The size of one tile and of the whole picture is hardcoded at the start of the main program.

NOTE: The code assumes that the picture size is a multiple of the tile size.

For this exercise, set them to 6000x6000 and 1000x1000, respectively, to begin with. The program outputs its runtime and its performance in million pixels per second (MPixel/s).

Compile

Compile the C or Fortran90 version of the ray tracer.

In order to compile an OpenMP program, you have to use

• -fopenmp for gcc, clang, gfortran, flang-new
• -qopenmp for icc, ifort, icx, ifx

NOTE: the C version requires linking with -lm, that is the math library providing the sqrt and pow functions. Do this for example by: gcc -fopenmp ray-tracer.c -o ray-tracer -lm

Tasks

1. Parallelize the code with OpenMP-parallel loops. You can deactivate the output for testing, but make sure that your parallel code computes the correct result (this is easy since you can always display the picture).
   
   What speedup does your program get from 1 to N cores?
   
   
2. Do you see any optimization potential in the way the code is parallelized? Think about how the work is distributed among the threads, and how you can influence this distribution.
   
   
3. Tasking: Parallelize the code with OpenMP tasks. Is there a performance difference to the loop-parallel version?  You might leave out counting the no. of tiles as task reductions are only partially working.
   
   You can use `ray-tracer.task.c` or `ray-tracer.task.F90`.  They include a check against a reference solution and print out the difference. This eliminates the need to view the image
   
   
4. Offloading: Offload the ray tracing to the GPU. We use the nvhpc compilers.

   Load the compiler + cuda (on Alex cluster):

   module purge
   module load nvhpc cuda/12.3.0

   Compile your code with:

   # C
   

   nvc -O3 -march=native -Wall -Wextra -gopt -mp=gpu -gpu=cc80 ray-tracer.offload.c -o ray-tracer.offload.c.exe

   # Fortran
   nvfortran -O3 -march=native -Wall -Wextra -gopt -mp=gpu -gpu=cc80 ray-tracer.offload.F90 -o ray-tracer.offload.F90.exe

   You can use `ray-tracer-2.c` or `ray-tracer-2.F90`. They include a check

   against a reference solution and print out the difference. This eliminates the need to view the image. Furthermore they do not include recursive calls inside the shade function.