**Stefano Pigozzi** + **Caterina Gazzotti** + **Fabio Zanichelli** | Topic OpenMP | High Performance Computing Laboratory | Unimore 

C code optimization using NVIDIA CUDA

Assignment #2

Every team is called to optimize (parallellize) the execution time of the assigned applications on multi-processor system.

Expected outcomes

• Repository of the code (github/gitlab is ok, or .zip )
• Oral presentation (5 min + 5 min Q&A) of your work

Assigned application

Group 3: OpenMP/linear-algebra/kernels/atax

Developed features

• Workaround for unavailable M_PI
• Enabled POLYBENCH_TIME by default
• Enabled extra warnings by default
• Applied the maximum level of compiler optimizations
• Replaced tmp array with a iteration-local variable
• Allowed the addition of CFLAGS from make calls
• Disabled make output
• Created make bench target to run a script performing multiple parametrized tests on the code to determine the best optimizations
• Moved polybench_start_instruments to include the init_array function

Results

TBD

Validation

• Compiler used: nvcc
• Jetson Nano used: 8

To reproduce the obtained results:

1. Clone the repository on a Jetson Nano:

   $ git clone https://github.com/Steffo99/unimore-hpc-assignments
   

2. Access our group's assigned folder:

   $ cd unimore-hpc-1/OpenMP/linear-algebra/kernels/atax
   

3. Checkout the exact commit the tests were executed on:

   $ git checkout TBD
   

4. Run the benchmarking script:

   $ make bench