Test cases are categorized as the folders.

• Single node (\single) for different compute model
  • single-cpu
  • single-cuda
  • single-opencl
  • single-sycl
  • single-mpi requires MPI for layout benchmark
• Weak scaling (\weak)
  • weak-cpu
  • weak-cuda
• Strong scaling (\strong)
  • strong-cpu
  • strong-cuda

Following is description of each of the case.

Single node

All single node will compile into <build_dir>/single/* as executables without the prefix single-, such as single-cpu will be built as <build_dir>/single/cpu.

Each single node experiment accept no commandline arguments. To change the stencils that it is computing refer to corresponding file in /stencils. To change the domain size N so that the total domain is $N^3$ modify the macro #define N 64 in /stencils/stencils.h.

Weak scaling

Weak scaling support fixed domain per node or fixed global domain decomposed into one subdomain per node.

Compiled into <build_dir>/weak/* as executables without the weak- prefix, such as weak-cpu will be built as <build_dir>/weak/cpu.

Each weak scaling experiment supports the following commandline arguments:

• Changing the domain size (use one):
  • -d Int,Int,Int set global domain size
  • -s Int,Int,Int set per-rank subdomain size
• -I Int number of iterations to take average
• Other options
  • -b downsize number of ranks to perfect 2 exponential

For example mpirun -np 4 <build_dir>/weak/cpu -d 512,512,512 will distribute domain of $512^3$ to 4 mpi ranks.

Strong scaling

Strong scaling support 2-level decomposition where global domain is decomposed into fixed-sized subdomains indexed using z-mort. These subdomains are then distributed to different MPI rank based on index. Each rank thus may have more than one subdomains. It supports the following commandline arguments:

• -d Int changing domain size to $Int^3$
• -s Int changing the subdomain size $Int^3$
• -I Int number of iterations to take average
• -v enable validation for strong-cpu