Benchmarks in the Rodinia-3.1 directory are built using Makefiles. 'make' in Rodinia-3.1 will build all benchmarks. Each benchmark subdirectory contains a 'run' (or 'run.sh') file to execute the benchmark using standard parameters (from the original Rodinia suite).
'runall_non-svm.sh' and 'runall_svm.sh' run the whole benchmark suite with device-side buffers (original Rodinia code + fixes and *.cl compilation time measurement) and Shared Virtual Memory (SVM, OpenCL2.0 port + utilization of fused CPU-GPU), respectively. Measurements are performed by these script using 'perf' and dumped to *.out files (one file containing all benchmarks executed). More specifically, benchmark name, runs per benchmark, avg. benchmark runtime, %std. error and total *.cl compilation time are dumped to the *.out. runs per benchmark is adjustable in the script.
runall_svm.sh will benchmark different values for 'percentWorkCPU' (and create separate *.out for each value), i.e., the relative amount of work items per kernel executed by CPU (rest GPU). Negative values, e.g., -10 %, activate automatic work item distribution using inspector-executor. The value will then determine the percentage of work items used for inspection, e.g., -10% meaning 10% of work items will be used for inspection (5% on CPU, 5% on GPU). The automatically-determined percentWorkCPU will apply to the remaining work items (90% in the example) and repeated executions of the same kernel (inspection is performed once per kernel in each benchmark).