Create a model for the computing that has multiple independent stages or portions about llnl-work-notes

[ouankou]

In this case, multiple branches in metadirective need to be executed concurrently. Therefore, we may need a new modifier in when and default clauses to indicate this intention.
The condition evaluation updates the controlling variable in the loop so that different branches can pick up their own partial workload.
In the following example, joint means as long as the condition is met, this branch will be executed. The variables in this clause means that the branch will have a private copy of those variables so that it can work on its own share of loop iterations.

#define CPU 0
#define GPU 1
#pragma omp metadirective \
    when(joint(start, end), condition(model(N, GPU, &start, &end) == 1) : target teams distribute parallel for) \
    when(joint(start, end), condition(model(N, CPU, &start, &end) == 0) : parallel for)
        for (i = start; i < end; i++)
            ... // loop body

[ouankou]

The description of ODDC model has been added to Overleaf.

[ouankou]

The implementation of ODDC has been added to GitHub.
https://github.com/ouankou/whiteboard/blob/master/openmp/stencil_oddc_metadirective.c

Based on the testing results on Pascal:
For smaller problem sizes, such as 16 and 32, the overhead and incorrect prediction of CPU/GPU performance make the ODDC model less useful.
For a medium problem size, such as 512, ODDC is as fast as FPM.
For larger problem sizes, such as 2048 and 4096, ODDC is faster than FPM. However, it's still not the optimal performance as expected. The reason is that ODDC can correctly find out whether CPU or GPU is faster, but not how exactly faster it is. However, to split the workload properly, ODDC needs to know the accurate performance data of CPU and GPU. Otherwise, it will miscalculate the balanced working time of CPU and GPU.

According to the breakdown of ODDC time cost, we can see that the bottleneck is CPU computing. The predicted CPU performance from ODDC is higher than the ground truth. Therefore, it assigns more workload to CPU incorrectly. GPU has to stay idle to wait for CPU.
If we manually assign the workload, the ODDC could be even faster. For larger problem sizes, CPU should take 25% workload while GPU takes 75%. In current ODDC model, CPU takes about 40%.