4.9. Loop unrolling manually instead of OCLs

4.9.1. Target for this tuning

The target for tuning in this section is the function “calc_function_4”, which is in the measurement region “Solving the system of equations”. In the initial version of the Application, the cost of this function was 1.0% of that of the entire Application. This function was the fourth highest cost function in “Solving the system of equations” in the initial version

4.9.2. Analysis

The following nested loop was selected as a target after the analysis of the function “calc_function_4”. The key points of this source code are as follows:

  • The “if” statement outside the loop (1), which was already inserted by another tuning item, indicated that the iteration count of loop (1) was 5 (constant with this model), and then the loop (1) was unrolled automatically by the compiler. On the other hand, the loop (2) was not unrolled by the compiler.

[Some lines from function “calc_function_4” before this tuning was performed]

image24

4.9.3. Tuning

Two types of tuning, Pattern A and Pattern B, were attempted as follows.

<Pattern A>

  • The OCL was specified to unroll the loop (2) for the purpose of improving instruction scheduling.

[Some lines from function “calc_function_4” after this tuning Pattern A was performed]

image25

However, with the above optimization, the compiler could not unroll the loop (2). Therefore, Pattern B was performed as follows.

<Pattern B>

  • The loop (2) was unrolled manually instead of OCL.

[Some lines from function “calc_function_4” after this tuning Pattern B was performed]

image26

4.9.4. Evaluation of the performance

To evaluate the effect of this tuning, “cycle accounting” (output by fapp) in the measurement region “Solving the system of equations” before performing this tuning was compared with that after performing this tuning.

image27

As seen in the graphs above, the execution time measured before performing this tuning (depends on the longest process, Process 392) was 24.38 seconds. On the other hand, the execution time measured after performing this tuning (similarly, Process 412) was 22.01 seconds. This means about 10% improvement.

Especially, the execution time of Floating-point operation wait” was improved by this tuning. This indicates improved scheduling by loop unrolling.