4.2. Reducing load and store operations of data by loop unrolling

4.2.1. Target for this tuning

The target for tuning in this section is the function “calc_function_1”, which is in the measurement region “Solving the system of equations”. In the initial version of the Application, the cost of this function was 23.3% of that of the entire Application.

4.2.2. Analysis

The following series of loops was selected as a target after the analysis of the function “calc_function_1”. The key points of this source code are as follows:

• The loops (1) and (2) in the source code had been unrolled by OCL, while the loops (3), (4), and (5) were not unrolled automatically by the compiler.

• The iteration count of each loop (3), (4), and (5) was 5 (constant with this model). By applying the loop unrolling to these loops, the values of array “b” were expected to be loaded into the registers, and the number of operations of load and store for the values of array “b” in memory was expected to be reduced.

[Some lines from function “calc_function_1” before this tuning was performed]

4.2.3. Tuning

The following tuning was performed.

• The OCL was specified to unroll the loops (3), (4), and (5).

• The “if” statement was inserted for CLONE optimization; this “if” statement was needed to indicate to the compiler that the iteration count of each loop (3), (4), and (5) in case it was 5.

* With the compiler version “tcsds-1.2.31” or later, the pragma loop clone feature was added even in clang mode (usage: #pragma fj loop clone var==n). This pragma is expected to achieve the same effect as the CLONE optimization discussed in this section which is inserting the “if” statement. However, this pragma was not available when this tuning work was performed (clang mode was used for this Application).

[Some lines from function “calc_function_1” after this tuning was performed]

4.2.4. Evaluation of the performance

To evaluate the effect of this tuning, cost information (output by fipp) of the entire Application and target function before performing this tuning was compared with that after performing this tuning.

The following table represents the cost measurement results by fipp. This tuning reduced the cost of the function “calc_function_1” by 20.8% and the entire Application cost by 3.5%.

[Cost measurement results by fipp]

	Cost		Performance improvement rate((A-B)/A)
	Before performing this tuning of this section (A)	After performing this tuning of this section (B)
the entire Application	5889171	5683484	3.5%
calc_function_1	1272637	1008226	20.8%