4.3. SIMDization by loop collapse

4.3.1. Target for this tuning

The target for tuning in this section is the function “make_function_6”, which is in the measurement region “Making the system of equations”. In the initial version of the Application, the cost of this function was 1.2% of that of the entire Application.

4.3.2. Analysis

The following nested loop was selected as a target after the analysis of the function “make_function_6”. The key points of this source code are as follows:

• The loop (1) in the following source code was not SIMDized.

• The iteration count of each loop (1) and (2) was 5 (constant with this model).

[Some lines from function “make_function_6” before this tuning was performed]

4.3.3. Tuning

The following tuning was performed.

• The nested loop was divided into two loops (3) and (4) as seen in the following source code, whose iteration count was 5 and 25 respectively. The latter one was composed of the nested loop by “i” and “j”, but collapsed manually.

• In order to facilitate the SIMDization, the OCL for suppressing loop unrolling was specified for the collapsed loop.

[Some lines from function “make_function_6” after this tuning was performed]

4.3.4. 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 “make_function_6” by 31.2% and the entire Application cost by 2.1%.

[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	1577140	1544039	2.1%
make_function_6	113984	78416	31.2%