Appendix A. Performance Impact

SPRAA56

A.1 Overhead of Performance Measurement Techniques

Because most of the benchmarking APIs are called once every 30 frames, the additional CPU load expected after adding the instrumentation is low. The measured performance of the benchmarking techniques is given in Table 3. A spreadsheet containing the expected and actual timing values is provided with the software distribution.

Table 3. Measured Performance of Benchmarking Techniques

	Execution Time	Execution Time		Execution Rate
Benchmark	(Avg) [instr]	(Max) [instr]	CPU Load	[per N frames]
MBX check in process Task	3641	17112	0.00018205	1
LOAD module call	1182	2432	0.00000197	30
Single Call to UTL_stsStart	517	13968	0.00043945	16
Single Call to UTL_stsStop	325	488	0.00027625	16
Capture Task benchmarking	1848	15064	0.00000308	30
Display Task benchmarking	2288	7824	3.81333E-06	30
Process Task benchmarking	3196	18568	5.32667E-06	30
Control Task	1533	2856	0.00007665	?
SubTotal Load (Task bchmrk)	7332		0.0003666	30
SubTotal Load (UTL calls)	13472		0.0006736	30
Total Load of benchmarking	17357.4		0.00086787	1

These benchmarks are given in instructions, and the individual CPU load of each function is calculated by dividing the benchmark by 20M instructions per frame, the number of cycles available on a 600 MHz 64x device in a 30 fps NTSC system.

These benchmarks were measured using UTL_stsStart and UTL_stsStop API calls bracketing the regions of code to be benchmarked. For example, to benchmark the LOAD_getcpuload function, the measurement code was the following:

UTL_stsStart( stsBenchmark1 ); benchVid.cpuLoad.current = LOAD_getcpuload();

UTL_stsStop( stsBenchmark1 );

This method of benchmarking allows execution time to be measured in real-time, although if an interrupt or context switch occurs between the UTL calls, the time spent executing the interrupt or out-of-context code would also be included in the benchmark.

A.2 RTA Effects on CPU Load

The CPU load was measured with RTA debugging turned off and the UTL_DBGLEVEL set to

40.The total CPU load of the application with the instrumentation turned off was 93% average and 95% peak. The CPU load of the instrumented application was 93% average and 95% peak when using the same video content, a repeating high-motion sequence from a DVD. The benchmarking did not make a statistically significant impact on the CPU load.

DSP/BIOS Real-Time Analysis (RTA) and Debugging Applied to a Video Application