OMAP3530 Performance

OMAP3530 - DVSDK 3.00 (Linux)

Environment

• The results were taken using a Mistral OMAP3530 EVM E.S. 3.0 board using DVSDK 3.0.0.39.
• The settings for the system are ARM running at 500MHz and the IVA running at 360MHz.
• Several sources were used for the input file to validate the impact: MMC/SD, RAM & NFS with no noticeable differences.
• The internal LCD was used as the display output.
• No quality analysis was performed for the results.
• H.264 Video Codec: Up to Level 3.0 features of the Baseline Profile [BP] supported.
• MPEG4 Video Decoder: MPEG4 visual Simple Profile [SP] Level 0, 1, 2, 3, and 4A Compliant.
• MPEG4 Video Encoder: MPEG4 visual Simple Profile [SP] Level 0, 1, 2, 3, 4A, 5 and 6 Compliant.

Results

Table 1: Decoding Performance results. (Click on a footnote to read its content)

DECODING PERFORMANCE STATISTICS [ms/frame]
TEST METHOD(*)				DSP ONLY		FILE to FILE		OOB DEMO			GSTREAMER		CONDITIONS
CODEC	VIDEO SIZE	RATE1	MV2	AVG	MAX	AVG	MAX	FPS3	AVG	MAX	AVG	MAX	Clip Used4	PRESET
H264DEC	VGA	5Mbps	4	24.95	28.85	28.26	31.88	N/A	28.58	31.47	N/A	N/A	foreman	Default
H264DEC	D1	5Mbps	16	28.09	29.33	30.91	32.32	N/A	32.77	35.39	N/A	N/A	fire	Default
H264DEC	D1	1.55Mbps	1	20.09	26.66	22.54	28.76	30	24.56	28.73	22.23	28.35	shields	Default
MPEG4DEC	VGA	3Mbps	N/A	11.61	12.47	14.91	21.89	N/A	14.86	16.44	N/A	N/A	foreman	Default
MPEG4DEC	D1	10Mbps	N/A	19.61	26.66	23.67	31.45	30	24.67	181.634	N/A	N/A	HPotter	Default

Table 2: Encoding Performance results. (Click on a footnote to read its content)

ENCODING PERFORMANCE STATISTICS [ms/frame]
TEST METHOD(*)				DSP ONLY		FILE to FILE		OOB DEMO			GSTREAMER		CONDITIONS
CODEC	VIDEO SIZE	RATE1	MV2	AVG	MAX	AVG	MAX	FPS3	AVG	MAX	AVG	MAX	Clip Used4	PRESET
H264ENC	D1	4Mbps	1	30.55	30.83	32.92	33.84	N/A	N/A	N/A	N/A	N/A	fire	Default
H264ENC	D1	4Mbps	4	36.38	36.57	37.64	38.20	N/A	N/A	N/A	N/A	N/A	fire	High Speed

Analysis

Figure 1, depicts the analysis for the different measurement methods used to produce the reports based on the shields_720x480_1MV_1_5Mbps.264 clip.

Figure 1: Analysis of the different applications used for shields_720x480_1MV_1_5Mbps.264.

The orange and cyan lines represent respectively the average and maximum processing frame times when running the analysis to decode a H.264 video clip from the DSP perspective only. These indicators serve as a reference to assess the burden of having a multiprocessor system with shared resources; in this particular case, the File-to-File and GStreamer (in yellow) shows an average increment of 12% in processing time; while the DEMO^* adds a 22% increment with respect to the average baseline.

Figure 2, show the analysis made from the Decode Demonstration application only and present the impact of playing an audio clip (AAC-HE @ 62.2Kbps/16 bit resolution/24KHz Sample Rate) concurrently; the purpose in this case is to show the behavior of the decode demonstration application, and the results of using a buffer window.

Figure 2: DEMO analysis for Video, Audio and Average Window for shields_720x480_1MV_1_5Mbps.264.

For this analysis scenario the Blue line represents the results for the processing time for a video frame when running the OOB demo applications. (As in the previous case, only 100 frames are extracted for visual clarity). The pink data represents the effects of decoding audio concurrently from the point of view of the video processing (for the DEMO^*). Even though a few of the frames go beyond the limit of the allowed time to present the NTSC standard of 29.97FPS, a queue (window) of 8 frames is used after the decoder to smooth out this occurrences and to take advantages of the easy and fast processed frames, the results of this process is represented by the yellow line.

Conclusions

• H.264 presents the highest performance burden from the IVA perspective as it is seen in other processors, and as it was expected. Given the planned application for this platform, decoding should perform acceptably in the current processing model using the CE and clock settings. The out of box decode demonstration application performs fairly well to show the video along audio and OSD with the current clock and window settings.
• H.264 Encoding, doesn’t seem feasible the current settings, since when performing file-to-file transactions the system is already close to the edge (without having the display enabled). Different clock settings might alleviate this problem. As it is stated and expected from the H.264 Encoder Release Notes the High-speed settings perform worse than the default preset due to an increase in the number of Motion Vectors.
• The difference between forcing the Chroma standard between 420 & 422 or using the OSD does not seem to be significant when evaluating the performance in any scenario.
• A queue of 8 buffered frames from the DEMO perspective seems to be a fair balance between latency and frame rate.

Notes

• The DVSDK software version used for this test is 3.00.00.39. The results should be the same for the GA version, in which some documentation changes were introduced (DVSDK 3.00.00.40).