View AN1443_842652.PDF datasheet online --- IC-ON-LINE

Datasheet File OCR Text:

AN1443 APPLICATION NOTE
ST100 VIDEO LIBRARY Baseline JPEG Encoding and Decoding on the ST120DSP
By Maurizio Colombo
ABSTRACT This application note shows the results of porting the JPEG application for a SW implementation on ST120DSP. JPEG standard is very generic and includes different techniques. The Baseline method is by far the most widely implemented JPEG method and has thus been chosen for this study. Baseline JPEG includes DCT, Quantization and Huffman encoding from the encoding side, IDCT, Inverse Quantization and Huffman decoding from the decoding side. CONCLUSION The number of clock cycles required to encode one 8x8 block of pixels is 1895, while its decoding requires 2467 cycles. It means that the encoding and the decoding of one QCIF image (176x144) takes respectively 1.12 MCycles and 1.465 MCycles. This analysis is based on very good quality encoding/ decoding, so this number of MCycles can be reduced accordingly to the quality and compression ratio required for the specific application.
October 2001
1/8
AN1443 - APPLICATION NOTE
1 - OVERVIEW ON JPEG ENCODING The processing steps of a Baseline JPEG encoder and decoder are respectively shown in Figures 1 and 2. These are basically a subset of the MPEG functions. Figure 1 : JPEG Encoder
DCT
Quantization
Huffman encoding
Figure 2 : JPEG Decoder
Huffman decoding
Inverse Quantization
IDCT
The Discrete Cosine Transform (DCT) is the same algorithm as the one defined for MPEG. This is followed by Quantization, where each coefficient is basically divided by a corresponding weight in a matrix. Then, coefficients are ordered in zigzag scan and Run-length coding is applied (as for MPEG, zero sequences are replaced with their length and the first non-zero coefficient). Finally, Huffman encoding removes the statistical redundancy and the result is packed into a bitstream. The decoding process is represented exactly by the same inverse sequential steps. The quality of the compressed image is depending on the quantization matrix used. In this analysis a quantization matrix for very good quality has been chosen. The source C code used for this implementation has been downloaded from the site ftp.uu.net/graphics/ jpeg [1]. In the next paragraphs the performances of each one of the above functions are described.
2/8
AN1443 - APPLICATION NOTE
2 - THE ENCODER 2.1 - Wang's DCT The DCT algorithm is the same as the one developed for MPEG. Visual quality tests show that there are no visible differences between this algorithm and the one proposed in the source code (jfdctint.c). Zigzag scan is directly performed in column DCT to allow run-length calculations into quantization (see next paragraph). The results of the profiling are shown in Table 1. Table 1 : Number of clock cycles and code size for one 8x8 block DCT
Function DCT Cycles 314 Code Size 2692
Regarding data memory usage, DCT needs 128 bytes for the block (however, to store intermediate results, between row and column processing, 128 bytes of stack are needed). For constants 72 bytes are needed. 2.2 - Quantization Quantization consists in dividing each coefficient by the corresponding weight in the quantization table. In this step another operation is performed, to prepare the information for run-length coding. Table 2 shows the performances and code size. Quantization uses 136 bytes for the block (8 bytes are for run-length information) and 256 bytes for the quantization matrices (one for luminance, one for chrominance). Table 2 : Number of clock cycles for Quantization+run-length of one 8x8 block
Function Quantization Cycles 177 Code Size 664
2.3 - Huffman Encoding It consists in coding the run/length couples into Huffman variable length codes. The results, obtained from a simulation on the test image provided with the sources are shown in Table 3. More intensive simulations, executed on real high quality photos[2] are shown in Table 4. About memory usage, VLC needs four look-up tables of 1280 bytes each (5120 bytes total) and a buffer to store the output bitstream (in the implementation a size of 2 KBytes is used but it depends on the DMA policy used). Table 3 : Number of cycles and code size for one 8x8 block encoding (test image)
Function Huffman encoding Cycles 682 Code Size 1452
3/8
AN1443 - APPLICATION NOTE
Table 4 : Number of cycles for one 8x8 block encoding (simulations)
Image Cfive Cucorn Icestorm Seagull Palm Duneprint Glasses Flags Average Cycles 886 1226 2031 992 1331 1847 1043 992 1404
2.4 - The Complete Application The blocks described in the previous paragraphs represent the core of the application. Another part that is worth to be considered is the routine writing headers at the beginning of the image. This part takes 24401 cycles and is executed only once. The rest of the program contains some additional parts which enable to read different image formats (in this case only ppm is considered, but this is not a constraint) to convert chroma formats (RGB to YUV and vice versa...), to write the bitstream into a physical file, to set up global configuration variables. All this has not been optimized for speed, but it has been compiled with the"Optimization for size" option in GP16 (16-bit instruction set) mode to guarantee code density. Depending on the embedded application targeted this can be customized: a MIPS-intensive application like a Printer one takes care only of performances, whereas a Wireless application targets low power consumption and small memory requirements without looking at speed (a typical application is encoding/ decoding of static images taken from the web on a 3G cellular phone). The following Table 5 summarizes the program/data memory requirements for the whole application. Table 5 : Memory requirements for the whole application
Section Startup Code Libraries Read-only data Data Zero init data Stack Heap Size 256 31234 32450 8880 412 9100 3072 application-dependent
4/8
AN1443 - APPLICATION NOTE
3 - THE DECODER 3.1 - Huffman Decoding This function reads data from the input bitstream and decodes the DCT quantized coefficients (they are written in raster order). Table 7 represents the results obtained by a simulation on the test image provided with the sources. The results of intensive simulations made on real photos[2] are shown in Table 6. Table 6 : Cycles for one 8x8 block (real photos simulations)
Image Cfive Cucorn Icestorm Seagull Palm Duneprint Glasses Flags Average Cycles 1375 1902 3086 438 2121 2806 1619 1565 1864
Table 7 : Cycles and code size for one 8x8 block (test image)
Function Huffman decoding Cycles 1073 Code SIze 2412
About data memory requirements, 4KB are needed for the internal buffer (It is customizable) and four look-up tables of 1564 bytes each. It means 10352 bytes. 128 bytes are needed to write the output block to be de quantized. 3.2 - Inverse Quantization This consists simply in multiplying each coefficient by its corresponding weight in the quantization matrix. Each coefficient is represented with 12-bit precision, while the weights are represented by 8 bits. So, all multiplications are 16x16. The results are shown in Table 8. As data memory128 bytes are needed for the block and 128 bytes for the two quantization tables. Table 8 : Cycles for one 8x8 block inverse quantization
Function Inverse Quant Cycles 71 Code Size 148
3.3 - Wang's Inverse DCT The same algorithm used for MPEG2 has been implemented here. This includes IDCT, clipping and add128. The input data are read in raster order, because zigzag is performed directly into Huffman decoding. The results are reported in Table 9. The required data memory is 128 bytes for the block and 30 bytes for the constants. Table 9 : Clock cycles and code size for one 8x8 block IDCT
Function IDCT Cycles 532 Code Size 1108 5/8
AN1443 - APPLICATION NOTE
3.4 - The Complete Application As for the encoder, the complete application includes other sections which are tools around the application core, represented by IDCT, IQ and Huffman decoding. Again, these modules have been compiled with the "Optimization for size" (-OS) option in GP16 mode and thus do not represent the maximum performance achievable by the ST120 DSP. A summary of memory requirements for the whole application is reported in the following Table 10. Table 10 : Memory requirements for the whole application
Section Startup Code Libraries Read-only data Data Zero init data Stack Heap Size 256 31234 31890 7552 328 12236 3072 application-dependent
6/8
AN1443 - APPLICATION NOTE
4 - CONCLUSIONS In the following Table 11 is reported a summary of the results presented in this document. For each function are shown the number of clock cycles for one 8x8 block, the code size and the data memory usage. Table 11 : Summary
Function DCT Quantization Huffman encoding IDCT Inv quant Huffman decoding Cycles 314 177 1404 532 71 1864 Code size 2692 664 1452 1108 148 2412 Data Memory 200 392 7168 158 256 10352
REFERENCES [1] Thomas G. Lane, Independent JPEG Group's Software, ftp.uu.net/graphics/jpeg [2] Kodak Digital Images Offering, www.kodak.com/digitalImaging/samples/
7/8
AN1443 - APPLICATION NOTE
Information furnished is believed to be accurate and reliable. However, STMicroelectronics assumes no responsibility for the consequences of use of such information nor for any infringement of patents or other rights of third parties which may result from its use. No license is granted by implication or otherwise under any patent or patent rights of STMicroelectronics. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all information previously supplied. STMicroelectronics products are not authorized for use as critical components in life support devices or systems without express written approval of STMicroelectronics. The ST logo is a registered trademark of STMicroelectronics (c) 2001 STMicroelectronics - All Rights Reserved STMicroelectronics GROUP OF COMPANIES Australia - Brazil - Canada - China - Finland - France - Germany - Hong Kong - India - Israel - Italy - Japan - Malaysia - Malta - Morocco Singapore - Spain - Sweden - Switzerland - United Kingdom - United States http://www.st.com
8/8

▲Up To Search▲

Price & Availability of AN1443

	To Download AN1443 Datasheet File
If you can't view the Datasheet, Please click here to try to view without PDF Reader .