Reversible variable length coding

Abstract

An efficient decoding of reversible variable length codes of the type implemented for H.263++ and MPEG-4 DCT coefficients using lookup tables. Simple-profile MPEG-4 also has the option of using reversible variable length codewords (RVLC) that can be decoded in either the forward or backward direction, for error recovery. Because of the special structure of RVLC codewords, the decoding techniques for regular VLC are less efficient with RVLC. A new method for RVLC decoding is described.

Description

BACKGROUND

[0002] The present application relates to information encoding for transmission over noisy channels and/or storage, and more particularly to error resilient coding.

[0003] The current rapid expansion of digital communication (speech, video, data) relies on increasingly economical digital signal processing. For example, video communication has general functionality as illustrated in FIG. 8a, and increasingly includes a link through the air interface (FIG. 8b) which can introduce noise and bit errors to the digital signal. Attempts to mitigate bit errors include the use of reversible codes as described in the following.

[0004] Commonly used video compression methods (e.g., MPEG) have block-based motion compensation to remove temporal redundancy (code only (macro)block motion vectors plus the corresponding quantized DCT residuals (texture) as in FIG. 8c) and use variable length coding (VLC) to increase coding efficiency. However, variable length coding often is highly susceptible to transmission channel errors, and a decoder easily loses synchronization with the encoder when uncorrectable errors arise. Further, the predictive nature of motion compensation makes matters much worse because the uncorrectable errors in one video frame quickly propagate across the entire video sequence and rapidly degrade the decoded video quality.

[0005] The typical approach to such uncorrectable errors includes the steps of: error detection (e.g., out-of-range motion vectors, invalid VLC table entry, or invalid number of residuals in a block), resynchronization of the decoder with the encoder, and error concealment by repetition of previously transmitted correct data in place of the uncorrectable data. For example, video compressed using MPEG1-2 has a resynchronization marker (start code) at the start of each slice of macroblocks (MBs) of a frame, and an uncorrectable error results in all of the data between correctly decoded resynchronization markers being discarded. This implies degradation in quality of the video stream.

[0006] VLC tables prove to be particularly sensitive to bit errors because bit errors can make one codeword be incorrectly interpreted to be another codeword of a different length, and the error is not detected. This makes the decoder lose synchronization with the encoder. Although the error may finally be detected due to an invalid VLC table entry, usually the location in the bitstream where the error is detected is not the same as the location where the error occurred. Hence, when the decoder detects an error, it has to seek the next resynchronization marker and discard all the data between this and the previous resynchronization marker. Thus, even a single bit error can sometimes result in a loss of a significant amount of data, and this is a problem of the known coding schemes.

[0007] Enhanced error concealment properties for motion compensated compression, such as MPEG, can be achieved by using data partitioning. Consider a “video packet” to consist of the data between two consecutive resynchronization markers. In a data partitioning approach, the motion data and the texture (DCT) data within each of the video packets are separately encoded in the bitstream. Another resynchronization word (Motion Resync. Word) is imbedded between the motion data and the DCT data to signal the end of the motion data and the beginning of the DCT data. This data partitioning allows the decoder to use the motion data even if the DCT data is corrupted by undetectable errors. This provides advantages including partial recovery over uncorrectable error in a packet of compressed video data with little additional overhead. The error concealment that is made possible by the use of motion compensation by applying decoded motion vectors results in a much better decoded video quality. And this extends to video packets for intra-coded frames in that the DCT dc coefficients can be separated from the other, less important texture data (DCT ac coefficients) by a DC resynchronization word.

[0008] When using data partitioning the data within the video packet is organized to look as shown in FIGS. 6a-c: FIG. 6a shows the fields between two resynchronization markers and FIGS. 6b-c illustrate the motion data field and the texture data field in more detail by an example. In particular, the first field (“Resynch Marker”) is a resynchronization marker, the second field (“MB No.”) is the the number in the frame of the first macroblock (16×16 block of pixels) in the video packet, the third field (“QP”) is the default quantization parameter used to quantize the texture data (DCT coefficients) in the video packet, the fourth field (“Motion Data”) is the motion data, the fifth field (“Motion Resynch Word”) is the resynchronization marker between the motion data and the texture data, the sixth field (“DCT Data”) is the texture data, and the last field (“Resynch Marker”) is the ending resynchronization marker.

[0009] FIG. 6 b shows the motion data field consisting of a COD field, an MCBPC field, and an MV field for each of the macroblocks in the packet. The COD field indicates whether the macroblock is coded or skipped (COD=0 macroblock is coded, COD=1 macroblock is skipped). The MCBPC field indicates (1) the mode of the macroblock and (2) which of the chrominance blocks in the macroblock are coded and which are skipped: the mode indicates whether the current macroblock is coded INTRA (no motion compensation), INTER (motion compensated with one 16×16 motion vector), or INTER4V (motion compensated with four 8×8 motion vectors). Of course, if COD indicates the macroblock is not coded, then the MCBPC field is not present. The MV field is the actual motion vector data; either one vector or four vectors. Again, if COD indicates that the macroblock is not coded, then the MV field is not present. FIG. 6c shows the texture (DCT Data) field as consisting of a CBPY field and a DQUANT field for each of the macroblocks followed by the DCT data for each of the macroblocks. The CBPY field indicates which of the luminance blocks of the macroblock are coded and which are skipped. The DQUANT field indicates the differential increment to the default quantizer value (QP) to compute the quantization value for the macroblock. The DCT fields are run-length-encoded quantized DCT coefficient values of the macroblock.

[0010] MPEG-4 has three kinds of VLCs to encode the DCT coefficients: Table B-16 for encoding INTRA macroblocks, Table B-17 for coding INTER macroblocks, and Table B-23 which is used for coding macroblocks if reversible variable length codes (RVLC) are used. In contrast, H.263 uses only one table for encoding both the INTER and INTRA MBs: Table 13/H263. Table 13/H263 is identical to Table B-17 of MPEG-4.

[0011] Decoding normal VLCs (Table B-16/B-17 of MPEG-4 and Table 13/H263 of H.263) is done using identical techniques, thus consider the decoding of VLCs from Table 13/H263 of H.263. The length of VLC codewords in Table 13/H263 varies from 3 to 13 bits. The last bit is always a sign bit and is not used in variable length decoding, so the number of decodable bits varies up to 12. Variable length decoding is typically carried out in most of the standard decoders by using two different tables: DCT3DtabXval which contains the entries of Table 13/H263 and DCT3DtabXlen which contains the length of the corresponding codewords. Since the length of the VLC codeword is not known in advance, the fastest way to decode a VLC would be by using 212 (4096) entries in DCT3DtabXval. If 212 elements are used in DCT3DtabXval, then 12 bits from the bitstream can be read and be directly used to index into DCT3DtabXval to obtain the decoded values. The same 12 bits are indexed into DCT3DtabXlen to obtain the length of the codeword. The initial bits in the bitstream corresponding to the length of the codeword are then discarded and the process is repeated on the remaining bits plus the next bits in the bitstream up to 12 bits. Note that there are only 102 entries in Table 13/H263. Hence the DCT3DtabXval table and the DCT3DtabXlen table in sequential memory with index addresses would contain many duplicate entries. To conserve on memory, one may split DCT3DtabXval/DCT3DtabXlen based on the number of leading zeros present in the 12 bits, but this increases index complexity.

[0012] Reversible variable length codes (RVLC) are designed such that they can be decoded in either the forward or backward direction; see FIGS. 1-4 illustrating error detection possibilities. These codes are useful when the data may be corrupted by errors. MPEG-4 video includes the option to use RVLC for the DCT data. The RVLC for MPEG-4 is formed by concatenating a variable-length code (VLC) with a fixed-length code (FLC) where the FLC part has 2 bits, including the sign bit. The VLC part either starts and ends with a 1 with all, if any, zeroes in between; or starts and ends with a 0 with exactly one 0 in between (the rest of the bits, if any, equal 1). That is, the VLC part is either 100 . . . 001 (with possibly no 0s) or 01 . . . 101 . . . 10 with position of the interior 0 anywhere among the 1s. The longest valid RVLC codeword is 15 bits plus a sign bit, so the VLC part can be as long as 14 bits. Contrarily, the VLC part can be as short as 2 bits (11 if starting with a 1) or 3 bits (000 if starting with a 0). Because of the way the RVLC is designed, the codebook is very sparse. Also, not all RVLC codewords with this structure are used; see the following table which shows codewords 01 1111 1101 1110 through 01 1111 1111 1100 are not used among the 14-bit VLC part. Codewords are defined for the 169 most commonly occurring events (combinations of last, level, and run), and an escape codeword for all other cases. The RVLC structure does not lend itself to the type of table lookup strategy that is used for regular VLCs. With the typical VLC lookup strategy, 15 bits (without the sign bit) would be read from the bitstream. A single table would require 215=32K (0x0000 to 0x7FFF in hexadecimal) entries, or the table could be partitioned according to the number of leading zeroes. However, for the MPEG-4 RVLC, this doesn't help much: there are 24 codewords beginning with a 1, 22 codewords with two leading 0's, only two codewords with three or more leading 0's, and all other 122 codewords have exactly one leading 0. The lookup table for codewords with exactly one leading 0 would have to cover indices 010 00xx xxxx xxxx through 011 1111 0111 1101, requiring 0x3F7D−0x2000=0x1F7D (=8061 decimal) entries for the 122 codewords that begin with a single leading 0. Thus, the normal VLC decoding approach is very inefficient for RVLC decoding. Because RVLC codewords require so much memory to use the typical VLC decoding approach, the known MoMuSys software uses a brute force approach for decoding RVLC of MPEG-4 Table B-23 as follows. A test is made, comparing with each possible codeword (not counting the sign bit). When the match is found, the index into a densely-packed lookup table is hard-coded. The MoMuSys RVLC decode is implemented using one gigantic case statement. For efficiency, it is best to test for the shorter, more common codewords first. However, the worst-case cycles can be quite high, if all 169 cases are tested before finding a match. And this is not sufficiently efficient for effective use of RVLC decoding with MPEG-4.

SUMMARY OF THE INVENTION

[0013] The present invention uses a codeword hashing index to access reversible VLC (RVLC) tables such as in MPEG and H.263..

[0014] This has the advantage of better performance and smaller memory requirements for RVLC decoding.

BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings are heuristic for clarity.

[0016] FIGS. 1-4 shows error detection with reversible coding.

[0017] FIG. 5 illustrates the fixed length code for events outside of the preferred embodiment code tables.

[0018] FIGS. 6 a -6c indicate video bitstream structure.

[0019] FIG. 7 is a flow diagram.

[0020] FIGS. 8 a -8c illustrate systems.

[0021] FIG. 9 shows a header structure.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0022] Overview

[0023] The preferred embodiments store/decode the reversible variable length codes of MPEG-4 Table B-23 (used for DCT coefficients) with a hash-type function of the codeword providing an index (address) for compact memory storage and efficient decoding. Thus preferred embodiment systems such as teleconferencing or Internet video reception would have a receiver with a memory storing the Table B-23 in a preferred embodiment indexed manner and available for use for decoding received video.

[0024] Note that these video systems and decoding methods may be implemented on special integrated circuits or on programmable microprocessors or digital signal processors such as the TMS320C5000 or C6000 families from Texas Instruments.

[0025] First Preferred Embodiments

[0026] Preliminarily consider modification of the MoMuSys code for RVLC decoding to reduce the worst-case computation. Because of the RVLC structure, it is possible to find the end of the codeword during parsing. The codeword can be extracted without any trailing “don't care” bits before it is decoded. The worst-case computation can be reduced by testing for ranges of code values before testing for matches. For instance, the first preferred embodiments modify the MoMuSys code to test according to the position of the MSB (after finding the end of the codeword and storing it in a separate memory location), i.e.,

if (code < 0 × 8) { }
else if (code < 0 × 10) { }
else if (code < 0 × 20) { }
else if (code < 0 × 40) { }
. . .
else if (code < 0 × 2000) { }
else { }

[0027] This approach breaks the codewords into 12 classes of similar size. But even after the codewords are broken into these 12 classes, a table lookup is not memory-efficient. For example, the 24 codewords between 0x1000 and 0×2000 range from 0x1002 to 0x1FF9, which would require a table of size 0x1FF9−0x1002=4087 (decimal) entries. Thus, the code is written to test for matches against every possible codeword, but some matches are skipped by dividing the codewords into classes. The worst case requires 35 “if” tests (compared to 169), or one more, for an invalid codeword. Note, similar techniques could be applied for backward decoding of RVLC data.

[0028] Note that the events coded for an 8×8 block of DCT coefficients are a combination of three fields: a 1-bit LAST field which indicates whether the coefficient is the last nonzero coefficient in the block, a 6-bit RUN which indicates the number of successive zeros preceding the coded coefficient, and an 11-bit LEVEL which is the nonzero value of the coded coefficient. As illustrated in FIG. 5 a 5-bit ESCAPE code (0000x) begins and ends the event code with the last bit of the end ESCAPE code reflecting the sign of the nonzero coefficient. Table B-23 provides RVLC for the 169 most common events plus the ESCAPE for less common events.

[0029] The sparse nature of RVLC codewords makes it inefficient to use table lookups indexed by the codeword itself, but the structure of the codewords is used by the preferred embodiments to define a sort of hashing function to access a table that can be more densely packed. In particular, the preferred embodiments define an index (for a codeword table) based on the codeword itself as follows.

[0030] (1) For codewords starting with 1 (VLC part of form 100 . . . 001), form an index into a separate table of such codewords:

[0031] Index=2*(number of 0's in VLC part)+first bit (non-sign bit) of FLC part. Thus this index runs from 0 to 23. The index can also be expressed as a 5-bit concatenation of 4 bits to indicate the up-to-eleven 0's in the VLC part followed by the 1 bit equal to the non-sign bit of the FLC part. And the indexing running from 0 to 23 implies a 24-word block of memory can contain each of the inter and intra tables.

[0032] (2) For codewords starting with 0 (VLC part of form 01 . . . 101 . . . 10), similarly form an index into a table for such codewords:

[0033] Index=32*(number of 1's in VLC part)+2*(number of 1's preceding the interior 0 in VLC part)+first bit (non-sign bit) of FLC.

[0034] Thus this index runs from 0 to 365 and can be expressed as a 9-bit concatenation of 4 bits indicating the up to eleven interior 1's of the VLC part followed by 4 bits indicating the up to ten 1's preceding the interior 0 in the VLC part and completed by the 1 bit equal to the non-sign bit of the FLC part. The index running from 0 to 365 implies a 366-word block of memory can contain each of the inter and intra tables.

[0035] The computation of the appropriate index can be done fairly efficiently as a byproduct of the parsing, especially on devices (such as DSPS) which offer a single-cycle EXP-like instruction to count the number of leading 0's or 1's (after stripping off the starting bit). The following table shows the MPEG-4 RVLC table B-23, with the corresponding preferred embodiment “hash” indices shown beside the codeword for the 0-starting and 1-starting codewords.

B-23	intra	inter		1-start	0-start
index	LAST	RUN	LEVEL	LAST	RUN	LEVEL	BITS	hash Index	hash Index	RVLC_CODE
0	0	0	1	0	0	1	4	0		110s
1	0	0	2	0	1	1	4	1		111s
2	0	1	1	0	0	2	5		1	0001s
3	0	0	3	0	2	1	5	2		1010s
4	1	0	1	1	0	1	5	3		1011s
5	0	2	1	0	0	3	6		32	00100s
6	0	3	1	0	3	1	6		33	00101s
7	0	1	2	0	4	1	6		34	01000s
8	0	0	4	0	5	1	6		35	01001s
9	1	1	1	1	1	1	6	4		10010s
10	1	2	1	1	2	1	6	5		10011s
11	0	4	1	0	1	2	7		64	001100s
12	0	5	1	0	6	1	7		65	001101s
13	0	0	5	0	7	1	7		66	010100s
14	0	0	6	0	8	1	7		67	010101s
15	1	3	1	1	3	1	7		68	011000s
16	1	4	1	1	4	1	7		69	011001s
17	1	5	1	1	5	1	7	6		100010s
18	1	6	1	1	6	1	7	7		100011s
19	0	6	1	0	0	4	8		96	0011100s
20	0	7	1	0	2	2	8		97	0011101s
21	0	2	2	0	9	1	8		98	0101100s
22	0	1	3	0	10	1	8		99	0101101s
23	0	0	7	0	11	1	8		100	0110100s
24	1	7	1	1	7	1	8		101	0110101s
25	1	8	1	1	8	1	8		102	0111000s
26	1	9	1	1	9	1	8		103	0111001s
27	1	10	1	1	10	1	8	8		1000010s
28	1	11	1	1	11	1	8	9		1000011s
29	0	8	1	0	0	5	9		128	00111100s
30	0	9	1	0	0	6	9		129	00111101s
31	0	3	2	0	1	3	9		130	01011100s
32	0	4	2	0	3	2	9		131	01011101s
33	0	1	4	0	4	2	9		132	01101100s
34	0	1	5	0	12	1	9		133	01101101s
35	0	0	8	0	13	1	9		134	01110100s
36	0	0	9	0	14	1	9		135	01110101s
37	1	0	2	1	0	2	9		136	01111000s
38	1	12	1	1	12	1	9		137	01111001s
39	1	13	1	1	13	1	9	10		10000010s
40	1	14	1	1	14	1	9	11		10000011s
41	0	10	1	0	0	7	10		160	001111100s
42	0	5	2	0	1	4	10		161	001111101s
43	0	2	3	0	2	3	10		162	010111100s
44	0	3	3	0	5	2	10		163	010111101s
45	0	1	6	0	15	1	10		164	011011100s
46	0	0	10	0	16	1	10		165	011011101s
47	0	0	11	0	17	1	10		166	011101100s
48	1	1	2	1	1	2	10		167	011101101s
49	1	15	1	1	15	1	10		168	011110100s
50	1	16	1	1	16	1	10		169	011110101s
51	1	17	1	1	17	1	10		170	011111000s
52	1	18	1	1	18	1	10		171	011111001s
53	1	19	1	1	19	1	10	12		100000010s
54	1	20	1	1	20	1	10	13		100000011s
55	0	11	1	0	0	8	11		192	0011111100s
56	0	12	1	0	0	9	11		193	0011111101s
57	0	6	2	0	1	5	11		194	0101111100s
58	0	7	2	0	3	3	11		195	0101111101s
59	0	8	2	0	6	2	11		196	0110111100s
60	0	4	3	0	7	2	11		197	0110111101s
61	0	2	4	0	8	2	11		198	0111011100s
62	0	1	7	0	9	2	11		199	0111011101s
63	0	0	12	0	18	1	11		200	0111101100s
64	0	0	13	0	19	1	11		201	0111101101s
65	0	0	14	0	20	1	11		202	0111110100s
66	1	21	1	1	21	1	11		203	0111110101s
67	1	22	1	1	22	1	11		204	0111111000s
68	1	23	1	1	23	1	11		205	0111111001s
69	1	24	1	1	24	1	11	14		1000000010s
70	1	25	1	1	25	1	11	15		1000000011s
71	0	13	1	0	0	10	12		224	00111111100s
72	0	9	2	0	0	11	12		225	00111111101s
73	0	5	3	0	1	6	12		226	01011111100s
74	0	6	3	0	2	4	12		227	01011111101s
75	0	7	3	0	4	3	12		228	01101111100s
76	0	3	4	0	5	3	12		229	01101111101s
77	0	2	5	0	to	2	12		230	01110111100s
78	0	2	6	0	21	1	12		231	01110111101s
79	0	1	8	0	22	1	12		232	01111011100s
80	0	1	9	0	23	1	12		233	01111011101s
81	0	0	15	0	24	1	12		234	01111101100s
82	0	0	16	0	25	1	12		235	01111101101s
83	0	0	17	0	26	1	12		236	01111110100s
84	1	0	3	1	0	3	12		237	01111110101s
85	1	2	2	1	2	2	12		238	01111111000s
86	1	26	1	1	26	1	12		239	01111111001s
87	1	27	1	1	27	1	12	16		10000000010s
88	1	28	1	1	28	1	12	17		10000000011s
89	0	10	2	0	0	12	13		256	001111111100s
90	0	4	4	0	1	7	13		257	0011111111010
91	0	5	4	0	2	5	13		258	010111111100s
92	0	6	4	0	3	4	13		259	010111111101s
93	0	3	5	0	6	3	13		260	011011111100s
94	0	4	5	0	7	3	13		261	011011111101s
95	0	1	10	0	11	2	13		262	011101111100s
96	0	0	18	0	27	1	13		263	011101111101s
97	0	0	19	0	28	1	13		264	011110111100s
98	0	0	22	0	29	1	13		265	011110111101s
99	1	1	3	1	1	3	13		266	011111011100s
100	1	3	2	1	3	2	13		267	0111110111010
101	1	4	2	1	4	2	13		268	011111101100s
102	1	29	1	1	29	1	13		269	011111101101s
103	1	30	1	1	30	1	13		270	011111110100s
104	1	31	1		31	1	13		271	011111110101s
105	1				32	1			272	011111111000s
106	1	33	1	1	33	1	13		273	011111111001s
107	1	34	1	1	34	1	13	18		100000000010s
108	1	35	1	1	77	1	13	19		100000000011s
109	0	14	1	0	0	13	14		288	0011111111100s
110	0	15	1	0	0	14	14		289	0011111111101s
111	0	11	2	0	0	15	14		290	0101111111100s
112	0	8	3	0	0	16	14		291	0101111111101s
113	0	9	3	0	1	8	14		292	0110111111100s
114	0	7	4	0	3	5	14		293	0110111111101s
115	0	3	6	0	4	4	14		294	0111011111100s
116	0	2	7	0	5	4	14		295	0111011111101s
117	0	2	8	0	8	3	14		296	0111101111100s
118	0	2	9	0	12	2	14		297	0111101111101s
119	0	1	11	0	30	1	14		298	0111110111100s
120	0	0	20	0	31	1	14		299	0111110111101s
121	0	0	21	0	32	1	14,		300	0111111011100s
122	0	0	23	0	33	1	14		301	0111111011101s
123	1	0	4	1	0	4	14		302	0111111101100s
124	1	5	2	1	5	2	14		303	0111111101101s
125	1	6	2	1	6	2	14		304	0111111110100s
126	1	7	2	1	7	2	14		305	0111111110101s
127	1	8	2	1	8	2	14		306	0111111111000s
128	1	9	2	1	9	2	14		307	011111111,100ts
129	1	36	1	1	36	1	14	20		1000000000010s
130	1	37	1	1	37	1	14	21		1000000000011s
131	0	16	1	0	0	17	15		320	00111111111100s
132	0	17	1	0	0	18	15		321	00111111111101s
133	0	18	1	0	1	9	15		322	01011111111100s
134	0	8	4	0	1	10	15		323	01011111111101s
135	0	5	5	0	2	6	15		324	01101111111100s
136	0	4	6	0	2	7	15		325	01101111111101s
137	0	5	6	0	3	6	15		326	01110111111100s
138	0	3	7	0	6	4	15		327	01110111111101s
139	0	3	8	0	9	3	15		328	01111011111100s
140	0	2	10	0	13	2	15		329	01111011111101s
141	0	2	11	0	14	2	15		330	01111101111100s
142	0	1	12	0	15	2	15		331	01111101111101s
143	0	1	13	0	16	2	15		332	01111110111100s
144	0	0	24	0	34	1	15		333	01111110111101s
145	0	0	25	0	35	1	15		334	01111111011100s
146	0	0	26	0	36	1	15		335	01111111011101s
147	1	0	5	1	0	5	15		336	01111111101100s
148	1	1	4	1	1	4	15		337	0111111110110ts
149	1	10	2	1	10	2	15		338	01111111110100s
150	1	11	2	1	11	2	15		339	01111111110101s
151	1	12	2	1	12	2	15		340	01111111111000s
152	1	38	1	1	38	1	15		341	01111111111001s
153	1	39	1	1	39	1	15	22		10000000000010s
154	1	40	1	1	40	1	15	23		10000000000011s
155	0	0	27	0	0	19	16		352	00111111111110
156	0	3	9	0	3	7	16		353	001111111111101s,
157	0	6	5	0	4	5	16		354	010111111111100s
158	0	7	5	0	7	4	16		355	010111111111101s
159	0	9	4	0	17	2	16		356	011011111111100s
160	0	12	2	0	37	1	16		357	011011111111101s
161	0	19	1	0	38	1	16		358	011101111111100s
162	1	1	5	1	1	5	16		359	011101111111101s
163	1	2	3	1	2	3	16		360	011110111111100s
164	1	13	2	1	13	2	16		361	011110111111101s
165	1	41	1	1	41	1	16		362	011111011111100s
166	1	42	1	1	42	1	16		363	011111011111101s
167	1	43	1	1	43	1	16		364	011111101111100s
168	1	44	1	1	44	1	16		365	011111101111101s
169	ESCAPE	5		0	0000s

[0036] The foregoing table would be stored as two separate tables: one for the 1-starting codewords and having a 5-bit index running from 0 to 23 and one for the 0-starting codewords with a 9-bit index running from 0 to 366.

[0037] FIG. 7 illustrates the decoding flow.

[0038] Other Preferred Embodiments

[0039] Second preferred embodiments modify the foregoing hash indices in the MPEG-4 B-23 table by combining the 0-start and 1-start indices to a single index:

[0040] Index=32*(number of bits in VLC part)+2*(the position number of the first repeat of the first bit in the VLC part)+non-sign bit of FLC−68.

[0041] This index runs from 0 to 409: the number of bits in the VLC part ranges from 2 to 14 (so the first number can be expressed with four bits), the position of the first repeat is also in the range from 2 to 14 (another four bits), and the last number is one bit. Thus a concatenated 9-bit index suffices. Note that the number of bits in the VLC part equals the position of the first repeat for the codewords starting with a 1; but this is not possible for the codewords starting with a 0. Also, the number of bits and the position number both have a minimum of 2, so the −68 normalizes the index to have a minimum of 0.

[0042] Third preferred embodiments fold the 0-starting codeword table (subtable of foregoing MPEG-4 table) over onto itself to fill in the unused spaces. That is, the Index has an unused gap between 1 and 32, so the 14 Index values 352-365 can be mapped into this gap; similarly the unused gap between 35 and 64 can be filled with the 22 Index values 320-341, the unused gap between 69 and 96 can be filled with the 20 Index values 288-307, and so forth up to the gap between 171 and 192 filled with the 14 Index values 192-205. This requires a little more computation:

[0043] Index=32*(number of 1's in VLC part)+2*(number of 1's before interior 0 of VLC part)+non-sign bit of FLC.

[0044] if (Index>=192) Index=383−Index.

[0045] Thus Index values 352-365 map to 18-31, 320-341 map to 42-63, 288-307 map to 76-95, and so forth up to 192-205 map to 178-191. That is, the folded 0-starting codeword table has 146 entries, but the index for the folded table runs from 0 to 191 and thus requires a 192-word memory block for each of the intra and inter tables.

[0046] Fourth Fourth preferred embodiments further reduce the gaps in the folded 0-starting codeword table, using the index: 1$\begin{array}{c}\text{Index}={26}^{*}\text{(number of 1's in}\text{VLC}\text{part)}+\\ 2^{*}\text{(number of 1's before interior 0 of}\text{VLC}\text{part)}+\\ \text{non-sign bit of}\mathrm{FLC}.\end{array}$

[0047] If (Index>=156) Index=311−Index.

[0048] The intermediate Index (before folding) can no longer be formed as the concatenation of three fields, because 26 is not a power of 2, so a little more computation is required. The Index values 286-299 map to 12-25, 260-281 map to 30-51, 234-253 map to 58-77, and so forth up to 156-160 map to 142-155. That is, the compact folded 0-starting codeword table has 146 entries, but the index for the compact folded table runs from 0 to 155 and thus requires a 156-word memory block for each of the intra and inter tables.

[0049] Fifth preferred embodiments further reduce the gaps in the nonfolded 0-starting codeword table, using the index: 2$\begin{array}{c}\text{Index}={22}^{*}\text{(number of 1's in}\mathrm{VLC}\mathrm{part})+\\ 2^{*}\text{(number of 1's before interior 0 of}\mathrm{VLC}\text{part)}+\\ \text{non-sign bit of}\mathrm{FLC}.\end{array}$

[0050] This Index can not be formed as the concatenation of 3 fields, because 22 is not a power of 2, so a little more computation is required, (but not as much as the compact folded approach). The index running from 0 to 255 implies a 256-word block of memory can contain each of the inter and intra tables.

[0051] In summary, the preferred embodiment indexings and table(s) for MPEG4 B-23 (RVLC for DCT coefficients) reduce the code size significantly, as well as the number of cycles required for execution compared to using a gigantic case statement or nested if-statements, particularly in the worst case. The C code reduces from roughly 900 lines to about 150 lines, at the expense of increasing table storage from 2*170=340 words (inter and intra tables) to 2*24+2*366=780 words. If the 0-start table is folded over, only 2*24+2*192=432 words are required. If the 0-start table is compacted (without folding), only 2*24+2*256=560 words are required. If the 0-start table is folded over and compacted, only 2*24+2*156=360 words are required.

[0052] H.263 preferred embodiments

[0053] The H.263++ standard has an optional data partitioning mode in which the MBs in a video picture segment (packet) are rearranged analogously to MPEG4 so that header information for all MBs in the segment is followed by MVs for all MBs in the segment and then followed by DCT coefficients for all MBs. The header, MV, and DCT partitions are separated by markers, allowing for resynchronization when the end of the partition in which an error occurred is reached; see FIG. 9 illustrating the HD (header data), HM (header marker), MVD (motion vector data), LMW (last MV value), MM (MV marker), and DCT (DCT data) fields.

[0054] The HD field includes reversible variable length code (RVLC) coded data for combined COD and MCBPC for all MBs in a packet. The standard has five tables which apply to the cases of INTRA MBs, INTER MBs, B MBs, EP MBs, and EI MBs, respectively. The INTER MBs table (see below) is the largest of the five and has 26 codewords; the other four tables are subsets. The codewords of the INTER MBs table have the symmetrical form 0 . . . 01 . . . 10 . . . 0 for coded MBs, and for a non-coded MB the codeword is simply 1. Thus both the first and last bits of the codeword can be interpreted as the COD field. The codewords vary from one to eleven bits, with 1 the smallest and 01111111110 the largest. Thus indexing by the codeword itself would spread the codewords out in a block of 1023 memory addresses. Partitioning into classes according to the number of initial 0s does not help much because the codewords starting with a single 0 run from 010 to 01111111110, a span of 1022.

[0055] The preferred embodiment Index for the INTER MBs table again is a hash type function of the codeword:

[0056] (1) For the codeword starting with 1, take the Index equal to 0.

[0057] (2) For codewords starting with 0, form an index as:

[0058] Index=10*(number of initial 0's minus 1)+(number of 1's).

[0059] Thus this index runs from 0 to 41. The index running from 0 to 41 implies a 42-word block of memory can contain the table. Indeed, the 26 Index values are 0-9, 11-17, 21-25, 31-33, and 41.

		Codeword		Number
MB type	CBPC	COD + MCBPC	Index	of bits
		1	0	1
0 (INTER)	00	010	1	3
0	10	00100	11	5
0	01	011110	4	6
0	11	0011100	13	7
1(INTER+Q)	00	01110	3	5
1	10	00011000	22	8
1	01	011111110	7	9
1	11	01111111110	9	11
2(INTER4V)	00	0110	2	4
2	10	01111110	6	8
2	01	00111100	14	8
2	11	000010000	31	9
3(INTRA)	00	001100	12	6
3	11	0001000	21	7
3	10	001111100	15	9
3	01	000111000	23	9
4(INTRA+Q)	00	0111110	5	7
4	11	0011111100	16	10
4	10	0001111000	24	10
4	01	0000110000	32	10
5(INTER4V+Q)	00	00111111100	17	11
5	01	00011111000	25	11
5	10	00001110000	33	11
5	11	00000100000	41	11
Stuffing		0111111110	8	10

[0060] Another preferred embodiment folds the codeword table over onto itself to fill in the unused spaces. That is, the Index has an unused gap between 9 and 11, so the Index values 41 can be mapped into this gap; and the unused gap between 17 and 21 can be filled with the Index values 31-33. This requires a little more computation:

[0061] if (initial bit is 1) Index=0

[0062] else Index=10*(number of initial 0's minus 1)+(number of 1').

[0063] if (Index>=25) Index=51−Index.

[0064] That is, the folded codeword table has 26 entries, but the index for the folded table runs from 0 to 25 and thus just fills the 26 word memory block with addresses corresponding to Index.

[0065] More generally, the various RVLC tables of H.263++ all have symmetric codewords of the type 0 . . . 01 . . . 10 . . . 0 plus the codeword 1, and indexes of the form:

[0066] if (initial bit is 1) Index=0

[0067] else Index=N*(number of initial 0's minus 1)+(number of 1's).

[0068] can be used for any N such that N>=(maximum number of 1's). Thus for INTRA MBs use N>=5, for INTER MBs use N>=9, for B MBs use N>=7, for EP MBs use N>=6, and for El MBs use N>=8. Of course, a larger N implies a larger stretch of memory but also more opportunity for folding.

[0069] For example, the foregoing table (the table for INTER MBs) lists the Index with N=10 where the maximum Index=41 for codeword 00000100000; with the Index based on N=9 this codeword would have the maximum Index=37. But the N=10 Index can fold the Index=41 codeword into the gap between Index=9 (codeword 01111111110) and Index=11 (codeword 00100) by using Index=51−Index if Index>=25. If instead the Index had been based on N=9, then this gap closes because codeword 01111111110 would still have Index=9, but codeword 00100 would now have Index=10. Thus the maximum folding would be to the first gap at Index=17 and would use Index=54−Index when Index>=28. With this folding the maximum Index then decreases from 37 to 26 (for codeword 000010000), which implies a memory block of 27 words and all addresses correspond to Index values except for Index=18.

Claims

1. A method of decoding reversible variable-length codewords, comprising: (a) providing a table of entries, each of said entries including an information portion and a corresponding codeword, said entries indexed by a hash function of the codewords; and (b) upon receipt of a first codeword, applying said hash function to said codeword to yield a first index; and (c) using said first index to access said table and recover the information portion of the entry including said first codeword.

2. The method of claim 1, wherein: (a) said table is an MPEG-4 table.

3. The method of claim 1, wherein: (a) said table is an H.263++ table.