.H264 File Format – CABAC binarisation and contexts

Introduction

When entropy_coding_mode_flag is 1, every syntax element marked ae(v) in the slice data is coded by context-adaptive binary arithmetic coding (clause 9.3 of ITU-T H.264, 06/2026, version 16). The process has four parts: a binarisation turns the value into a string of bins; each bin gets a context index ctxIdx; the arithmetic engine decodes the bin using the probability state stored for that context; and the state is updated. The starting states of all contexts are on the initial values and 4:4:4 values pages. This page gives the other three parts. The slice data starts after cabac_alignment_one_bits, and the engine is initialised again after the PCM samples of an I_PCM macroblock.

Arithmetic decoding engine (9.3.1.2, 9.3.3.2)

Initialisation. codIRange is 510 and codIOffset is the next 9 bits, most significant first. The stream never produces an offset of 510 or 511.

DecodeBin(ctxIdx). A bypass bin uses DecodeBypass; ctxIdx 276 uses DecodeTerminate; all others use DecodeDecision.

DecodeDecision. Each context holds pStateIdx (0 to 63; 0 is an LPS probability of 0.5, falling with the index) and valMPS.

1. qCodIRangeIdx = (codIRange >> 6) & 3 and codIRangeLPS = rangeTabLPS[pStateIdx][qCodIRangeIdx] (Table 9-44, 9-25, 9-26).
2. codIRange −= codIRangeLPS. If codIOffset ≥ codIRange the bin is 1 − valMPS, codIOffset −= codIRange and codIRange = codIRangeLPS; otherwise the bin is valMPS.
3. State update (9-27): after an MPS, pStateIdx = transIdxMPS[pStateIdx]; after an LPS, if pStateIdx is 0 then valMPS flips, and pStateIdx = transIdxLPS[pStateIdx] (Table 9-45).
4. Renormalisation: while codIRange < 256, codIRange is doubled and codIOffset = (codIOffset << 1) | read_bits(1). After it codIOffset < codIRange always holds.

DecodeBypass. codIOffset = (codIOffset << 1) | read_bits(1); if codIOffset ≥ codIRange the bin is 1 and codIOffset −= codIRange, otherwise 0. No renormalisation.

DecodeTerminate. Used for end_of_slice_flag and for the bin that signals I_PCM. codIRange −= 2. If codIOffset ≥ codIRange the bin is 1, there is no renormalisation and decoding ends; the last bit read into codIOffset is then 1 and, for end_of_slice_flag, is the rbsp_stop_one_bit. Otherwise the bin is 0 and the engine renormalises.

Table 9-44: rangeTabLPS

Columns are qCodIRangeIdx 0 to 3. The same numbers, as far as rows 23, 31 and the transition table's ends were compared, are on the H.265 CABAC page as Table 9-52.

Table 9-45: state transitions

Binarisation processes (9.3.2)

Elements with a prefix and a suffix (UEGk elements, mb_type, coded_block_pattern) index binIdx from 0 in each part.

Binarisation and context offsets by element (Table 9-34)

The residual elements depend on ctxBlockCat and on whether the block is frame or field coded. All are FL with cMax 1 and maxBinIdxCtx 0, except coeff_abs_level_minus1, which is UEG0, unsigned, uCoff 14, maxBinIdxCtx 1 for the prefix and bypass for the suffix. Their offsets:

The ctxBlockCat values (Table 9-42; maxNumCoeff in brackets): 0 luma DC of Intra_16x16 (16), 1 luma AC of Intra_16x16 (15), 2 luma 4×4 (16), 3 chroma DC for 4:2:0 or 4:2:2 (4 × NumC8x8), 4 chroma AC (15), 5 luma 8×8 (64), 6 to 9 the Cb forms of 0, 1, 2 and 5 (4:4:4: DC 16, AC 15, 4×4 16, 8×8 64), 10 to 13 the Cr forms (the same order). Rows with the 8×8 categories (5, 9, 13) use the coded_block_flag offset 1012 and the contexts of ctxIdx 1012 to 1023.

Macroblock type bin strings

I slice mb_type (Table 9-36): bin 0 is 0 for I_NxN; bin 1 is the terminate bin (ctxIdx 276), where 1 means I_PCM. SI slices send the prefix bin 0 for SI itself; any other value sends 1 followed by this table indexed by mb_type − 1. In P and SP slices an intra type sends the prefix 1 and then this table indexed by mb_type − 5; in B slices the prefix 111101 (Table 9-37) and this table indexed by mb_type − 23.

The name I_16x16_a_b_c is the prediction mode a, the chroma coded block pattern b and the luma flag c (Table 7-11). P, SP and B macroblock types (Table 9-37; in P and SP type 4, P_8x8ref0, never occurs):

Sub-macroblock types (Table 9-38):

coded_block_pattern. The prefix is CodedBlockPatternLuma as 4 bins (FL, cMax 15; bin k is bit k, the 8×8 block with luma8x8BlkIdx k). When ChromaArrayType is neither 0 nor 3 a TU suffix with cMax 2 follows for CodedBlockPatternChroma, always, whatever the prefix. mb_qp_delta. The signed value is mapped as in the se(v) code (Table 9-3: 0 → 0, 1 → 1, −1 → 2, 2 → 3, −2 → 4, …) and sent in U binarisation.

Choosing the context index (9.3.3.1)

ctxIdx = ctxIdxOffset + ctxIdxInc. For the elements of Table 9-39, ctxIdxInc depends on binIdx; every bin beyond maxBinIdxCtx uses the context of maxBinIdxCtx. A value with several entries in the table is chosen by the clause named. For the residual elements ctxIdx = ctxIdxOffset + ctxIdxBlockCatOffset(ctxBlockCat) + ctxIdxInc. Bins marked bypass use no context, and ctxIdx 276 is decoded with DecodeTerminate.

Table 9-39: ctxIdxInc by bin

"(B)" means the value comes from the earlier bins (Table 9-41, below). In the rows for offsets 17 and 32 bin 1 is the I_PCM terminate bin of the intra suffix, so the table's ctxIdx 276 stands in the place of a context.

Table 9-41: ctxIdxInc from earlier bins

Neighbour rules

Neighbours A (left) and B (above) are the macroblocks or blocks found by the derivation processes of clause 6.4. Unavailable means outside the picture or slice. For clauses (A) to (K), condTermFlagN is as follows; ctxIdxInc is their sum unless stated.

coded_block_flag. For the neighbouring transform block of the same kind (transBlockN): the luma, Cb or Cr DC block of an Intra_16x16 neighbour (ctxBlockCat 0, 6, 10); for 4×4 and AC categories the block with the same index in the neighbour when its coded block pattern bit is set, taking the 8×8 block when the neighbour used the 8×8 transform; the chroma DC block when CodedBlockPatternChroma is not 0; chroma AC when it is 2; for the 8×8 categories (5, 9, 13) the 8×8 block when the neighbour used the 8×8 transform. Skipped and I_PCM neighbours have none. Then condTermFlagN is 0 when N is unavailable and the current macroblock is inter, when N is available but has no such block and is not I_PCM, or when the current macroblock is intra with constrained_intra_pred_flag 1, N is inter and data partitioning is in use. It is 1 when N is unavailable and the current macroblock is intra, or N is I_PCM. Otherwise it is the coded_block_flag that was decoded for transBlockN.

Residual contexts (9.3.3.1.3)

ctxIdxBlockCatOffset (Table 9-40), by ctxBlockCat:

significant_coeff_flag and last_significant_coeff_flag. ctxIdxInc = levelListIdx (0 to maxNumCoeff − 2) for categories other than 3, 5, 9 and 13 (9-21); for category 3, Min(levelListIdx ÷ NumC8x8, 2) (9-22); for the 8×8 categories 5, 9 and 13, Table 9-43. The list index is the position in the scan of the coefficient levels, and the last coefficient of a block needs no flag.

coeff_abs_level_minus1. Let numDecodAbsLevelEq1 and numDecodAbsLevelGt1 count the levels already decoded in the same block with absolute value 1 and above 1. Bin 0: ctxIdxInc = (numDecodAbsLevelGt1 != 0) ? 0 : Min(4, 1 + numDecodAbsLevelEq1) (9-23). Other prefix bins: ctxIdxInc = 5 + Min(4 − (ctxBlockCat == 3 ? 1 : 0), numDecodAbsLevelGt1) (9-24). The UEG0 suffix and coeff_sign_flag are bypass bins.

Not covered on this page

The informative arithmetic encoder (9.3.4), the flowchart figures, and the macroblock neighbour derivations of clause 6.4 that the context rules call (availability, MBAFF neighbours). The alternative CABAC parsing of the scalable and 3D extensions has its added contexts on the SVC and 3D-AVC pages. Source: ITU-T H.264 (06/2026, version 16), clause 9.3.