Joint Video Experts Team (jvet) of itu-t sg 6 wp and iso/iec jtc 1/sc 29/wg 11
CE5 related – Arithmetic coding engine (5)
Yüklə 4,04 Mb.
|
- Bu səhifədəki naviqasiya:
- 7.6CE6 related – Transforms and transform signalling (19)
- 7.6.2Secondary transforms
- 7.6.3Shrink transform
7.5CE5 related – Arithmetic coding engine (5)Contributions in this category were discussed Friday 13 July 1600–1700 (chaired by GJS). A CE for the arithmetic coding engine will be done. Throughput issues should be understood and sufficient gain should be shown to justify the change of the engine. JVET-K0273 CE5-related: Implementation considerations for entropy coding engine [F. . Bossen] Modifications are proposed to the entropy coding core engine such as to enable a wider variety of implementations. These newly enabled implementations may be beneficial in both software and hardware. For example, it is claimed that software implementations with reduced per bin cycle counts are enabled. It is asserted that the proposed changes do not noticeably impact compression efficiency. Two changes are proposed, relative to techniques studied in CE5.
Throughput improvements and cycle count reductions on the order of 10-20% were reported for these tricks. The coding efficiency impact of the first technique was estimated at 0.00% and for the second was 0.04%. JVET-K0385 CE5-related: Context state memory reduction [A. . Said, H. . Egilmez, Y.-H. Chao, M. . Karczewicz, V. . Seregin (Qualcomm)] [late] This contribution proposes a method to reduce the amount of RAM and ROM needed for binary arithmetic coding tools tested in CE 5.1 and CE 5.1A (JVET-K381 and JVET-K380). This is done by using a single adaptation window per context. The proposed method reportedly reduces the RAM and ROM memory requirements by 15 bits and 1 bit per-context, respectively. The proposed method reportedly provides average BD-rate gains of 0.8%, 0.8% and 0.9% for AI, RA and LD over the BMS, respectively, and the gains over the VTM are 0.6% for AI, RA and LD coding. The proposed single window solution leads to a BD-rate loss about 0.4% for AI, RA and LD coding as compared to the results in CE5.1 and CE5.1A.
An extension of the state-based probability estimator of VTM-1.0 to two states per context model is proposed. The transition table size is reduced from 64 to 32 elements and the two states per context model require 8 and 12 bit, respectively. Experimental results for the VTM configuration reportedly show overall luma BD rate reductions of 0.67%, 0.45%, and 0.41% for AI, RA, and LB, respectively. The BMS configuration reportedly show overall luma BD rate reductions of 0.71%, 0.46%, and 0.44% for AI, RA, and LB, respectively. This proposal consists of two core elements. First, a state-based probability estimator is presented. The derivation of the subinterval range is the second part. It was commented that for software it may be preferable to use a multiply rather than a table-lookup and that the bit width of storage is not critical unless it affects multiples of 8, 16, or 32 bits. With custom window sizes there would be somewhat more gain, but this contribution did not consider how to combine the concept of custom window sizes with this.
7.6.1Primary transformsJVET-K0113 CE6-related: EMT signalling [C. . Rosewarne, A. . Dorrell (Canon)] Not needed to review, since it is not relevant to the adopted AMT scheme. JVET-K0130 CE6-related: Type4 only AMT [K. . Abe, T. . Toma (Panasonic)] This proposes a simplified AMT that uses a DCT4 instead of a DST7. A DCT4 is part of what is needed for a DCT2, which is said to make this easier to implement. About 0.2% loss is reported relative to using a DST7. K0265 and K0394 are said to be similar, and K0292 also has a similar spirit but a different approach. These are to be further studied in a CE.
See notes for K0130. JVET-K0462 Cross-check of JVET-K0265: CE6-related: Reduction of the number of core transforms in AMT [S. . Bandyopadhyay, Y. . He, Y. . Ye (InterDigital)] [late] JVET-K0394 CE6-related: AMT with only Type2/Type4 DCT/DST [T. . Tsukuba, M. . Ikeda, T. . Suzuki (Sony)] [late] See notes for K0130. JVET-K0426 Cross-check of JVET-K0394: CE6-related: AMT with only Type2/Type4 DCT/DST [X. . Zhao (Tencent)] [late] JVET-K0292 CE6-related: Compound orthonormal transform [X. . Zhao, Z. . Zhang, X. . Li, S. . Liu (Tencent)] See notes for K0130. This contribution keeps a DST7 for 4 point and 8 point transforms and embeds a DST7 into a DCT2 for larger sizes. JVET-K0290 CE6-related: On 8-bit primary transform core [X. . Zhao, X. . Li, S. . Liu (Tencent)] The current VTM has 10 bit coefficients for the 64-point transform, but has 8 bit coefficients for shorter transform. This proposes an 8 bit transform design that uses the shorter HEVC transforms as a component of the 64-point transform. The performance impact was said to be negligible (0.00%). Low QP usage was suggested to be especially important. For very low QP with AMT, a penalty of about 0.1% for AI and 0.05% for RA was reported. Without AMT, the penalty is said to be 0.00%. It was commented that the basis functions have repeated values that could potentially cause plateaus visually, and so suggested visual testing. Another participant noted that for high bit-depth data we have historically used a high-precision forward transform that is matched – i.e., designed as a high-precision inverse of the lower-precision inverse transform, and using it in the encoder gives measurable gain. It was asked how this would interact with such a scheme. Further study is needed to study this and other potential approaches to transform simplification. JVET-K0419 Cross-check of JVET-K0290: CE6-related: On 8-bit primary transform core [T. . Tsukuba (Sony)] [late] JVET-K0291 CE6-related: Fast DST-7/DCT-8 with dual implementation support [Z. . Zhang, X. . Zhao, X. . Li, S. . Liu (Tencent)] This reports on a partial butterfly implementation of a DST7 that is compatible with a matrix multiply approach if a couple of number are changed by 1. About a 7-8% reduction in encoder runtime is reported by using it (with no loss in coding efficiency) in an implementation that is written in ordinary C without SIMD optimization. 40% to 50% operation count reduction is reported. Only 3 numbers are affected (by 1). It was commented by some participants that there would not really be a benefit expected for this. The proponent has both 8 bit and 10 bit variations available. To be further studied with other potential ways of simplifying the transform.
This is a proposed way of reducing the implementation complexity of the inverse transform process for AMT. This should be further studied along with other complexity reduction methods for the inverse transform. JVET-K0126 CE6-related: Simplified multiple-core transform for intra residual coding [Y. . Lin, Q. . Yu, J. . Zheng (HiSilicon), X. . Cao, C. . Zhu (UESTC)] This contribution was discussed Saturday 14 July 1215 (chaired by GJS). This contribution presents two simplified versions of the adaptive multiple-core transform (AMT) in BMS. On the one hand, the number of transform cores for intra residual is reduced from 5 to 3, as a unified transform design of AMT for intra and inter residual coding. On the other hand, encoding is accelerated by reducing the number of signalled transform pairs. It is reported that the proposed transform versions achieve better trade-off between coding performance and encoding/decoding complexity. This is similar to what is proposed in K0171. The proposal is to not support one of the 5 transform combinations used in AMT. The one it proposes to not include is having a DCT8 style transform in both dimensions. The coding efficiency impact of omitting this combination is reported to be negligible (0.06% for AI). It was asked whether there is any significant impact on the decoder for whether this combination is supported or not. It was not clear whether there is a benefit for prohibiting the combination. If the benefit is intended to be saving encoder complexity, the scheme should be tested relative to an encoder-only optimization (the most obvious being simply not checking this combination). Generally, when considering potential syntax restrictions, if there is no benefit for decoders, testing should consider a good encoder-only alternative. Further study was encouraged. JVET-K0499 Crosscheck of JVET-K0126 (Simplified multiple-core transform for intra residual coding) [M.-S. Chiang (MediaTek)] [late] 7.6.2Secondary transformsThis topic remained open after the discussions of Friday 13 July. This was further discussed on Saturday 14 July 1000 (chaired by GJS). A CE will be done to measure the available gain and complexity of methods of secondary transforms relative to the VTM (which will now include AMT).
A non-separable secondary transform (NSST) called reduced secondary transform (RST) was proposed and investigated in CE 6.2.6. A direct matrix multiplication NSST for 4x4 NSST (16x16 direct matrix multiplications) and 8x8 NSST (16x64 direct matrix multiplication) is introduced and investigated in this contribution. relative to a full secondary transform, this reduces the multiplication and multilayer complexity. Also results with memory reduction also reported. This uses 16 secondary transform kernels instead of ~100 as used in the CE test. Ordinarily, implementing a secondary transform larger than 4x4 has high complexity. This proposal use a sparse matrix decomposition to simplify the computation. The number of transform kernels is also reduced.
This proposal involves having a set of secondary transforms and selecting a candidate set of secondary transforms using a LUT based on the transform size and intra mode, then sending an index to select the transform to apply (e.g., among 5 candidates). The secondary transform sizes are 4x4 and 8x8. JVET-K0501 Crosscheck of Section 2.3 in JVET-K0306 (CE6 - related: “Set of Transforms” selection and signalling scheme tested with different types of secondary transforms sets) [M.-S. Chiang (MediaTek)] [late] JVET-K0405 CE6-related: Secondary Transforms Coupled with a Simplified Primary Transformation [H. . Egilmez, A. . Said, Y.-H. Chao, M. . Karczewicz, V. . Seregin (Qualcomm)] [late] JVET-K0110 CE6-related: NSST restriction [C. . Rosewarne, A. . Dorrell (Canon)] This proposes prohibiting NSST when the block aspect ratio is greater than 2:1. However, this does have some coding efficiency penalty. No action was taken on this. 7.6.3Shrink transformThis topic was discussed Friday 13 July 1840 (chaired by GJS). JVET-K0399 CE6-related: Simplification of Shrink Transform (CE6.1.9) [K. . Kawamura, Y. . Kidani, Sei Naito (KDDI)] [late] This is applied to transform length 64 (only). From the decoder perspective, the decoder does a length-32 inverse transform and then upscales the result to form the final residual. An encoder may be designed to perform a transform of input length 64 and keep the lower frequency coefficients or to downsample and perform a shorter transform. In the CE the upscaling used an 8 tap filter. In this contribution it used value replication. Text was not available. The BMS uses a 64 point inverse transform with only the 32 lowest-frequency transform coefficients. This processing treats this particular block length with a different processing in a way that did not seem clearly better and potentially inconsistent with the rest of the design.
Yüklə 4,04 Mb. Dostları ilə paylaş:
|