International organisation for standardisation organisation internationale de normalisation

18.10IBDI and memory compression

See also the section on TE2.

18.10.1.1.1.1.1.1.1JCTVC-C077 A framework for standardization of memory compression [T. Chujoh, T. Yamakage (Toshiba)]

A framework for standardization of memory compression was proposed. A purpose of the contribution was to discuss how to standardize memory compression. Memory compression can reduce the bandwidth of memory access; however, most memory compression techniques depend on hardware architectures and the process of memory compression causes a degradation of coding efficiency and an increase of computational complexity for most implementations. This contribution does not propose a common memory compression algorithm, but rather proposes a framework of memory compression. Since the essence of a conformance point is the output of the decoding process, it was suggested that a bitstream format and a bit rate not be specified; rather, a specification of the distortion and a system to control the distortion would be specified in this framework. The purpose of this framework is to enable introducing arbitrary memory compression effects into the encoder and the decoder in a matching fashion.

It was discussed whether encoder/decoder mismatch should be allowed. Perhaps not, but it may be a question worth discussing.

In the discussion, the creation of a related AHG was suggested.

18.10.1.1.1.1.1.1.2JCTVC-C094 An extension of DPCM-based memory compression to 2-D structure with ADPCM [H. Aoki, K. Chono, K. Senzaki, J. Tajime, Y. Senda (NEC)]

In this contribution, a 2-D extension of the previously-proposed 1-D DPCM-based memory compression method was presented. According to some previous reports, there are conditions where 2-D image mapping to frame memory provides better memory access bandwidth performance. In such cases, memory compression with a 2-D structure may be more straightforward and may improve coding efficiency. In the proposed 2-D extension, the top-most and left-most samples in every memory compression unit are coded in the same manner as the 1-D case, and other samples are coded with ADPCM. The 1-D DPCM-based method can be regarded as a specialized version of the proposed 2-D ADPCM-based method, where its unit height is one. For each sample coded with ADPCM, the prediction value is selected in the same manner as the DC and AC prediction of the MPEG-4 part 2 standard. Experimental results were reported to have shown that the proposed 2-D ADPCM-based memory compression method has higher coding efficiency than the previously-proposed 1-D DPCM-based memory compression method. Average coding losses were reported to be 7.9% for test cases without IBDI and 1.5% for test cases 12-bit IBDI, while those of the 1-D methods are 10.0% and 1.9%, respectively. The average increase of decoding time was reported to be 3.8%. Memory bandwidth reduction ratios were reported as 45.1% for test cases without IBDI and 44.3% for test cases with 12-bit IBDI. It was proposed for the method to be adopted in TMuC software with the previously-proposed method in JCTVC-C093, and to be further evaluated in various aspects.

18.10.1.1.1.1.1.1.3JCTVC-C095 Performance improvement of DPCM-based memory compression by adaptive quantization [H. Aoki, K. Chono, K. Senzaki, J. Tajime, Y. Senda (NEC)]

This contribution presented a method for improving coding efficiency of DPCM-based memory compression. In the proposed method, a nonlinear quantization matrix for each memory compression unit (MCU) was adaptively selected from multiple quantization matrices so as to fit quantizer characteristics to image characteristics of each MCU. Each index to the selected quantization matrix is embedded into the compressed data and signaled to the frame memory decompressor. The actual frame memory size and memory access bandwidth reportedly does not increase as the result of this additional signaling, since the indices can reportedly be embedded into a gap area for memory alignment. Experimental results have reportedly shown that the proposed adaptive quantization can reduce the coding loss from 10.7% to 6.9% for test cases without IBDI, and from 2.0% to 1.3% for test cases with 12-bit IBDI, respectively. When applied to the 2-D version presented in JCTVC-C094, more gains were reported to have been achieved and the coding loss can reportedly be reduced to 4.8% for test cases without IBDI and to 1.0% for test cases with 12-bit IBDI, respectively. The average increase of decoding times against anchors was reported to be 14.7% for LC cases and 12.4% for HCE cases when four matrices were used. It can reportedly be shown that the proposed adaptive quantization provides flexibility on the trade-off between coding efficiency and complexity to the DPCM-based memory compression scheme.

Anomolous results were reported for Class E coding effects of memory rounding. This should be investigated.

18.11Complexity analysis

JCTVC-C228 A simple cache model for measuring motion compensation bandwidth [M. Budagavi (TI)]

Motion compensation reference data accesses dominate the external memory bandwidth requirements in video decoder system. External memory access bandwidth requirements are expected to increase with HEVC since HEVC motion compensation interpolation filters are longer than those in AVC and since HEVC is expected to be used with larger resolution video such as 4Kx2K and 8Kx4K. External memory bandwidth used directly impacts power consumed and cost; in fact, power consumed due to external memory accesses in a video decoder can reportedly equal the power consumed by computational processing in video decoding. Hence techniques that reduce external memory bandwidth are desirable. Memory compression is one technique that is being studied in TE2 and Memory compression AhG for reducing external memory accesses. This contribution proposed a cache model for use in measuring external memory access bandwidth. The cache model is built on top of NEC’s memory bandwidth measurement module in TMuC-0.7. The main asserted goal of this work was to have a simple cache model that captures the most relevant parameters for comparing bandwidth consumed with and without memory compression and that is fair to all block structures used in memory compression. The goal was reportedly not to develop a very sophisticated cache model. This cache model can also be used to compare memory bandwidth requirements of various interpolation filters. The cache model was reportedly developed as a part of Memory Compression AhG.

It was agreed to include this in the reference software.

The AHG on complexity analysis should perform an example study e.g. for multi-pass and single-pass loop filter optimization to assess the validity of the measurement.

JCTVC-C072 TE2: Improved memory bandwidth measurement model for reference frame compression [J. Chen, X. Shen, L. Yu (Zhejiang Univ.)]

A memory-bandwidth measurement model for reference frame compression algorithm was proposed. This proposal analyses reference frame decompression processes and divides them into parsing process and decoding process, in which three new concepts, parsing unit, decoded area and decoding area, are defined. Based on this, a memory access area conduction method had been designed to determine the memory access area.

It was asserted that the proposed model can still be efficient when memory compression unit is not equal to decompressible unit. It expands the scope of application and can measure memory bandwidth for more general reference frame compression algorithms.

The contribution was noted.

It was remarked that this is based on software that was provided to TE2 participants.

It was suggested to work with TE coordinators to determine its appropriateness for upcoming experiments.