Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when bench-marked with an array of well-established perceptual quality scores. These "neural wrapper" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with -18.5% average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding.

We present Dithen, a novel computation-as-a-service (CaaS) cloud platform specifically tailored to the parallel execution of large-scale multimedia tasks. Dithen handles the upload/download of both multimedia data and executable items, the assignment of compute units to multimedia workloads, and the reactive control of the available compute units to minimize the cloud infrastructure cost under deadline-abiding execution. Dithen combines three key properties: (i) the reactive assignment of individual multimedia tasks to available computing units according to availability and predetermined time-to-completion constraints; (ii) optimal resource estimation based on Kalman-filter estimates; (iii) the use of additive increase multiplicative decrease (AIMD) algorithms (famous for being the resource management in the transport control protocol) for the control of the number of units servicing workloads. The deployment of Dithen over Amazon EC2 spot instances is shown to be capable of processing more than 80,000 video transcoding, face detection and image processing tasks (equivalent to the processing of more than 116 GB of compressed data) for less than $1 in billing cost from EC2. Moreover, the proposed AIMD-based control mechanism, in conjunction with the Kalman estimates, is shown to provide for more than 27 percent reduction in EC2 spot instance cost against methods based on reactive resource estimation. Finally, Dithen is shown to offer a 38 to 500 percent reduction of the billing cost against the current state-of-the-art in CaaS platforms on Amazon EC2 (Amazon Lambda and Amazon Autoscale). A baseline version of Dithen is currently available at dithen.com under the “AutoScale” option.

The generic matrix multiply (GEMM) routine comprises the compute and memory-intensive parts of many information retrieval, machine learning, and object recognition systems that process integer inputs. Therefore, it is of paramount importance to ensure that integer GEMM computations remain robust to silent data corruptions (SDCs), which stem from accidental voltage or frequency overscaling, or other hardware nonidealities. In this paper, we introduce a new method for SDC mitigation based on the concept of numerical packing. The key difference between our approach and all existing methods is the production of redundant resultswithinthe numerical representation of the outputs, rather than as a separate set of checksums. Importantly, unlike well-known algorithm-based fault tolerance approaches for GEMM, the proposed approach can reliably detect the locations of the vast majority of all possible SDCs in the results of GEMM computations. An experimental investigation of voltage-scaled integer GEMM computations for visual descriptor matching within state-of-the-art image and video retrieval algorithms running on an Intel i7-4578U 3 GHz processor shows that SDC mitigation based on numerical packing leads to comparable or lower execution and energy-consumption overhead in comparison with all other alternatives.

Computation-as-a-Service (CaaS) offerings have gained traction in the last few years due to their effectiveness in balancing between the scalability of Software-as-a-Service and the customisation possibilities of Infrastructure-as-a-Service platforms. To function effectively, a CaaS platform must have three key properties: (i) reactive assignment of individual processing tasks to available cloud instances (compute units) according to availability and predetermined time-to-completion (TTC) constraints, (ii) accurate resource prediction, (iii) efficient control of the number of cloud instances servicing workloads, in order to optimize between completing workloads in a timely fashion and reducing resource utilization costs. In this paper, we propose three approaches that satisfy these properties (respectively): (i) a service rate allocation mechanism based on proportional fairness and TTC constraints, (ii) Kalman-filter estimates for resource prediction, and (iii) the use of additive increase multiplicative decrease (AIMD) algorithms (famous for being the resource management in the transport control protocol) for the control of the number of compute units servicing workloads. The integration of our three proposals into a single CaaS platform is shown to provide for more than 27% reduction in Amazon EC2 spot instance cost against methods based on reactive resource prediction and 38% to 60% reduction of the billing cost against the current state-of-the-art in CaaS platforms (Amazon Lambda and Autoscale).

A new roll-forward technique is proposed that recovers from any single fail-stop failure in M integer data streams (M ≥ 3) when undergoing linear, sesquilinear or bijective (LSB) operations, such as: scaling, additions/subtractions, inner or outer vector products and permutations. In the proposed approach, the M input integer data streams are linearly superimposed to form M numerically entangled integer data streams that are stored inplace of the original inputs. A series of LSB operations can then be performed directly using these entangled data streams. The output results can be extracted from any M-1 entangled output streams by additions and arithmetic shifts, thereby guaranteeing robustness to a fail-stop failure in any single stream computation. Importantly, unlike other methods, the number of operations required for the entanglement, extraction and recovery of the results is linearly related to the number of the inputs and does not depend on the complexity of the performed LSB operations. We have validated our proposal in an Intel processor (Haswell architecture with AVX2 support) via convolution operations. Our analysis and experiments reveal that the proposed approach incurs only 1.8% to 2.8% reduction in processing throughput in comparison to the failure-intolerant approach. This overhead is 9 to 14 times smaller than that of the equivalent checksum-based method. Thus, our proposal can be used in distributed systems and unreliable processor hardware, or safety-critical applications, where robustness against fail-stop failures becomes a necessity.

Generic matrix multiplication (GEMM) and convolution (CONV)/cross-correlation kernels often constitute the bulk of the compute- and memory-intensive processing within image/audio recognition and matching systems. We propose a novel method to scale the energy and processing throughput of GEMM and CONV kernels for such error-tolerant multimedia applications by adjusting the precision of computation. Our technique employs linear projections to the input matrix or signal data during the top-level GEMM and CONV blocking and reordering. The GEMM and CONV kernel processing then uses the projected inputs and the results are accumulated to form the final outputs. Throughput and energy scaling takes place by changing the number of projections computed by each kernel, which in turn produces approximate results, i.e., changes the precision of the performed computation. Results derived from a voltage- and frequency-scaled ARM Cortex A15 processor running face recognition and music-matching algorithms demonstrate that the proposed approach allows for a 280%-440% increase of processing throughput and a 75%-80% decrease of energy consumption against the optimized GEMM and CONV kernels without any impact on the obtained recognition or matching accuracy. Even higher gains can be obtained, if one is willing to tolerate some reduction in the accuracy of the recognition and matching applications.

The generic matrix multiply (GEMM) routine comprises the compute and memory-intensive part of many information retrieval, relevance ranking and object recognition systems. Because of the prevalence of GEMM in these applications, ensuring its robustness to transient hardware faults is of paramount importance for highly-efficientlhighly-reliable systems. This is currently accomplished via error control coding (ECC) or via dual modular redundancy (DMR) approaches that produce a separate set of “parity” results to allow for fault detection in GEMM. We introduce a third family of methods for fault detection in integer matrix products based on the concept of numerical packing. The key difference of the new approach against ECC and DMR approaches is the production of redundant results within the numerical representation of the inputs rather than as a separate set of parity results. In this way, high reliability is ensured within integer matrix products while allowing for: (i) in-place storage; (ii) usage of any off-the-shelf 64-bit floating-point GEMM routine; (iii) computational overhead that is independent of the GEMM inner dimension. The only detriment against a conventional (i.e. fault-intolerant) integer matrix multiplication based on 32-bit floating-point GEMM is the sacrifice of approximately 30.6% of the bitwidth of the numerical representation. However, unlike ECC methods that can reliably detect only up to a few faults per GEMM computation (typically two), the proposed method attains more than “12 nines” reliability, i.e. it will only fail to detect 1 fault out of more than 1 trillion arbitrary faults in the GEMM operations. As such, it achieves reliability that approaches that of DMR, at a very small fraction of its cost. Specifically, a single-threaded software realization of our proposal on an Intel i7-3632QM 2.2GHz processor (Ivy Bridge architecture with AVX support) incurs, on average, only 19% increase of execution time against an optimized, fault-intolerant, 32-bit GEMM routine over a range of matrix sizes and it remains more than 80% more efficient than a DMR-based GEMM.

Convolution and cross-correlation are the basis of filtering and pattern or template matching in multimedia signal processing. We propose two throughput scaling options for any one-dimensional convolution kernel in programmable processors by adjusting the imprecision (distortion) of computation. Our approach is based on scalar quantization, followed by two forms of tight packing in floating-point (one of which is proposed in this paper) that allow for concurrent calculation of multiple results. We illustrate how our approach can operate as an optional pre- and post-processing layer for off-the-shelf optimized convolution routines. This is useful for multimedia applications that are tolerant to processing imprecision and for cases where the input signals are inherently noisy (error tolerant multimedia applications). Indicative experimental results with a digital music matching system and an MPEG-7 audio descriptor system demonstrate that the proposed approach offers up to 175% increase in processing throughput against optimized (full-precision) convolution with virtually no effect in the accuracy of the results. Based on marginal statistics of the input data, it is also shown how the throughput and distortion can be adjusted per input block of samples under constraints on the signal-to-noise ratio against the full-precision convolution.