Dynamic power consumption estimation and reduction for full search motion estimation hardware

Official URL: http://192.168.1.20/record=b1293643 (Table of Contents)

Motion Estimation (ME) is the most computationally intensive and most power consuming part of video compression and video enhancement systems. ME is used in video compression standards such as MPEG4, H.264 and it is used in video enhancement algorithms such as frame rate conversion and de-interlacing. Since portable devices operate with battery, it is important to reduce power consumption so that the battery life can be increased. In addition, consuming excessive power degrades the performance of integrated circuits, increases packaging and cooling costs, reduces the reliability and may cause device failures. Therefore, estimating and reducing power consumption of motion estimation hardware is very important. In this thesis, we propose a novel dynamic power estimation technique for full search ME hardware. We estimated the power consumption of two full search ME hardware implementations on a Xilinx Virtex II FPGA using several existing high and low level dynamic power estimation techniques and our technique. Gate-level timing simulation based power estimation of full search ME hardware for an average frame using Xilinx XPower tool takes 6 - 18 hours in a state-of-the-art PC, whereas estimating the power consumption of the same ME hardware for the same frame takes a few seconds using our technique. The average and maximum difference between the power consumptions estimated by our technique and the power consumptions estimated by XPower tool for four different video sequences are %3 and %13 respectively. We also propose a novel dynamic power reduction technique for ME hardware. We quantified the impact of glitch reduction, clock gating and the proposed technique on the power consumption of two full search ME hardware implementations on a Xilinx Virtex II FPGA using Xilinx XPower tool. Glitch reduction and clock gating together achieved an average of 21% dynamic power reduction. The proposed technique achieved an average of 23% dynamic power reduction with an average of 0.4dB PSNR loss. The proposed technique achieves better power reduction than pixel truncation technique with a similar PSNR loss. We also showed that our dynamic power estimation technique can be used for developing novel dynamic power reduction techniques. To do this, we used our technique to estimate the dynamic power consumption of the ME hardware when two different dynamic power reduction techniques are used. The results show that if a power reduction technique only changes the input data order of the ME hardware, the proposed dynamic power estimation technique can be used to quickly estimate the effectiveness of that technique. However, if the architecture of the ME hardware is modified, the accuracy of the power consumption estimations decrease. Therefore the proposed power estimation technique should be improved for this case.