GPU平台上ADL算法的实现

Implementation of ADL Algorithm on GPU

自适应方向提升小波变换(ADL)利用图像纹理特征进行变换编码,从而获得更高的编码质量,但同时也增加了计算复杂度。为了提高图像编码速率,在统一计算设备架构(CUDA)的图形处理器(GPU)上,提出一种并行实现ADL中的插值和方向变换计算的新方案,对插值部分同时采用粗粒度和细粒度的并行,即把图像数据分成若干个块进行粗粒度的并行,而对块中的每个像素点采用细粒度的并行。对变换部分中的9个变换方向采用粗粒度的并行。实验表明,在GPU上并行实现ADL变换是CPU实现的4倍左右,CPU-GPU整体架构下的ADL变换编码的速度是CPU平台下的3倍左右。

In order to gain much better image quality, ADL （ adaptive directional lifting） wavelet transform takes use of the texture property of image to implement the transform coding at the cost of high computation complexity. Implement the interpolation and directional lifting transform of ADL in parallel on GPU （graphic processing unit） with CUDA （compute unified device architecture） to speed up the image encoding. Both fine-grained and coarse-grained granularity parallelization are used for data block and pixels respectively in interpolation, while only coarse-grained granularity is used in nine directions for transform. Experiments results show that implementation of ADL on GPU is 4 times faster than that on CPU. The total time of ADI, transform image coding on CPU-GPU framework is almost 4 times faster than on CPU.