GPU-based 3D Wavelet Transform
Vicente Galiano, Otoniel López-Granado, Manuel Perez Malumbres, Héctor Migallón · 2013
Wide amount of applications like volumetric medical data compression, video watermarking and video coding use the three-dimensional wavelet transform (3D-DWT) in their algorithms. In this work, we present GPU algorithms, based on both global and shared memory, to compute the 3D-DWT transform on both the GTX280 and the GMT540 platforms. The results obtained show that speed-ups of 19.7 and 10.65 on average can be obtained for the GTX280 and GMT540 platforms respectively when only the GPU’s global memory is used. Moreover, Speed-ups increase considerably to 87 and 25 when the shared memory in the device is used optimizing the memory access to avoid idle threads. Futhermore, we discuss speed-up evolution depending on the group of pictures size (GOP).