Anuprit
My Past Work: True Motion Estimation
Optimal Implementation of True Motion Estimation (TME) on GPU compute. I had loved this period of my life.
Under the guidance of Dr. Philips Koshy, Phd. IIT Madras and Murali Krishna Kamisetty, M.Tech. IIT Guwahati, Software Engineer Nvidia Graphics Private Limited.
Motion estimation being a compute intensive problem TME is supposed to be an offline task and that is why we felt the need to port this problem on GPU to make it real time. A speedup of 80X on NVIDIA’s GT200 GPU over Intel’s 3.0 GHZ processor was gained making TME a real time application. We extended Multi Pass Motion Vector Propagation (MPMVP) to suit GPU architecture. We came up with efficient memory access patterns giving us coalesced access to memory. Optimal use of on chip and off chip memory gave us fast and efficient memory access. Lowering the number of registers used per thread helped us in increasing GPU occupancy. The result was a linear graph between pixels vs. time which indicates successfully hidden memory latency. We achieved high calculation to global memory access ratio giving us high execution speed in kernels.
It is evident from the findings that we have well addressed the compute-intensive problem using GPGPU.
Architecture used to program GPU: NVIDIA’s CUDA
Role played in the project: Team member in a team of three students.
Comments: 0
There are no comments yet, be the first to write a comment!