Accelerating 3D convolution using streaming architectures on FPGAs |
Our target application is a 512 by 512 by 512 finite difference problem, with 6th to 8th order in space and 2nd order in time accuracy. Each time step of the computation takes the current wave-field state, the wave-field state from the previous time step and the velocity model as inputs, and produces the next wave-field state as the output.