Fig. 3
From: ADEPT: a domain independent sequence alignment strategy for gpu architectures
The overall pipeline of the ADEPT strategy. ADEPT’s driver detects all the available GPUs and their available memory. Based on this, it determines the amount of work that can be dispatched to each GPU. A separate CPU context is created for each GPU. On the CPU side, the batch size is determined based on the GPU’s available memory or the batch size can also be fixed by the user. To overlap the data preparation step, data transfers and the GPU computation, multiple CUDA streams are used. Immediately after making the GPU to CPU transfer call, the CPU returns to the data preparation step while kernel execution is still going on at the GPU side. This asynchronous behavior overlaps the CPU and GPU parts of the computation