Quantum Computer Simulation on Multi-GPU Incorporating Data Locality

Quantum computer simulation (QCS) provides an effective platform for the development and validation of quantum algorithms. The exponential runtime overhead limits the simulation scale on classical computers which makes advisable the use of Graphics Processing Units. However, simulating quantum computers on multi-GPU has poor performance due to low data locality and frequent data transfer. Here, we propose a novel implemental scheme for QCS on multi-GPU. Our implementation addresses the aforementioned challenges by (i) an efficient data distribution method enhancing high data locality on each GPU global memory and (ii) an assignment function for the threads mapping to each GPU memory space achieving high bandwidth and data reuse for multiple quantum gates. Experimental results show that the simulation of 29-qubit Quantum Fourier Transform algorithm using four NVIDIA K20c GPUs gains a performance ratio of 358, compared to the sequential implementation of released libquantum, along with a parallel efficiency of 0.92.