AMT: asynchronous in-place matrix transpose mechanism for sunway many-core processor

Matrix multiplication is widely used in a variety of application domains. When the input matrices and the product differ in the memory format, matrix transpose is required. The efficiency of matrix transpose has a non-negligible impact on performance. However, the state-of-the-art software solution and its optimizations suffer from low efficiency due to frequent interference to main pipeline and their inability to achieve parallel matrix transpose and multiplication. To address this issue, we propose AMT, an asynchronous and in-place matrix transpose mechanism based on C2R algorithm, to efficiently perform matrix transpose. AMT performs matrix transpose in an asynchronous processing module and uses two customized asynchronous matrix transpose instructions to facilitate processing. We implement the logic design of AMT using RTL and verify its correctness. Simulation results show that AMT achieves an average of 1.27x (up to 1.48x) speedup over a state-of-the-art software baseline, and is within 95.4% of an ideal method. Overhead analysis shows that AMT only incurs small area overhead and power consumption.