Optimizing wait-states in the synthesis of memory references with unpredictable latencies

Abstract

In here we consider the problem of automatic synthesizing, from C to Verilog, circuits that are optimized to handle unpredictable latencies of memory operations. Unpredictable memory latencies can occur due to the use of on chip caches, DRAM memory modules, buffers/queues or multi-port memories. Typically highlevel synthesis compilers assume fixed and known memory latencies thus the technique presented here expands the use of HLS. Assume that in a current state we have k active memory references that can terminate in this clock cycle or possibly continue for an unknown number of clock cycles. Thus, the scheduler must emit 2k new states, one state for each possible combination of memory requests that can continue in the next clock cycle. Synthesizing a state machine with exponential number of states is not practical thus we show a simple technique of synthesizing a compact state machine which is a compromise between the fast full exponential state machine and the linear state machine that would have been generated had we waited for the termination of all the active memory references in every state. Our results shows that the compact state machine obtains similar performances as the full state machine but with a significant less number of resources.1