Kernel Merger - Automatic Dataflow Design Optimizations Sample Clauses

Kernel Merger - Automatic Dataflow Design Optimizations. In this deliverable we consider designs that use Maxeler Dataflow Engines (DFEs) as hardware accelerators. Such designs are typically developed by porting an existing software implementation into a dataflow implementation using Maxeler’s MaxCompiler. After the porting the design goes through an optimization phase to maximize the performance. While it is very challenging to automate the translation of a sequential software implementation into a dataflow implementation (or any other parallel implementation for that matter) we can target a number of typically time- consuming optimizations for tool automation to improve the overall productiveness of the translation process. The purpose of the Kernel Merger tool is to provide a number of automatic optimizations to a dataflow design that would otherwise take considerable time and expertise by the developer. The main focus is merging different parts of a dataflow graph that are not active at the same time, but a number of further optimizations are also included. As an example, consider an algorithm that computes in hardware either a * b or a + b depending on a condition c. Both, the adder and the multiplier must be instantiated even though only one of the results will be selected by a multiplexer (MUX) at the output. However, if the computation is either a * b or a * d then instead of creating two multipliers we can use a single multiplier with a MUX selecting between inputs b or d. This is a more efficient hardware implementation as a MUX uses far fewer resources than the second multiplier. In practice, this situation can be found when applications offload several compute intensive parts, often referred to as kernels, to the DFE. Since only one kernel is active at any given time, the other kernels remain idle. However, since they still have to be implemented on the DFE, a large portion of the chip is not in use most of the time. The solution for this problem is to reuse arithmetic units between kernels to minimize the area overhead introduced by simultaneously implementing multiple kernels. Manually identifying the arithmetic units that can be shared and creating the required control logic would be very time consuming. For this reason, we created a dedicated pre-processing optimization tool called Kernel Merger. It is implemented as a software library for the Maxeler high-level synthesis tool-chain MaxCompiler and it is able to automatically merge multiple calculations and thereby significantly reduces the hardw...