A Novel Approach to Software-Defined FPGA Computing

2015年12月16日 | By News | Filed in: News.

Source: https://forums.xilinx.com/t5/Xcell-Daily-Blog/A-Novel-Approach-to-Software-Defined-FPGA-Computing/ba-p/671503

By Stephane Monboisset, PLDA

Recent methodologies for FPGA design, centered on high-level synthesis (HLS) tools and leveraging software programming languages such as OpenCL, C, and C++ have provided a sandbox for software developers to reap the benefits of FPGA-based hardware acceleration in numerous applications. But the methodologies often fall short in one essential respect: enabling software developers to define and configure, on their own, the hardware infrastructure best suited for their application. The industry has continued to pursue the holy grail of a high-level workflow for implementing applications on FPGA-based platforms that does not require specific FPGA expertise.

Over the past five years, PLDA has developed just such a workflow. Called QuickPlay, it efficiently addresses the implementation complexity challenge and enables multiple use models for FPGA development. But one of its core sources of value is the way in which it lets software developers take applications intended for CPUs and implement them, partially or fully, on FPGA hardware. QuickPlay leverages all of the FPGA resources, turning these powerful but complex devices into software-defined platforms that yield the benefits of FPGAs without the pain of hardware design.

Consider a software algorithm that can be broken down into two functions: Data is processed into one function and is then sent to another for further processing. From a software perspective, this implementation is as simple as a call to Function1() followed by a separate call to Function2(), using pointers to the location of the data to be processed.

Implementing such an algorithm on an FPGA-based hardware platform without the right hardware abstraction tool flow would require the software developer to come up with a hardware design resembling that in Figure 1 (where Kernel 1 and Kernel 2 are the respective hardware implementations of Function 1 and Function 2).

PLDA QuickPlay Design.jpg

The hardware design would need to include two elements: the control plane and the data plane. The control plane is the execution engine that generates clocks and resets, manages system startup, orchestrates data plane operations, and performs all housekeeping functions. The data plane instantiates and connects the processing elements, Kernel 1 and Kernel 2, as well as the necessary I/O interfaces required to read data in and write processed data out. In our example, those interfaces are Ethernet and PCI Express (PCIe), as Figure 1 shows, though different application requirements will call for different I/O interfaces.

Beyond the design work—and equally challenging—is the implementation work, which includes mapping the design onto the resources of the selected FPGA platform, generating the appropriate constraints, and confirming that those constraints are met after logic synthesis and implementation on the FPGA hardware. It can take even an experienced hardware designer weeks to achieve a working design on a new piece of FPGA hardware.

Thus, any tool that aims to enable software developers to augment their applications with custom hardware must be able to:

  • create functional hardware from pure software code
  • incorporate existing hardware IP blocks if needed
  • infer and create all of the support hardware (interfaces, control, clocks, etc.)
  • support the use of commercial, off-the-shelf boards and custom platforms
  • ensure that the generated hardware is correct by construction so that it requires no hardware debug
  • support debug of functional blocks using standard software debug tools only

PLDA engineered QuickPlay from the ground up to meet all of those requirements, thereby enabling pure software developers to specify, build and integrate FPGAs into their software architectures with minimal effort.

Note: This blog was adapted from an article by the same name that appeared in the recently published issue of Xcell Software Journal. To read the full article online, click here. To download a PDF of the entire issue, click here.


发表评论

邮箱地址不会被公开。 必填项已用*标注