Developer Guide

FPGA Optimization Guide for Intel® oneAPI Toolkits

ID 767853
Date 3/31/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

Global Memory

If the kernel code accesses a host-allocated buffer, the compiler creates a hardware interface through which the datapath accesses the buffer in global memory. A host-allocated buffer resides in device global memory off-chip. The code snippet in the Kernel Memory section shows a device global memory and its accesses within the kernel.

Unlike kernel memory, the compiler does not define the structure of a buffer in global memory. The compiler instantiates a specialized LSU for each access site based on the memory access pattern to maximize the efficiency of data accesses.

All accesses to global memory must go through the hardware interface. The compiler connects every LSU to an existing hardware interface through which it transacts with device global memory. Since the compiler cannot alter that interface or create more such interfaces, it must share the interface between multiple datapath reads or writes, which can limit the throughput of the design. The strategies used by the compiler to maximize efficient use of available memory interface bandwidth include (but are not limited to) the following:

  • Eliminating unnecessary accesses.
  • Statically coalescing contiguous accesses.
  • Generating specialized LSUs that can perform the following:
    • Dynamically coalesced accesses that fall within the same memory word (as defined by the interface).
    • Prefetch and cache memory contents.