Developer Guide

FPGA Optimization Guide for Intel® oneAPI Toolkits

ID 767853
Date 12/16/2022
Public

A newer version of this document is available.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

Document Table of Contents

N-Way Buffering to Overlap Kernel Execution

N-way buffering is a generalization of the double buffering optimization technique. This system-level optimization enables kernel execution to occur in parallel with host-side processing and buffer transfers between host and device, improving application performance. N-way buffering can achieve this overlap even when the host-processing time exceeds kernel execution time.

In an application where the FPGA kernel is executed multiple times, the host must perform the following processing and buffer transfers before each kernel invocation.

  1. The output data from the previous invocation must be transferred from the device to the host and then processed by the host. Examples of this processing include the following:
    • Copying the data to another location
    • Rearranging the data
    • Verifying it in some way
  2. The input data for the next invocation must be processed by the host and then transferred to the device. Examples of this processing include:
    • Copying the data from another location
    • Rearranging the data for kernel consumption
    • Generating the data in some way

Without the N-way buffering, host processing and buffer transfers occur between kernel executions. Therefore, there is a gap in time between kernel executions, which you can refer as kernel downtime (See Figure 1). If these operations overlap with kernel execution, the kernels can execute back-to-back with minimal downtime, thereby increasing the overall application performance.

Determine When is N-Way Buffering Possible

N-way buffering is frequently referred to as double buffering in the most common case where N=2.

Consider the following illustration:

N-Way Buffering Host

The following are the definitions of the required variables:

  • R: Time to transfer the kernel's output buffer from device to host.
  • Op: Host-side processing time of kernel output data (output processing).
  • Ip: Host-side processing time for kernel input data (input processing).
  • W: Time to transfer the kernel's input buffer from host to device.
  • K: Kernel execution time
  • N: The number of buffer sets used.
  • C: The number of host-side CPU cores.

In general, R, Op, Ip, and W operations must all complete before the next kernel is launched. To maximize performance, while one kernel is executing on the device, these operations must run in parallel and operate on a separate set of buffer locations. They must complete before the current kernel completes, thus allowing the next kernel to be launched immediately with no downtime. In general, to maximize performance, the host must launch a new kernel every K.

If these host-side operations are executed serially, this leads to the following constraint:

R + Op + Ip + W <= K, to minimize kernel downtime

In the above example, if the constraint is satisfied, the application requires two sets of buffers. In this case, N=2. However, the above constraint may not be satisfied in some applications if host-processing takes longer than the kernel execution time.

NOTE:

A performance improvement may still be observed because kernel downtime may still be reduced (though perhaps not maximally reduced).

In this case, to further improve performance, reduce host-processing time through multithreading. Instead of executing the above operations serially, perform the input and output-processing operations in parallel using two threads, leading to the following constraint:

Max (R+Op, Ip+W) <= K

R + W <= K, to minimize kernel downtime

If the above constraint is still unsatisfied, the technique can be extended beyond two sets of buffers to N sets of buffers to help improve the degree of overlap. In this case, the constraint becomes:

Max (R + Op, Ip + W) <= (N-1)*K

R + W <= K, to minimize kernel downtime.

The idea of N-way buffering is to prepare N sets of kernel input buffers, launch N kernels, and when the first kernel completes, begin the subsequent host-side operations. These operations may take a long time (longer than K), but they do not cause kernel downtime because an additional N-1 kernels have already been queued and can launch immediately. By the time these first N kernels complete, the aforementioned host-side operations would have also completed, and the N+1 kernel can be launched with no downtime. As additional kernels complete, corresponding host-side operations are launched on the host, using multiple threads in a parallel fashion. Although the host operations take longer than K, if N is chosen correctly, they complete with a period of K, which is required to ensure you can launch a new kernel every K. To reiterate, this scheme requires multi-threaded host-operations because the host must perform processing for up to N kernels in parallel to keep up.

Use the above formula to calculate the N required to minimize downtime. However, there are some practical limits:

  • N sets of buffers are required on both the host and device. Therefore, both must have the capacity for this many buffers.
  • If the input and output processing operations are launched in separate threads, then (N-1)*2 cores are required so that C can become the limiting factor.

Measure the Impact of N-Way Buffering

You must get a sense of the kernel downtime to identify the degree to which this technique can help improve performance.

You can do this by querying the total kernel execution time from the runtime and comparing it to the overall application execution time. In an application where kernels execute with minimal downtime, these two numbers are close. However, if kernels have a lot of downtime, overall execution time notably exceeds the kernel execution time.

NOTE:

For additional information, refer to the FPGA tutorial sample "N-Way Buffering" listed in the Intel® oneAPI Samples Browser on Linux* or Windows*, or access the code sample on GitHub.