Developer Guide

FPGA Optimization Guide for Intel® oneAPI Toolkits

Download PDF
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A newer version of this document is available.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
Document Table of Contents x
FPGA Optimization Guide for Intel® oneAPI Toolkits
FPGA Optimization Guide for Intel® oneAPI Toolkits x
Introduction To FPGA Design Concepts Analyze Your Design Optimize Your Design FPGA Optimization Flags, Attributes, Pragmas, and Extensions Quick Reference Additional Information Document Revision History for the FPGA Optimization Guide for Intel® oneAPI Toolkits Notices and Disclaimers
Introduction To FPGA Design Concepts x
FPGA Architecture Overview Concepts of FPGA Hardware Design Methods of Hardware Design How Source Code Becomes a Custom Hardware Datapath Scheduling Mapping Parallelism Models to FPGA Hardware Memory Types
FPGA Architecture Overview x
Adaptive Logic Module (ALM) Lookup Table (LUT) Register Digital Signal Processing (DSP) Block Random Access Memory (RAM) Blocks
Concepts of FPGA Hardware Design x
Maximum Frequency (fMAX) Latency Pipelining Throughput Datapath Control Path Occupancy
How Source Code Becomes a Custom Hardware Datapath x
Mapping Source Code Instructions to Hardware Mapping Arrays and Their Accesses to Hardware
Scheduling x
Dynamic Scheduling Clustering the Datapath Handshaking Between Clusters
Mapping Parallelism Models to FPGA Hardware x
Data Parallelism Task Parallelism
Data Parallelism x
Executing Independent Operations Simultaneously Pipelining
Memory Types x
Kernel Memory Global Memory
Analyze Your Design x
Analyze the FPGA Early Image Analyze the FPGA Image
Analyze the FPGA Early Image x
Review the FPGA Optimization Report Access HLD FPGA Reports in JSON Format
Review the FPGA Optimization Report x
Loop Analysis Bottlenecks Viewer Area Estimates System Viewer Kernel Memory Viewer Schedule Viewer
Analyze the FPGA Image x
Quartus (Static) Summary Intel® FPGA Dynamic Profiler for DPC++ System-level Profiling Using the Intercept Layer for OpenCL* Applications
Quartus (Static) Summary x
Timing Failures
Intel® FPGA Dynamic Profiler for DPC++ x
Measure Kernel Performance Instrument the Kernel Pipeline with Performance Counters (<span class='codeph'>-Xsprofile</span>) Obtain Profiling Data During Runtime Reduce Area Resource Use While Profiling Profiler Analyses of Example SYCL* Design Scenarios Limitations
Obtain Profiling Data During Runtime x
Invoke the Profiler Runtime Wrapper to Obtain Profiling Data Use Intel® VTune™ Profiler
Use Intel® VTune™ Profiler x
Interpret Performance Counter Data
System-level Profiling Using the Intercept Layer for OpenCL* Applications x
Set Up the Intercept Layer for OpenCL* Applications
Optimize Your Design x
Throughput Resource Use
Throughput x
Single Work-item Kernels NDRange Kernels Memory Accesses Pipes Host
Single Work-item Kernels x
Single Work-item Kernel Design Guidelines Loops Single-Cycle Floating-Point Accumulator for Single Work-Item Kernels
Loops x
Refactor the Loop-Carried Data Dependency Relax Loop-Carried Dependency Transfer Loop-Carried Dependency to Local Memory Minimize the Memory Dependencies for Loop Pipelining Unroll Loops Fuse Loops to Reduce Overhead and Improve Performance Optimize Loops With Loop Speculation Remove Loop Bottlenecks Shannonization to Improve FMAX/II Optimize Inner Loop Throughput Improve Loop Performance by Caching On-Chip Memory
Single-Cycle Floating-Point Accumulator for Single Work-Item Kernels x
Strategies for Inferring the Accumulator
Memory Accesses x
Load-Store Units Global Memory Accesses Optimization Perform Kernel Computations Using Local or Private Memory Local and Private Memory Accesses Optimization Annotating Unified Shared Memory Pointers Zero-Copy Memory Access Additional Recommendations
Load-Store Units x
Load-Store Unit Styles Load-Store Unit Modifiers Load-Store Unit Controls
Global Memory Accesses Optimization x
Global Memory Bandwidth Use Calculation Manual Partition of Global Memory Partitioning Buffers Across Different Memory Types (Heterogeneous Memory) Partitioning Buffers Across Memory Channels of the Same Memory Type Ignoring Dependencies Between Accessor Arguments Contiguous Memory Accesses Static Memory Coalescing
Host x
Multi-Threaded Host Application Utilizing Hardware Kernel Invocation Queue Double Buffering Host Utilizing Kernel Invocation Queue N-Way Buffering to Overlap Kernel Execution Prepinning Memory Simple Host-Device Streaming Offload Processing Host-device Streaming Processing Multiple Kernel Pipeline Limitations Buffered Host-Device Streaming
Double Buffering Host Utilizing Kernel Invocation Queue x
Applying Double-Buffering Using the Intercept Layer for OpenCL* Applications
Resource Use x
Data Types and Operations Kernel Variable Accesses
Data Types and Operations x
Optimize Floating-point Operation Avoid Expensive Functions Variable-Precision Integer and Floating-Point Support
Variable-Precision Integer and Floating-Point Support x
Advantages and Limitations of Arbitrary Precision Data Types Declare and Use the AC Data Types
Declare and Use the AC Data Types x
Declare the <span class='codeph'>ac_int</span> Data Type Declare the <span class='codeph'>ac_fixed</span> Data Type Declare the <span class='codeph'>ac_complex</span> Data Type Declare the <span class='codeph'>ap_float</span> Data Type
Declare the <span class='codeph'>ap_float</span> Data Type x
Conversion Rules for <span class='codeph'>ap_float</span> Operations with Explicit Precision Controls Comparison Operators Additional <span class='codeph'>ap_float</span> Functions Additional Data Types Provided by the <span class='codeph'>ap_float.hpp</span> Header File Quality of Results and the ap_float Data Type
FPGA Optimization Flags, Attributes, Pragmas, and Extensions x
Optimization Flags Kernel Attributes Kernel Controls Kernel Variables Memory Attributes Loop Directives Floating-Point Pragmas Latency Controls (Beta) System of Tasks Extension (<span class='codeph'>task_sequence</span>)
Optimization Flags x
Specify Schedule FMAX Target for Kernels (<span class='codeph'>-Xsclock=<clock target>) Disable Burst-Interleaving of Global Memory (<span class='codeph'>-Xsno-interleaving=<global_memory_type></span>) Force Ring Interconnect for Global Memory (<span class='codeph'>-Xsglobal-ring</span>) Force a Single Store Ring to Reduce Area (<span class='codeph'>-Xsforce-single-store-ring</span>) Force Fewer Read Data Reorder Units to Reduce Area (<span class='codeph'>-Xsnum-reorder</span>) Disable Hardware Kernel Invocation Queue (<span class='codeph'>-Xsno-hardware-kernel-invocation-queue</span>) Modify the Handshaking Protocol Between Clusters (<span class='codeph'>-Xshyper-optimized-handshaking</span>) Disable Automatic Fusion of Loops (<span class='codeph'>-Xsdisable-auto-loop-fusion</span>) Fuse Adjacent Loops With Unequal Trip Counts (<span class='codeph'>-Xsenable-unequal-tc-fusion</span>) Pipeline Loops in Non-task Kernels (<span class='codeph'>-Xsauto-pipeline</span>) Control Semantics of Floating-Point Operations (<span class='codeph'>-fp-model=<var><value></var> </span>) Modify the Rounding Mode of Floating-point Operations (<span class='codeph'>-Xsrounding=<rounding_type></span>) Global Control of Exit FIFO Latency of Stall-free Clusters (<span class='codeph'>-Xssfc-exit-fifo-type=<var><value></var> </span>) Enable the Read-Only Cache for Read-Only Accessors (<span class='codeph'>-Xsread-only-cache-size=<var><N></var>)</span> Control Hardware Implementation of the Supported Data Types and Math Operations (<span class='codeph'>-Xsdsp-mode=<var><option></var> </span>)
Kernel Attributes x
Specify Schedule FMAX Target for Kernels Specify a Workgroup Size Specify Number of SIMD WorkItems Omit Hardware that Generates and Dispatches Kernel IDs Omit Hardware to Support the <span class='codeph'>no_global_work_offset</span> Attribute in <span class='codeph'>parallel_for</span> Kernels Reduce Kernel Area and Latency
Kernel Controls x
Pipes Extension
Pipes Extension x
Key Properties of a Pipe Accessing Pipes The <span class='codeph'>pipe</span> Class and its Use I/O Pipes Characteristics of Pipes Restrictions of Pipes Guidelines for Designing Pipes Pipe and Atomic Fence
Loop Directives x
<span class='codeph'>disable_loop_pipelining</span> Attribute <span class='codeph'>initiation_interval</span> Attribute <span class='codeph'>ivdep</span> Attribute <span class='codeph'>loop_coalesce</span> Attribute <span class='codeph'>max_concurrency</span> Attribute <span class='codeph'>max_interleaving</span> Attribute <span class='codeph'>speculated_iterations</span> Attribute <span class='codeph'>unroll</span> Pragma Loop Fuse Functions and <span class='codeph'>nofusion</span> Attribute
System of Tasks Extension (<span class='codeph'>task_sequence</span>) x
Task Functions <span class='codeph'>task_sequence</span> Use Cases
Quick Reference x
Algorithmic C Data Types Floating Point Pragmas FPGA Accessor Properties FPGA Extensions FPGA Kernel Attributes FPGA Local Memory Function Latency Control Properties (Beta) FPGA LSU Controls FPGA Loop Directives FPGA Memory Attributes FPGA Optimization Flags Pipe API <span class='codeph'>task_sequence</span> Template Parameters and Function APIs
FPGA Optimization Guide for Intel® oneAPI Toolkits

Simple Host-Device Streaming

You can take advantage of SYCL Unified Shared Memory host allocation and zero-copy host memory to implement a streaming host-device design with low latency and high throughput.

TIP:

Before you begin learning how to stream data, review Pipes and Zero-Copy Memory Access concepts.

NOTE:

SYCL USM host allocations are only supported by some BSPs, such as the Intel® FPGA Programmable Acceleration Card (PAC) D5005 (previously known as Intel® FPGA Programmable Acceleration Card (PAC) with Intel® Stratix® 10 SX FPGA). Check with your BSP vendor to see if they support SYCL USM host allocations.

To understand the concept, consider the following design example:

Offload Processing

Typical SYCL designs perform offload processing. All of the input data is prepared by the CPU, and then transferred to an FPGA device. Kernels are started on the device to process the data. When the kernels finish, the CPU copies the output data from the FPGA back to its memory. Memory synchronization is achieved on the host after the device kernel signals its completion.

Offload processing achieves excellent throughput when the memory transfers and kernel computation are performed on large data sets, as the CPU's kernel management overhead is minimized. You can conceal data transfer overhead using double buffering or N-way buffering to maximize kernel throughput. However, a significant shortcoming of this design pattern is latency. The coarse-grain synchronization of waiting for the entire set of data to be processed results in a latency that is equal to the processing time of the entire data.

Host-device Streaming Processing

The method for achieving lower latency between the host and device is to break data set into smaller chunks and, instead of enqueueing a single long-running kernel, launch a set of shorter-running kernels. Together, these shorter-running kernels process the data in smaller batches. As memory synchronization occurs upon kernel completion, this strategy makes the output data available to the CPU in a more granular way. This is illustrated in the following figure, where the orange lines illustrate the time when the first set of data is available in the host:

Host-device Streaming Processing

In the streaming version, the first piece of data is available in the host earlier than in the offload version. To understand how much earlier, consider that you have total_size elements of data to process and you break the computation into chunks chunks of size chunk_size=total_size/chunks (as is the case in Figure 1). Then, in an ideal situation, the streaming design achieves a latency that is chunks times better than the offload version.

Setting the chunk_size

You might wonder if you can set the chunk_size to 1 (i.e., chunks=total_size) to minimize the latency. In the figure above, you may notice small gaps between the kernels in the streaming design (for example, between K0 and K1). This is caused by the overhead of launching kernels and detecting kernel completion on the host. These gaps increase the total processing time and therefore, decrease the throughput of the design (that is, when compared to the offload design, it takes more time to process the same amount of data). If these gaps are negligible, then the throughput is negligibly affected.

In the streaming design, the choice of the chunk_size is thus a tradeoff between latency (a smaller chunk size results in a smaller latency) and throughput (a smaller chunk size increases the relevance of the inter-kernel latency).

Lower Bounds on Latency

Lowering the chunk_size can reduce latency, sometimes at the expense of throughput. However, even if you are not concerned with throughput, there still exists a lower-bound on the latency of a kernel. Review the following figure:

Lower Bounds on Latency

In the Figure 2:

  • tlaunch is the time for a kernel launch signal to go from the host to the device.
  • tkernel is the time for the kernel to execute on the device.
  • tfinish is the time for the finished signal to go from the device to the host.

Even if you set chunk_size to 0 (i.e., launch a kernel that does nothing) and therefore tkernel ~= 0, the latency is still tlaunch + tfinish. In other words, the lower bound on kernel latency is the time needed for the start signal to go from the host to the device and for the finished signal to go from the device to the host.

In the Setting the chunk_size section above, you learned how gaps between kernel invocations can degrade throughput. In the Figure 2, there appears to be a minimum tlaunch + tfinish gap between kernel invocations. This is illustrated as the Naive Relaunch timeline in the following figure:

Naive Relaunch Timeline

Fortunately, this overhead is circumvented by an automatic runtime kernel launch scheme that buffers kernel arguments on the device before the previous kernel finishes. This enables kernels to queue on the device and to begin execution without waiting for the previous kernel's finished to propagate back to the host. You can refer to this as Fast Kernel Relaunch, and it is also illustrated in the Figure 3. Fast kernel relaunch reduces the gap between kernel invocations and allows you to achieve lower latency while maintaining throughput.

Multiple Kernel Pipeline

More complicated FPGA designs often instantiate multiple kernels connected by SYCL pipes (for examples, see FPGA Reference Designs). Suppose you have a kernel system of N kernels connected by pipes, as shown in the following figure:
Multiple Kernels Connected by Pipes

With the goal of achieving lower latency, you use the Host-device Streaming Processing technique to launch multiple invocations of your N kernels to process chunks of data. This gives you a timeline as shown in the following figure:

Multiple Invocation of Kernels to Process Chunks of Data

Notice the gaps between the start times of the N kernels for a single chunk. This is the tlaunch time discussed in the Host-device Streaming Processing. However, the multi-kernel design introduces a potential new lower bound on the latency for a single chunk because processing a single chunk of data requires launching N kernels, which takes N x tlaunch. If N (the number of kernels in your system) is sufficiently large, this limits your achievable latency.

For designs with N > 2, Intel recommends a different approach. The idea is to enqueue your system of N kernels once, and to introduce Producer (P) and Consumer (C) kernels to handle the production and consumption of data from and to the host, respectively. This method is illustrated in the following figure:

Enqueuing a System of Kernels

The Producer streams data from the host and presents it to the kernel system through a SYCL pipe. The output of the kernel system is consumed by the Consumer and written back into host memory. To achieve low latency, you still process the data in chunks, but instead of having to enqueue N kernels for each chunk, you must only enqueue a single Producer and Consumer kernel per chunk. This enables you to reduce the lower bound on the latency to MAX(2 x tlaunch, tlaunch + tfinish). Observe that this lower bound does not depend on the number of kernels in your system (N).

Use this method only when N > 2. FPGA area is sacrificed to implement the Producer, Consumer, and their pipes to achieve lower overall processing latency.


Limitations

Fundamentally, the ability to stream data between the host and device is built around SYCL USM host allocations. The underlying problem is how to efficiently synchronize between the host and device to signal that some data is ready to be processed, or has been processed. In other words, how does the host signal to the device that some data is ready to be processed? Conversely, how does the device signal to the host that some data is done being processed?

One method to achieve this signaling is to use the start of a kernel to signal to the device that data is ready to be processed, and the end of a kernel to signal to the host that data has been processed. This is the approach taken in the examples discussed above. However, this method has two notable drawbacks. First, the latency to start and end kernels is high. To maintain high throughput, you must size the chunk_size sufficiently large to hide the inter-kernel latency, resulting in a latency increase. Second, the programming model to achieve this performance is non-trivial as you must intelligently manage the SYCL device queue.

NOTE:

For additional information, refer to the FPGA tutorial sample Simple Host Streaming listed in the Intel® oneAPI Samples Browser on Linux* or Windows*, or access the code sample in GitHub.

Parent topic: Host
Level Two Title