Skip to content
This repository has been archived by the owner on Oct 27, 2022. It is now read-only.

Latest commit

 

History

History
135 lines (84 loc) · 5.82 KB

F1Guide.md

File metadata and controls

135 lines (84 loc) · 5.82 KB
Raw

Optimization and Deployment Guide

THIS IS CURRENTLY A PILE OF MISCELLANEOUS STUFF.

TODO: Instructions for using a specific AMI version and checking the latest version.

These instructions were last debugged with SDx v2018.3. Check to see what version you have. If it differs, you might get to help us debug a new platform.

sdx -version

-j8

To the best of our understanding the actual AFI is stored permanently by AWS at no cost.

sdaccel_setup breaks git.

Tracing in hw_emu mode is TBD (see simon's answer to steve's xilinx post).

Instances for Development

These instances allow TCP/IP traffic through on port 80 for running a production web server and ports 8880-8889 for development. Information about this instance is , and you can find all the instance parameters in terraform.tfstate. Note that this file contains the private TLS key.

For access to these machines, TLS key pairs are placed into ~/.ssh/<instance-name>/ (locally). The public IP address of each instance is reported on the terminal upon completion of each command, though, when stopped and restarted, instances will get new IP addresses.

The following commands terminate the instance:

Note that this also deletes the created storage. (This step can be disabled in the ec2_instance.tf file.)

Remmina

The Remmina configuration file used by make commands is in <repo>/framework/build/template.remmina. It can be copied into ~/.remmina for interactive use, or modified to alter the options (such as adding SSH tunneling)

Dynamic Start/Stop of Static F1 Instances

1st CLaaS has support for deploying applications with an associated F1 instance. This instance is created statically using:

make static_accelerated_instance

REST calls to the Web Server are responsible for starting and stopping the instance. Stopping is performed by an "EC2 Time Bomb" process that is run in the background. If it is not pinged with a certain frequency, it will shut down the instance, so the client application is responsible for sending this periodic ping. (This approach ensures that the instance will stopped when clients fail to stop them explicitly.) See <repo>/framework/aws/ec2_time_bomb comments for usage details.

SSH keys

In case you want to set up passwordless authentication from your Development Instance for git access or any other reason, you may need to generate ssh keys for your instance.

ssh-keygen -o -t rsa -b 4096 -C "<machine-identifying comment>"
sudo yum install xclip -y
xclip -sel clip < ~/.ssh/id_rsa.pub

And paste this SSH key into the settings of your other account (e.g. gitlab/github).

Development of this Framework

These are WIP notes:

SDAccel

This is how I was able to generate an RTL kernel in SDAccel manually and run hardware emulation. This is scripted as part of the build process, but not with the ability to run in SDAccel. These are notes to help figure that out.

  • cd ~/workdisk/1st-CLaaS
  • source sdaccel_setup
  • sdx
  • echo $AWS_PLATFORM (You will need to know this path.)
  • On Welcome screen "Add Custon Platform".
    • Add (+) $AWS_PLATFORM/.. .
  • "New SDxProject":
    • Application.
    • name it, select only platform, accept linux on x86 and OpenCL runtime, and "Empty Application". "Finish".
  • Menu "Xilinx", "Create RTL Kernel".
    • name it; 2 clocks (independent kernel clock); 1 reset?
    • default scalars (or change them)
    • default AXI master (Global Memory) options (or change them)
    • wait for Vivado
  • In Vivado:
    • "Generate RTL Kernel"
    • wait
    • Close
  • Back in SDAccel:
    • in projects.sdx:
      • Add binary container under "Hardware Functions" by clicking the lightning bolt icon. There should be only one function identified as the target for the kernel. Click "OK". You will get "binary_container_1".
      • Select "Emulation-HW"
  • Menu "Run", "Run Configuration"
    • Tab "Main", "Kernel Debug", Check "Use RTL waveform...", and check "Launch live waveform".
    • Tab "Arguments", check "Automatically add binary container(s) to arguments"
    • Click "Run" (or "Apply" and click green circle w/ play arrow).

Tips and Tricks

Verilog Development

Use Vivado(TM) to develop your Verilog code (make edit_kernel). Vivado will show syntax errors (and more) as you type. If your code is clean in the editor you have a good chance of getting through synthesis and place-and-route. We've seen very misleading error messages on code that is not clean. For example, for hw_emu builds, we often see in vivado.log: Cannot find design unit xil_defaultlib.emu_wrapper, for which we have no explanation.

TL-Verilog

TL-Verilog development cannot be done using Vivado, but SandPiper(TM) tends to generate clean Verilog, so you are unlikely to see downstream errors. As a last resort, you can always (make edit_kernel) to get feedback from Vivado about the generated Verilog. Of course, the Verilog is not source code, but if you are careful, you can make experimental edits and compile with that edited Verilog code. The code will be overwritten by a build if the .tlv source is edited.