BSG Replicant: Cosimulation and Emulation Infrastructure for HammerBlade

Contents

This repository contains the following folders:

• build: Vivado scripts for building FPGA Design Checkpoint Files to upload to AWS-F1
• hardware: HDL sources, ROM scripts, and package files
• libraries: C/C++ driver and CUDA-lite Runtime library sources
• examples: Example C/C++ applications and regression tests.
• machines: Customized Makefile.machine.include file for different HammerBlade designs.

This repository contains the following files:

• README.md: This file
• Makefile: Contains targets for Bitstream Generation, and Regression
• platform.mk: Defines the path to the current exeuction Platform (BSG_PLATFORM_PATH)
• machine.mk: Defines the path to the current Machine Configuration (BSG_MACHINE_PATH)
• environment.mk: A makefile fragment for deducing the build environment.
• cadenv.mk: A makefile fragment for deducing the CAD tool environment (e.g. VCS_HOME)
• hdk.mk: A makefile fragment for deducing the AWS-FPGA HDK build environment.

Platforms

HammerBlade applications can be run on multiple platforms. These platforms could simulate the manycore and run the host code natively (VCS, Verilator), emuate the manycore (AWS F1) and run the host code natively, or simulate the manycore and host (coming soon!).

We currently support four platforms:

• aws-fpga: Native (x86) host execution, Emulated Manycore (with an FPGA),
• aws-vcs: Native (x86) host execution, Simulated Manycore (with VCS, on a simulated FPGA)
• dpi-verilator: Native (x86) host execution, Simulated Manycore (with Verilator, using DPI for IO)
• dpi-vcs: Native (x86) host execution, Simulated Manycore (with VCS, using DPI for IO)

Each platform has different advantages and drawbacks. Simulated platforms support an in-depth profiling infrastructure and emulated memory systems via non-synthesizable constructs. VCS is a 4-state simulator, but requires Synopsys and VCS licenses. Verilator does not require licenses, but it cannot simulated an FPGA system with encrypted HDL. The aws-fpga systems is much faster, but has a limited size, and no introspection or profiling tools.

To select the execution platform, set the BSG_PLATFORM variable in platform.mk. Most users will use dpi-vcs or dpi-verilator. Users with Vivado installed can use aws-vcs. Users with access to AWS F1 images should use aws-fpga.

Machines

Each HammerBlade configuration is called a Machine and defines a size, memory type, memory hierarchy, cache type, and many other parameters. Each machine is defined by Makefile.machine.include file in machines.

To switch machines set BSG_MACHINE_PATH, defined in machine.mk. To switch machines, modify the value of BSG_MACHINE_PATH to point to any subdirectory of the machines directory.

See machines/README.md for more documentation.

Quick-Start

The simplest way to use this project is to clone its meta-project: BSG Bladerunner.

BSG Bladerunner tracks this repository, BSG Manycore, and BaseJump STL repositories as submodules and maintains a monotonic versionining scheme.

Once the setup instructions in BSG Bladrunner have been completed, run:

make regression

This will run all of the example programs in examples. See the README in that directory for more information.

Dependencies

To use the aws-vcs platform, users will need:

1. Vivado 2019.1
2. A clone of aws-fpga (Cloned by bsg_bladerunner)
3. Synopsys VCS (We use O-2018.09-SP2, but others would work)

To use the dpi-verilator platform, users will need:

1. A recent version of Verilator

Users should use the Verilator installation provided by bsg_bladerunner.

This repository depends on the following repositories:

1. BSG Manycore
2. BaseJump STL
3. AWS FPGA