Debugging & Testing with Metasimulation

When discussing RTL simulation in FireSim, we are generally referring to metasimulation: simulating the FireSim simulator’s RTL, typically using VCS or Verilator. In contrast, we’ll refer to simulation of the target’s unmodified (by GoldenGate decoupling, host and target transforms) RTL as target-level simulation. Target-level simulation in Chipyard is described at length here.

Metasimulation is the most productive way to catch bugs before generating an AGFI, and a means for reproducing bugs seen on the FPGA. By default, metasimulation uses an abstract but fast model of the host: the FPGA’s DRAM controllers are modeled with a single-cycle memory system, the PCI-E subsystem is not simulated, instead the driver presents DMA and MMIO traffic directly on the FPGATop interfaces. Since FireSim simulations are robust against timing differences across hosts, target behavior observed in an FPGA-hosted simulation should be exactly reproducible in a metasimulation.

As a final note, metasimulations are generally only slightly slower than target-level simulations. Example performance numbers can be found at Metasimulation vs. Target simulation performance.

Supported Host Simulators

Currently, the following host simulators are supported for metasimulation:

  • Verilator

    • FOSS, automatically installed during FireSim setup.

    • Referred to throughout the codebase as verilator.

  • Synopsys VCS

    • License required.

    • Referred to throughout the codebase as vcs.

Pull requests to add support for other simulators are welcome.

Running Metasimulations using the FireSim Manager

The FireSim manager supports running metasimulations using the standard firesim {launchrunfarm, infrasetup, runworkload, terminaterunfarm} flow that is also used for FPGA-accelerated simulations. Rather than using FPGAs, these metasimulations run within one of the aforementioned software simulators (Supported Host Simulators) on standard compute hosts (i.e. those without FPGAs). This allows users to write a single definition of a target (configured design and software workload), while seamlessly moving between software-only metasimulations and FPGA-accelerated simulations.

As an example, if you have the default config_runtime.yaml that is setup for FPGA-accelerated simulations (e.g. the one used for the 8-node networked simulation from the :ref:cluster-sim section), a few modifications to the configuration files can convert it to running a distributed metasimulation.

First, modify the existing metasimulation mapping in config_runtime.yaml to the following:

metasimulation:
    metasimulation_enabled: true
    # vcs or verilator. use vcs-debug or verilator-debug for waveform generation
    metasimulation_host_simulator: verilator
    # plusargs passed to the simulator for all metasimulations
    metasimulation_only_plusargs: "+fesvr-step-size=128 +max-cycles=100000000"
    # plusargs passed to the simulator ONLY FOR vcs metasimulations
    metasimulation_only_vcs_plusargs: "+vcs+initreg+0 +vcs+initmem+0"

This configures the manager to run Verilator-hosted metasimulations (without waveform generation) for the target specified in config_runtime.yaml. When in metasimulation mode, the default_hw_config that you specify in target_config references an entry in config_build_recipes.yaml instead of an entry in config_hwdb.yaml.

As is the case when the manager runs FPGA-accelerated simulations, the number of metasimulations that are run is determined by the parameters in the target_config section, e.g. topology and no_net_num_nodes. Many parallel metasimulations can then be run by writing a FireMarshal workload with a corresponding number of jobs.

In metasimulation mode, the run farm configuration must be able to support the required number of metasimulations (see run_farm for details). The num_metasims parameter on a run farm host specification defines how many metasimulations are allowed to run on a particular host. This corresponds with the num_fpgas parameter used in FPGA-accelerated simulation mode. However num_metasims does not correspond as tightly with any physical property of the host; it can be tuned depending on the complexity of your design and the compute/memory resources on a host.

For example, in the case of the AWS EC2 run farm (aws_ec2.yaml), we define three instance types (z1d.{3, 6, 12}xlarge) by default that loosely correspond with f1.{2, 4, 16}xlarge instances, but instead have no FPGAs and run only metasims (of course, the f1.* instances could run metasims, but this would be wasteful):

run_farm_hosts_to_use:
    - z1d.3xlarge: 0
    - z1d.6xlarge: 0
    - z1d.12xlarge: 1

run_farm_host_specs:
    - z1d.3xlarge:
        num_fpgas: 0
        num_metasims: 1
        use_for_switch_only: false
    - z1d.6xlarge:
        num_fpgas: 0
        num_metasims: 2
        use_for_switch_only: false
    - z1d.12xlarge:
        num_fpgas: 0
        num_metasims: 8
        use_for_switch_only: false

In this case, the run farm will use a z1d.12xlarge instance to host 8 metasimulations.

To generate waveforms in a metasimulation, change metasimulation_host_simulator to a simulator ending in -debug (e.g. verilator-debug). When running with a simulator with waveform generation, make sure to add waveform.vpd to the common_simulation_outputs area of your workload JSON file, so that the waveform is copied back to your manager host when the simulation completes.

A last notable point is that unlike the normal FPGA simulation case, there are two output logs in metasimulations. There is the expected uartlog file that holds the stdout from the metasimulation (as in FPGA-based simulations). However, there will also be a metasim_stderr.out file that holds stderr coming out of the metasimulation, commonly populated by printf calls in the RTL, including those that are not marked for printf synthesis. If you want to copy metasim_stderr.out to your manager when a simulation completes, you must add it to the common_simulation_outputs of the workload JSON.

Other than the changes discussed in this section, manager behavior is identical between FPGA-based simulations and metasimulations. For example, simulation outputs are stored in deploy/results-workload/ on your manager host, FireMarshal workload definitions are used to supply target software, etc. All standard manager functionality is supported in metasimulations, including running networked simulations and using existing FireSim debugging tools (i.e. AutoCounter, TracerV, etc).

Once the configuration changes discussed thus far in this section are made, the standard firesim {launchrunfarm, infrasetup, runworkload, terminaterunfarm} set of commands will run metasimulations.

If you are planning to use FireSim metasimulations as your primary simulation tool while developing a new target design, see the (optional) firesim builddriver command, which can build metasimulations through the manager without requiring run farm hosts to be launched or accessible. More about this command is found in the firesim builddriver section.

Understanding a Metasimulation Waveform

Module Hierarchy

To build out a simulator, Golden Gate adds multiple layers of module hierarchy to the target design and performs additional hierarchy mutations to implement bridges and resource optimizations. Metasimulation uses the FPGATop module as the top-level module, which excludes the platform shim layer (F1Shim, for EC2 F1). The original top-level of the input design is nested three levels below FPGATop:

../../_images/metasim-module-hierarchy.png

The module hierarchy visible in a typical metasimulation.

Note that many other bridges (under FPGATop), channel implementations (under SimWrapper), and optimized models (under FAMETop) may be present, and vary from target to target. Under the FAMETop module instance you will find the original top-level module (FireSimPDES_, in this case), however it has now been host-decoupled using the default LI-BDN FAME transformation and is referred to as the hub model. It will have ready-valid I/O interfaces for all of the channels bound to it, and internally containing additional channel enqueue and clock firing logic to control the advance of simulated time. Additionally, modules for bridges and optimized models will no longer be found contained in this submodule hierarchy. Instead, I/O for those extracted modules will now be as channel interfaces.

Clock Edges and Event Timing

Since FireSim derives target clocks by clock gating a single host clock, and since bridges and optimized models may introduce stalls of their own, timing of target clock edges in a metasimulation will appear contorted relative to a conventional target-simulation. Specifically, the host-time between clock edges will not be proportional to target-time elapsed over that interval, and will vary in the presence of simulator stalls.

Finding The Source Of Simulation Stalls

In the best case, FireSim simulators will be able to launch new target clock pulses on every host clock cycle. In other words, for single-clock targets the simulation can run at FMR = 1. In the single clock case delays are introduced by bridges (like FASED memory timing models) and optimized models (like a multi-cycle Register File model). You can identify which bridges are responsible for additional delays between target clocks by filtering for *sink_valid and *source_ready on the hub model. When <channel>_sink_valid is deasserted, a bridge or model has not yet produced a token for the current timestep, stalling the hub. When <channel>_source_ready is deasserted, a bridge or model is back-pressuring the channel.

Scala Tests

To make it easier to do metasimulation-based regression testing, the ScalaTests wrap calls to Makefiles, and run a limited set of tests on a set of selected designs, including all of the MIDAS examples and a handful of Chipyard-based designs. This is described in greater detail in the Developer documentation.

Running Metasimulations through Make

Warning

This section is for advanced developers; most metasimulation users should use the manager-based metasimulation flow described above.

Metasimulations are run out of the firesim/sim directory. If you are running a metasim for Chipyard, ensure you properly add the TARGET_PROJECT_MAKEFRAG variable to point to Chipyard’s makefrag. Generally this is set to TARGET_PROJECT_MAKEFRAG=${CY_DIR}/generators/firechip/src/main/makefrag/firesim.

[in firesim/sim]
make <verilator|vcs>

To compile a simulator with full-visibility waveforms, type:

make <verilator|vcs>-debug

As part of target-generation, Rocket Chip emits a make fragment with recipes for running suites of assembly tests. MIDAS puts this in firesim/sim/generated-src/f1/<DESIGN>-<TARGET_CONFIG>-<PLATFORM_CONFIG>/firesim.d. Make sure your $RISCV environment variable is set by sourcing firesim/sourceme-manager.sh or firesim/env.sh, and type:

make run-<asm|bmark>-tests EMUL=<vcs|verilator>

To run only a single test, the make target is the full path to the output. Specifically:

make EMUL=<vcs|verilator> $PWD/output/f1/<DESIGN>-<TARGET_CONFIG>-<PLATFORM_CONFIG>/<RISCV-TEST-NAME>.<vpd|out>

A .vpd target will use (and, if required, build) a simulator with waveform dumping enabled, whereas a .out target will use the faster waveform-less simulator.

Additionally, you can run a unique binary in the following way:

make SIM_BINARY=<PATH_TO_BINARY> run-<vcs|verilator>
make SIM_BINARY=<PATH_TO_BINARY> run-<vcs|verilator>-debug

Examples

Run all RISCV-tools assembly and benchmark tests on a Verilated simulator.

[in firesim/sim]
make
make -j run-asm-tests
make -j run-bmark-tests

Run all RISCV-tools assembly and benchmark tests on a Verilated simulator with waveform dumping.

make verilator-debug
make -j run-asm-tests-debug
make -j run-bmark-tests-debug

Run rv64ui-p-simple (a single assembly test) on a Verilated simulator.

make
make $(pwd)/output/f1/FireSim-FireSimRocketConfig-BaseF1Config/rv64ui-p-simple.out

Run rv64ui-p-simple (a single assembly test) on a VCS simulator with waveform dumping.

make vcs-debug
make EMUL=vcs $(pwd)/output/f1/FireSim-FireSimRocketConfig-BaseF1Config/rv64ui-p-simple.vpd

Metasimulation vs. Target simulation performance

Generally, metasimulations are only slightly slower than target-level simulations. This is illustrated in the chart below.

Type

Waves

VCS

Verilator -O1

Verilator -02

Verilator

Target

Off

4.8 kHz

3.9 kHz

6.6 kHz

N/A

Target

On

0.8 kHz

3.0 kHz

5.1 kHz

N/A

Meta

Off

3.8 kHz

2.4 kHz

4.5 kHz

5.3 KHz

Meta

On

2.9 kHz

1.5 kHz

2.7 kHz

3.4 KHz

Note that using more aggressive optimization levels when compiling the Verilated-design dramatically lengthens compile time:

Type

Waves

VCS

Verilator -O1

Verilator -02

Verilator

Meta

Off

35s

48s

3m32s

4m35s

Meta

On

35s

49s

5m27s

6m33s

Notes: Default configurations of a single-core, Rocket-based instance running rv64ui-v-add. Frequencies are given in target-Hz. Presently, the default compiler flags passed to Verilator and VCS differ from level to level. Hence, these numbers are only intended to give ball park simulation speeds, not provide a scientific comparison between simulators. VCS numbers collected on a local Berkeley machine, Verilator numbers collected on a c4.4xlarge. (metasimulation Verilator version: 4.002, target-level Verilator version: 3.904)