fpga-verification 0.2.5

Warning:

• The project is currently under development
• I try to release stable versions, but there may be bugs until the code has been fully tested and completed.

Reusable FPGA verification helpers. The package provides shared data formats, wire-protocol codecs, cocotb simulation utilities, and an Intel System Console HIL transport.

Install

Install the package:

python -m pip install fpga-verification

This installs the numeric format helpers, protocol codecs, cocotb simulation utilities, pyuvm agents, and Intel System Console HIL helpers together.

Public API

from fpga_verification.formats import QFormat, UIntFormat
from fpga_verification.protocols.avalon_st.intel_video import (
    IntelVIPFrameCodec,
    VIPControlPacket,
    VIPFrame,
    VIPInterlacing,
    VIPPacketType,
    VIPProtocolChecker,
    VIPProtocolError,
    VIPUserPacket,
    VIPVideoPacket,
    vip_packet_from_symbols,
)
from fpga_verification.sim.buses import (
    AvalonFormat,
    AvalonMMBus,
    AvalonMMMasterBFM,
    AvalonSTBeat,
    AvalonSTBus,
    AvalonSTFrame,
    AvalonSTMonitor,
    AvalonSTSink,
    AvalonSTSource,
)
from fpga_verification.sim.bfms.intel_dma import (
    DMAAddressRegion,
    IntelDMABFM,
    IntelDMACommandMonitor,
    SparseByteMemory,
)
from fpga_verification.sim.agents import VIPAgent, VIPItem, VIPMonitor, VIPSequence
from fpga_verification.sim.models import BaseVIPPredictor
from fpga_verification.sim.scoreboards import BaseVIPScoreboard
from fpga_verification.sim.platform_designer import platform_test_cocotb
from fpga_verification.sim.runners import intel_component_test_cocotb, rtl_test_cocotb
from fpga_verification.hil.intel import IntelSystemConsoleSession

Numeric Formats

UIntFormat and QFormat convert between Python/numpy values and raw integer words used by hardware buses, memories, and scoreboards.

UIntFormat

from fpga_verification.formats import UIntFormat

pixel = UIntFormat(width=10)
raw_pixels = pixel.array([0, 1023, 1024, -1])

assert raw_pixels.tolist() == [0, 1023, 0, 1023]
assert raw_pixels.dtype == pixel.dtype

Inputs:

• width: unsigned word width in bits, from 1 to 64.
• zeros(shape): creates a zero-filled numpy array.
• wrap(values): masks values to the configured width.
• array(values, shape=None): masks, casts to the smallest unsigned storage dtype, and optionally reshapes.

Outputs:

• dtype: numpy unsigned dtype selected from uint8, uint16, uint32, or uint64.
• mask: integer bit mask for the configured width.

QFormat

from fpga_verification.formats import QFormat

sample = QFormat(qi=3, qf=2, signed=True)
raw = sample.float_to_qraw([1.25, -1.0])
back = sample.qraw_to_float(raw)

assert raw.tolist() == [5, 28]
assert back.tolist() == [1.25, -1.0]

Inputs:

• qi: integer width. For signed formats, this includes the sign bit.
• qf: fractional width.
• signed: True for two's-complement signed values, False for unsigned.
• float_to_qraw(x, saturate=True): converts floats to raw fixed-point words.
• int_to_qraw(raw, saturate=True): converts signed integer values to raw stored words.
• qraw_to_int(raw): converts raw words to signed or unsigned integers.
• qraw_to_float(raw): converts raw words to floating-point values.
• multiply(left_raw, right_format, right_raw, out_qf=None): multiplies two raw fixed-point arrays. If out_qf is provided, the result is shifted to the requested fractional width.
• zeros(size=None), ones(size=None), full(size, value, raw=False), and randomize(...): create test data.

Outputs:

• width: total raw word width, qi + qf.
• scale: 2 ** qf.
• mask: integer bit mask for the raw word.
• min_float, max_float: representable numeric range.
• dtype: numpy unsigned storage dtype for the raw word.

Video frames

The neutral video layer is independent of cocotb and protocol-specific packet formats:

from fpga_verification.video import (
    FrameSize,
    ImageGenerator,
    VideoFormat,
    VideoPayloadCodec,
    compare_frames,
)

fmt = VideoFormat(
    bits_per_symbol=10,
    number_of_color_planes=3,
    color_planes_are_in_parallel=True,
    pixels_in_parallel=2,
)
size = FrameSize(width=640, height=480)
generator = ImageGenerator(fmt, rng=1)
frame = generator.random(size)

codec = VideoPayloadCodec(fmt)
payload_beats = codec.pack_frame(frame, size)
decoded = codec.unpack_frame(payload_beats, size)
compare_frames(decoded, frame)

VideoFormat contains only static AV-ST sample layout. FrameSize contains the width and height of one frame and is an explicit argument to every generation and conversion operation. One codec can therefore process frames with different resolutions without retaining hidden state.

Canonical frame shapes are (height, width) for one color plane and (height, width, planes) for multiple planes. Sample zero occupies the least significant payload bits. In parallel-plane mode each pixel's planes are adjacent; in serial-plane mode each beat carries one plane for pixels_in_parallel adjacent pixels.

Row-oriented adapters (row_to_symbols, pack_row, pack_frame) pad each incomplete row to the configured interface beat width and validate that padding on decode. Frame-symbol adapters (frame_to_symbols, symbols_to_frame) use a continuous raster stream with no per-row padding; protocols such as Intel VIP carry any final partial beat with Avalon-ST empty.

ImageGenerator provides constant, linspace, random, and horizontal_ramp. VideoPayloadCodec provides frame/row/symbol/beat round-trips and strict shape, range, payload-length, and padding validation.

Avalon-ST Protocols

Avalon-ST helpers follow the Avalon interface terminology used by Intel/Altera. The protocol reference is: https://docs.altera.com/r/docs/683091/22.3/avalon-interface-specifications/introduction-to-the-avalon-interface-specifications

The protocol codec layer is independent of cocotb and simulator state. It accepts and returns Python lists of symbols.

Intel Avalon-ST Video Packets

from fpga_verification.protocols.avalon_st.intel_video import (
    VIPControlPacket,
    VIPFrame,
    VIPInterlacing,
    VIPUserPacket,
    vip_packet_from_symbols,
)

control = VIPControlPacket(
    width=1920,
    height=1080,
    interlacing=VIPInterlacing.PROGRESSIVE_FRAME,
)
symbols = control.to_symbols()
decoded = vip_packet_from_symbols(symbols)

assert decoded.width == 1920
assert decoded.height == 1080

frame = VIPFrame(
    width=2,
    height=2,
    pixels=[0x10, 0x20, 0x30, 0x40],
    user_packets=[VIPUserPacket(1, [0xA, 0xB])],
)
packets = frame.packets()

Inputs:

• VIPControlPacket(width, height, interlacing=...): frame dimensions and interlacing metadata. Width and height must fit in 16 bits.
• VIPVideoPacket(payload): video payload symbols.
• VIPUserPacket(user_type, payload): user packet type 1..8 and payload symbols.
• VIPFrame(width, height, pixels, interlacing=..., user_packets=...): a black-box container that produces user, control, and video packets.
• vip_packet_from_symbols(symbols, symbols_per_beat=1): decodes one packet from raw symbols. symbols_per_beat controls how many symbols belong to the first Avalon-ST beat; payload starts after that first beat.

Outputs:

• to_symbols(): returns a list of 4-bit packet symbols.
• VIPFrame.control_packet(): returns a VIPControlPacket.
• VIPFrame.video_packet(): returns a VIPVideoPacket.
• VIPFrame.packets(): returns user packets followed by control and video packets.
• VIPInterlacing.description: human-readable interlacing mode.

Ancillary packets are currently reported as unsupported by the decoder.

VIP Protocol Checker

VIPProtocolChecker validates the packet order and active frame size for one observed Intel VIP stream. The codec stays stateless; the checker owns the stream state:

from fpga_verification.protocols.avalon_st.intel_video import VIPProtocolChecker

checker = VIPProtocolChecker(fmt)

for packet in observed_packets:
    checker.observe(packet)

The checker enforces these wire-visible rules:

• a video packet must follow a control packet;
• a reset clears the active control resolution when used through VIPMonitor.

It can also validate video payload length against the most recent control packet resolution and VideoFormat. In the current default mode a mismatch is logged as a warning; set checker.check_video_packet_size = True to raise VIPProtocolError instead.

Equal-area frame-size changes cannot be detected by a passive stream checker, because Intel VIP video packets do not carry width or height.

VIP pyuvm Agent

VIPAgent wraps Intel VIP source, monitor, sink, sequencer, and driver pieces for pyuvm environments. When a bus is provided, the corresponding monitor is created automatically and publishes decoded VIPPacket objects through its analysis port. Each monitor also runs VIPProtocolChecker before publishing.

from fpga_verification.sim.agents import VIPAgent, VIPSequence

vip_agent = VIPAgent(
    "vip_agent",
    parent=self,
    clock=dut.clk,
    reset=dut.reset,
    source_bus=din_bus,
    sink_bus=dout_bus,
    source_fmt=fmt,
    sink_fmt=fmt,
    packet_logging=True,
)

packets = codec.frame_to_packets(frame, size)
sequence = VIPSequence.from_packets(packets, name="input_frame")
await sequence.start(vip_agent.sequencer)

Inputs:

• source_bus and source_fmt: stream driven by the active agent and observed by source_monitor.
• sink_bus and sink_fmt: stream observed by sink_monitor; in active mode the sink monitor also drives ready/backpressure.
• is_active: active agents create a sequencer and source driver when source_bus is present. Passive agents only monitor provided buses.
• packet_logging and packet_log_level: optional packet summaries such as nuc_component.dout: got vip video packet (2048 symbols).
• set_packet_logging(enable, level=None): updates logging after build.
• randomize: enables randomized source pauses and sink backpressure.

Outputs:

• source_monitor.analysis_port: decoded packets observed on source_bus.
• sink_monitor.analysis_port: decoded packets observed on sink_bus.
• sequencer: accepts VIPSequence items in active source mode.

Base VIP Predictor And Scoreboard

BaseVIPPredictor and BaseVIPScoreboard split Intel VIP checking into two parts. The predictor consumes input packets and produces expected output packet descriptions. The scoreboard receives both DUT input and output packet streams, queues predictor expectations, and compares observed output packets against those expectations.

Packet-flow diagram:

from fpga_verification.sim.models import BaseVIPPredictor
from fpga_verification.sim.scoreboards import BaseVIPScoreboard


class MyVIPPredictor(BaseVIPPredictor):
    def get_tolerance(self):
        return 1


scoreboard = BaseVIPScoreboard(
    "vip_scoreboard",
    self,
    source_fmt=source_fmt,
    sink_fmt=sink_fmt,
)
scoreboard.predictor = MyVIPPredictor(
    core=model,
    input_codec=scoreboard.vip_input_codec,
    output_codec=scoreboard.vip_output_codec,
)

vip_agent.source_monitor.analysis_port.connect(scoreboard.data_in_export)
vip_agent.sink_monitor.analysis_port.connect(scoreboard.data_out_export)

Predictor behavior:

• process_packet(packet): dispatches input control, video, and user packets.
• Control packets update the active input FrameSize and emit an expected output control packet using expected_output_size(input_size).
• Video packets are decoded to neutral frames, passed to process_frame(frame, size), and encoded back to expected output video packets.
• support_passthrough=True allows set_mode(BaseVIPPredictor.IpMode.PASSTHROUGH), where input frames are expected unchanged.
• Override expected_output_size(input_size), is_supported_frame_size(size), get_tolerance(), process_frame(frame, size), or _process_user_packet() for IP-specific behavior.

Scoreboard behavior:

• predictor must be assigned before packets arrive.
• data_in_export: connect packets observed before the DUT.
• data_out_export: connect packets observed after the DUT.
• expected_queue: stores predictor expectations until matching output packets arrive.
• Control packets compare width, height, and interlacing.
• Video packets compare decoded frames with compare_frames() using the expectation tolerance.
• enable_compare=False keeps packet ordering checks but disables frame content comparison.
• wait_frame_checked(timeout=1, timeout_unit="ms"): wait until one more output frame has been checked or raise the stored failure.

Avalon-ST Cocotb Bus Helpers

The cocotb bus helpers drive and observe Avalon-ST interfaces through cocotb handles. They support scalar valid/ready, optional packet signals, optional empty, error, and channel, and ready modes ready_latency=0 or ready_latency=1.

Instantiating A Source And Sink

import cocotb
from cocotb.clock import Clock
from cocotb.triggers import RisingEdge

from fpga_verification.sim.buses import (
    AvalonFormat,
    AvalonSTBus,
    AvalonSTFrame,
    AvalonSTSink,
    AvalonSTSource,
)


@cocotb.test()
async def stream_loopback_test(dut):
    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())

    dut.reset.value = 1
    await RisingEdge(dut.clk)
    dut.reset.value = 0

    fmt = AvalonFormat(bits_per_symbol=8, symbols_per_beat=1)

    source = AvalonSTSource(
        AvalonSTBus.from_prefix(dut, "sink"),
        fmt,
        dut.clk,
        reset=dut.reset,
        packets=True,
    )
    sink = AvalonSTSink(
        AvalonSTBus.from_prefix(dut, "source"),
        fmt,
        dut.clk,
        reset=dut.reset,
        packets=True,
    )

    await source.send(AvalonSTFrame([0x11, 0x22, 0x33]))
    received = await sink.recv()

    assert received.data == [0x11, 0x22, 0x33]

Inputs:

• AvalonSTBus.from_prefix(dut, prefix): binds signals named like <prefix>_data, <prefix>_valid, <prefix>_ready, <prefix>_startofpacket, and <prefix>_endofpacket.
• AvalonSTFrame(data, channel=None, error=None, empty=None, tx_complete=None): frame payload and optional sideband metadata.
• AvalonFormat(bits_per_symbol=8, symbols_per_beat=1, first_symbol_in_high_order_bits=False): static symbol layout for the stream data word.
• AvalonSTSource(bus, fmt, clock, reset=None, reset_active_level=True, ready_latency=0, ready_allowance=None, packets=None, idle_value="x").
• AvalonSTSink(...) and AvalonSTMonitor(...): use the same AvalonFormat and timing options as AvalonSTSource.
• send(frame) / send_nowait(frame): queue transmit data.
• recv() / recv_nowait(): receive complete frames.
• recv_beat() / recv_beat_nowait(): receive one transferred beat.
• set_pause_generator(generator): apply backpressure or idle insertion from an iterable of booleans.

Outputs:

• AvalonSTFrame.data: list of symbols.
• AvalonSTFrame.channel, error, empty: captured sideband metadata.
• AvalonSTFrame.sim_time_start, sim_time_end: simulation timestamps.
• AvalonSTBeat: one handshake beat with data, decoded symbols, sop, eop, empty, error, channel, and sim_time.
• wait(): waits for a source to become idle or a monitor/sink to see activity, depending on the helper type.

Avalon-MM Cocotb Bus Helpers

AvalonMMMasterBFM is a lightweight Avalon-MM host BFM for register-style cocotb tests. It issues one transaction at a time and is intentionally simpler than the full Avalon-MM protocol surface.

import cocotb
from cocotb.clock import Clock
from cocotb.triggers import RisingEdge

from fpga_verification.sim.buses import AvalonMMMasterBFM


@cocotb.test()
async def control_register_test(dut):
    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())

    mm = AvalonMMMasterBFM.from_prefix(
        dut,
        "control",
        dut.clk,
        reset=dut.reset,
        default_byteenable=0xF,
    )
    mm.init_idle()

    dut.reset.value = 1
    await RisingEdge(dut.clk)
    dut.reset.value = 0
    await mm.wait_reset_release(active_value=1)

    await mm.write(0x00, 0x00000001, timeout_cycles=32)
    status = await mm.read(0x04, timeout_cycles=32)
    await mm.wait_set(0x04, 0x1, timeout_cycles=256)

AvalonMMMemoryBFM is a slave-side BFM for full-IP tests where the DUT exposes Avalon-MM master ports. It can connect read-only, write-only, or read/write master ports to any byte-addressed memory object with read(address, length) and write(address, data) methods, including SparseByteMemory.

from fpga_verification.sim.buses import AvalonMMMemoryBFM
from fpga_verification.sim.bfms.intel_dma import SparseByteMemory


memory = SparseByteMemory()
memory.write(0x1000, b"\x01\x02\x03\x04")

rd_mem = AvalonMMMemoryBFM.from_prefix(
    dut,
    "mem_master_rd",
    dut.mem_clk,
    reset=dut.mem_reset,
    memory=memory,
).start()

wr_mem = AvalonMMMemoryBFM.from_prefix(
    dut,
    "mem_master_wr",
    dut.mem_clk,
    reset=dut.mem_reset,
    memory=memory,
).start()

Inputs:

• AvalonMMBus.from_prefix(dut, prefix): binds required <prefix>_address plus optional <prefix>_writedata, <prefix>_write, <prefix>_read, <prefix>_readdata, <prefix>_waitrequest, <prefix>_readdatavalid, <prefix>_byteenable, <prefix>_burstcount, <prefix>_beginbursttransfer, <prefix>_response, <prefix>_writeresponsevalid, <prefix>_lock, and <prefix>_debugaccess.
• AvalonMMMasterBFM(bus, clock, reset=None, read_response_latency=0, default_byteenable=None, packet_logging=False, packet_log_level=logging.INFO): creates a single-beat Avalon-MM host.
• AvalonMMMemoryBFM(bus, clock, reset=None, memory=..., read_latency=1, byteorder="little"): creates a slave-side byte-addressed memory BFM.
• init_idle(): drives host outputs to idle values.
• write(address, data, byteenable=None, timeout_cycles=None): issues one write and waits until waitrequest is deasserted, when present.
• read(address, byteenable=None, timeout_cycles=None): issues one read and waits for readdatavalid when present, otherwise waits the configured fixed read_response_latency.
• read_modify_write(address, update, ...): convenience read/update/write.
• poll(address, predicate, ...), wait_set(address, mask, ...), and wait_clear(address, mask, ...): register polling helpers.
• AvalonMMMemoryBFM.read_transactions and write_transactions: observed memory-side transfer beats.

Supported Avalon-MM features:

• Master BFM: single-beat read and write transfers for register access.
• Memory BFM: read and write bursts via burstcount.
• Optional waitrequest backpressure.
• Optional readdatavalid variable-latency read completion.
• Optional fixed read response latency when readdatavalid is absent.
• Optional byteenable, defaulting to all byte lanes asserted when present.
• Separate read-only and write-only master ports sharing one backing memory.
• Intel mSGDMA-style write bursts where address and burstcount remain constant while each accepted write beat advances the memory address.
• Width validation for address, data, and byteenable values.

Unsupported features:

• Master BFM burst generation.
• Out-of-order read responses.
• Read/write response status behavior beyond idle driving of optional response and writeresponsevalid.
• waitrequestAllowance, active-low role variants, reset-interface timing, and Platform Designer address-unit/alignment property modeling.

Reference: https://docs.altera.com/r/docs/683091/current

Intel DMA BFM

IntelDMABFM is a cocotb black-box model for Intel read and write DMA streaming interfaces. It consumes DMA command descriptors, emits DMA responses, sources read data from memory, and stores write data into memory.

import cocotb
from cocotb.clock import Clock
from cocotb.triggers import RisingEdge

from fpga_verification.sim.buses import AvalonSTBus
from fpga_verification.sim.bfms.intel_dma import (
    DMAAddressRegion,
    IntelDMABFM,
    IntelDMACommandMonitor,
    SparseByteMemory,
)


@cocotb.test()
async def dma_component_test(dut):
    cocotb.start_soon(Clock(dut.clk, 10, units="ns").start())

    memory = SparseByteMemory()
    memory.write(0x1000, b"\x01\x02\x03\x04")

    dma = IntelDMABFM(
        dut,
        clock=dut.clk,
        reset=dut.reset,
        memory=memory,
        mode="full",
    ).start()

    command_monitor = IntelDMACommandMonitor(
        clock=dut.clk,
        reset=dut.reset,
        rdma_cmd_bus=AvalonSTBus.from_prefix(dut, "rdma_cmd"),
        wdma_cmd_bus=AvalonSTBus.from_prefix(dut, "wdma_cmd"),
        read_address_regions=[DMAAddressRegion("input", 0x1000, 0x4000)],
        write_address_regions=[DMAAddressRegion("output", 0x8000, 0x4000)],
    ).start()

    dut.reset.value = 1
    await RisingEdge(dut.clk)
    dut.reset.value = 0

    # Drive the DUT here. The BFM responds on the DMA Avalon-ST interfaces.
    # Later, inspect memory or descriptor logs as black-box outputs.
    written_bytes = memory.read(0x8000, 16)
    read_descriptors = command_monitor.read_descriptors

    dma.stop()
    command_monitor.stop()

Inputs:

• IntelDMABFM(dut, clock, reset, memory=None, read_response_delay_cycles=2, write_response_delay_cycles=2, ..., mode="full").
• mode: "full", "read"/"read_only", or "write"/"write_only".
• memory: optional SparseByteMemory shared by read and write paths.
• Optional bus overrides: rdma_cmd_bus, rdma_resp_bus, wdma_cmd_bus, wdma_resp_bus, din_bus, and dout_bus. If omitted, buses are discovered from DUT prefixes with the same names.
• SparseByteMemory.write(address, data): initializes byte-addressed memory.
• DMAAddressRegion(name, start, size): allowed address interval for passive checking. End address is exclusive.
• IntelDMACommandMonitor(...): pass command buses or existing AvalonSTMonitor instances and optional allowed address regions.

Outputs:

• SparseByteMemory.read(address, length): returns bytes stored by the BFM.
• IntelDMABFM.read_commands, write_commands: descriptor queues observed by the model.
• IntelDMABFM.read_responses, write_responses: queues of descriptors whose responses were issued.
• IntelDMACommandMonitor.read_descriptors, write_descriptors: decoded descriptor history.
• ReadDMADescriptor.decode(value) and WriteDMADescriptor.decode(value): convert raw descriptor words into address, length, and control fields.

HIL Session

IntelSystemConsoleSession opens one persistent system-console process and uses it sequentially for Avalon-MM memory access and JTAG UART commands. Intel Quartus system-console must be available on PATH.

import numpy as np

from fpga_verification.hil.intel import IntelSystemConsoleSession

frame = np.arange(1024 * 1280, dtype=np.uint16).reshape(1024, 1280)

with IntelSystemConsoleSession(
    system_console="system-console",
    master_index=0,
    uart_index=0,
    startup_timeout=30.0,
    work_dir=".",
) as hw:
    hw.write_memory(frame, address=0x01E84800)
    response = hw.command("g\n", timeout=3.0)
    frame_out = hw.read_memory((1024, 1280), address=0x02DC6C00)

print(response)
print(frame_out.shape)

Inputs:

• system_console: executable name or path.
• master_index: System Console Avalon-MM master index.
• uart_index: JTAG UART service index.
• startup_timeout: seconds to wait for the Tcl worker to become ready.
• work_dir: directory used for temporary binary transfer files.
• write_memory(data, address, chunk_size=4096): writes numpy-compatible data as little-endian 16-bit words.
• read_memory(shape, address, chunk_size=4096): reads little-endian 16-bit words and reshapes them.
• command(command, timeout=3.0, debug=False): sends a UTF-8 command over JTAG UART and waits for the first non-empty response line.

Outputs:

• read_memory(...): numpy array with the requested shape.
• command(...): response string.
• Methods raise TimeoutError or RuntimeError if System Console stops or reports a protocol error.

Simulation Runners

The simulation helpers cover three levels of generated and non-generated designs:

rtl_runner
  RTL sources -> cocotb build/test

intel_component_runner
  *_hw.tcl -> ip-generate -> generated composition HDL + original RTL -> rtl_runner

platform_runner
  already generated Platform Designer sim dir/msim_setup.tcl -> simulator flow

rtl_test_cocotb is the direct RTL path. Pass it explicit HDL sources or source directories, and it delegates build/test to the selected cocotb simulator runner.

intel_component_test_cocotb is for Platform Designer component .tcl files. It generates only the HDL needed for simulation, keeps composition HDL that has no source equivalent, replaces generated copies of project RTL with exact matches from source_dirs, and then calls rtl_test_cocotb.

Its generated-catalog flow is:

source_dirs
  -> ip-make-ipx --thorough-descent --source-directory=<source_dirs>
  -> components.ipx in generated temp dir
  -> ip-generate --search-path=<components.ipx>,$
  -> parse .spd
  -> replace generated RTL copies with original source files
  -> rtl_test_cocotb

Pass generate_only=True to retain and return the generated composition directory without running simulation.

platform_test_cocotb is for already generated Platform Designer simulation trees. The expected layout is:

project_root/
  <hdl_toplevel>/
    <hdl_toplevel>/
      testbench/
        mentor/
          msim_setup.tcl

For Questa, the platform runner compiles through msim_setup.tcl and runs cocotb against the generated simulator libraries. For Verilator, it reads Verilog/SystemVerilog sources from msim_setup.tcl and builds them directly.