6. Model Tuning with microTVM

Authors: Andrew Reusch, Mehrdad Hessar

This tutorial explains how to autotune a model using the C runtime.

Install microTVM Python dependencies

TVM does not include a package for Python serial communication, so we must install one before using microTVM. We will also need TFLite to load models.

%%shell
pip install pyserial==3.5 tflite==2.1
# You can skip the following section (installing Zephyr) if the following flag is False.
# Installing Zephyr takes ~20 min.
import os

use_physical_hw = bool(os.getenv("TVM_MICRO_USE_HW"))

Install Zephyr

%%shell
# Install west and ninja
python3 -m pip install west
apt-get install -y ninja-build

# Install ZephyrProject
ZEPHYR_PROJECT_PATH="/content/zephyrproject"
export ZEPHYR_BASE=${ZEPHYR_PROJECT_PATH}/zephyr
west init ${ZEPHYR_PROJECT_PATH}
cd ${ZEPHYR_BASE}
git checkout v3.2-branch
cd ..
west update
west zephyr-export
chmod -R o+w ${ZEPHYR_PROJECT_PATH}

# Install Zephyr SDK
cd /content
ZEPHYR_SDK_VERSION="0.15.2"
wget "https://github.com/zephyrproject-rtos/sdk-ng/releases/download/v${ZEPHYR_SDK_VERSION}/zephyr-sdk-${ZEPHYR_SDK_VERSION}_linux-x86_64.tar.gz"
tar xvf "zephyr-sdk-${ZEPHYR_SDK_VERSION}_linux-x86_64.tar.gz"
mv "zephyr-sdk-${ZEPHYR_SDK_VERSION}" zephyr-sdk
rm "zephyr-sdk-${ZEPHYR_SDK_VERSION}_linux-x86_64.tar.gz"

# Install python dependencies
python3 -m pip install -r "${ZEPHYR_BASE}/scripts/requirements.txt"

Import Python dependencies

import json
import numpy as np
import pathlib

import tvm
from tvm.relay.backend import Runtime
import tvm.micro.testing

Defining the model

To begin with, define a model in Relay to be executed on-device. Then create an IRModule from relay model and fill parameters with random numbers.

data_shape = (1, 3, 10, 10)
weight_shape = (6, 3, 5, 5)

data = tvm.relay.var("data", tvm.relay.TensorType(data_shape, "float32"))
weight = tvm.relay.var("weight", tvm.relay.TensorType(weight_shape, "float32"))

y = tvm.relay.nn.conv2d(
    data,
    weight,
    padding=(2, 2),
    kernel_size=(5, 5),
    kernel_layout="OIHW",
    out_dtype="float32",
)
f = tvm.relay.Function([data, weight], y)

relay_mod = tvm.IRModule.from_expr(f)
relay_mod = tvm.relay.transform.InferType()(relay_mod)

weight_sample = np.random.rand(
    weight_shape[0], weight_shape[1], weight_shape[2], weight_shape[3]
).astype("float32")
params = {"weight": weight_sample}

Defining the target

Now we define the TVM target that describes the execution environment. This looks very similar to target definitions from other microTVM tutorials. Alongside this we pick the C Runtime to code generate our model against.

When running on physical hardware, choose a target and a board that describe the hardware. There are multiple hardware targets that could be selected from PLATFORM list in this tutorial. You can chose the platform by passing –platform argument when running this tutorial.

RUNTIME = Runtime("crt", {"system-lib": True})
TARGET = tvm.micro.testing.get_target("crt")

# Compiling for physical hardware
# --------------------------------------------------------------------------
#  When running on physical hardware, choose a TARGET and a BOARD that describe the hardware. The
#  STM32L4R5ZI Nucleo target and board is chosen in the example below.
if use_physical_hw:
    BOARD = os.getenv("TVM_MICRO_BOARD", default="nucleo_l4r5zi")
    SERIAL = os.getenv("TVM_MICRO_SERIAL", default=None)
    TARGET = tvm.micro.testing.get_target("zephyr", BOARD)

Extracting tuning tasks

Not all operators in the Relay program printed above can be tuned. Some are so trivial that only a single implementation is defined; others don’t make sense as tuning tasks. Using extract_from_program, you can produce a list of tunable tasks.

Because task extraction involves running the compiler, we first configure the compiler’s transformation passes; we’ll apply the same configuration later on during autotuning.

pass_context = tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True})
with pass_context:
    tasks = tvm.autotvm.task.extract_from_program(relay_mod["main"], {}, TARGET)
assert len(tasks) > 0

Configuring microTVM

Before autotuning, we need to define a module loader and then pass that to a tvm.autotvm.LocalBuilder. Then we create a tvm.autotvm.LocalRunner and use both builder and runner to generates multiple measurements for auto tunner.

In this tutorial, we have the option to use x86 host as an example or use different targets from Zephyr RTOS. If you choose pass –platform=host to this tutorial it will uses x86. You can choose other options by choosing from PLATFORM list.

module_loader = tvm.micro.AutoTvmModuleLoader(
    template_project_dir=pathlib.Path(tvm.micro.get_microtvm_template_projects("crt")),
    project_options={"verbose": False},
)
builder = tvm.autotvm.LocalBuilder(
    n_parallel=1,
    build_kwargs={"build_option": {"tir.disable_vectorize": True}},
    do_fork=True,
    build_func=tvm.micro.autotvm_build_func,
    runtime=RUNTIME,
)
runner = tvm.autotvm.LocalRunner(number=1, repeat=1, timeout=100, module_loader=module_loader)

measure_option = tvm.autotvm.measure_option(builder=builder, runner=runner)

# Compiling for physical hardware
if use_physical_hw:
    module_loader = tvm.micro.AutoTvmModuleLoader(
        template_project_dir=pathlib.Path(tvm.micro.get_microtvm_template_projects("zephyr")),
        project_options={
            "board": BOARD,
            "verbose": False,
            "project_type": "host_driven",
            "serial_number": SERIAL,
        },
    )
    builder = tvm.autotvm.LocalBuilder(
        n_parallel=1,
        build_kwargs={"build_option": {"tir.disable_vectorize": True}},
        do_fork=False,
        build_func=tvm.micro.autotvm_build_func,
        runtime=RUNTIME,
    )
    runner = tvm.autotvm.LocalRunner(number=1, repeat=1, timeout=100, module_loader=module_loader)

    measure_option = tvm.autotvm.measure_option(builder=builder, runner=runner)

Run Autotuning

Now we can run autotuning separately on each extracted task on microTVM device.

autotune_log_file = pathlib.Path("microtvm_autotune.log.txt")
if os.path.exists(autotune_log_file):
    os.remove(autotune_log_file)

num_trials = 10
for task in tasks:
    tuner = tvm.autotvm.tuner.GATuner(task)
    tuner.tune(
        n_trial=num_trials,
        measure_option=measure_option,
        callbacks=[
            tvm.autotvm.callback.log_to_file(str(autotune_log_file)),
            tvm.autotvm.callback.progress_bar(num_trials, si_prefix="M"),
        ],
        si_prefix="M",
    )

Timing the untuned program

For comparison, let’s compile and run the graph without imposing any autotuning schedules. TVM will select a randomly-tuned implementation for each operator, which should not perform as well as the tuned operator.

with pass_context:
    lowered = tvm.relay.build(relay_mod, target=TARGET, runtime=RUNTIME, params=params)

temp_dir = tvm.contrib.utils.tempdir()
project = tvm.micro.generate_project(
    str(tvm.micro.get_microtvm_template_projects("crt")),
    lowered,
    temp_dir / "project",
    {"verbose": False},
)

# Compiling for physical hardware
if use_physical_hw:
    temp_dir = tvm.contrib.utils.tempdir()
    project = tvm.micro.generate_project(
        str(tvm.micro.get_microtvm_template_projects("zephyr")),
        lowered,
        temp_dir / "project",
        {
            "board": BOARD,
            "verbose": False,
            "project_type": "host_driven",
            "serial_number": SERIAL,
            "config_main_stack_size": 4096,
        },
    )

project.build()
project.flash()
with tvm.micro.Session(project.transport()) as session:
    debug_module = tvm.micro.create_local_debug_executor(
        lowered.get_graph_json(), session.get_system_lib(), session.device
    )
    debug_module.set_input(**lowered.get_params())
    print("########## Build without Autotuning ##########")
    debug_module.run()
    del debug_module
########## Build without Autotuning ##########
Node Name                                     Ops                                           Time(us)  Time(%)  Shape              Inputs  Outputs  Measurements(us)
---------                                     ---                                           --------  -------  -----              ------  -------  ----------------
tvmgen_default_fused_nn_contrib_conv2d_NCHWc  tvmgen_default_fused_nn_contrib_conv2d_NCHWc  296.6     98.761   (1, 2, 10, 10, 3)  2       1        [296.6]
tvmgen_default_fused_layout_transform_1       tvmgen_default_fused_layout_transform_1       2.796     0.931    (1, 6, 10, 10)     1       1        [2.796]
tvmgen_default_fused_layout_transform         tvmgen_default_fused_layout_transform         0.924     0.308    (1, 1, 10, 10, 3)  1       1        [0.924]
Total_time                                    -                                             300.321   -        -                  -       -        -

Timing the tuned program

Once autotuning completes, you can time execution of the entire program using the Debug Runtime:

with tvm.autotvm.apply_history_best(str(autotune_log_file)):
    with pass_context:
        lowered_tuned = tvm.relay.build(relay_mod, target=TARGET, runtime=RUNTIME, params=params)

temp_dir = tvm.contrib.utils.tempdir()
project = tvm.micro.generate_project(
    str(tvm.micro.get_microtvm_template_projects("crt")),
    lowered_tuned,
    temp_dir / "project",
    {"verbose": False},
)

# Compiling for physical hardware
if use_physical_hw:
    temp_dir = tvm.contrib.utils.tempdir()
    project = tvm.micro.generate_project(
        str(tvm.micro.get_microtvm_template_projects("zephyr")),
        lowered_tuned,
        temp_dir / "project",
        {
            "board": BOARD,
            "verbose": False,
            "project_type": "host_driven",
            "serial_number": SERIAL,
            "config_main_stack_size": 4096,
        },
    )

project.build()
project.flash()
with tvm.micro.Session(project.transport()) as session:
    debug_module = tvm.micro.create_local_debug_executor(
        lowered_tuned.get_graph_json(), session.get_system_lib(), session.device
    )
    debug_module.set_input(**lowered_tuned.get_params())
    print("########## Build with Autotuning ##########")
    debug_module.run()
    del debug_module
########## Build with Autotuning ##########
Node Name                                     Ops                                           Time(us)  Time(%)  Shape              Inputs  Outputs  Measurements(us)
---------                                     ---                                           --------  -------  -----              ------  -------  ----------------
tvmgen_default_fused_nn_contrib_conv2d_NCHWc  tvmgen_default_fused_nn_contrib_conv2d_NCHWc  99.5      97.376   (1, 6, 10, 10, 1)  2       1        [99.5]
tvmgen_default_fused_layout_transform_1       tvmgen_default_fused_layout_transform_1       1.728     1.691    (1, 6, 10, 10)     1       1        [1.728]
tvmgen_default_fused_layout_transform         tvmgen_default_fused_layout_transform         0.953     0.933    (1, 1, 10, 10, 3)  1       1        [0.953]
Total_time                                    -                                             102.181   -        -                  -       -        -

Total running time of the script: ( 1 minutes 25.188 seconds)

Gallery generated by Sphinx-Gallery