Auto-scheduling a Neural Network for ARM CPU

Author: Thierry Moreau, Lianmin Zheng, Chengfan Jia

Auto-tuning for specific devices and workloads is critical for getting the best performance. This is a tutorial on how to tune a whole neural network for ARM CPU with the auto-scheduler via RPC.

To auto-tune a neural network, we partition the network into small subgraphs and tune them independently. Each subgraph is treated as one search task. A task scheduler slices the time and dynamically allocates time resources to these tasks. The task scheduler predicts the impact of each task on the end-to-end execution time and prioritizes the one that can reduce the execution time the most.

For each subgraph, we use the compute declaration in tvm/python/topi to get the computational DAG in the tensor expression form. We then use the auto-scheduler to construct a search space of this DAG and search for good schedules (low-level optimizations).

Different from the template-based autotvm which relies on manual templates to define the search space, the auto-scheduler does not require any schedule templates. In other words, the auto-scheduler only uses the compute declarations in tvm/python/topi and does not use existing schedule templates.

Note that this tutorial will not run on Windows or recent versions of macOS. To get it to run, you will need to wrap the body of this tutorial in a if __name__ == "__main__": block.

import numpy as np
import os

import tvm
from tvm import relay, auto_scheduler
from tvm.relay import data_dep_optimization as ddo
import tvm.relay.testing
from tvm.contrib import graph_executor
from tvm.contrib.utils import tempdir

Define a Network

First, we need to define the network with relay frontend API. We can load some pre-defined network from tvm.relay.testing. We can also load models from MXNet, ONNX, PyTorch, and TensorFlow (see front end tutorials).

For convolutional neural networks, although auto-scheduler can work correctly with any layout, we found the best performance is typically achieved with NHWC layout. We also implemented more optimizations for NHWC layout with the auto-scheduler. So it is recommended to convert your models to NHWC layout to use the auto-scheduler. You can use ConvertLayout pass to do the layout conversion in TVM.

def get_network(name, batch_size, layout="NHWC", dtype="float32", use_sparse=False):
    """Get the symbol definition and random weight of a network"""

    # auto-scheduler prefers NHWC layout
    if layout == "NHWC":
        image_shape = (224, 224, 3)
    elif layout == "NCHW":
        image_shape = (3, 224, 224)
    else:
        raise ValueError("Invalid layout: " + layout)

    input_shape = (batch_size,) + image_shape
    output_shape = (batch_size, 1000)

    if name.startswith("resnet-"):
        n_layer = int(name.split("-")[1])
        mod, params = relay.testing.resnet.get_workload(
            num_layers=n_layer,
            batch_size=batch_size,
            layout=layout,
            dtype=dtype,
            image_shape=image_shape,
        )
    elif name.startswith("resnet3d-"):
        n_layer = int(name.split("-")[1])
        mod, params = relay.testing.resnet.get_workload(
            num_layers=n_layer,
            batch_size=batch_size,
            layout=layout,
            dtype=dtype,
            image_shape=image_shape,
        )
    elif name == "mobilenet":
        mod, params = relay.testing.mobilenet.get_workload(
            batch_size=batch_size, layout=layout, dtype=dtype, image_shape=image_shape
        )
    elif name == "squeezenet_v1.1":
        assert layout == "NCHW", "squeezenet_v1.1 only supports NCHW layout"
        mod, params = relay.testing.squeezenet.get_workload(
            version="1.1",
            batch_size=batch_size,
            dtype=dtype,
            image_shape=image_shape,
        )
    elif name == "inception_v3":
        input_shape = (batch_size, 3, 299, 299) if layout == "NCHW" else (batch_size, 299, 299, 3)
        mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
    elif name == "mxnet":
        # an example for mxnet model
        from mxnet.gluon.model_zoo.vision import get_model

        assert layout == "NCHW"

        block = get_model("resnet50_v1", pretrained=True)
        mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
        net = mod["main"]
        net = relay.Function(
            net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
        )
        mod = tvm.IRModule.from_expr(net)
    elif name == "mlp":
        mod, params = relay.testing.mlp.get_workload(
            batch_size=batch_size, dtype=dtype, image_shape=image_shape, num_classes=1000
        )
    else:
        raise ValueError("Network not found.")

    if use_sparse:
        from tvm.topi.sparse.utils import convert_model_dense_to_sparse

        mod, params = convert_model_dense_to_sparse(mod, params, random_params=True)

    return mod, params, input_shape, output_shape

Start RPC Tracker

TVM uses RPC session to communicate with ARM boards. During tuning, the tuner will send the generated code to the board and measure the speed of code on the board.

To scale up the tuning, TVM uses RPC Tracker to manage distributed devices. The RPC Tracker is a centralized controller node. We can register all devices to the tracker. For example, if we have 10 phones, we can register all of them to the tracker, and run 10 measurements in parallel, accelerating the tuning process.

To start an RPC tracker, run this command on the host machine. The tracker is required during the whole tuning process, so we need to open a new terminal for this command:

python -m tvm.exec.rpc_tracker --host=0.0.0.0 --port=9190

The expected output is

INFO:RPCTracker:bind to 0.0.0.0:9190

Register Devices to RPC Tracker

Now we can register our devices to the tracker. The first step is to build the TVM runtime for the ARM devices.

• For Linux: Follow this section Build TVM Runtime on Device to build the TVM runtime on the device. Then register the device to tracker by

  python -m tvm.exec.rpc_server --tracker=[HOST_IP]:9190 --key=rasp4b-64
  

  (replace [HOST_IP] with the IP address of your host machine)

• For Android: Follow this readme page to install the TVM RPC APK on the android device. Make sure you can pass the android rpc test. Then you have already registered your device. During tuning, you have to go to developer option and enable “Keep screen awake during changing” and charge your phone to make it stable.

After registering devices, we can confirm it by querying rpc_tracker

python -m tvm.exec.query_rpc_tracker --host=0.0.0.0 --port=9190

For example, if we have 2 Huawei mate10 pro, 11 Raspberry Pi 4B with 64bit OS, and 2 rk3399, the output can be

Queue Status
----------------------------------
key          total  free  pending
----------------------------------
mate10pro    2      2     0
rk3399       2      2     0
rasp4b-64    11     11    0
----------------------------------

You can register multiple devices to the tracker to accelerate the measurement in tuning.

Set Tuning Options

Before tuning, we should apply some configurations. Here I use a Raspberry Pi 4b 4GB board as example with a 64bit OS (Ubuntu 20.04). In your setting, you should modify the target and device_key accordingly. set use_ndk to True if you use android phone.

#### DEVICE CONFIG ####

# Replace "aarch64-linux-gnu" with the correct target of your board.
# This target is used for cross compilation. You can query it by :code:`gcc -v` on your device.
# FIXME(tmoreau89, merrymercy): We leave '-device=arm_cpu' out of the target string
#                               because we're sharing x86 op strategy.
target = tvm.target.Target("llvm -mtriple=aarch64-linux-gnu -mattr=+neon")

# Also replace this with the device key, rpc host and rpc port in your tracker
device_key = "rasp4b-64"
rpc_host = "127.0.0.1"
rpc_port = 9190

# Set this to True if you use ndk tools for cross compiling
# And also set the environment variable below to point to the cross compiler
use_ndk = False
# os.environ["TVM_NDK_CC"] = "/usr/bin/aarch64-linux-gnu-g++"

#### TUNING OPTION ####
network = "mobilenet"
use_sparse = False
batch_size = 1
layout = "NHWC"
dtype = "float32"
log_file = "%s-%s-B%d-%s.json" % (network, layout, batch_size, target.kind.name)

Extract Search Tasks

Next, we extract the search tasks and their weights from a network. The weight of a task is the number of appearances of the task’s subgraph in the whole network. By using the weight, we can approximate the end-to-end latency of the network as sum(latency[t] * weight[t]), where latency[t] is the latency of a task and weight[t] is the weight of the task. The task scheduler will just optimize this objective.

# Extract tasks from the network
print("Get model...")
mod, params, input_shape, output_shape = get_network(
    network, batch_size, layout, dtype=dtype, use_sparse=use_sparse
)
print("Extract tasks...")
tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target)

for idx, task in enumerate(tasks):
    print("========== Task %d  (workload key: %s) ==========" % (idx, task.workload_key))
    print(task.compute_dag)

Tuning and Evaluation

Now, we set some options for tuning and launch the search tasks

• num_measure_trials is the number of measurement trials we can use during the tuning. You can set it to a small number (e.g., 200) for a fast demonstrative run. In practice, we recommend setting it around 800 * len(tasks), which is typically enough for the search to converge. For example, there are 29 tasks in resnet-50, so we can set it as 20000. You can adjust this parameter according to your time budget.

• In addition, we use RecordToFile to dump measurement records into a log file, The measurement records can be used to query the history best, resume the search, and do more analyses later.

• see auto_scheduler.TuningOptions, auto_scheduler.LocalRunner for more parameters.

After auto-tuning, we can compile the network with the best schedules we found. All measurement records are dumped into the log file during auto-tuning, so we can read the log file and load the best schedules.

def tune_and_evaluate():
    print("Begin tuning...")
    tuner = auto_scheduler.TaskScheduler(tasks, task_weights)
    tune_option = auto_scheduler.TuningOptions(
        num_measure_trials=200,  # change this to 20000 to achieve the best performance
        builder=auto_scheduler.LocalBuilder(build_func="ndk" if use_ndk else "default"),
        runner=auto_scheduler.RPCRunner(
            device_key,
            host=rpc_host,
            port=rpc_port,
            timeout=30,
            repeat=1,
            min_repeat_ms=200,
            enable_cpu_cache_flush=True,
        ),
        measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
    )

    tuner.tune(tune_option)

    # Compile with the history best
    print("Compile...")
    with auto_scheduler.ApplyHistoryBest(log_file):
        with tvm.transform.PassContext(
            opt_level=3, config={"relay.backend.use_auto_scheduler": True}
        ):
            lib = relay.build(mod, target=target, params=params)

    # Export library
    tmp = tempdir()
    if use_ndk:
        from tvm.contrib import ndk

        filename = "net.so"
        lib.export_library(tmp.relpath(filename), fcompile=ndk.create_shared)
    else:
        filename = "net.tar"
        lib.export_library(tmp.relpath(filename))

    # Upload module to device
    print("Upload...")
    remote = auto_scheduler.utils.request_remote(device_key, rpc_host, rpc_port, timeout=10000)
    remote.upload(tmp.relpath(filename))
    rlib = remote.load_module(filename)

    # Create graph executor
    dev = remote.cpu()
    module = graph_executor.GraphModule(rlib["default"](dev))
    data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))
    module.set_input("data", data_tvm)

    # Evaluate
    print("Evaluate inference time cost...")
    print(module.benchmark(dev, repeat=3, min_repeat_ms=500))


# We do not run the tuning in our webpage server since the server doesn't have a Raspberry Pi,
# or device tracker running.
# Uncomment the following line to run it by yourself.

# tune_and_evaluate()

Note

Explaining the printed information during tuning

During the tuning, a lot of information will be printed on the console. They are used for debugging purposes. The most important info is the output of the task scheduler. The following table is a sample output.

----------------------------------------------------------------------
------------------------------  [ Task Scheduler ]
----------------------------------------------------------------------
|  ID  | Latency (ms) | Speed (GFLOPS) | Trials |
-------------------------------------------------
|    0 |        0.013 |           0.31 |     64 |
|    1 |        0.845 |           2.43 |    448 |
|    2 |        0.046 |          -0.00 |     64 |
|    3 |        4.194 |          24.53 |   2112 |
|    4 |        0.109 |           9.21 |     64 |
|    5 |        1.759 |          29.27 |    896 |
|    6 |        0.083 |           6.01 |     64 |
|    7 |        3.084 |          33.38 |   7680 |
|    8 |        0.136 |          14.78 |    384 |
|    9 |        1.349 |          38.23 |    768 |
|   10 |        0.133 |           7.55 |    128 |
|   11 |        2.747 |          37.56 |   1536 |
|   12 |        0.338 |          11.87 |    192 |
|   13 |        1.295 |          40.00 |    704 |
|   14 |        0.482 |           4.16 |    256 |
|   15 |        2.686 |          38.56 |   1344 |
|   16 |        0.884 |           9.08 |    448 |
|   17 |        1.332 |          39.18 |    704 |
|   18 |        1.045 |           3.84 |    576 |
|   19 |        1.391 |          38.09 |    704 |
|   20 |        0.777 |          10.34 |    448 |
|   21 |        0.739 |          30.97 |    448 |
-------------------------------------------------
 Estimated total latency: 38.347 ms      Trials: 19992   Used time : 19260 s     Next ID: 3

This table lists the latency and (estimated) speed of all tasks. It also lists the allocation of measurement trials for all tasks. The last line prints the total weighted latency of these tasks, which can be a rough estimation of the end-to-end execution time of the network. The last line also prints the total number of measurement trials, total time spent on auto-tuning and the id of the next task to tune.

There will also be some “dmlc::Error”s errors, because the auto-scheduler will try some invalid schedules. You can safely ignore them if the tuning can continue, because these errors are isolated from the main process.

Note

Terminate the tuning earlier

You can terminate the tuning earlier by forcibly killing this process. As long as you get at least one valid schedule for each task in the log file, you should be able to do the compilation (the secion below).

Other Tips

1. During the tuning, the auto-scheduler needs to compile many programs and extract feature from them. This part is CPU-intensive, so a high-performance CPU with many cores is recommended for faster search.

2. You can use python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json to distill the large log file and only save the best useful records.

3. You can resume a search from the previous log file. You just need to add a new argument load_log_file when creating the task scheduler in function run_tuning. Say, tuner = auto_scheduler.TaskScheduler(tasks, task_weights, load_log_file=log_file)

4. If you have multiple target CPUs, you can use all of them for measurements to parallelize the measurements. Check this section to learn how to use the RPC Tracker and RPC Server. To use the RPC Tracker in auto-scheduler, replace the runner in TuningOptions with auto_scheduler.RPCRunner.