Auto-tuning a Convolutional Network for NVIDIA GPU

Author: Lianmin Zheng, Eddie Yan

Auto-tuning for specific devices and workloads is critical for getting the best performance. This is a tutorial on how to tune a whole convolutional network for NVIDIA GPU.

The operator implementation for NVIDIA GPU in TVM is written in template form. The template has many tunable knobs (tile factor, unrolling, etc). We will tune all convolution and depthwise convolution operators in the neural network. After tuning, we produce a log file which stores the best knob values for all required operators. When the TVM compiler compiles these operators, it will query this log file to get the best knob values.

We also released pre-tuned parameters for some NVIDIA GPUs. You can go to NVIDIA GPU Benchmark to see the results.

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.

Install dependencies

To use the autotvm package in tvm, we need to install some extra dependencies. (change “3” to “2” if you use python2):

pip3 install --user psutil xgboost tornado cloudpickle

To make TVM run faster during tuning, it is recommended to use cython as FFI of tvm. In the root directory of tvm, execute:

pip3 install --user cython
sudo make cython3

Now return to python code. Import packages.

import os

import numpy as np

import tvm
from tvm import relay, autotvm
import tvm.relay.testing
from tvm.autotvm.tuner import XGBTuner, GATuner, RandomTuner, GridSearchTuner
import tvm.contrib.graph_executor as runtime

Define Network

First we need to define the network in relay frontend API. We can load some pre-defined network from tvm.relay.testing. We can also load models from MXNet, ONNX and TensorFlow.

def get_network(name, batch_size):
    """Get the symbol definition and random weight of a network"""
    input_shape = (batch_size, 3, 224, 224)
    output_shape = (batch_size, 1000)

    if "resnet" in name:
        n_layer = int(name.split("-")[1])
        mod, params = relay.testing.resnet.get_workload(
            num_layers=n_layer, batch_size=batch_size, dtype=dtype
        )
    elif "vgg" in name:
        n_layer = int(name.split("-")[1])
        mod, params = relay.testing.vgg.get_workload(
            num_layers=n_layer, batch_size=batch_size, dtype=dtype
        )
    elif name == "mobilenet":
        mod, params = relay.testing.mobilenet.get_workload(batch_size=batch_size, dtype=dtype)
    elif name == "squeezenet_v1.1":
        mod, params = relay.testing.squeezenet.get_workload(
            batch_size=batch_size, version="1.1", dtype=dtype
        )
    elif name == "inception_v3":
        input_shape = (batch_size, 3, 299, 299)
        mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
    else:
        raise ValueError("Unsupported network: " + name)

    return mod, params, input_shape, output_shape

Set Tuning Options

Before tuning, we apply some configurations.

#### DEVICE CONFIG ####
target = tvm.target.cuda()

#### TUNING OPTION ####
network = "resnet-18"
log_file = "%s.log" % network
dtype = "float32"

tuning_option = {
    "log_filename": log_file,
    "tuner": "xgb",
    "n_trial": 2000,
    "early_stopping": 600,
    "measure_option": autotvm.measure_option(
        builder=autotvm.LocalBuilder(timeout=10),
        runner=autotvm.LocalRunner(number=20, repeat=3, timeout=4, min_repeat_ms=150),
    ),
}

/workspace/python/tvm/target/target.py:446: UserWarning: Try specifying cuda arch by adding 'arch=sm_xx' to your target.
  warnings.warn("Try specifying cuda arch by adding 'arch=sm_xx' to your target.")

Note

How to set tuning options

In general, the default value provided here works well.

If you have large time budget, you can set n_trial, early_stopping larger, which makes the tuning runs longer.

If you have multiple devices, you can use all of them for measurement to accelerate the tuning process. (see the ‘Scale up measurement` section below).

Begin Tuning

Now we can extract tuning tasks from the network and begin tuning. Here, we provide a simple utility function to tune a list of tasks. This function is just an initial implementation which tunes them in sequential order. We will introduce a more sophisticated tuning scheduler in the future.

# You can skip the implementation of this function for this tutorial.
def tune_tasks(
    tasks,
    measure_option,
    tuner="xgb",
    n_trial=1000,
    early_stopping=None,
    log_filename="tuning.log",
    use_transfer_learning=True,
):
    # create tmp log file
    tmp_log_file = log_filename + ".tmp"
    if os.path.exists(tmp_log_file):
        os.remove(tmp_log_file)

    for i, tsk in enumerate(reversed(tasks)):
        prefix = "[Task %2d/%2d] " % (i + 1, len(tasks))

        # create tuner
        if tuner == "xgb":
            tuner_obj = XGBTuner(tsk, loss_type="reg")
        elif tuner == "xgb_knob":
            tuner_obj = XGBTuner(tsk, loss_type="reg", feature_type="knob")
        elif tuner == "xgb_itervar":
            tuner_obj = XGBTuner(tsk, loss_type="reg", feature_type="itervar")
        elif tuner == "xgb_curve":
            tuner_obj = XGBTuner(tsk, loss_type="reg", feature_type="curve")
        elif tuner == "xgb_rank":
            tuner_obj = XGBTuner(tsk, loss_type="rank")
        elif tuner == "xgb_rank_knob":
            tuner_obj = XGBTuner(tsk, loss_type="rank", feature_type="knob")
        elif tuner == "xgb_rank_itervar":
            tuner_obj = XGBTuner(tsk, loss_type="rank", feature_type="itervar")
        elif tuner == "xgb_rank_curve":
            tuner_obj = XGBTuner(tsk, loss_type="rank", feature_type="curve")
        elif tuner == "xgb_rank_binary":
            tuner_obj = XGBTuner(tsk, loss_type="rank-binary")
        elif tuner == "xgb_rank_binary_knob":
            tuner_obj = XGBTuner(tsk, loss_type="rank-binary", feature_type="knob")
        elif tuner == "xgb_rank_binary_itervar":
            tuner_obj = XGBTuner(tsk, loss_type="rank-binary", feature_type="itervar")
        elif tuner == "xgb_rank_binary_curve":
            tuner_obj = XGBTuner(tsk, loss_type="rank-binary", feature_type="curve")
        elif tuner == "ga":
            tuner_obj = GATuner(tsk, pop_size=100)
        elif tuner == "random":
            tuner_obj = RandomTuner(tsk)
        elif tuner == "gridsearch":
            tuner_obj = GridSearchTuner(tsk)
        else:
            raise ValueError("Invalid tuner: " + tuner)

        if use_transfer_learning:
            if os.path.isfile(tmp_log_file):
                tuner_obj.load_history(autotvm.record.load_from_file(tmp_log_file))

        # do tuning
        tsk_trial = min(n_trial, len(tsk.config_space))
        tuner_obj.tune(
            n_trial=tsk_trial,
            early_stopping=early_stopping,
            measure_option=measure_option,
            callbacks=[
                autotvm.callback.progress_bar(tsk_trial, prefix=prefix),
                autotvm.callback.log_to_file(tmp_log_file),
            ],
        )

    # pick best records to a cache file
    autotvm.record.pick_best(tmp_log_file, log_filename)
    os.remove(tmp_log_file)

Finally, we launch tuning jobs and evaluate the end-to-end performance.

def tune_and_evaluate(tuning_opt):
    # extract workloads from relay program
    print("Extract tasks...")
    mod, params, input_shape, out_shape = get_network(network, batch_size=1)
    tasks = autotvm.task.extract_from_program(
        mod["main"], target=target, params=params, ops=(relay.op.get("nn.conv2d"),)
    )

    # run tuning tasks
    print("Tuning...")
    tune_tasks(tasks, **tuning_opt)

    # compile kernels with history best records
    with autotvm.apply_history_best(log_file):
        print("Compile...")
        with tvm.transform.PassContext(opt_level=3):
            lib = relay.build_module.build(mod, target=target, params=params)

        # load parameters
        dev = tvm.device(str(target), 0)
        module = runtime.GraphModule(lib["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, number=1, repeat=600))


# We do not run the tuning in our webpage server since it takes too long.
# Uncomment the following line to run it by yourself.

# tune_and_evaluate(tuning_option)

Sample Output

The tuning needs to compile many programs and extract feature from them. So a high performance CPU is recommended. One sample output is listed below. It takes about 4 hours to get the following output on a 32T AMD Ryzen Threadripper. The tuning target is NVIDIA 1080 Ti. (You can see some errors during compilation. If the tuning is not stuck, it is okay.)

Extract tasks...
Tuning...
[Task  1/12]  Current/Best:  541.83/3570.66 GFLOPS | Progress: (960/2000) | 1001.31 s Done.
[Task  2/12]  Current/Best:    0.56/ 803.33 GFLOPS | Progress: (704/2000) | 608.08 s Done.
[Task  3/12]  Current/Best:  103.69/1141.25 GFLOPS | Progress: (768/2000) | 702.13 s Done.
[Task  4/12]  Current/Best: 2905.03/3925.15 GFLOPS | Progress: (864/2000) | 745.94 sterminate called without an active exception
[Task  4/12]  Current/Best: 2789.36/3925.15 GFLOPS | Progress: (1056/2000) | 929.40 s Done.
[Task  5/12]  Current/Best:   89.06/1076.24 GFLOPS | Progress: (704/2000) | 601.73 s Done.
[Task  6/12]  Current/Best:   40.39/2129.02 GFLOPS | Progress: (1088/2000) | 1125.76 s Done.
[Task  7/12]  Current/Best: 4090.53/5007.02 GFLOPS | Progress: (800/2000) | 903.90 s Done.
[Task  8/12]  Current/Best:    4.78/1272.28 GFLOPS | Progress: (768/2000) | 749.14 s Done.
[Task  9/12]  Current/Best: 1391.45/2325.08 GFLOPS | Progress: (992/2000) | 1084.87 s Done.
[Task 10/12]  Current/Best: 1995.44/2383.59 GFLOPS | Progress: (864/2000) | 862.60 s Done.
[Task 11/12]  Current/Best: 4093.94/4899.80 GFLOPS | Progress: (224/2000) | 240.92 sterminate called without an active exception
[Task 11/12]  Current/Best: 3487.98/4909.91 GFLOPS | Progress: (480/2000) | 534.96 sterminate called without an active exception
[Task 11/12]  Current/Best: 4636.84/4912.17 GFLOPS | Progress: (1184/2000) | 1381.16 sterminate called without an active exception
[Task 11/12]  Current/Best:   50.12/4912.17 GFLOPS | Progress: (1344/2000) | 1602.81 s Done.
[Task 12/12]  Current/Best: 3581.31/4286.30 GFLOPS | Progress: (736/2000) | 943.52 s Done.
Compile...
Evaluate inference time cost...
Mean inference time (std dev): 1.07 ms (0.05 ms)

As a reference baseline, the time cost of MXNet + TensorRT on resnet-18 is 1.30ms. So we are a little faster.

Note

Experiencing Difficulties?

The auto tuning module is error-prone. If you always see ” 0.00/ 0.00 GFLOPS”, then there must be something wrong.

First, make sure you set the correct configuration of your device. Then, you can print debug information by adding these lines in the beginning of the script. It will print every measurement result, where you can find useful error messages.

import logging
logging.getLogger('autotvm').setLevel(logging.DEBUG)

Finally, always feel free to ask our community for help on https://discuss.tvm.apache.org

Scale up measurement by using multiple devices

If you have multiple devices, you can use all of them for measurement. TVM uses the 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 GPU cards, 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

Then open another new terminal for the RPC server. We need to start one dedicated server for each device. We use a string key to distinguish the types of devices. You can pick a name you like. (Note: For rocm backend, there are some internal errors with the compiler, we need to add –no-fork to the argument list.)

python -m tvm.exec.rpc_server --tracker=127.0.0.1:9190 --key=1080ti

After registering devices, we can confirm it by querying rpc_tracker

python -m tvm.exec.query_rpc_tracker --host=127.0.0.1 --port=9190

For example, if we have four 1080ti, two titanx and one gfx900, the output can be

Queue Status
----------------------------------
key          total  free  pending
----------------------------------
1080ti       4      4     0
titanx       2      2     0
gfx900       1      1     0
----------------------------------

Finally, we need to change the tuning option to use RPCRunner. Use the code below to replace the corresponding part above.

tuning_option = {
    "log_filename": log_file,
    "tuner": "xgb",
    "n_trial": 2000,
    "early_stopping": 600,
    "measure_option": autotvm.measure_option(
        builder=autotvm.LocalBuilder(timeout=10),
        runner=autotvm.RPCRunner(
            "1080ti",  # change the device key to your key
            "127.0.0.1",
            9190,
            number=20,
            repeat=3,
            timeout=4,
            min_repeat_ms=150,
        ),
    ),
}