Note
This tutorial can be used interactively with Google Colab! You can also click here to run the Jupyter notebook locally.
Auto-tuning a convolutional network on VTA¶
Author: Lianmin Zheng, Thierry Moreau
Auto-tuning for a specific accelerator design is critical for getting the best performance for any given operator. This is a tutorial showcases how to tune a whole convolutional network on VTA.
The operator implementation for VTA in TVM is written in template form. The template has many tunable knobs (tile factor, virtual threads, etc). We will tune all convolution operators in the neural network. After tuning, we produce a log file which stores the best schedule parameters for all tuned operators. When the TVM compiler compiles these operators, it will query this log file to get the best knob parameters.
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 mxnet requests "Pillow<7" 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 (change “3” to “2” if you use python2):
pip3 install --user cython
sudo make cython3
Now return to python code. Import packages.
import os
from mxnet.gluon.model_zoo import vision
import numpy as np
from PIL import Image
from tvm import topi
import tvm
from tvm import te
from tvm import rpc, autotvm, relay
from tvm.contrib import graph_executor, utils, download
from tvm.autotvm.measure.measure_methods import request_remote
from tvm.autotvm.tuner import XGBTuner, GATuner, RandomTuner, GridSearchTuner
import vta
from vta.testing import simulator
from vta.top import graph_pack
Compile network¶
Perform vta-specific compilation with Relay from a Gluon model
def compile_network(env, target, model, start_pack, stop_pack):
# Populate the shape and data type dictionary
dtype_dict = {"data": "float32"}
shape_dict = {"data": (env.BATCH, 3, 224, 224)}
# Get off the shelf gluon model, and convert to relay
gluon_model = vision.get_model(model, pretrained=True)
mod, params = relay.frontend.from_mxnet(gluon_model, shape_dict)
# Update shape and type dictionary
shape_dict.update({k: v.shape for k, v in params.items()})
dtype_dict.update({k: str(v.dtype) for k, v in params.items()})
# Perform quantization in Relay
# Note: We set opt_level to 3 in order to fold batch norm
with tvm.transform.PassContext(opt_level=3):
with relay.quantize.qconfig(global_scale=8.0, skip_conv_layers=[0]):
mod = relay.quantize.quantize(mod, params=params)
# Perform graph packing and constant folding for VTA target
if target.device_name == "vta":
assert env.BLOCK_IN == env.BLOCK_OUT
relay_prog = graph_pack(
mod["main"],
env.BATCH,
env.BLOCK_OUT,
env.WGT_WIDTH,
start_name=start_pack,
stop_name=stop_pack,
)
return relay_prog, params
Start RPC Tracker¶
TVM uses an RPC session to communicate with Pynq 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 tuning, TVM uses an RPC Tracker to manage multiple devices. The RPC Tracker is a centralized controller node. We can register all devices to the tracker. For example, if we have 10 Pynq boards, 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 Pynq devices.
Follow VTA: Versatile Tensor Accelerator to build the TVM runtime on the device. Then register the device to the tracker with:
python -m tvm.exec.rpc_server --tracker=[HOST_IP]:9190 --key=pynq
(replace [HOST_IP] with the IP address of your host machine)
After registering devices, we can confirm it by querying the rpc_tracker:
python -m tvm.exec.query_rpc_tracker --host=0.0.0.0 --port=9190
For example, if we have 6 Pynq boards and 11 Raspberry Pi 3B, the output can be
Queue Status
----------------------------------
key total free pending
----------------------------------
pynq 6 6 0
rpi3b 11 11 0
----------------------------------
You can register multiple devices to the tracker to accelerate tuning.
Set Tuning Options¶
Before tuning, we should apply some configurations. Here we use an Pynq-Z1 board as an example.
# Tracker host and port can be set by your environment
tracker_host = os.environ.get("TVM_TRACKER_HOST", "127.0.0.1")
tracker_port = int(os.environ.get("TVM_TRACKER_PORT", 9190))
# Load VTA parameters from the 3rdparty/vta-hw/config/vta_config.json file
env = vta.get_env()
# This target is used for cross compilation. You can query it by :code:`gcc -v` on your device.
# Set ``device=arm_cpu`` to run inference on the CPU
# or ``device=vta`` to run inference on the FPGA.
device = "vta"
target = env.target if device == "vta" else env.target_vta_cpu
# Name of Gluon model to compile
# The ``start_pack`` and ``stop_pack`` labels indicate where
# to start and end the graph packing relay pass: in other words
# where to start and finish offloading to VTA.
network = "resnet18_v1"
start_pack = "nn.max_pool2d"
stop_pack = "nn.global_avg_pool2d"
# Tuning option
log_file = "%s.%s.log" % (device, network)
tuning_option = {
"log_filename": log_file,
"tuner": "random",
"n_trial": 1000,
"early_stopping": None,
"measure_option": autotvm.measure_option(
builder=autotvm.LocalBuilder(),
runner=autotvm.RPCRunner(
env.TARGET,
host=tracker_host,
port=tracker_port,
number=5,
timeout=60,
module_loader=vta.module_loader(),
# check_correctness=True, # TODO: re-enable when check_correctness works again.
),
),
}
Note
How to set tuning options
In general, the default values provided here work well.
If you have enough time budget, you can set n_trial, early_stopping
to larger values, makes the tuning run for longer.
If your device is under-powered or your conv2d operators are large, consider
setting a longer timeout.
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.
Given that the tuning will be done on Pynq FPGA boards, make sure that
the `TARGET entry in the vta_config.json file is set to pynq.
# 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=50)
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)
Register VTA-specific tuning tasks
def register_vta_tuning_tasks():
from tvm.autotvm.task import TaskExtractEnv
@tvm.te.tag_scope(tag=topi.tag.ELEMWISE)
def my_clip(x, a_min, a_max):
"""Unlike topi's current clip, put min and max into two stages."""
const_min = tvm.tir.const(a_min, x.dtype)
const_max = tvm.tir.const(a_max, x.dtype)
x = te.compute(x.shape, lambda *i: tvm.te.min(x(*i), const_max), name="clipA")
x = te.compute(x.shape, lambda *i: tvm.te.max(x(*i), const_min), name="clipB")
return x
# init autotvm env to register VTA operator
TaskExtractEnv()
@autotvm.template("conv2d_packed.vta")
def _topi_nn_conv2d(*args, **kwargs):
assert not kwargs, "Do not support kwargs in template function call"
A, W = args[:2]
with tvm.target.vta():
res = vta.top.conv2d_packed(*args, **kwargs)
res = topi.right_shift(res, 8)
res = my_clip(res, 0, 127)
res = topi.cast(res, "int8")
if tvm.target.Target.current().device_name == "vta":
s = vta.top.schedule_conv2d_packed([res])
else:
s = te.create_schedule([res.op])
return s, [A, W, res]
Finally, we launch tuning jobs and evaluate the end-to-end performance.
def tune_and_evaluate(tuning_opt):
# Register VTA tuning tasks
register_vta_tuning_tasks()
# Perform task extraction on Relay program
print("Extract tasks...")
relay_prog, params = compile_network(env, target, network, start_pack, stop_pack)
mod = tvm.IRModule.from_expr(relay_prog)
tasks = autotvm.task.extract_from_program(
mod,
params=params,
ops=(relay.op.get("nn.conv2d"),),
target=target,
target_host=env.target_host,
)
# filter out non-packed conv2d task
tasks = list(filter(lambda t: len(t.args[0][1]) > 4 and "conv" in t.name, tasks))
# We should have extracted 10 convolution tasks
assert len(tasks) == 10
print("Extracted {} conv2d tasks:".format(len(tasks)))
for tsk in tasks:
inp = tsk.args[0][1]
wgt = tsk.args[1][1]
batch = inp[0] * inp[4]
in_filter = inp[1] * inp[5]
out_filter = wgt[0] * wgt[4]
height, width = inp[2], inp[3]
hkernel, wkernel = wgt[2], wgt[3]
hstride, wstride = tsk.args[2][0], tsk.args[2][1]
hpad, wpad = tsk.args[3][0], tsk.args[3][1]
print(
"({}, {}, {}, {}, {}, {}, {}, {}, {}, {}, {})".format(
batch,
height,
width,
in_filter,
out_filter,
hkernel,
wkernel,
hpad,
wpad,
hstride,
wstride,
)
)
# We do not run the tuning in our webpage server since it takes too long.
# Comment the following line to run it by yourself.
return
# run tuning tasks
print("Tuning...")
tune_tasks(tasks, **tuning_opt)
# evaluate with tuning history
if env.TARGET != "sim":
# Get remote from fleet node
remote = autotvm.measure.request_remote(
env.TARGET, tracker_host, tracker_port, timeout=10000
)
# Reconfigure the JIT runtime and FPGA.
vta.reconfig_runtime(remote)
vta.program_fpga(remote, bitstream=None)
else:
# In simulation mode, host the RPC server locally.
remote = rpc.LocalSession()
# compile kernels with history best records
with autotvm.tophub.context(target, extra_files=[log_file]):
# Compile network
print("Compile...")
if target.device_name != "vta":
with tvm.transform.PassContext(opt_level=3, disabled_pass={"AlterOpLayout"}):
lib = relay.build(
relay_prog, target=target, params=params, target_host=env.target_host
)
else:
with vta.build_config(opt_level=3, disabled_pass={"AlterOpLayout"}):
lib = relay.build(
relay_prog, target=target, params=params, target_host=env.target_host
)
# Export library
print("Upload...")
temp = utils.tempdir()
lib.export_library(temp.relpath("graphlib.tar"))
remote.upload(temp.relpath("graphlib.tar"))
lib = remote.load_module("graphlib.tar")
# Generate the graph executor
ctx = remote.ext_dev(0) if device == "vta" else remote.cpu(0)
m = graph_executor.GraphModule(lib["default"](ctx))
# upload parameters to device
image = tvm.nd.array((np.random.uniform(size=(1, 3, 224, 224))).astype("float32"))
m.set_input("data", image)
# evaluate
print("Evaluate inference time cost...")
timer = m.module.time_evaluator("run", ctx, number=1, repeat=10)
tcost = timer()
prof_res = np.array(tcost.results) * 1000 # convert to millisecond
print(
"Mean inference time (std dev): %.2f ms (%.2f ms)"
% (np.mean(prof_res), np.std(prof_res))
)
# Run the tuning and evaluate the results
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 2 hours on a 16T CPU, and 6 Pynq boards.
Extract tasks...
[Warning] Invalid shape during AutoTVM task creation
Extracted 10 conv2d tasks:
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 16, 14, 14, 1, 16), 'int8'), ('TENSOR', (32, 16, 1, 1, 16, 16), 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 16, 14, 14, 1, 16, 'int8'), (32, 16, 1, 1, 16, 16, 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 8, 28, 28, 1, 16), 'int8'), ('TENSOR', (16, 8, 1, 1, 16, 16), 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 8, 28, 28, 1, 16, 'int8'), (16, 8, 1, 1, 16, 16, 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 4, 56, 56, 1, 16), 'int8'), ('TENSOR', (8, 4, 1, 1, 16, 16), 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 4, 56, 56, 1, 16, 'int8'), (8, 4, 1, 1, 16, 16, 'int8'), (2, 2), (0, 0), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 4, 56, 56, 1, 16), 'int8'), ('TENSOR', (4, 4, 3, 3, 16, 16), 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 4, 56, 56, 1, 16, 'int8'), (4, 4, 3, 3, 16, 16, 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 8, 28, 28, 1, 16), 'int8'), ('TENSOR', (8, 8, 3, 3, 16, 16), 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 8, 28, 28, 1, 16, 'int8'), (8, 8, 3, 3, 16, 16, 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 4, 56, 56, 1, 16), 'int8'), ('TENSOR', (8, 4, 3, 3, 16, 16), 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 4, 56, 56, 1, 16, 'int8'), (8, 4, 3, 3, 16, 16, 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 16, 14, 14, 1, 16), 'int8'), ('TENSOR', (16, 16, 3, 3, 16, 16), 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 16, 14, 14, 1, 16, 'int8'), (16, 16, 3, 3, 16, 16, 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 8, 28, 28, 1, 16), 'int8'), ('TENSOR', (16, 8, 3, 3, 16, 16), 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 8, 28, 28, 1, 16, 'int8'), (16, 8, 3, 3, 16, 16, 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 32, 7, 7, 1, 16), 'int8'), ('TENSOR', (32, 32, 3, 3, 16, 16), 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 32, 7, 7, 1, 16, 'int8'), (32, 32, 3, 3, 16, 16, 'int8'), (1, 1), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Task(func_name=topi_nn_conv2d, args=(('TENSOR', (1, 16, 14, 14, 1, 16), 'int8'), ('TENSOR', (32, 16, 3, 3, 16, 16), 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'), kwargs={}, workload=('conv2d', (1, 16, 14, 14, 1, 16, 'int8'), (32, 16, 3, 3, 16, 16, 'int8'), (2, 2), (1, 1), (1, 1), 'NCHW1n16c', 'int32'))
Tuning...
[Task 1/10] Current/Best: 0.72/ 23.24 GFLOPS | Progress: (480/1000) | 640.31 s Done.
[Task 2/10] Current/Best: 0.00/ 27.69 GFLOPS | Progress: (576/1000) | 810.09 s Done.
[Task 3/10] Current/Best: 0.00/ 22.97 GFLOPS | Progress: (1000/1000) | 1125.37 s Done.
[Task 4/10] Current/Best: 0.00/ 31.26 GFLOPS | Progress: (1000/1000) | 1025.52 s Done.
[Task 5/10] Current/Best: 0.00/ 15.15 GFLOPS | Progress: (1000/1000) | 1236.58 s Done.
[Task 6/10] Current/Best: 0.00/ 22.74 GFLOPS | Progress: (1000/1000) | 906.60 s Done.
[Task 7/10] Current/Best: 0.00/ 15.27 GFLOPS | Progress: (1000/1000) | 1056.25 s Done.
[Task 8/10] Current/Best: 0.00/ 2.18 GFLOPS | Progress: (1000/1000) | 2275.29 s Done.
[Task 9/10] Current/Best: 2.23/ 3.99 GFLOPS | Progress: (1000/1000) | 2527.25 s Done.
[Task 10/10] Current/Best: 1.56/ 6.32 GFLOPS | Progress: (480/1000) | 1304.84 s Done.
Compile...
Upload...
Evaluate inference time cost...
Mean inference time (std dev): 621.79 ms (0.14 ms)
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