TVM 是一个开源深度学习编译器，可适用于各类 CPUs, GPUs 及其他专用加速器。它的目标是使得我们能够在任何硬件上优化和运行自己的模型。不同于深度学习框架关注模型生产力，TVM 更关注模型在硬件上的性能和效率。

本文只简单介绍 TVM 的编译流程，及如何自动调优自己的模型。更深入了解，可见 TVM 官方内容：

• 文档: https://tvm.apache.org/docs/
• 源码: https://github.com/apache/tvm

编译流程

TVM 文档 Design and Architecture 讲述了实例编译流程、逻辑结构组件、设备目标实现等。其中流程见下图：

从高层次上看，包含了如下步骤：

• 导入（Import）：前端组件将模型提取进 IRModule，其是模型内部表示（IR）的函数集合。
• 转换（Transformation）：编译器将 IRModule 转换为另一个功能等效或近似等效（如量化情况下）的 IRModule。大多转换都是独立于目标（后端）的。TVM 也允许目标影响转换通道的配置。
• 目标翻译（Target Translation）：编译器翻译（代码生成） IRModule 到目标上的可执行格式。目标翻译结果被封装为 runtime.Module，可以在目标运行时环境中导出、加载和执行。
• 运行时执行（Runtime Execution）：用户加载一个 runtime.Module 并在支持的运行时环境中运行编译好的函数。

调优模型

TVM 文档 User Tutorial 从怎么编译优化模型开始，逐步深入到 TE, TensorIR, Relay 等更底层的逻辑结构组件。

这里只讲下如何用 AutoTVM 自动调优模型，实际了解 TVM 编译、调优、运行模型的过程。原文见 Compiling and Optimizing a Model with the Python Interface (AutoTVM)。

准备 TVM

首先，安装 TVM。可见文档 Installing TVM，或笔记「TVM 安装」。

之后，即可通过 TVM Python API 来调优模型。我们先导入如下依赖：

import onnx
from tvm.contrib.download import download_testdata
from PIL import Image
import numpy as np
import tvm.relay as relay
import tvm
from tvm.contrib import graph_executor

准备模型，并加载

获取预训练的 ResNet-50 v2 ONNX 模型，并加载：

model_url = "".join(
    [
        "https://github.com/onnx/models/raw/",
        "main/vision/classification/resnet/model/",
        "resnet50-v2-7.onnx",
    ]
)

model_path = download_testdata(model_url, "resnet50-v2-7.onnx", module="onnx")
onnx_model = onnx.load(model_path)

准备图片，并前处理

获取一张测试图片，并前处理成 224x224 NCHW 格式：

img_url = "https://s3.amazonaws.com/model-server/inputs/kitten.jpg"
img_path = download_testdata(img_url, "imagenet_cat.png", module="data")

# Resize it to 224x224
resized_image = Image.open(img_path).resize((224, 224))
img_data = np.asarray(resized_image).astype("float32")

# Our input image is in HWC layout while ONNX expects CHW input, so convert the array
img_data = np.transpose(img_data, (2, 0, 1))

# Normalize according to the ImageNet input specification
imagenet_mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
imagenet_stddev = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
norm_img_data = (img_data / 255 - imagenet_mean) / imagenet_stddev

# Add the batch dimension, as we are expecting 4-dimensional input: NCHW.
img_data = np.expand_dims(norm_img_data, axis=0)

编译模型，用 TVM Relay

TVM 导入 ONNX 模型成 Relay，并创建 TVM 图模型：

target = input("target [llvm]: ")
if not target:
    target = "llvm"
    # target = "llvm -mcpu=core-avx2"
    # target = "llvm -mcpu=skylake-avx512"

# The input name may vary across model types. You can use a tool
# like Netron to check input names
input_name = "data"
shape_dict = {input_name: img_data.shape}

mod, params = relay.frontend.from_onnx(onnx_model, shape_dict)

with tvm.transform.PassContext(opt_level=3):
    lib = relay.build(mod, target=target, params=params)

dev = tvm.device(str(target), 0)
module = graph_executor.GraphModule(lib["default"](dev))

其中 target 是目标硬件平台。llvm 指用 CPU，建议指明架构指令集，可更优化性能。如下命令可查看 CPU：

$ llc --version | grep CPU
  Host CPU: skylake
$ lscpu

或直接上厂商网站（如 Intel® Products）查看产品参数。

运行模型，用 TVM Runtime

用 TVM Runtime 运行模型，进行预测：

dtype = "float32"
module.set_input(input_name, img_data)
module.run()
output_shape = (1, 1000)
tvm_output = module.get_output(0, tvm.nd.empty(output_shape)).numpy()

收集优化前的性能数据

收集优化前的性能数据：

import timeit

timing_number = 10
timing_repeat = 10
unoptimized = (
    np.array(timeit.Timer(lambda: module.run()).repeat(repeat=timing_repeat, number=timing_number))
    * 1000
    / timing_number
)
unoptimized = {
    "mean": np.mean(unoptimized),
    "median": np.median(unoptimized),
    "std": np.std(unoptimized),
}

print(unoptimized)

之后，用以对比优化后的性能。

后处理输出，得知预测结果

输出的预测结果，后处理成可读的分类结果：

from scipy.special import softmax

# Download a list of labels
labels_url = "https://s3.amazonaws.com/onnx-model-zoo/synset.txt"
labels_path = download_testdata(labels_url, "synset.txt", module="data")

with open(labels_path, "r") as f:
    labels = [l.rstrip() for l in f]

# Open the output and read the output tensor
scores = softmax(tvm_output)
scores = np.squeeze(scores)
ranks = np.argsort(scores)[::-1]
for rank in ranks[0:5]:
    print("class='%s' with probability=%f" % (labels[rank], scores[rank]))

调优模型，获取调优数据

于目标硬件平台，用 AutoTVM 自动调优，获取调优数据：

import tvm.auto_scheduler as auto_scheduler
from tvm.autotvm.tuner import XGBTuner
from tvm import autotvm

number = 10
repeat = 1
min_repeat_ms = 0  # since we're tuning on a CPU, can be set to 0
timeout = 10  # in seconds

# create a TVM runner
runner = autotvm.LocalRunner(
    number=number,
    repeat=repeat,
    timeout=timeout,
    min_repeat_ms=min_repeat_ms,
    enable_cpu_cache_flush=True,
)

tuning_option = {
    "tuner": "xgb",
    "trials": 10,
    "early_stopping": 100,
    "measure_option": autotvm.measure_option(
        builder=autotvm.LocalBuilder(build_func="default"), runner=runner
    ),
    "tuning_records": "resnet-50-v2-autotuning.json",
}

# begin by extracting the tasks from the onnx model
tasks = autotvm.task.extract_from_program(mod["main"], target=target, params=params)

# Tune the extracted tasks sequentially.
for i, task in enumerate(tasks):
    prefix = "[Task %2d/%2d] " % (i + 1, len(tasks))
    tuner_obj = XGBTuner(task, loss_type="rank")
    tuner_obj.tune(
        n_trial=min(tuning_option["trials"], len(task.config_space)),
        early_stopping=tuning_option["early_stopping"],
        measure_option=tuning_option["measure_option"],
        callbacks=[
            autotvm.callback.progress_bar(tuning_option["trials"], prefix=prefix),
            autotvm.callback.log_to_file(tuning_option["tuning_records"]),
        ],
    )

上述 tuning_option 选用的 XGBoost Grid 算法进行优化搜索，数据记录进 tuning_records。

重编译模型，用调优数据

重新编译出一个优化模型，依据调优数据：

with autotvm.apply_history_best(tuning_option["tuning_records"]):
    with tvm.transform.PassContext(opt_level=3, config={}):
        lib = relay.build(mod, target=target, params=params)

dev = tvm.device(str(target), 0)
module = graph_executor.GraphModule(lib["default"](dev))


# Verify that the optimized model runs and produces the same results

dtype = "float32"
module.set_input(input_name, img_data)
module.run()
output_shape = (1, 1000)
tvm_output = module.get_output(0, tvm.nd.empty(output_shape)).numpy()

scores = softmax(tvm_output)
scores = np.squeeze(scores)
ranks = np.argsort(scores)[::-1]
for rank in ranks[0:5]:
    print("class='%s' with probability=%f" % (labels[rank], scores[rank]))

对比调优与非调优模型

收集优化后的性能数据，与优化前的对比：

import timeit

timing_number = 10
timing_repeat = 10
optimized = (
    np.array(timeit.Timer(lambda: module.run()).repeat(repeat=timing_repeat, number=timing_number))
    * 1000
    / timing_number
)
optimized = {"mean": np.mean(optimized), "median": np.median(optimized), "std": np.std(optimized)}

print("optimized: %s" % (optimized))
print("unoptimized: %s" % (unoptimized))

调优模型，整个过程的运行结果，如下：

$ time python autotvm_tune.py
# TVM 编译运行模型
## Downloading and Loading the ONNX Model
## Downloading, Preprocessing, and Loading the Test Image
## Compile the Model With Relay
target [llvm]: llvm -mcpu=core-avx2
One or more operators have not been tuned. Please tune your model for better performance. Use DEBUG logging level to see more details.
## Execute on the TVM Runtime
## Collect Basic Performance Data
{'mean': 44.97057118016528, 'median': 42.52320024970686, 'std': 6.870915251002107}
## Postprocess the output
class='n02123045 tabby, tabby cat' with probability=0.621104
class='n02123159 tiger cat' with probability=0.356378
class='n02124075 Egyptian cat' with probability=0.019712
class='n02129604 tiger, Panthera tigris' with probability=0.001215
class='n04040759 radiator' with probability=0.000262
# AutoTVM 调优模型 [Y/n]
## Tune the model
[Task  1/25]  Current/Best:  156.96/ 353.76 GFLOPS | Progress: (10/10) | 4.78 s Done.
[Task  2/25]  Current/Best:   54.66/ 241.25 GFLOPS | Progress: (10/10) | 2.88 s Done.
[Task  3/25]  Current/Best:  116.71/ 241.30 GFLOPS | Progress: (10/10) | 3.48 s Done.
[Task  4/25]  Current/Best:  119.92/ 184.18 GFLOPS | Progress: (10/10) | 3.48 s Done.
[Task  5/25]  Current/Best:   48.92/ 158.38 GFLOPS | Progress: (10/10) | 3.13 s Done.
[Task  6/25]  Current/Best:  156.89/ 230.95 GFLOPS | Progress: (10/10) | 2.82 s Done.
[Task  7/25]  Current/Best:   92.33/ 241.99 GFLOPS | Progress: (10/10) | 2.40 s Done.
[Task  8/25]  Current/Best:   50.04/ 331.82 GFLOPS | Progress: (10/10) | 2.64 s Done.
[Task  9/25]  Current/Best:  188.47/ 409.93 GFLOPS | Progress: (10/10) | 4.44 s Done.
[Task 10/25]  Current/Best:   44.81/ 181.67 GFLOPS | Progress: (10/10) | 2.32 s Done.
[Task 11/25]  Current/Best:   83.74/ 312.66 GFLOPS | Progress: (10/10) | 2.74 s Done.
[Task 12/25]  Current/Best:   96.48/ 294.40 GFLOPS | Progress: (10/10) | 2.82 s Done.
[Task 13/25]  Current/Best:  123.74/ 354.34 GFLOPS | Progress: (10/10) | 2.62 s Done.
[Task 14/25]  Current/Best:   23.76/ 178.71 GFLOPS | Progress: (10/10) | 2.90 s Done.
[Task 15/25]  Current/Best:  119.18/ 534.63 GFLOPS | Progress: (10/10) | 2.49 s Done.
[Task 16/25]  Current/Best:  101.24/ 172.92 GFLOPS | Progress: (10/10) | 2.49 s Done.
[Task 17/25]  Current/Best:  309.85/ 309.85 GFLOPS | Progress: (10/10) | 2.69 s Done.
[Task 18/25]  Current/Best:   54.45/ 368.31 GFLOPS | Progress: (10/10) | 2.46 s Done.
[Task 19/25]  Current/Best:   78.69/ 162.43 GFLOPS | Progress: (10/10) | 3.29 s Done.
[Task 20/25]  Current/Best:   40.78/ 317.50 GFLOPS | Progress: (10/10) | 4.52 s Done.
[Task 21/25]  Current/Best:  169.03/ 296.36 GFLOPS | Progress: (10/10) | 3.95 s Done.
[Task 22/25]  Current/Best:   90.96/ 210.43 GFLOPS | Progress: (10/10) | 2.28 s Done.
[Task 23/25]  Current/Best:   48.93/ 217.36 GFLOPS | Progress: (10/10) | 2.87 s Done.
[Task 25/25]  Current/Best:    0.00/   0.00 GFLOPS | Progress: (0/10) | 0.00 s Done.
[Task 25/25]  Current/Best:   25.50/  33.86 GFLOPS | Progress: (10/10) | 9.28 s Done.
## Compiling an Optimized Model with Tuning Data
class='n02123045 tabby, tabby cat' with probability=0.621104
class='n02123159 tiger cat' with probability=0.356378
class='n02124075 Egyptian cat' with probability=0.019712
class='n02129604 tiger, Panthera tigris' with probability=0.001215
class='n04040759 radiator' with probability=0.000262
## Comparing the Tuned and Untuned Models
optimized: {'mean': 34.736288779822644, 'median': 34.547542000655085, 'std': 0.5144378649382363}
unoptimized: {'mean': 44.97057118016528, 'median': 42.52320024970686, 'std': 6.870915251002107}

real    3m23.904s
user    5m2.900s
sys     5m37.099s

对比性能数据，可以发现：调优模型的运行速度更快、更平稳。

参考

• 笔记: start-ai-compiler
• 资料:
  • 2020 / The Deep Learning Compiler: A Comprehensive Survey
    • [译] 深度学习编译器综述
  • 2018 / TVM: An Automated End-to-End Optimizing Compiler for Deep Learning
    • [译] TVM: 一个自动的端到端深度学习优化编译器

GoCoding 个人实践的经验分享，可关注公众号！