FlexTensor

FlexTensor is moving to https://github.com/pku-liang

Introductions

Tensor computation plays a paramount role in a broad range of domains, including machine learning, data analytics, and scientific computing. The wide adoption of tensor computation and its huge computation cost has led to high demand for flexible, portable, and high-performance library implementation on heterogeneous hardware accelerators such as GPUs and FPGAs. However, the current tensor library implementation mainly requires programmers to manually design low-level implementation and optimize from the algorithm, architecture, and compilation perspectives. Such a manual development process often takes months or even years, which falls far behind the rapid evolution of the application algorithms.

We introduce FlexTensor, which is a schedule exploration and optimization framework for tensor computation on heterogeneous systems. FlexTensor can optimize tensor computation programs without human interference, allowing programmers to only work on high-level programming abstraction without considering the hardware platform details. FlexTensor systematically explores the optimization design spaces that are composed of many different schedules for different hardware. Then, FlexTensor combines different exploration techniques, including heuristic method and machine learning method to find the optimized schedule configuration. Finally, based on the results of exploration, customized schedules are automatically generated for different hardware. In the experiments, we test 12 different kinds of tensor computations with totally hundreds of test cases and FlexTensor achieves average 1.83x performance speedup on NVIDIA V100 GPU compared to cuDNN; 1.72x performance speedup on Intel Xeon CPU compared to MKL-DNN for 2D convolution; 1.5x performance speedup on Xilinx VU9P FPGA compared to OpenCL baselines; 2.21x speedup on NVIDIA V100 GPU compared to the state-of-the-art.

Installation

Requires: Python 3.5+, Numpy, tvm: https://github.com/KnowingNothing/tvm/tree/mirror

1. Install TVM, follow the instructions.
2. Clone this repo:

   git clone https://github.com/KnowingNothing/FlexTensor.git

3. Set the environments: export AUTO_HOME=path/to/FlexTensor export PYTHONPATH=$AUTO_HOME:$PYTHONPATH

To run the baselines, PyTorch is required.

Usage

1. Write compute in Python

For example, write a gemm kernel. This uses TVM DSL in Python.

import tvm

def gemm(A, B):
    """Matrix multiplies matrix

    Args:
    -----------------------------
    A: tvm.tensor.Tensor
        shape [height, width]
    B: tvm.tensor.Tensor
        shape [width, length]
    -----------------------------

    Returns:
    -----------------------------
    tvm.tensor.Tensor
        shape [height, length]
    -----------------------------
    """
    k = tvm.reduce_axis((0, B.shape[0]))
    return tvm.compute((A.shape[0], B.shape[1]), lambda i, j: tvm.sum(A[i, k] * B[k, j], axis=k))

Existing computes for some common tensor computations are in nn/ops.py.

2. Register optimization tasks

Take the gemm kernel above as example. First wrap the kernel.

def wrap_gemm(N, K, M):
    A = tvm.placeholder((N, K))
    B = tvm.placeholder((K, M))
    Output = gemm(A, B)
    return [Output.op], [A, B, Output]

Then register a task.

from flextensor.task import register_task, Task

'''
To create a task, the parameters are:
1. type of operator: str
2. name of this operator: str
3. the wrapper for tensor computation
4. arguments to the wrapper, i.e. input shapes
5. target device: str ("llvm" or "cuda" currently)
6. device number: int
'''
task = Task(
    "gemm", 
    "gemm", 
    wrap_gemm, 
    (1024, 1024, 1024), 
    "llvm", 
    0)
# register the task
register_task(task)

Existing tasks are registered in task.py.

3. Push the button

from flextensor.scheduler import schedule

s, bufs, configs = schedule(
            task.key, # give the key of target task
            slevel=4,
            rlevel=3,
            op_trial=100, 
            timeout=10, 
            op_stop=30, 
            method="searching", 
            parallel=8,
            )

Wait patiently for the scheudling process to finish. This usually cost tens of miniutes on GPU and several hours on CPU. But it depends on the op_trial, timeout, parallel settings and input shapes.

4. Enjoy the results

The resulting s and bufs can be directly used to generate codes, the resulting configs can be saved to retrieve schedules.

# directly use the results
func = tvm.build(s, bufs, task.target)
# use the configs
from flextensor.scheduler import schedule_with_config

s, bufs = schedule_with_config(task_key, configs)
func = tvm.build(s, bufs, task.target)

Citing FlexTensor

If you find FlexTensor useful for your project, please cite the following paper：

@inproceedings{FlexTensor, author = {Size Zheng and Yun Liang and Shuo Wang and Renze Chen and Kaiwen Sheng}, editor = {James R. Larus and Luis Ceze and Karin Strauss}, title = {FlexTensor: An Automatic Schedule Exploration and Optimization Framework for Tensor Computation on Heterogeneous System}, booktitle = {{ASPLOS} '20: Architectural Support for Programming Languages and Operating Systems, Lausanne, Switzerland, March 16-20, 2020 {[ASPLOS} 2020 was canceled because of {COVID-19]}}, pages = {859--873}, publisher = {{ACM}}, year = {2020}, url = {https://doi.org/10.1145/3373376.3378508}, doi = {10.1145/3373376.3378508}, timestamp = {Mon, 16 Mar 2020 11:14:36 +0100}, biburl = {https://dblp.org/rec/conf/asplos/Zheng0WCS20.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} }