Skip to content

Latest commit

 

History

History
99 lines (67 loc) · 3.27 KB

README.md

File metadata and controls

99 lines (67 loc) · 3.27 KB
Raw

EOP

now

For high level RelayIR, EOP can construct computation graph (in deprecated/relayIR/relay_graph.py) and calculate each op's execution time consumption separately (in deprecated/relayIR/op_statistics.py). Detailed data are stored in csv format.

Yuanjia Xu, Heng Wu, Wenbo Zhang, and Yi Hu. 2022. EOP: Efficient Operator Partition for Deep Learning Inference over Edge Servers. In Proceedings of the 18th ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments (VEE ’22), March 1, 2022, Virtual, Switzerland. ACM, New York, NY, USA, 13 pages. https://doi.org/10.1145/3516807.3516820

The paper can be found here

This work was partially supported by National Key Research and Development Program of China (2018YFB1402803) and Provincial Key Research and Development Program of Shandong, China (2021CXGC010101).

prerequest

    llvm 10.0
    cuda 11.0 (or 10.2)
    nvidia drive 450.51.05
    pytorch 1.7.1
    torchvision 0.8.2+cu110
    torchaudio 0.7.2
    huggingFace (Transformer) needRecheck
    darkNet needRecheck
    OnnX needRecheck
    MxNet needRecheck

how to use

  1. set proper path:
    export TVM_HOME=/root/github/tvm
    export PYTHONPATH=/root/github/TVMProfiler/memory_profiler:/root/github/TVMProfiler/relayIR:$TVM_HOME/python:${PYTHONPATH}
  1. demo (profile CPU/GPU time, memory and CUDA memory)
from std_memory_profiler import operation_memory_profile, operation_time_profile,operation_cuda_memory_profile, profile

def create_data(ra):
    ret = []
    for n in range(ra):
        ret.append(np.random.randn(1, 70, 71, 72))
    return ret

dict = {'a': 1, 'b': 2, 'b': '3'}

@operation_cuda_memory_profile(stream = sys.stdout,operation_meta=dict)
def process_data(data):
    data = np.concatenate(data)
    detrended = scipy.signal.detrend(data, axis=0)
    return detrended

roadmap

  1. support static ananlysis and get necessary input/out put informaction (provide both python and C++ debug modes).
  2. support catogerized compile optimization mechanisms.
  3. support operator optimizations on heterogeneous resources with reduced cost.

current support DL model

from TVM test, Pytorch, MxNet, Keras and Tensorflow (CNN classification)

resnet-, resnet3d-, mobilenet, squeezenet_v1.1, inception_v3, InceptionV1, vgg-,densenet-

from darknet or yolort.models, pytorch classification (RCNN detection)

yolov2, yolov3 or yolov3-tiny

from Onnx (RNN language)

LSTM,GRU,transformers(e.g., Bert)

from TVM test (GAN)

DCGAN

from Pytorch (GCN)

DGL-PyTorch,

debug tools

  1. codelldb (for C++/C): https://github.com/vadimcn/vscode-lldb (may manually download the vsix file)
  2. ff-navigator (for python): https://github.com/tqchen/ffi-navigator (may encounter some issues sovlved in tqchen/ffi-navigator#45)

target optimized tvm codes

  1. all are from te.compute with dynamic shape, calculation process, tag and name.

  2. te.reduce_axis: values on related axises are reduced to sum/avg.

  3. condition expression and true value testing: to do some calculation like any and some.

  4. index and shape expression: may denote dynamic shapes and expressions.

  5. data types: from int8 to float32.

  6. basic topi operators can be used as meta-operators.

data path

Profile data are stored in /data.