Skip to main content

Explore the energy-efficient dataflow scheduling for neural networks.

Project description

Neural Network Dataflow Scheduling

This Python tool allows you to explore the energy-efficient dataflow scheduling for neural networks (NNs), including array mapping, loop blocking and reordering, and (coarse-grained) parallel processing within and across layers.

For hardware, we assume an Eyeriss-style NN accelerator [Chen16], i.e., a 2D array of processing elements (PEs) with a local register file in each PE, and a global SRAM buffer shared by all PEs. We further support a tiled architecture with multiple nodes that can partition and process the NN computations in parallel. Each node is an Eyeriss-style engine as above.

In software, we decouple the dataflow scheduling into three subproblems:

  • Array mapping, which deals with mapping one 2D convolution computation (one 2D ifmap convolves with one 2D filter to get one 2D ofmap) onto the hardware PE array. We support row stationary mapping [Chen16].
  • Loop blocking and reordering, which decides the order between all 2D convolutions by blocking and reordering the nested loops. We support exhaustive search over all blocking and reordering schemes [Yang16], and analytical bypass solvers [Gao17].
  • Parallel processing, which partitions the NN computations across the multiple tiled engines. We support both intra-layer and inter-layer parallelism. For intra-layer, we support batch partitioning, fmap partitioning, output partitioning, input partitioning, and the combination between them (hybrid) [Gao17]. We also explore various dataflow optimizations including access forwarding and buffer sharing [Gao19]. We use exhaustive search within each layer. For inter-layer, we support spatial pipelining (inter-layer pipelining) and temporal pipelining (time multiplexing without writing back intermediate data) as well as their optimized scheduling [Gao19]. We use layer-wise greedy beam search across layers.

See the details in our ASPLOS’17 [Gao17] and ASPLOS’19 [Gao19] papers.

If you use this tool in your work, we kindly request that you reference our paper(s) below, and send us a citation of your work.

  • Gao et al., “TETRIS: Scalable and Efficient Neural Network Acceleration with 3D Memory”, in ASPLOS, April 2017.
  • Gao et al., “TANGRAM: Optimized Coarse-Grained Dataflow for Scalable NN Accelerators”, in ASPLOS. April 2019.


nn_dataflow supports Python 3.6 and above.

nn_dataflow can be directly used without installation if you have first defined the environment variable PYTHONPATH to include the top directory path. See the Usage section below for details.

nn_dataflow has been registered on PyPI, so it can be installed through pip as:

> pip install nn-dataflow

And pip will take care of all dependencies.

To only install nn_dataflow in local user install directory (without sudo), and/or to install in editable mode, at the top directory do:

> pip install --user -e .


First, define the NN structure in nn_dataflow/nns. We already defined several popular NNs for you, including AlexNet, VGG-16, GoogLeNet, ResNet-152, etc.

Then, use nn_dataflow/tools/ to search for the optimal dataflow for the NN. For detailed options, type:

> python ./nn_dataflow/tools/ -h

You can specify NN batch size and word size, PE array dimensions, number of tile nodes, register file and global buffer capacity, and the energy cost of all components. Note that, the energy cost of array bus should be the average energy of transferring the data from the buffer to one PE, not local neighbor transfer; the unit static energy cost should be the static energy of all nodes in one clock cycle.

Other options include:

  • -g, --goal: E, D, or ED. the optimization goal, e(nergy), d(elay), or ED product.
  • --mem-type: 2D or 3D. With 2D memory, memory channels are only on the four corners of the chip; with 3D memory, memory channels are on the top of all tile nodes (one per each).
  • --bus-width: the multicast bus bit width in the PE array for one data type. Set to 0 to ignore multicast overheads.
  • --dram-bw: float or inf. Total DRAM bandwidth for all tile nodes, in bytes per cycle.
  • --disable-bypass: a combination of i, o, f, whether to disallow global buffer bypass for ifmaps, ofmaps, and weights.
  • --solve-loopblocking: whether to use analytical bypass solvers for loop blocking and reordering. See [Gao17].
  • --hybrid-partitioning: whether to use hybrid partitioning in [Gao17]. If not enabled, use naive partitioning, i.e., fmap partitioning for CONV layers, and output partitioning for FC layers.
  • --batch-partitioning and --ifmap-partitioning: whether the hybrid partitioning also explores batch and input partitioning.
  • --enable-access-forwarding: access forwarding, where the nodes fetch disjoint subsets of data and forward them to other nodes. See [Gao19].
  • --enable-gbuf-sharing: buffer sharing, where the global buffer capacity is shared across nodes through NoC. See [Gao19].
  • --enable-save-writeback: allow to elide the intermediate data writeback to memory when switching between layers if it is possible to store the entire data set in on-chip buffers.
  • --interlayer-partition: whether to use inter-layer pipelining to partition resources across multiple layers and process them simultaneously.
  • --layer-pipeline-time-overhead, --layer-pipeline-max-degree: constrain the configuration space of inter-layer pipelining, by specifying the maximum execution time overhead, or the maximum pipelining degree.
  • --disable-interlayer-opt: disable optimizations and only allow basic inter-layer pipelining.

Code Structure

  • nn_dataflow
    • core
      • Top-level dataflow exploration: nn_dataflow, nn_dataflow_scheme.
      • Layer scheduling: scheduling.
      • Array mapping: map_strategy.
      • Loop blocking and reordering: loop_blocking, loop_blocking_scheme, loop_blocking_solver.
      • Intra-layer partitioning: partition, partition_scheme, buf_shr_scheme.
      • Inter-layer pipelining: inter_layer_pipeline, pipeline_segment.
      • Network and layer: network, layer.
    • nns: example NN definitions.
    • tests: unit tests.
    • tools: executables.

Verification and Testing

To verify the tool against the Eyeriss result [Chen16], see nn_dataflow/tests/dataflow_test/

To run (unit) tests, do one of the following:

> python -m unittest discover

> python -m pytest

> pytest

To check code coverage with pytest-cov plug-in:

> pytest --cov=nn_dataflow


[Gao19](1, 2, 3, 4, 5) Gao, Yang, Pu, Horowitz, and Kozyrakis, TANGRAM: Optimized Coarse-Grained Dataflow for Scalable NN Accelerators, in ASPLOS. April, 2019.
[Gao17](1, 2, 3, 4, 5) Gao, Pu, Yang, Horowitz, and Kozyrakis, TETRIS: Scalable and Efficient Neural Network Acceleration with 3D Memory, in ASPLOS. April, 2017.
[Chen16](1, 2, 3) Chen, Emer, and Sze, Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks, in ISCA. June, 2016.
[Yang16]Yang, Pu, Rister, Bhagdikar, Richardson, Kvatinsky, Ragan-Kelley, Pedram, and Horowitz, A Systematic Approach to Blocking Convolutional Neural Networks, arXiv preprint, 2016.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for nn-dataflow, version 2.1
Filename, size File type Python version Upload date Hashes
Filename, size nn_dataflow-2.1-py3-none-any.whl (234.2 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size nn_dataflow-2.1.tar.gz (163.9 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page