Using the Winograd algorithm for convolution execution can provide increased performance compared to common implementation. However, it may be difficult to understand which algorithm would be faster due to dependency on convolution layer parameters and hardware configuration. The Winograd Algorithmic Tuner solves this problem automatically.
For more detailed information about the algorithm, refer to this whitepaper.
Winograd Algorithmic Tuner Workflow
- Input: Original IR
- Collecting per-layer performance data for the original model.
- Collecting per-layer performance numbers for the modified model, in which the Winograd algorithm was set as the priority for all convolution layers.
- Determining convolutions to be executed with the Winograd algorithm. The tuner uses heuristic approaches in this step. It considers the performance of each particular convolution layer and analyzes changes in the execution graph that can lead to performance overhead.
- Output: Modified IR with preassigned algorithm priority for each convolution layer
NOTE: OpenVINO™ toolkit supports the Winograd algorithm only for a limited set of convolution parameters. Some convolution layers with prioritization of the Winograd algorithm cannot be executed with this approach.
Winograd Algorithmic Tuner Prerequisites and Limitations
- Supports only CPU targets with Intel® Advanced Vector Extensions 512 (Intel® AVX-512) instruction set.
- Supports only convolution layers.
- Applicable only to a certain subset of convolution parameters, such as 3x3 kernel. Strides are not supported.
- Mostly applicable to highly compute-bound convolutions, because the algorithm includes large overhead on data permutations.
- Is not available for configurations with remote machines.
Configure Winograd Optimization Settings
Go to the Perform tab on the Projects page and open the Optimize subtab. Select Winograd and click ****.
NOTE: Using Winograd optimization, you can tune an original (top-level) model, or a model that has already been tuned.
View Winograd Calibration
Click Optimize and a new row for your model appears.
Once the job is complete, click on it to view inference results.