Research map / Architectures and efficiency / CNN design
DWM: A Decomposable Winograd Method for Convolution Acceleration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Decompose large kernels into smaller kernels before applying Winograd | Winograd convolution · 5x5 to 11x11 stride 1 convolutions | speedup: 0.33 -> 2.04; 0.10 -> 1.96; 0.05 -> 2.25; 0.02 -> 2.15 | Helped |
| Split kernel and input by stride groups to enable Winograd | Winograd convolution · stride 2 convolution | speedup: N/A -> 1.44; N/A -> 2.04; N/A -> 1.96; N/A -> 1.92; N/A -> 2.15 | Helped |
| Use DWM instead of traditional Winograd | Convolution runtime · 5x5 to 11x11 convolution, input 14x14 | runtime (ms): 69.70 -> 11.37; 133.34 -> 22.83; 248.71 -> 30.67; 349.33 -> 48.29 | Helped |
| Use DWM instead of cuDNN | Convolution runtime · 9x9 and 11x11 convolution, input 14x14 | runtime (ms): 50.92 -> 30.67; 94.63 -> 48.29 | Helped |
| Use DWM instead of cuDNN | Convolution runtime · 3x3 convolution, input 14x14 | runtime (ms): 2.80 -> 3.67 | Hurt |
| Use DWM instead of Winograd | CNN inference · ImageNet 2012 networks | top-1 accuracy: AlexNet: 56.51 -> 56.51; GoogLeNet: 69.79 -> 69.77; Inception-V3: 69.47 -> 69.46; ResNet-152: 78.31 -> 78.31; DenseNet-1 | No clear effect |
| Use DWM rather than traditional Winograd for larger kernels | Single-layer convolution · 5x5 to 11x11 FP32 convolutions | MSE versus FP64: 7.14E-08 -> 1.47E-09; 1.47E-03 -> 2.97E-09; 5.31E-03 -> 3.67E-09; 1.48E-02 -> 5.30E-09 | Helped |
| Use DWM rather than traditional Winograd in FP16 on large kernels | Single-layer convolution · 7x7 to 11x11 FP16 convolutions | MSE versus FP64: NaN -> 1.97E-01; NaN -> 5.88E-02; NaN -> 2.36E-01; NaN -> 7.33E-02; NaN -> 3.46E-01; NaN -> 1.15E-01 | Helped |
Closest papers by meaning
- Efficient Sparse-Winograd Convolutional Neural Networks (ICLR 2018 · similarity 0.90)
- Towards Efficient and Accurate Winograd Convolution via Full Quantization (NeurIPS 2023 · similarity 0.88)
- Fast Algorithms for Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.85)
- SFC: Achieve Accurate Fast Convolution under Low-precision Arithmetic (ICML 2024 · similarity 0.82)
- FlashFFTConv: Efficient Convolutions for Long Sequences with Tensor Cores (ICLR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.