Cross-Feature Transfer Learning for Efficient Tensor Program Generation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper proposes a transfer learning approach to generate more efficient tensor programs with less tuning time and fewer Kernel measurements on heterogeneous hardware. The paper approaches a new method that learns the neural network and the hardware feature space, facilitating knowledge transfer to new, unseen target hardware. A comprehensive analysis is performed on the existing state-of-the-art dataset, TenSet. From the tests presented, the proposed method has good effectiveness and can be analyzed with interest by researchers in the field.

I would recommend a more detailed description of the method that adapts tensor programs to incorporate neural networks and hardware-specific features.

Comments on the Quality of English Language

There are some spelling mistakes in the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Please see the attached file.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear authors,

Hello, I am honored to have the opportunity to read your paper. I personally believe that your achievements have a certain degree of innovation, the experiments are sufficient, and the results are convincing, with some instructive significance. However, there are still some areas that need modification:

1. The introduction of Cross-device Learning is relatively brief compared to the extensive coverage of Machine Learning-based Auto Tuners in the related work section. This imbalance might be addressed by the author through further elaboration on Cross-device Learning and potentially reducing some less crucial details about Machine Learning-based Auto Tuners.

2. The labels in Figure 1 do not align perfectly with the seven labels starting from line 289, as label 8 is missing. Additionally, the next paragraph should be indented.

3. In Section 4.1 under "Dataset and Model" for the three models, please provide the references for the models and clarify the dataset used.

4. In Section 4.1, please provide a detailed explanation of the specific meanings of all the evaluation metrics used in the upcoming experiments, including terms like "Occurrence."

5. I suggest adding a few more references to enhance the persuasiveness of your paper, such as:

1) https://doi.org/10.1016/j.compeleceng.2022.108401

2) https://doi.org/10.1109/TPDS.2020.3030548

3) https://doi.org/10.1016/j.eswa.2023.122289

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The paper has undergone a very serious edition and all reviewer's remarks were taken into account. In its corrected form he paper is ready for publication.