[AdvancedCompiler]Optimize argmin

[AdvancedCompiler]

PR Category

Operator

Type of Change

Performance Optimization

Description

Issue

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Performance

[gemini-code-assist]

Summary of Changes

Hello @AdvancedCompiler, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the performance of the argmin operator by introducing specialized Triton kernels. It refactors the existing argmin implementation to dynamically select an optimized kernel based on the input tensor's K dimension, alongside new heuristics to fine-tune execution parameters for improved efficiency on NVIDIA GPUs.

Highlights

• Kernel Specialization: The argmin Triton kernel has been refactored into two specialized kernels: argmin_kernel_non_inner and argmin_kernel_inner. This allows for tailored optimizations based on the input tensor's dimensions.
• Optimized Heuristics: New heuristic functions have been introduced to provide fine-grained control over tiling, CTA configuration, and warp counts for the newly specialized argmin kernels, aiming for improved performance on NVIDIA GPUs.
• Dynamic Kernel Dispatch: The Python argmin function now includes logic to dynamically dispatch to either argmin_kernel_non_inner (when K > 1) or argmin_kernel_inner (when K == 1), ensuring the most efficient kernel is used for a given input.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request optimizes the argmin operator by splitting the implementation into two specialized Triton kernels for inner and non-inner reduction dimensions, which is a solid performance optimization strategy. The new kernels and their corresponding heuristics are well-structured. However, I've identified a significant issue in the heuristic functions where the input data type is not correctly inferred, potentially leading to suboptimal performance. Additionally, I've provided suggestions to improve code clarity and maintainability in one of the new kernels and a heuristic function.

[gemini-code-assist]

The dtype is fetched using args.get("dtype", "fp16"), but the kernel call from argmin does not pass a dtype argument. This will cause the heuristic to always use the default value "fp16", leading to suboptimal performance for other data types like float32. You should get the dtype from the input tensor inp which is available in args.

Suggested change

	dtype = args.get("dtype", "fp16")
	dtype = "fp32" if args["inp"].dtype == torch.float32 else "fp16"

[StrongSpoon]

it seems there is no argument named dtype

[gemini-code-assist]

The dtype is fetched using args.get("dtype", "fp16"), but the kernel call from argmin does not pass a dtype argument. This will cause the heuristic to always use the default value "fp16", leading to suboptimal performance for other data types like float32. You should get the dtype from the input tensor inp which is available in args.

Suggested change

	dtype = args.get("dtype", "fp16")
	dtype = "fp32" if args["inp"].dtype == torch.float32 else "fp16"

[gemini-code-assist]

This while loop to find the largest power of two less than or equal to K can be simplified. Using 1 << (K.bit_length() - 1) is more concise and achieves the same result for K > 0.

Suggested change

	if K <= 128:
	tile_k = 1
	while tile_k * 2 <= K:
	tile_k *= 2
	return tile_k
	if K <= 128:
	return 1 << (K.bit_length() - 1) if K > 0 else 1

[StrongSpoon]

the suggested code is better.

[StrongSpoon]

I believe the points mentioned above also apply to the optimization of argmax. Please update the code in accordance with the suggestions provided in both pull requests.