Skip to content

Conversation

@apanteleev
Copy link
Contributor

This is a follow-up for #5959

The SPIR-V spec requires that OpSDot and OpUDot have an extra Packed Vector Format operand when Vector 1 and Vector 2 are scalar integer types. My previous implementation didn't add that operand - sorry about that. The generated code worked on NVIDIA GPUs but produced incorrect ouptut on AMD and resulted in a crash on Intel ARC. This PR fixes the generated code so that it works on all three vendors' GPUs.

@apanteleev apanteleev changed the title [SPIRV] Fix the code generated for dot4add_8packed and dot4add_i8packed [SPIRV] Fix the code generated for dot4add_u8packed and dot4add_i8packed Dec 6, 2023
@s-perron s-perron requested a review from Keenuts January 15, 2024 16:22
@s-perron s-perron enabled auto-merge (squash) January 15, 2024 16:22
@s-perron s-perron requested a review from sudonatalie January 15, 2024 16:22
@s-perron s-perron merged commit bdcbd7b into microsoft:main Jan 15, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Archived in project

Development

Successfully merging this pull request may close these issues.

3 participants