Repository navigation
feat(rocfft): twiddle table 5-step decomposition - #12054
Merged
eng-flavio-teixeira merged 92 commits intoOct 8, 2026
Merged
eng-flavio-teixeira merged 92 commits into
eng-flavio-teixeira merged 92 commits into
Conversation
- Add 64 bit index check to generate_node().
- Fix issue with rtc test harness.
…avio-teixeira/rocm-libraries into transpose_64_bit_indexing
…table_64_bit_length
…ts on large enough sizes.
…table_64_bit_length
eng-flavio-teixeira
marked this pull request as ready for review
September 22, 2026 20:58
eng-flavio-teixeira
changed the title
eng-flavio-teixeira
changed the title
eng-flavio-teixeira
changed the title
malcolmroberts
approved these changes
Sep 29, 2026
malcolmroberts
left a comment
malcolmroberts
left a comment
Contributor
There was a problem hiding this comment.
I got some performance deltas, but a priori this looks like the only code change is a minor change in some integer types. The performance deltas are mostly in the right direction though, and it may just be noise in any case, so this looks good.
… it 0.1 * test_prob of running.
evetsso
approved these changes
Oct 1, 2026
eng-flavio-teixeira
enabled auto-merge (squash)
October 5, 2026 16:03
1 task done
regan-amd
added a commit
that referenced
this pull request
Oct 9, 2026
## Motivation Self-explanatory title, drift introduced in #12054 JIRA ID: AIFFT-532 ## Technical Details Self-explanatory title ## Test Plan Usual suite ## Test Result Tests should pass. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.
Contributor
TheRock Submodule Bump ActivityNewest first
|
1 task done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Very large FFTs, e.g., lengths greater than 2^32, can still fail on MI455X due to the twiddle table kernel code missing a 5-step decomposition for retrieving the twiddle values.
JIRA ID: AIFFT-532
Technical Details
This PR adds support for the 5-step decomposition, which theoretically enables support for FFTs of lengths up to 2^40, or 16 TiB of VRAM in single-precision.
This PR also expands the capability of the multi-kernel Bluestein plan for very large FFTs. Launching one of the Bluestein kernels could overflow the grid X dim if the length was greater than 2^32. The kernel now uses the grid Y dim if the length is big enough.
Test Plan
New tests have been added to the adhoc test suite to cover the improvements.
Test Result
All tests should pass without issues. Some of the test cases can run on MI300x and MI355x, but the very large length Bluestein plan (token complex_forward_len_2147483659_single_ip_batch_1_istride_1_CI_ostride_1_CI_idist_2147483659_odist_2147483659_ioffset_0_0_ooffset_0_0) can only run on a MI455x due to the VRAM constraints.
Submission Checklist