Repository navigation
fix(tensilelite): fix TDM pointer-array batch handling - #11334
Merged
Merged
Conversation
github-actions
Bot
added
project: hipblaslt
project: hipsparselt
ci:hipsparselt-fast
labels
nakajee
reviewed
Aug 26, 2026
jaopaulolc
added
the
rocm:gemm algos
Codecov Report❌ Patch coverage is ❌ Your project check has failed because the head coverage (38.52%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## develop #11334 +/- ##
===========================================
+ Coverage 69.64% 69.65% +0.02%
===========================================
Files 2826 2815 -11
Lines 469878 466858 -3020
Branches 69394 68907 -487
===========================================
- Hits 327211 325187 -2024
+ Misses 118978 118101 -877
+ Partials 23689 23570 -119
*This pull request uses carry forward flags. Click here to find out more.
🚀 New features to boost your workflow:
|
Mailbox ds_store/ds_load carry memTokenLdsBuffer0 so the scheduler places a publish fence between them and a release before the next LDS0 write. WG barriers stay untokened. SK5 extracts HybridMode from MagicShift bit 30, then clears that bit by XOR with (HybridMode<<30). The magic add + 5-bit shift is masked into a temp so the SKTiles overlay in MagicShiftItersPerTile is left intact. sk_hybrid_quick includes a TDMInst=3 HGEMM group (SK5, SIA 0/4).
JIRA ID : N/A
ArgType 3 supplies A and B as pointer arrays. TDM loads the selected batch
matrix, then applies tile offsets to that matrix base. A/B strided-batch WG2
offsets are suppressed after the pointer-array load.
Motivation
- Dereference the selected A and B batch pointers before applying TDM tile offsets on classic and subtile paths.
- Suppress strided-batch offsets after selecting a pointer-array matrix base, and only for A/B.
- Wait after the conventional A/B pointer-array SRD load before reusing that temp for batchOffset (gfx1250 XNACK replay).
- Route pointer tables into output-conversion C/D, and reject sparse, grouped, MBSK, and non-MultipleBuffer GSU>1.
- Initialize non-wave-separated TDM descriptors before applying their global offsets.
- Add standalone-client BatchMode plumbing for the pointer-array ABI.
- Add gfx1250 runtime regressions for GSU1/GSU2, static SK3/SK5, and subtile TDM, sweeping SIA 0 and 4.
Test Plan
- Characterization: r3 datamover, StreamK TDM SGPR-budget, ClientConfigIni.
- `tdm_general_batched_gfx1250.yaml`: non-StreamK pointer-array TDM, SIA 0 and 4.
- `sk_tdm_general_batched.yaml`: SK {3,4,5} x SIA {0,4} x hybrid {0,1}, BatchMode 1.
- `sk_tdm_subtile_general_batched.yaml`: subtile SK3, SIA 0 and 4.
- Rebuild gfx1250 device lib / hipBLASLt gtest; nightly general-batched BF16.
- tox on gfx1250
Test Results
- Characterization tests listed above passed.
- Non-StreamK pointer-array TDM passed GSU1/GSU2, SIA 0 and 4.
- SK3, SK4, and SK5 pointer-array TDM passed SIA 0 and 4.
- Subtile pointer-array TDM (SK3, SIA 0 and 4) passed.
- gfx1250 device-library and hipBLASLt gtest builds passed.
- The original nightly general-batched BF16 regression passed.
ThanHenderson
force-pushed
the
users/nhenders/fix-tdm-sk-hybrid-extract
branch
from
August 27, 2026 20:51
7ba08b2 to
b2f7586
Compare
ThanHenderson
force-pushed
the
users/nhenders/fix-tdm-general-batched
branch
from
August 27, 2026 20:51
76e3a0d to
f9a04e6
Compare
Contributor
|
Thanks for your comments. |
yenong-amd
approved these changes
Aug 28, 2026
Base automatically changed from
users/nhenders/fix-tdm-sk-hybrid-extract
to
develop
August 28, 2026 22:11
✅ All Policy Checks Passed
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🎉 All checks passed! This PR is ready for review. |
Contributor
Author
|
Fixes items added to |
jaopaulolc
approved these changes
Aug 31, 2026
Contributor
|
@mengzcai Could you help to review this? |
1 task done
jaopaulolc
reviewed
Sep 2, 2026
Contributor
|
This function should work for complex TDM as well (ROCM-30021: gfx1250 CGEMM via the pointer-array batched API). |
1 task done
Contributor
TheRock Submodule Bump ActivityNewest first
|
mengzcai
added a commit
that referenced
this pull request
Sep 8, 2026
JIRA ID : ROCM-30021 ## Motivation Add a complex general-batch edge case covering the pointer-array (batch_mode=1) path over degenerate/boundary sizes and conjugate transposes, and toggle batch_mode [0, 1] on matmul_batch_medium_complex, to guard [ROCM-30021](https://amd-hub.atlassian.net/browse/ROCM-30021). <!-- Explain the purpose of this PR and the goals it aims to achieve. --> ## Technical Details Add batch_mode: [0, 1] to matmul_batch_medium_complex and one gtest case `matmul_batch_complex_general_batch_edge` reusing existing anchors (`one_matrix_size_range` + `complex_transA_transB_range`), covering batch_mode [0, 1]. No source changes. <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan `hipblaslt-test --gtest_filter=*matmul*complex*` on gfx1250. <!-- Explain any relevant testing done to verify this PR. --> ## Test Result 8684 complex tests pass, 0 failures (verified on develop branch that includes #11334). <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. [ROCM-30021]: https://amd-hub.atlassian.net/browse/ROCM-30021?atlOrigin=eyJpIjoiNWRkNTljNzYxNjVmNDY3MDlhMDU5Y2ZhYzA5YTRkZjUiLCJwIjoiZ2l0aHViLWNvbS1KU1cifQ Co-authored-by: mengzcai <mengzcai@amd.com>
1 of 2 tasks
This was referenced Sep 14, 2026
marbre
pushed a commit
that referenced
this pull request
Sep 15, 2026
…12074) #11653 Source commit: `a873b5eb2ab4cd7d58b4e2f3e3a42e64640e14fa` JIRA ID : ROCM-30021 ## Motivation Add a complex general-batch edge case covering the pointer-array (batch_mode=1) path over degenerate/boundary sizes and conjugate transposes, and toggle batch_mode [0, 1] on matmul_batch_medium_complex, to guard [ROCM-30021](https://amd-hub.atlassian.net/browse/ROCM-30021). <!-- Explain the purpose of this PR and the goals it aims to achieve. --> ## Technical Details Add batch_mode: [0, 1] to matmul_batch_medium_complex and one gtest case `matmul_batch_complex_general_batch_edge` reusing existing anchors (`one_matrix_size_range` + `complex_transA_transB_range`), covering batch_mode [0, 1]. No source changes. <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan `hipblaslt-test --gtest_filter=*matmul*complex*` on gfx1250. <!-- Explain any relevant testing done to verify this PR. --> ## Test Result 8684 complex tests pass, 0 failures (verified on develop branch that includes #11334). <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. [ROCM-30021]: https://amd-hub.atlassian.net/browse/ROCM-30021?atlOrigin=eyJpIjoiNWRkNTljNzYxNjVmNDY3MDlhMDU5Y2ZhYzA5YTRkZjUiLCJwIjoiZ2l0aHViLWNvbS1KU1cifQ Co-authored-by: mengzcai <152131106+mengzcai@users.noreply.github.com> Co-authored-by: mengzcai <mengzcai@amd.com>
marbre
pushed a commit
that referenced
this pull request
Sep 15, 2026
…#12072) #11334 Source commit: `6f0c83a90f7b00f873c1b51e82f54dd4e0e136bb` JIRA ID : AIHPBLAS-4515 ## Motivation ArgType 3 supplies A and B as pointer arrays. TDM loads the selected batch matrix, then applies tile offsets to that matrix base. A/B strided-batch WG2 offsets are suppressed after the pointer-array load. ## Technical Details - Dereference the selected A and B batch pointers before applying TDM tile offsets on classic and subtile paths. - Suppress strided-batch offsets after selecting a pointer-array matrix base, and only for A/B. - Wait after the conventional A/B pointer-array SRD load before reusing that temp for batchOffset (gfx1250 XNACK replay). - Route pointer tables into output-conversion C/D. - Initialize non-wave-separated TDM descriptors before applying their global offsets. - Add standalone-client BatchMode plumbing for the pointer-array ABI. - Add gfx1250 runtime regressions for GSU1/GSU2, static SK3/SK5, and subtile TDM, sweeping SIA 0 and 4. ## Test Plan - Characterization: r3 datamover, StreamK TDM SGPR-budget, ClientConfigIni. - `tdm_general_batched_gfx1250.yaml`: non-StreamK pointer-array TDM, SIA 0 and 4. - `sk_tdm_general_batched.yaml`: SK {3,4,5} x SIA {0,4} x hybrid {0,1}, BatchMode 1. - `sk_tdm_subtile_general_batched.yaml`: subtile SK3, SIA 0 and 4. - Rebuild gfx1250 device lib / hipBLASLt gtest; nightly general-batched BF16. - tox on gfx1250 ## Test Results - Characterization tests listed above passed. - Non-StreamK pointer-array TDM passed GSU1/GSU2, SIA 0 and 4. - SK3, SK4, and SK5 pointer-array TDM passed SIA 0 and 4. - Subtile pointer-array TDM (SK3, SIA 0 and 4) passed. - gfx1250 device-library and hipBLASLt gtest builds passed. - The original nightly general-batched BF16 regression passed. Co-authored-by: Nathan Henderson <nathan.henderson@amd.com>
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.
JIRA ID : AIHPBLAS-4515, AIHPBLAS-4770
AIHPBLAS-4770
Motivation
ArgType 3 supplies A and B as pointer arrays. TDM loads the selected batch matrix, then applies tile offsets to that matrix base. A/B strided-batch WG2 offsets are suppressed after the pointer-array load.
Technical Details
Test Plan
tdm_general_batched_gfx1250.yaml: non-StreamK pointer-array TDM, SIA 0 and 4.sk_tdm_general_batched.yaml: SK {3,4,5} x SIA {0,4} x hybrid {0,1}, BatchMode 1.sk_tdm_subtile_general_batched.yaml: subtile SK3, SIA 0 and 4.Test Results