Skip to content

feat: Support for BLAS methods on AMD GPUs, CPU libraries, and improvements on CUDA APIs - #9

Open
sanjibansg wants to merge 29 commits into
mainfrom
dev
Open

feat: Support for BLAS methods on AMD GPUs, CPU libraries, and improvements on CUDA APIs#9
sanjibansg wants to merge 29 commits into
mainfrom
dev

Conversation

@sanjibansg

@sanjibansg sanjibansg commented Mar 21, 2026

Copy link
Copy Markdown
Member

This PR adds configurable leading dimensions for the matrix-mul functions targeted for cuBLASLt

@sanjibansg sanjibansg linked an issue Mar 21, 2026 that may be closed by this pull request
Configurable Leading dimensions in Gemm functions for cublaslt #7
Open
sanjibansg and others added 13 commits April 17, 2026 12:50
@sanjibansg
fix: use explicit data type for signatures with pointer arguments
c995464
@sanjibansg
fix: non transpose axis for matmul method
75f75a4
@sanjibansg
fix (experimental): layout shape order
ed303fb
@sanjibansg
feat: cuda cleanup and cpu blas api
54a9401
@sanjibansg
feat: matmul method on raw cuda pointers
066fdae
@sanjibansg
fix: remove extra template parameter in cuda matmul method on raw poi…
1f1c4c7 
…nters
@sanjibansg
fix: method signature for matmul on cuda raw pointers
a867939
@sanjibansg
fix: dimension order for checkAddLayout method in cuda
ad41e99
@sanjibansg
fix: correct transpose values while adding layouts for cuda gemm methods
f86d299
@sanjibansg
feat: gemm apis for alpaka views
7b2133c
@sanjibansg
feat: gemm functions for cublas using raw pointers
6f2a9e1
@sanjibansg
feat: strided batch computation for gemm in cublaslt
fa108fb
@sanjibansg
feat: support for gemm functions on AMD hardware using hipBLASLt and …
97c12bc 
…cpu blas methods from different libraries
@sanjibansg sanjibansg changed the title feat: Configurable leading dimensions for Gemm functions with cuBLASLt feat: Support for BLAS methods on AMD GPUs, CPU libraries, and improvements on CUDA APIs Aug 11, 2026
@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

1 similar comment
@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): GPU Unit Tests ❌ failed - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): GPU Unit Tests ❌ failed - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): triggered - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest mi300x

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (mi300x): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (mi300x): GPU Unit Tests ✅ passed - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): GPU Unit Tests ✅ passed - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runtest h100

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runtest (h100): GPU Unit Tests ✅ passed - view run

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runbenchmark h100-47gb

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runbenchmark (h100-47gb): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026 •
edited
Loading

Copy link
Copy Markdown

Benchmark results (mi300x)

main does not yet support the benchmark harness - showing PR results only.

── PR ──────────────────────────────────────────────────────────────
sofieBLAS benchmark  (warmup=5, iterations=20)

Backend        M=N=K      avg (ms)       GFLOP/s
------------------------------------------------
hip              256         0.008       4305.66
hip              512         0.007      41122.20
hip             1024         0.024      87740.31
hip             2048         0.179      96138.75
hip             4096         1.681      81782.96

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runbenchmark (h100-47gb): Benchmark ✅ completed - view results

@sanjibansg

sanjibansg commented Aug 17, 2026

Copy link
Copy Markdown
Member Author

/runbenchmark mi300x

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runbenchmark (mi300x): triggered - view run

@github-actions

github-actions Bot commented Aug 17, 2026

Copy link
Copy Markdown

/runbenchmark (mi300x): Benchmark ✅ completed - view results

@harz05
Support runtime GEMM dimensions for dynamic shapes (#11)
7570dbf 
* fix: support runtime GEMM dimensions for dynamic shapes

* feat: resolve GEMM algorithms per call site instead of per shape

* test: cover multi-size use of one instance and bounded algorithm cache

* feat: optional LRU limit on the algorithm cache

* style: clang-format

* chore: cleanup, cache limit via constructor

* docs: layout and caching behavior

* feat: dynamic gemm shapes and algorithm cache for hipBLASLt

* docs: describe layout roles, envelope matching and shape naming

* docs: explain addLayoutConfig, algo resolution and stats counters

* fix: cache layouts and algorithms, one shared BLASLt backend

* feat: addOperationConfig with a epilogue, tests for runtime sizes

* docs: describe addOperationConfig, the epilogue enum and the cache limit

* docs: address review comments

* docs: README changes
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Configurable Leading dimensions in Gemm functions for cublaslt

2 participants

@sanjibansg @harz05