feat: Support for BLAS methods on AMD GPUs, CPU libraries, and improvements on CUDA APIs by sanjibansg · Pull Request #9 · ML4EP/sofieBLAS

sanjibansg · 2026-03-21T14:01:58Z

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

…nters

…cpu blas methods from different libraries

sanjibansg · 2026-08-17T10:09:22Z

/runtest h100

github-actions · 2026-08-17T12:19:07Z

/runtest (h100): triggered - view run

sanjibansg · 2026-08-17T12:19:12Z

/runtest mi300x

github-actions · 2026-08-17T12:19:26Z

/runtest (mi300x): triggered - view run

github-actions · 2026-08-17T12:21:14Z

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

sanjibansg · 2026-08-17T12:22:05Z

/runtest h100

github-actions · 2026-08-17T12:22:15Z

/runtest (h100): triggered - view run

github-actions · 2026-08-17T12:23:42Z

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

sanjibansg · 2026-08-17T12:33:28Z

/runtest h100

github-actions · 2026-08-17T12:33:37Z

/runtest (h100): triggered - view run

github-actions · 2026-08-17T12:35:14Z

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

sanjibansg · 2026-08-17T12:36:59Z

/runbenchmark h100-47gb

github-actions · 2026-08-17T12:37:08Z

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

github-actions · 2026-08-17T12:38:51Z

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 · 2026-08-17T12:38:53Z

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

sanjibansg · 2026-08-17T12:39:46Z

/runbenchmark mi300x

github-actions · 2026-08-17T12:39:56Z

/runbenchmark (mi300x): triggered - view run

github-actions · 2026-08-17T12:41:32Z

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

* 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

* test: deduplicate cuda and hip tests into a tag templated fixture test * build: detect gpu architectures and skip unavailable backends * tests: move cpu and gpu tests into unit_test.tpp files and share the dynamic shape test * tests: identify the gpu backend from the template tag instead of a name argument * ci: run the cpu tests against both OpenBLAS and MKL * ci: use the sequential MKL threading layer so libmkl_rt needs no OpenMP runtime

…unit tests for int8 methods

sanjibansg · 2026-10-06T13:28:26Z

/runtest l40s

github-actions · 2026-10-06T13:28:34Z

/runtest (l40s): triggered - view run

sanjibansg · 2026-10-06T13:28:35Z

/runtest w7900

github-actions · 2026-10-06T13:28:43Z

/runtest (w7900): triggered - view run

github-actions · 2026-10-06T13:29:48Z

/runtest (w7900): GPU Unit Tests ❌ failed - view run

github-actions · 2026-10-06T13:30:07Z

/runtest (l40s): GPU Unit Tests ✅ passed - view run

sanjibansg added 6 commits March 2, 2026 11:56

feat: configure leading dimensions explicitly

aaf2d08

fix: set stream while initializing handle

9aa4b88

feat: print requested workspace

2498197

fix: provide layout order during adding

901c5ae

fix: layout order configuration

8a27ffe

fix: remove look for requested workspace

54dce63

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 added 3 commits April 17, 2026 11:31

feat: matmul method for without bias matrix multiplication condition

c5128e6

feat: option to pass gpu pointers directly

9282ba9

fix: function signature for matmul method with void pointers

fc90737

sanjibansg force-pushed the dev branch from 5a62e6c to fc90737 Compare April 17, 2026 10:11

sanjibansg added 2 commits April 17, 2026 12:16

fix: template types for arguments to function with pointer signatures

b591c4a

fix: use float signatures for pointer arguments for blascuda

19b5b3f

sanjibansg force-pushed the dev branch from 5f0b1d6 to 19b5b3f Compare April 17, 2026 10:44

sanjibansg and others added 13 commits April 17, 2026 12:50

fix: use explicit data type for signatures with pointer arguments

c995464

fix: non transpose axis for matmul method

75f75a4

fix (experimental): layout shape order

ed303fb

feat: cuda cleanup and cpu blas api

54a9401

feat: matmul method on raw cuda pointers

066fdae

fix: remove extra template parameter in cuda matmul method on raw poi…

1f1c4c7

…nters

fix: method signature for matmul on cuda raw pointers

a867939

fix: dimension order for checkAddLayout method in cuda

ad41e99

fix: correct transpose values while adding layouts for cuda gemm methods

f86d299

feat: gemm apis for alpaka views

7b2133c

feat: gemm functions for cublas using raw pointers

6f2a9e1

feat: strided batch computation for gemm in cublaslt

fa108fb

feat: support for gemm functions on AMD hardware using hipBLASLt and …

97c12bc

…cpu blas methods from different libraries

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

feat: ci pipeline for tests and benchmarking

6c2b9a0

ci: avoid triggering cpu tests with gpu ones

38931d1

harz05 and others added 5 commits September 8, 2026 11:14

feat: support for Int8 matmul on hetergeneous architectures

c64ca73

fix: incorrect positioning of int8 methods, mkl impl of int8 matmul, …

a0357c9

…unit tests for int8 methods

feat: stride aware gemm calls for heterogeneous architectures

c3f1a2d

fix: restrict strided batched gemm to cuda only for now

1977fbb

Conversation

sanjibansg commented Mar 21, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Benchmark results (mi300x)

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

github-actions Bot commented Aug 17, 2026

Uh oh!

sanjibansg commented Oct 6, 2026

Uh oh!

github-actions Bot commented Oct 6, 2026

Uh oh!

sanjibansg commented Oct 6, 2026

Uh oh!

github-actions Bot commented Oct 6, 2026

Uh oh!

github-actions Bot commented Oct 6, 2026

Uh oh!

github-actions Bot commented Oct 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

github-actions Bot commented Aug 17, 2026 •

edited

Loading