[AMD] Pipeline small vector width but not thru shared memory

[sjw36]

Allow pipelining in register buffers for dot inputs with small vector width.

[chris002oak]

Ok

El jue, 21 de nov de 2024, 9:42 p. m., SJW @.***> escribió:

Allow pipelining in register buffers for dot inputs with small vector width.

You can view, comment on, or merge this pull request online at:

https://github.com/triton-lang/triton/pull/5227 Commit Summary

• b518e47 https://github.com/triton-lang/triton/pull/5227/commits/b518e4768799d786811190d650effeacb136b8c0 [AMD] Pipeline small vector width but not thru shared memory (even for dot input).

File Changes

(3 files https://github.com/triton-lang/triton/pull/5227/files)

• A test/TritonGPU/foo.mlir https://github.com/triton-lang/triton/pull/5227/files#diff-2d25e9984489f267b15d5eae0e6787b7051c92bddc8a3cda65ee78c186a09536 (57)
• M test/TritonGPU/loop-pipeline-hip.mlir https://github.com/triton-lang/triton/pull/5227/files#diff-1859c7f7dbc311a02b8fa31614bb124ac822496391eed1005aac5d207792d70c (41)
• M third_party/amd/lib/TritonAMDGPUTransforms/StreamPipelineV2.cpp https://github.com/triton-lang/triton/pull/5227/files#diff-49cdcce4a39ce246b69e23b15b5a2407227414209ad4b83ec57142b2383f3ed4 (14)

Patch Links:

• https://github.com/triton-lang/triton/pull/5227.patch
• https://github.com/triton-lang/triton/pull/5227.diff

— Reply to this email directly, view it on GitHub https://github.com/triton-lang/triton/pull/5227, or unsubscribe https://github.com/notifications/unsubscribe-auth/BKXFP3SNNPMAWJVHFQITTKD2B2RZPAVCNFSM6AAAAABSIN7QCKVHI2DSMVQWIX3LMV43ASLTON2WKOZSGY4DCNRWHAYTOMQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>