IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

MatmulGPU

struct MatmulGPU

Registers the mo.matmul graph op for GPU devices.

Two execute overloads under one registration: the graph compiler calls the fused one when the output has a fused epilogue and the plain one when it has none. The fused form takes a _LegacyComputeOutput because kernels that write output tiles with a bulk copy, such as TMA on SM100, apply the epilogue's value transform before the write; when the epilogue must own the store instead, the matmul gets only the output's shape. The plain form has the matmul write the output directly. The pair is grandfathered on the graph compiler's allow list; new kernels declare one execute with a StoreAccess output.

Implemented traitsโ€‹

AnyType, Deinitable, Movable

Methodsโ€‹

executeโ€‹

static def execute[transpose_b: Bool, packed_b: Bool, _trace_name: StaticString](c: T, a: T, b: T, ctx: DeviceContext)

static def execute[transpose_b: Bool, packed_b: Bool, _trace_name: StaticString](c: T, a: T, b: T, ctx: DeviceContext) where not conforms_to(T, Fused).__bool__()

Was this page helpful?