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 trait

StorageIO

Extends TensorEngine with element loads and stores that bypass unsafe_ptr.

Element-granular access normally goes through the raw pointer unsafe_ptr returns. An engine whose storage has no such pointer, such as a TPU memref, conforms to this trait instead, and TileTensor routes every offset-based load and store through it: coordinate loads and stores, element indexing, raw_load and raw_store, and fills. The engine decides which widths it supports. TileIO is the whole-tile counterpart, keyed by call shape rather than by trait: load() with no coordinate is TileIO, t[i] is StorageIO.

Implemented traits​

AnyType, TensorEngine

Required methods​

load​

static def load[mut: Bool, dtype: DType, origin: Origin[mut=mut], address_space: AddressSpace, //, width: SIMDLength, alignment: Int, invariant: Bool = False, non_temporal: Bool = False](storage: Self.StorageType[dtype, origin, address_space], offset: Int) -> SIMD[dtype, width]

Loads width elements starting at linear index offset.

Parameters:

  • ​mut (Bool): The mutability of the storage, inferred from origin.
  • ​dtype (DType): The element data type of the storage.
  • ​origin (Origin[mut=mut]): The origin tracking the lifetime of the storage.
  • ​address_space (AddressSpace): The address space the storage resides in.
  • ​width (SIMDLength): The number of elements to load.
  • ​alignment (Int): The alignment of the access, in bytes.
  • ​invariant (Bool): Whether the storage is unchanged for the kernel's lifetime, so the load may be hoisted or cached.
  • ​non_temporal (Bool): Whether the data is not reused soon, so the access may bypass caches.

Args:

  • ​storage (Self.StorageType[dtype, origin, address_space]): The storage handle.
  • ​offset (Int): The linear element index, as the layout computes it.

Returns:

SIMD[dtype, width]: The width elements at offset.

store​

static def store[dtype: DType, origin: MutOrigin, address_space: AddressSpace, width: SIMDLength, //, alignment: Int, non_temporal: Bool = False](storage: Self.StorageType[dtype, origin, address_space], offset: Int, value: SIMD[dtype, width])

Stores value starting at linear index offset.

Parameters:

  • ​dtype (DType): The element data type of the storage.
  • ​origin (MutOrigin): The origin tracking the lifetime of the storage.
  • ​address_space (AddressSpace): The address space the storage resides in.
  • ​width (SIMDLength): The number of elements to store, inferred from value.
  • ​alignment (Int): The alignment of the access, in bytes.
  • ​non_temporal (Bool): Whether the data is not reused soon, so the access may bypass caches.

Args:

  • ​storage (Self.StorageType[dtype, origin, address_space]): The storage handle, modified in place.
  • ​offset (Int): The linear element index, as the layout computes it.
  • ​value (SIMD[dtype, width]): The elements to store.

Was this page helpful?