namespace

`SushiAI`

Declared in: include/SushiAI/autograd/backward.hpp

Contains

Typedefs

using SushiAI::DType = SushiBLAS::Core::DataType

The element type of a tensor.

An alias, not a parallel enum: SushiAI performs no arithmetic of its own, so the set of types it can express is exactly the set SushiBLAS can compute on.

using SushiAI::Layout = SushiBLAS::Core::Layout

The memory layout of a tensor, aliased from SushiBLAS for the same reason.

using SushiAI::OpID = SushiRuntime::Graph::OpID

The runtime's operation-identity type, aliased for brevity.

Variables

constexpr int32_t MAX_RANK = static_cast<int32_t>(SushiBLAS::Core::MAX_TENSOR_RANK)

Maximum tensor rank, fixed by SushiBLAS's own storage of shape and strides.

Functions

SushiBLAS::Tensor allocate_tensor(SushiBLAS::Engine &engine, SushiRuntime::span< const int64_t > dims, DType dtype=DType::FLOAT32)

Allocates a tensor of dims with its own Storage.

Parameters

engine: The engine whose context owns the allocator.
dims: The extents, outermost first.
dtype: The element type.

Returns

The new tensor with standalone storage.

Exceptions

Error: If an extent is invalid or allocation fails.

SushiBLAS::Tensor allocate_tensor(SushiBLAS::Engine &engine, const Shape &shape, DType dtype=DType::FLOAT32)

Allocates a tensor of shape with its own Storage.

Parameters

engine: The engine whose context owns the allocator.
shape: The extents.
dtype: The element type.

Returns

The new tensor; nothing else references its Storage.

Exceptions

Error: As the span overload does.

SushiBLAS::Tensor allocate_like(SushiBLAS::Engine &engine, const SushiBLAS::Tensor &like)

Allocates a tensor with like's extents, dtype, and layout.

Parameters

engine: The engine whose context owns the allocator.
like: The tensor to copy the description from.

Returns

The new tensor; it shares nothing with like.

Exceptions

Error: If like holds no elements, or the allocation fails.

SushiRuntime::sushi_ptr< SushiBLAS::Storage > allocate_arena(SushiBLAS::Engine &engine, std::size_t bytes)

Allocates one arena buffer for memory planning.

Parameters

engine: The engine whose context owns the allocator.
bytes: Number of bytes the arena holds; must be positive.

Returns

The newly allocated storage.

Exceptions

Error: If bytes is zero or allocation fails.

SushiBLAS::Tensor view_tensor(SushiBLAS::Engine &engine, const SushiRuntime::sushi_ptr< SushiBLAS::Storage > &arena, std::size_t byte_offset, const Shape &shape, DType dtype)

Returns a tensor view into arena at byte_offset.

Parameters

engine: The engine whose default layout the view takes.
arena: Storage to carve from; must outlive the view.
byte_offset: Offset in bytes; must be aligned to element width.

Exceptions

Error: If offset is misaligned or out of bounds.

constexpr std::size_t dtype_size(DType dtype) noexcept

Returns the size in bytes occupied by one element of dtype.

Parameters

dtype: The element type to size.

Returns

The element's size in bytes; never zero.

constexpr const char * dtype_name(DType dtype) noexcept

A human-readable name for dtype, for diagnostics and error text.

Parameters

dtype: The element type to name.

Returns

A null-terminated literal owned by the library; never null.

constexpr bool is_real_float(DType dtype) noexcept

Reports whether dtype is a real floating-point type.

Parameters

dtype: The element type to test.

Returns

True for HALF, FLOAT32, and FLOAT64; false otherwise.

const char * version_string() noexcept

Returns the library version string.

Returns

A null-terminated literal owned by the library; never null.