namespace

`SushiBLAS::Internal`

Declared in: include/SushiBLAS/engine/blas/task_metadata.hpp

Contains

Functions

void get_vec_params(const Tensor &t, int64_t &n, int64_t &inc, const char *op, const char *operand)

Reads the element count and the positive increment of a vector operand.

Parameters

n: Receives the number of elements.
inc: Receives the distance between consecutive elements.

Exceptions

std::runtime_error: on a non-contiguous tensor of rank > 1.
std::invalid_argument: on missing storage or a rank-1 stride of zero or less.

void require_level2_matrix_rank(const Tensor &A, const char *op)

Throws unless a Level 2 matrix operand has rank 2 or less.

Exceptions

std::invalid_argument: on a rank above 2.

constexpr Core::DataType real_result_dtype(Core::DataType dtype)

Returns the dtype a norm or absolute-sum result holds for an operand of dtype.

Returns

The real component type of a complex dtype, dtype itself otherwise.

void require_scalar_result(const Tensor &result, Core::DataType expected, const char *op)

Throws unless a scalar result tensor has storage, the expected dtype and exactly one element.

Exceptions

std::invalid_argument: naming the first rule that fails.

ScratchBuffer allocate_level1_scratch(TaskRecorder &recorder, const Tensor &x, std::size_t elements, std::size_t partial_bytes, OperandSet &operands)

Leases the partials scratch a Level 1 reduction over x plans on x's device, and declares it in operands.

Parameters

elements: Values the reduction folds: n, or 2n for a complex nrm2.
partial_bytes: Bytes one partial takes.

Returns

The scratch, empty when x has no storage or nothing is planned.

template <Dispatch::Domain Dom = Dispatch::Domain::REAL_AND_COMPLEX, typename Func>
TaskHandle execute_blas_task(TaskRecorder &recorder, const SushiRuntime::Graph::TaskMetadata &meta, const OperandSet &operands, Core::DataType dtype, Func &&op_func)

Records one dtype-dispatched BLAS task whose body runs a reference and SYCL kernel pair.

Parameters

op_func: Callable taking a scalar of the element type, a const KernelPairDispatcher&, the queue and the dependencies.

Exceptions

At: record time as Dispatch::require_domain states, if dtype falls outside Dom.
std::invalid_argument: at record time, as TaskRecorder::record_pair states.

See also

include/SushiBLAS/engine/blas/README.md

template <Dispatch::Domain Dom = Dispatch::Domain::REAL_AND_COMPLEX, typename Func>
TaskHandle execute_typed_task(TaskRecorder &recorder, const SushiRuntime::Graph::TaskMetadata &meta, const OperandSet &operands, Core::DataType dtype, const char *family, Func &&op_func)

Records one dtype-dispatched task whose body runs a reference and SYCL kernel pair.

Parameters

family: What kind of operation this is, for diagnostics; a string literal.
op_func: Callable taking a scalar of the element type, a const KernelPairDispatcher&, the queue and the dependencies.

Exceptions

At: record time as Dispatch::require_domain states, if dtype falls outside Dom.
std::invalid_argument: at record time, as TaskRecorder::record_pair states.