HiCR::backend::ascend::ComputationKernel Class Reference

#include <computationKernel.hpp>

Inheritance diagram for HiCR::backend::ascend::ComputationKernel:

Classes
struct	tensorData_t

Public Member Functions
	ComputationKernel (const char *kernelName, const std::vector< tensorData_t > &inputs, const std::vector< tensorData_t > &outputs, const aclopAttr *kernelAttrs)
	ComputationKernel (const char *kernelPath, const char *kernelName, const std::vector< tensorData_t > &inputs, const std::vector< tensorData_t > &outputs, const aclopAttr *kernelAttrs)
__INLINE__ void	start (const aclrtStream stream) override
Public Member Functions inherited from HiCR::backend::ascend::Kernel
__INLINE__	~Kernel ()=default

Static Public Member Functions
static tensorData_t	createTensorData (const std::shared_ptr< HiCR::LocalMemorySlot > &memorySlot, aclTensorDesc *tensorDescriptor)

Detailed Description

This class represents a replicable Computation Kernel for the Ascend backend. A Computation Kernel enables the kernel execution in the HiCR runtime, and in particular enables the concatenation of kernel execution and memcpy operations in a common stream of operations.

Constructor & Destructor Documentation

ComputationKernel() [1/2]

HiCR::backend::ascend::ComputationKernel::ComputationKernel ( const char *  kernelName, const std::vector< tensorData_t > &  inputs, const std::vector< tensorData_t > &  outputs, const aclopAttr *  kernelAttrs  )

inline

Constructor for the Computation Kernel unit class of the Ascend backend. This will not perform any model loading so this aspect should be handled manually (e.g., with aclopSetModelDir())

Parameters

kernelName	name of the kernel
inputs	kernel input tensor data descriptors
outputs	kernel output tensor data descriptors
kernelAttrs	kernel attributes

ComputationKernel() [2/2]

HiCR::backend::ascend::ComputationKernel::ComputationKernel ( const char *  kernelPath, const char *  kernelName, const std::vector< tensorData_t > &  inputs, const std::vector< tensorData_t > &  outputs, const aclopAttr *  kernelAttrs  )

inline

Constructor for the Computation Kernel unit class of the Ascend backend. This will load an operator binary file located at the provided path with aclopLoad()

Parameters

kernelPath	path the the kernel .om file
kernelName	name of the kernel
inputs	kernel input tensor data descriptors
outputs	kernel output tensor data descriptors
kernelAttrs	kernel attributes

Member Function Documentation

createTensorData()

static tensorData_t HiCR::backend::ascend::ComputationKernel::createTensorData ( const std::shared_ptr< HiCR::LocalMemorySlot > &  memorySlot, aclTensorDesc *  tensorDescriptor  )

inlinestatic

Creates the Ascend-specific Tensor data to be used as input/output parameter to Ascend kernels

Parameters

memorySlot	The memory slot to be used as input/output
tensorDescriptor	Ascend-specific metadata about the passed memory slot

Returns

The new Ascend tensor data object

start()

__INLINE__ void HiCR::backend::ascend::ComputationKernel::start ( const aclrtStream  stream )

inlineoverridevirtual

Start the kernel using the given ACL stream

Parameters

stream

the ACL stream on which the kernel is to be executed

Implements HiCR::backend::ascend::Kernel.

---

The documentation for this class was generated from the following file:

• /home/runner/work/HiCR/HiCR/include/hicr/backends/ascend/computationKernel.hpp