Intel® oneAPI DPC++/C++ Compiler Developer Guide and Reference

ID 767253
Date 11/07/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

access_by

Enum to control how the memory layout will be accessed. #include <sdlt/access_by.h>

Syntax

enum access_by
{
    access_by_struct,
    access_by_stride
};

Description

The access_by_struct causes data access via structure member access. Nested structures will drill down through the structure members in a nested manner. For example an Axis Aligned Bounding Box (AABB) containing two Point3d objects (with x,y,z data members) will logically expand to something like:

AABB local;
local = accessor.mData[i];

access_by_stride will cause data access through pointers to built in types with a stride to account for the size of the primitive. For an Axis Aligned Bounding Box (AABB) containing two Point3d objects (with x,y,z data members) will logically expand to something like:

AABB local;
local.topLeft.x = *(accessor.mData + offsetof(AABB,topLeft) + offset(Point3d,x) + (sizeof(AABB)*i));
local.topLeft.y = *(accessor.mData + offsetof(AABB,topLeft) + offset(Point3d,y) + (sizeof(AABB)*i));
local.topLeft.z = *(accessor.mData + offsetof(AABB,topLeft) + offset(Point3d,z) + (sizeof(AABB)*i));
local.topRight.x = *(accessor.mData + offsetof(AABB,topRight) + offset(Point3d,x) + (sizeof(AABB)*i));
local.topRight.y = *(accessor.mData + offsetof(AABB,topRight) + offset(Point3d,y) + (sizeof(AABB)*i));
local.topRight.z = *(accessor.mData + offsetof(AABB,topRight) + offset(Point3d,z) + (sizeof(AABB)*i));

When vectorizing, access_by_struct can sometimes generate better code as the compiler could perform wide loads and use shuffle/insert instructions to move data into SIMD registers. However, depending on the complexity of the primitive, it can also fail to vectorize, especially when the primitive contains nested structures.

On the other hand access_by_stride has always vectorized successfully, because the data access is simplified to an array pointer with a stride. The compiler is able to handle any complexity of primitive, because it never sees the complexity and instead just sees the simple array pointer with strided access.

access_by_struct is probably the best choice as it offers a chance of better code generation especially when used outside of a SIMD loop. However if you run into issues when vectorizing, try access_by_stride to see if that alleviates the problem.

We leave this choice up to the developer and require they explicitly make a choice, so this is not hidden behavior.