SYCL^™ 2020 Specification (revision 7)

The SYCL generic programming model exposes a number of platforms, each of them exposing a number of devices. Each platform is bound to a certain SYCL backend. SYCL devices associated with said platform are associated with that SYCL backend.

Although the APIs in the SYCL generic programming model are defined according to this specification and their version is indicated by the macro SYCL_LANGUAGE_VERSION, this does not apply to APIs exposed by the SYCL backends. Each SYCL backend provides its own document that defines its APIs, and that document tells how to query for the device and platform versions.

The SYCL application defines the execution order of the kernels by grouping each kernel with its requirements into a command group function object. Command group function objects are submitted for execution via a queue object, which defines the device where the kernel will run. This specification sometimes refers to this as “submitting the kernel to a device”. The same command group object can be submitted to different queues. When a command group is submitted to a SYCL queue, the requirements of the kernel execution are captured. The implementation can start executing a kernel as soon as its requirements have been satisfied.

When a kernel is submitted for execution, an index space is defined. An instance of the kernel body executes for each point in this index space. This kernel instance is called a work-item and is identified by its point in the index space, which provides a global id for the work-item. Each work-item executes the same code but the specific execution pathway through the code and the data operated upon can vary by using the work-item global id to specialize the computation.

An index space of size zero is allowed. All aspects of kernel execution proceed as normal with the exception that the kernel function itself is not executed. Note this means the command queue will still schedule this kernel after satisfying the requirements and this satisfies requirements of any dependent enqueued kernels.

The application running on the host uses SYCL buffer objects using instances of the sycl::buffer class or USM allocation functions to allocate memory in the global address space, or can allocate specialized image memory using the sycl::unsampled_image and sycl::sampled_image classes.

In the SYCL application, memory objects are bound to all devices in which they are used, regardless of the SYCL context where they reside. SYCL memory objects (namely, buffer and image objects) can encapsulate multiple underlying SYCL backend memory objects together with multiple host memory allocations to enable the same object to be shared between devices in different contexts, platforms or backends. USM allocations uniquely identify a memory allocation and are bound to a SYCL context. They are only valid on the backend used by the context.

The order of execution of command group objects ensures a sequentially consistent access to the memory from the different devices to the memory objects. Accessing a USM allocation does not alter the order of execution. Users must explicitly inform the SYCL runtime of any requirements necessary for a legal execution.

To access a memory object, the user must create an accessor object which parameterizes the type of access to the memory object that a kernel or the host requires. The accessor object defines a requirement to access a memory object, and this requirement is defined by construction of an accessor, regardless of whether there are any uses in a kernel or by the host. An accessor object specifies whether the access is via global memory, constant memory or image samplers and their associated access functions. The accessor also specifies whether the access is read-only (RO), write-only (WO) or read-write (RW). An optional no_init property can be added to an accessor to tell the system to discard any previous contents of the data the accessor refers to, so there are two additional requirement types: no-init-write-only (NWO) and no-init-read-write (NRW). For simplicity, when a requisite represents an accessor object in a certain access mode, we represent it as MemoryObject_AccessMode. For example, an accessor that accesses memory object buf1 in RW mode is represented as buf1_RW. A command group object that uses such an accessor is represented as CG(buf1_RW). The action required to satisfy a requisite and the location of the latest copy of a memory object will vary depending on the implementation.

Table 3 illustrates an example where command group objects are enqueued to two separate SYCL queues executing in devices in different contexts. The requisites for the command group execution are the same, but the actions to satisfy them are different. For example, if the data is on the host before execution, A(b1_RW) and A(b2_RW) can potentially be implemented as copy operations from the host memory to context1 or context2 respectively. After CG_a and CG_b are executed, A'(b1_RW) will likely be an empty operation, since the result of the kernel can stay on the device. On the other hand, the results of CG_b are now on a different context than CG_c is executing, therefore A'(b2_RW) will need to copy data across two separate contexts using an implementation specific mechanism.

Table 3 shows actions performed when three command groups are submitted to two distinct queues, and potential implementation in an OpenCL SYCL backend by a SYCL runtime. Note that in this example, each SYCL buffer (b2,b2) is implemented as separate cl_mem objects per context.

Note that the order of the definition of the accessors within the command group is irrelevant to the requirements they define. All accessors always apply to the entire command group object where they are defined.

When multiple accessors in the same command group define different requisites to the same memory object these requisites must be resolved.

Firstly, any requisites with different access modes but the same access target are resolved into a single requisite with the union of the different access modes according to Table 4. The atomic access mode acts as if it was read-write (RW) when determining the combined requirement. The rules in Table 4 are commutative and associative.

The result of this should be that there should not be any requisites with the same access target.

Secondly, the remaining requisites must adhere to the following rule. Only one of the requisites may have write access (W or RW), otherwise the SYCL runtime must throw an exception. All requisites create a requirement for the data they represent to be made available in the specified access target, however only the requisite with write access determines the side effects of the command group, i.e. only the data which that requisite represents will be updated.

For example:

Where G and C correspond to a target::device and target::constant_buffer accessor and H corresponds to a host accessor.

A buffer created from a range of an existing buffer is called a sub-buffer. A buffer may be overlaid with any number of sub-buffers. Accessors can be created to operate on these sub-buffers. Refer to Section 4.7.2 for details on sub-buffer creation and restrictions. A requirement to access a sub-buffer is represented by specifying its range, e.g. CG(b1_RW,[0,5)) represents the requirement of accessing the range [0,5) buffer b1 in read write mode.

If two accessors are constructed to access the same buffer, but both are to non-overlapping sub-buffers of the buffer, then the two accessors are said to not overlap, otherwise the accessors do overlap. Overlapping is the test that is used to determine the scheduling order of command groups. Command-groups with non-overlapping requirements may execute concurrently.

It is permissible for command groups that only read data to not copy that data back to the host or other devices after reading and for the runtime to maintain multiple read-only copies of the data on multiple devices.

A special case of requirement is the one defined by a host accessor. Host accessors are represented with H(MemoryObject_AccessMode), e.g, H(b1_RW) represents a host accessor to b1 in read-write mode. Host accessors are a special type of accessor constructed from a memory object outside a command group, and require that the data associated with the given memory object is available on the host in the given pointer. This causes the runtime to block on construction of this object until the requirement has been satisfied. Host accessor objects are effectively barriers on all accesses to a certain memory object. Table 6 shows an example of multiple command groups enqueued to the same queue. Once the host accessor H(b1_RW) is reached, the execution cannot proceed until CG_a is finished. However, CG_b does not have any requirements on b1, therefore, it can execute concurrently with the barrier. Finally, CG_c will be enqueued after H(b1_RW) is finished, but still has to wait for CG_b to conclude for all its requirements to be satisfied. See Section 3.9.8 for details on synchronization rules.

The memory model for SYCL devices is based on the OpenCL 1.2 memory model. Work-items executing in a kernel have access to three distinct address spaces (memory regions) and a virtual address space overlapping some concrete address spaces:

The SYCL memory consistency model is based upon the memory consistency model of the C++ core language. Where SYCL offers extensions to classes and functions that may affect memory consistency, the default behavior when these extensions are not used always matches the behavior of standard C++.

A SYCL implementation must guarantee that the same memory consistency model is used across host and device code. Every device compiler must support the memory model defined by the minimum version of C++ described in Section 3.9.1; SYCL implementations supporting additional versions of C++ must also support the corresponding memory models.

Within a work-item, operations are ordered according to the sequenced before relation defined by the C++ core language.

Ensuring memory consistency across different work-items requires careful usage of group barrier operations, mem-fence operations and atomic operations. The ordering of operations across different work-items is determined by the happens before relation defined by the C++ core language, with a single relation governing all address spaces (memory regions).

On any SYCL device, local and global memory may be made consistent across work-items in a single group through use of a group barrier operation. On SYCL devices supporting acquire-release or sequentially consistent memory orderings, all memory visible to a set of work-items may be made consistent across the work-items in that set through the use of mem-fence and atomic operations.

Memory consistency between the host and SYCL device(s), or different SYCL devices in the same context, can be guaranteed through synchronization in the host application as defined in Section 3.9.8. On SYCL devices supporting concurrent atomic accesses to USM allocations and acquire-release or sequentially consistent memory orderings, cross-device memory consistency can be enforced through the use of mem-fence and atomic operations.

The C++ features used in SYCL are based on a specific version of C++. Implementations of SYCL must support this minimum C++ version, which defines the C++ constructs that can consequently be used by SYCL feature definitions (for example, lambdas).

The minimum C++ version of this SYCL specification is determined by the normative C++ core language defined in Section 3.3. All implementations of this specification must support at least this core language, and features within this specification are defined using features of the core language. Note that not all core language constructs are supported within SYCL kernel functions or code invoked by a SYCL kernel function, as detailed by Section 5.4.

Implementations may support newer C++ versions than the minimum required by SYCL. Code written using newer features than the SYCL requirement, though, may not be portable to other implementations that don’t support the same C++ version.

Some features of SYCL are aligned with the next C++ specification, as defined in Section 3.3.

The following features are pre-adopted by SYCL 2020 and made available in the sycl:: namespace: std::span, std::dynamic_extent, std::bit_cast. The implementations of pre-adopted features are compliant with the next C++ specification, and are expected to forward directly to standard C++ features in a future version of SYCL.

The following features of SYCL 2020 use syntax based on the next C++ specification: sycl::atomic_ref. These features behave as described in the next C++ specification, barring modifications to ensure compatibility with other SYCL 2020 features and heterogeneous programming. Any such modifications are documented in the corresponding sections of this specification.

Data-parallel kernels that execute as multiple work-items and where no local synchronization is required are enqueued with the sycl::parallel_for function parameterized by a sycl::range parameter. These kernels will execute the kernel function body once for each work-item in the specified range.

Functionality tied to groups of work-items, including group barriers and local memory, must not be used within these kernels.

Variables with reduction semantics can be added to basic data parallel kernels using the features described in Section 4.9.2.

Data parallel kernels can also execute in a mode where the set of work-items is divided into work-groups of user-defined dimensions. The user specifies the global range and local work-group size as parameters to the sycl::parallel_for function with a sycl::nd_range parameter. In this mode of execution, kernels execute over the nd-range in work-groups of the specified size. It is possible to share data among work-items within the same work-group in local or global memory and to synchronize between work-items in the same work-group by calling the group_barrier function. All work-groups in a given parallel_for will be the same size, and the global size defined in the nd-range must either be a multiple of the work-group size in each dimension, or the global size must be zero. When the global size is zero, the kernel function is not executed, the local size is ignored, and any dependencies are satisfied.

Work-groups may be further subdivided into sub-groups. The work-items that compose a sub-group are selected in an implementation-defined way, and therefore the size and number of sub-groups may differ for each kernel. Moreover, different devices may make different guarantees with respect to how sub-groups within a work-group are scheduled. The maximum number of work-items in any sub-group in a kernel is based on a combination of the kernel and its dispatch dimensions. The size of any sub-group in the dispatch is between 1 and this maximum sub-group size, and the size of an individual sub-group is invariant for the duration of a kernel’s execution. Similarly to work-groups, the work-items within the same sub-group can be synchronized by calling the group_barrier function.

Portable device code must not assume that work-items within a sub-group execute in any particular order, that work-groups are subdivided into sub-groups in a specific way, nor that the work-items within a sub-group provide specific forward progress guarantees.

Variables with reduction semantics can be added to work-group data parallel kernels using the features described in Section 4.9.2.

The SYCL compiler provides a way of specifying data parallel kernels that execute within work-groups via a different syntax which highlights the hierarchical nature of the parallelism. This mode is purely a compiler feature and does not change the execution model of the kernel. Instead of calling sycl::parallel_for the user calls sycl::parallel_for_work_group with a sycl::range value representing the number of work-groups to launch and optionally a second sycl::range representing the size of each work-group for performance tuning. All code within the parallel_for_work_group scope effectively executes once per work-group. Within the parallel_for_work_group scope, it is possible to call parallel_for_work_item which creates a new scope in which all work-items within the current work-group execute. This enables a programmer to write code that looks like there is an inner work-item loop inside an outer work-group loop, which closely matches the effect of the execution model. All variables declared inside the parallel_for_work_group scope are allocated in work-group local memory, whereas all variables declared inside the parallel_for_work_item scope are declared in private memory. All parallel_for_work_item calls within a given parallel_for_work_group execution must have the same dimensions.

Simple kernels for which only a single instance of the kernel function will be executed are enqueued with the sycl::single_task function. The kernel enqueued takes no “work-item id” parameter and will only execute once. The behavior is logically equivalent to executing a kernel on a single compute unit with a single work-group comprising only one work-item. Such kernels may be enqueued on multiple queues and devices and as a result may be executed in task-parallel fashion.

Some SYCL backends may expose pre-defined functionality to users as kernels. These kernels are not programmable, hence they are not bound by the SYCL C++ programming model restrictions, and how they are written is implementation-defined.

Synchronization of processing elements executing inside a device is handled by the SYCL device kernel following the SYCL kernel execution model. The synchronization of the different SYCL device kernels executing with the host memory is handled by the SYCL application via the SYCL runtime.

In SYCL, there are two types of errors: synchronous errors that can be detected immediately when an API call is made, and asynchronous errors that can only be detected later after an API call has returned. Synchronous errors, such as failure to construct an object, are reported immediately by the runtime throwing an exception. Asynchronous errors, such as an error occurring during execution of a kernel on a device, are reported via an asynchronous error-handler mechanism.

Asynchronous errors are not reported immediately as they occur. The asynchronous error handler for a context or queue is called with a sycl::exception_list object, which contains a list of asynchronously-generated exception objects, on the conditions described by Section 4.13.1.1 and Section 4.13.1.2.

Asynchronous errors may be generated regardless of whether the user has specified any asynchronous error handler(s), as described in Section 4.13.1.2.

Some SYCL backends can report errors that are specific to the platform they are targeting, or that are more concrete than the errors provided by the SYCL API. Any error reported by a SYCL backend must derive from the base sycl::exception. When a user wishes to capture specifically an error thrown by a SYCL backend, she must include the SYCL backend-specific headers for said SYCL backend.

A command group function object can be submitted either to a single queue to be executed on, or to a secondary queue. If a command group function object fails to be enqueued to the primary queue, then the system will attempt to enqueue it to the secondary queue, if given as a parameter to the submit function. If the command group function object fails to be queued to both of these queues, then a synchronous SYCL exception will be thrown.

It is possible that a command group may be successfully enqueued, but then asynchronously fail to run, for some reason. In this case, it may be possible for the runtime system to execute the command group function object on the secondary queue, instead of the primary queue. The situations where a SYCL runtime may be able to achieve this asynchronous fall-back is implementation-defined.

A command group function object takes a reference to a command group handler as a parameter and anything within that scope is immediately executed and takes the handler object as a parameter. The intention is that a user will perform calls to SYCL functions, member functions, destructors and constructors inside that scope. These calls will be non-blocking on the host, but enqueue operations to the queue that the command group is submitted to. All user functions within the command group scope will be called on the host as the command group function object is executed, but any commands it invokes will be added to the SYCL queue. All commands added to the queue will be executed out-of-order from each other, according to their data dependencies.

A SYCL application does not initialize any SYCL backend features until a sycl::context object is created. A user does not need to explicitly create a sycl::context object, but they do need to explicitly create a sycl::queue object, for which a sycl::context object will be implicitly created if not provided by the user.

All SYCL backend objects encapsulated in SYCL objects are reference-counted and will be destroyed once all references have been released. This means that a user needs only create a SYCL queue (which will automatically create an SYCL context) for the lifetime of their application to initialize and release any SYCL backend objects safely.

There is no global state specified to be required in SYCL implementations. This means, for example, that if the user creates two queues without explicitly constructing a common context, then a SYCL implementation does not have to create a shared context for the two queues. Implementations are free to share or cache state globally for performance, but it is not required.

Memory objects can be constructed with or without attached host memory. If no host memory is attached at the point of construction, then destruction of that memory object is non-blocking. The user may use C++ standard pointer classes for sharing the host data with the user application and for defining blocking, or non-blocking behavior of the buffers and images. If host memory is attached by using a raw pointer, then the default behavior is followed, which is that the destructor will block until any command groups operating on the memory object have completed, then, if the contents of the memory object is modified on a device those contents are copied back to host and only then does the destructor return.

In the case where host memory is shared between the user application and the SYCL runtime with a std::shared_ptr, then the reference counter of the std::shared_ptr determines whether the buffer needs to copy data back on destruction, and in that case the blocking or non-blocking behavior depends on the user application.

Instead of a std::shared_ptr, a std::unique_ptr may be provided, which uses move semantics for initializing and using the associated host memory. In this case, the behavior of the buffer in relation to the user application will be non-blocking on destruction.

As said in Section 3.9.8, the only blocking operations in SYCL (apart from explicit wait operations) are:

A user specifies which queue to submit a command group function object and each queue is targeted to run on a specific device (and context). A user can specify the actual device on queue creation, or they can specify a device selector which causes the SYCL runtime to choose a device based on the user’s provided preferences. Specifying a device selector causes the SYCL runtime to perform device discovery. No device discovery is performed until a SYCL device selector is passed to a queue constructor. Device topology may be cached by the SYCL runtime, but this is not required.

Device discovery will return all devices from all platforms exposed by all the supported SYCL backends.

There are two styles of developing a SYCL application:

When users follow 1., there is no assumption about what SYCL backend will be used during compilation or execution of the SYCL application. Therefore, the SYCL backend API is not assumed to be available to the developer. Only standard C++ types and interfaces are assumed to be available, as described in Section 3.9. Users only need to include the <sycl/sycl.hpp> header to write a SYCL generic application.

On the other hand, when users follow 2., they must know what SYCL backend APIs they are using. In this case, any header required for the normal programmability of the SYCL backend API is assumed to be available to the user. In addition to the <sycl/sycl.hpp> header, users must also include the SYCL backend-specific header as defined in Section 4.3. The SYCL backend-specific header provides the interoperability interface for the SYCL API to interact with native backend objects.

The interoperability API is defined in Section 4.5.1.

Some multi-dimensional objects can be viewed in a linear form. When this happens, the right-most term in the object’s range varies fastest in the linearization.

A three-dimensional element id{id0, id1, id2} within a three-dimensional object of range range{r0, r1, r2} has a linear position defined by:

A two-dimensional element id{id0, id1} within a two-dimensional range{r0, r1} follows a similar equation:

A one-dimensional element id{id0} within a one-dimensional range range{r0} is equivalent to its linear form.

Some multi-dimensional objects can be indexed using the subscript operator where consecutive subscript operators correspond to each dimension. The right-most operator varies fastest, as with standard C++ arrays. Formally, a three-dimensional subscript access a[id0][id1][id2] references the element at id{id0, id1, id2}. A two-dimensional subscript access a[id0][id1] references the element at id{id0, id1}. A one-dimensional subscript access a[id0] references the element at id{id0}.

With this technique, known as SMCP, there are separate host and device compilers. Each SYCL source file is compiled two times: once by the host compiler and once by the device compiler. An implementation could support more than one device compiler, in which case each SYCL source file is compiled more than two times. The host compiler in this technique could be an off-the-shelf compiler with no special knowledge of SYCL, but the device compiler must be SYCL aware. The device compiler parses the source file to identify each SYCL kernel function and any device functions it calls. SYCL is designed so that this analysis can be done statically. The device compiler then generates code only for the SYCL kernel functions and the device functions.

Typically, the device compilers generate header files which interface between the host compiler and the SYCL runtime. Therefore, the device compiler runs first, and then the host compiler consumes these header files when generating the host code.

The device compilers in this technique generate one or more device images for the SYCL kernel functions, which can be read by the SYCL runtime. Each device image could either contain native ISA for a device or it could contain an intermediate language such as SPIR-V. In the later case, the SYCL runtime must translate the intermediate language into native device ISA when the SYCL kernel function is submitted to a device.

Since this technique has separate host and device compilers, there needs to be some way to associate a SYCL kernel function (which is compiled by the device compiler) with the code that invokes it (which is compiled by the host compiler). Implementations conformant to the reduced feature set (Section B.2) can do this by using the C++ type of the SYCL kernel function. This type is specified via the kernel name template parameter if the SYCL kernel function is a lambda function, or it is obtained from the class type if the SYCL kernel function is an object. Implementations conformant to the full feature set (Section B.1) do not require a kernel name at the invocation site, so they must implement some other way to make the association.

With this technique, known as SSCP, the vendor implements a custom compiler that reads each SYCL source file only once, and that compiler generates the host code as well as the device images for the SYCL kernel functions. As in the SMCP case, each device image could either contain native device ISA or an intermediate language.

It is also possible to implement SYCL purely as a library, using an off-the-shelf host compiler with no special support for SYCL. In such an implementation, each kernel may run on the host system.

The SYCL implementation may need to copy data between the host and a device or between two devices. For example, this may occur when a command group has a requirement for the contents of a buffer or when the application passes certain arguments to a SYCL kernel function (as described in Section 4.12.4). Such data must have a type that is device copyable as defined below.

Any type that is trivially copyable (as defined by the C++ core language) is implicitly device copyable.

Although implementations are not required to support device code that calls library functions from the C++ core language, some implementations may provide device support for some of these functions. If the implementation provides device support for one of the following classes, that type is also implicitly device copyable:

If the implementation provides device support for one of the classes listed above, arrays of that class and cv-qualified versions of that class are also device copyable.

In addition, the implementation may allow the application to explicitly declare certain class types as device copyable. If the implementation has this support, it must predefine the preprocessor macro SYCL_DEVICE_COPYABLE to 1, and it must not predefine this preprocessor macro if it does not have this support. When the implementation has this support, a class type T is device copyable if all of the following statements are true:

When the application explicitly declares a class type to be device copyable, arrays of that type and cv-qualified versions of that type are also device copyable, and the implementation sets the is_device_copyable_v trait to true for these array and cv-qualified types.

As the identifiers defined in enum class backend are implementation-defined, and the associated backends not guaranteed to be available, a SYCL implementation must also define a preprocessor macro for each of these identifiers. If the SYCL backend is defined by the Khronos SYCL group, the name of the macro has the form SYCL_BACKEND_<backend_name>, where backend_name is the associated identifier from backend in all upper-case. See Chapter 6 for the name of the macro if the vendor defines the SYCL backend outside of the Khronos SYCL group.

If a backend listed in the enum class backend is not available, the associated macro must be left undefined.

Many of the SYCL runtime classes may be implemented such that they encapsulate an object unique to the SYCL backend that underpins the functionality of that class. Where appropriate, these classes may provide an interface for interoperating between the SYCL runtime object and the native backend object in order to support interoperability within an application between SYCL and the associated SYCL backend API.

There are three forms of interoperability with SYCL runtime classes: interoperability on the SYCL application with the SYCL backend API, interoperability within a SYCL kernel function with the equivalent kernel language types of the SYCL backend, and interoperability within a host task with the interop_handle.

SYCL application interoperability, SYCL kernel function interoperability and host task interoperability are provided via different interfaces and may have different behavior for the same SYCL object.

SYCL application interoperability may be provided for buffer, context, device, device_image, event, kernel, kernel_bundle, platform, queue, sampled_image, and unsampled_image.

SYCL kernel function interoperability may be provided for accessor, device_event, local_accessor, sampled_image_accessor, stream and unsampled_image_accessor inside kernel scope only and is not available outside of that scope.

host task interoperability may be provided for accessor, sampled_image_accessor, unsampled_image_accessor, queue, device, context inside the scope of a host task only, see Section 4.10.

Support for SYCL backend interoperability is optional and therefore not required to be provided by a SYCL implementation. A SYCL application using SYCL backend interoperability is considered to be non-generic SYCL.

Details on the interoperability for a given SYCL backend are available on the SYCL backend specification document for that SYCL backend.

Each of the following SYCL runtime classes: accessor, buffer, context, device, device_image, event, host_accessor, host_sampled_image_accessor, host_unsampled_image_accessor, kernel, kernel_id, kernel_bundle, local_accessor, platform, queue, sampled_image, sampled_image_accessor, stream, unsampled_image and unsampled_image_accessor must obey the following statements, where T is the runtime class type:

Some SYCL runtime classes will have additional behavior associated with copy, movement, assignment or destruction semantics. If these are specified they are in addition to those specified above unless stated otherwise.

Each of the runtime classes mentioned above must provide a common interface of special member functions in order to fulfill the copy, move, destruction requirements and hidden friend functions in order to fulfill the equality requirements.

A hidden friend function is a function first declared via a friend declaration with no additional out of class or namespace scope declarations. Hidden friend functions are only visible to ADL (Argument Dependent Lookup) and are hidden from qualified and unqualified lookup. Hidden friend functions have the benefits of avoiding accidental implicit conversions and faster compilation.

These common special member functions and hidden friend functions are described in Table 7 and Table 8 respectively.

Each of the following SYCL runtime classes: id, range, item, nd_item, h_item, group, sub_group and nd_range must follow the following statements, where T is the runtime class type:

Some SYCL runtime classes will have additional behavior associated with copy, movement, assignment or destruction semantics. If these are specified they are in addition to those specified above unless stated otherwise.

Each of the runtime classes mentioned above must provide a common interface of special member functions and member functions in order to fulfill the copy, move, destruction and equality requirements, following the rule of five and the rule of zero.

These common special member functions and hidden friend functions are described in Table 9 and Table 10 respectively.

Each of the following SYCL runtime classes: accessor, buffer, host_accessor, host_sampled_image_accessor, host_unsampled_image_accessor, context, local_accessor, queue, sampled_image, sampled_image_accessor, stream, unsampled_image, unsampled_image_accessor and usm_allocator provide an optional parameter in each of their constructors to provide a property_list which contains zero or more properties. Each of those properties augments the semantics of the class with a particular feature. Each of those classes must also provide has_property and get_property member functions for querying for a particular property.

The listing below illustrates the usage of various buffer properties, described in Section 4.7.2.2.

The example illustrates how using properties does not affect the type of the object, thus, does not prevent the usage of SYCL objects in containers.

Each property is represented by a unique class and an instance of a property is an instance of that type. Some properties can be default constructed while others will require an argument on construction. A property may be applicable to more than one class, however some properties may not be compatible with each other. See the requirements for the properties of the SYCL buffer class, SYCL unsampled_image class and SYCL sampled_image class in Table 41 and Table 48 respectively.

Properties can be passed to a SYCL runtime class via an instance of property_list. These properties get tied to the SYCL runtime class instance and copies of the object will contain the same properties.

A SYCL implementation or a SYCL backend may provide additional properties other than those defined here, provided they are defined in accordance with the requirements described in Section 4.3.

Since a system can have several SYCL-compatible devices attached, it is useful to have a way to select a specific device or a set of devices to construct a specific object such as a device (see Section 4.6.4) or a queue (see Section 4.6.5), or perform some operations on a device subset.

Device selection is done either by already having a specific instance of a device (see Section 4.6.4) or by providing a device selector which is a ranking function that will give an integer ranking value to all the devices on the system.

The SYCL platform class encapsulates a single SYCL platform on which SYCL kernel functions may be executed. A SYCL platform must be associated with a single SYCL backend.

A SYCL platform is also associated with one or more SYCL devices associated with the same SYCL backend.

All member functions of the platform class are synchronous and errors are handled by throwing synchronous SYCL exceptions.

The execution environment for a SYCL application has a fixed number of platforms which does not vary as the application executes. The application can get a list of all these platforms via platform::get_platforms(), and the order of the platform objects is the same each time the application calls that function. The platform class also provides constructors, but constructing a new platform instance merely creates a new object that is a copy of one of the objects returned by platform::get_platforms().

The SYCL platform class provides the common reference semantics (see Section 4.5.2).

The context class represents a SYCL context. A context represents the runtime data structures and state required by a SYCL backend API to interact with a group of devices associated with a platform.

The SYCL context class provides the common reference semantics (see Section 4.5.2).

The SYCL device class encapsulates a single SYCL device on which kernels can be executed.

All member functions of the device class are synchronous and errors are handled by throwing synchronous SYCL exceptions.

The execution environment for a SYCL application has a fixed number of root devices which does not vary as the application executes. The application can get a list of all these devices via device::get_devices(), and the order of the device objects is the same each time the application calls that function (assuming the parameter to that function is the same for each call). The device class also provides constructors, but constructing a new device instance merely creates a new object that is a copy of one of the objects returned by device::get_devices().

A SYCL device can be partitioned into multiple SYCL devices, by calling the create_sub_devices() member function template. The resulting SYCL devices are considered sub devices, and it is valid to partition these sub devices further. The range of support for this feature is SYCL backend and device specific and can be queried for through get_info().

The SYCL device class provides the common reference semantics (see Section 4.5.2).