Name

    ARM_core_id

Name Strings

    cl_arm_core_id

Contributors

    Robert Elliott, ARM Ltd. (robert.elliott 'at' arm.com)
    Hui Chen, ARM Ltd.       (hui.chen 'at' arm.com)
    Kevin Petit, ARM Ltd.    (kevin.petit 'at' arm.com)

Contacts

    Kevin Petit, ARM Ltd. (kevin.petit 'at' arm.com)

Status

    Shipping.

Version

    Revision: #2, Feb 26th, 2018

Number

    OpenCL Extension #26

Dependencies

    Requires OpenCL version 1.2 or later.

Overview

    This extension provides a built-in function which returns a unique ID
    for the compute unit that a work-group is running on.  This value is
    uniform for a work-group.

    This value can be used for a core-specific cache or atomic pool where the
    storage is required to be in global memory and persistent (but not ordered)
    between work-groups.  This does not provide any additional ordering on top
    of the existing guarantees between workgroups, nor does it provide any
    guarantee of concurrent execution.

    The IDs for the compute units may not be consecutive and applications must
    make sure they allocate enough memory to accommodate all the compute units
    present on the device.  A device info query allows the application to
    know the IDs associated with the compute units on a given device.

    The extension string cl_arm_core_id is returned for devices and platforms
    which support this extension.

Glossary

    No new terminology is introduced by this extension.

New Types

    None

New Procedures and Functions

    Device Info query

      CL_DEVICE_COMPUTE_UNITS_BITFIELD_ARM (return type cl_ulong) returns a
      bitfield where each bit set represents the presence of compute unit whose
      ID is the bit position.  The highest ID for any compute unit on the device
      is the position of the most significant bit set.  The total number of
      elements an application should allocate in an array indexed by core IDs is
      thus given by:

        ALLOC = sizeof(cl_ulong) * 8 - CLZ(CL_DEVICE_COMPUTE_UNITS_BITFIELD_ARM)

    Built-in Function

      uint arm_get_core_id( void )

    Description

      Returns the compute unit id as a uint, in the range [0;ALLOC-1]

    Example

      // Host code, size pool based on required_instance_size * ALLOC
      size_t required_instance_size = 1024; 
      cl_mem core_pool = clCreateBuffer(context,
                                        CL_MEM_READ_WRITE,
                                        required_instance_size * ALLOC, NULL, NULL);

      // Device/Kernel code, select memory instance
      kernel void test( global char *per_core_pool, global char *input, uint required_instance_size  )
      {
        global char *core_pool_instance = per_core_pool[ arm_get_core_id() * required_instance_size ];
        ...
      }

New Tokens

    OpenCL kernel code Now has access to:

      #pragma OPENCL EXTENSION cl_arm_core_id : enable
      The preprocessor macro cl_arm_core_id is also present

Revision History

    Revision: #1, Apr  2nd, 2013 - Initial revision
    Revision: #2, Feb 26th, 2018 - Added support for sparsely allocated compute
                                   unit IDs.