Name

    ARM_integer_dot_product

Name Strings

    cl_arm_integer_dot_product_int8
    cl_arm_integer_dot_product_accumulate_int8
    cl_arm_integer_dot_product_accumulate_int16

Contributors

    Kevin Petit, ARM Ltd.
    Abel Bernabeu, ARM Ltd.
    Giridhar Tammana, ARM Ltd.

Contact

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

IP Status

    No claims or disclosures are known to exist.

Version

    Revision: #3, November 17th, 2017

Number

    OpenCL Extension #52

Status

    Complete.

Extension Type

    OpenCL device extension

Dependencies

    Requires OpenCL version 1.2 or later.

Overview

    This extension adds built-in functions giving direct access to specialised
    integer dot product instructions that are supported on some devices.  An
    application wishing to support a fallback path for devices where a specific
    function is not available may use the following pattern:

    #ifdef cl_arm_integer_dot_product_XXX
    #pragma OPENCL EXTENSION cl_arm_integer_dot_product_XXX : enable
        code using arm_dot_xxx()
    #else
        alternative implementation
    #endif

Header File

    cl_ext.h

New Procedures and Functions

    Built-in Functions

      int arm_dot(char4 a, char4 b);
      uint arm_dot(uchar4 a, uchar4 b);

    Description

      These functions are available when cl_arm_integer_dot_product_int8 is reported.
      The cl_arm_integer_dot_product_int8 preprocessor macro is then defined.

      The value returned is:

        (a.x * b.x) + (a.y * b.y) + (a.z * b.z) + (a.w * b.w).

      Operands are zero- or sign-extended to 32-bit before the multiplications.

    Built-in Functions

      int arm_dot_acc(char4 a, char4 b, int acc);
      uint arm_dot_acc(uchar4 a, uchar4 b, uint acc);

    Description

      These functions are available when cl_arm_integer_dot_product_accumulate_int8
      is reported.  The cl_arm_integer_dot_product_accumulate_int8 preprocessor
      macro is then defined.

      The value returned is:

        acc + [ (a.x * b.x) + (a.y * b.y) + (a.z * b.z) + (a.w * b.w) ].

      Operands are zero- or sign-extended to 32-bit before the multiplications.

    Built-in Functions

      int arm_dot_acc(short2 a, short2 b, int acc);
      uint arm_dot_acc(ushort2 a, ushort2 b, uint acc);

    Description

      These functions are available when cl_arm_integer_dot_product_accumulate_int16
      is reported.  The cl_arm_integer_dot_product_accumulate_int16 preprocessor
      macro is then defined.

      The value returned is:

        acc + [ (a.x * b.x) + (a.y * b.y) ].

      Operands are zero- or sign-extended to 32-bit before the multiplications.