<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook MathML Module V1.1b1//EN"
              "http://www.oasis-open.org/docbook/xml/mathml/1.1CR1/dbmathml.dtd">

<refentry>
    <refentryinfo>
        <keywordset>
            <keyword>Function Qualifiers</keyword>
        </keywordset>
    </refentryinfo>

    <refmeta>
        <refentrytitle>Function Qualifiers</refentrytitle>

        <refmiscinfo>
            <copyright>
                <year>2007-2011</year>
                <holder>The Khronos Group Inc.
 Permission is hereby granted, free of charge, to any person obtaining a
copy of this software and/or associated documentation files (the
"Materials"), to deal in the Materials without restriction, including
without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Materials, and to
permit persons to whom the Materials are furnished to do so, subject to
the condition that this copyright notice and permission notice shall be included
in all copies or substantial portions of the Materials.</holder>
            </copyright>
        </refmiscinfo>
        <manvolnum>3</manvolnum>
    </refmeta>

<!-- ================================ SYNOPSIS -->

    <refnamediv id="FunctionQualifiers">
        <refname>Function Qualifiers</refname>

        <refpurpose>
            Qualifiers for kernel functions.
        </refpurpose>
    </refnamediv>

    <!-- ALTERNATIVE SYNTAX SYNOPSIS (NON-FUNCTION) -->
    <refsect2 id="synopsis">
        <title>
        </title>

        <informaltable frame="none">
            <tgroup cols="1" align="left" colsep="0" rowsep="0">
                <colspec colname="col1" colnum="1" />
                <tbody>
                    <row>
                        <entry>
__kernel
kernel

__attribute__((vec_type_hint(&lt;type<emphasis>n</emphasis>&gt;)))
__attribute__((work_group_size_hint(<emphasis>X</emphasis>, <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))
__attribute__((reqd_work_group_size(<emphasis>X</emphasis>, <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))
                        </entry>
                    </row>
                </tbody>
            </tgroup>
        </informaltable>
    </refsect2>

<!-- ================================ DESCRIPTION  -->

    <refsect1 id="description"><title>Description</title>
        <para>
          The <function>__kernel</function> (or <function>kernel</function>) qualifier declares a
          function to be a kernel that can be executed by an application on an OpenCL device(s).
          The following rules apply to functions that are declared with this qualifier:
        </para>

        <itemizedlist mark='bullet'>
            <listitem>
               <para>
                 It can be executed on the device only
               </para>
            </listitem>

            <listitem>
               <para>
                 It can be called by the host
               </para>
            </listitem>

            <listitem>
               <para>
                 It is just a regular function call if a <function>__kernel</function> function
                 is called by another kernel function.
               </para>
            </listitem>
        </itemizedlist>

        <para>
          Kernel functions with variables declared inside the function with the
          <citerefentry href="local"><refentrytitle>__local</refentrytitle></citerefentry>
          or <citerefentry><refentrytitle>local</refentrytitle></citerefentry>
          qualifier can be called by the host using appropriate APIs such as
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>,
          and <citerefentry><refentrytitle>clEnqueueTask</refentrytitle></citerefentry>.
        </para>

        <para>
          The behavior of calling kernel functions with variables declared inside the function with
          the <citerefentry href="local"><refentrytitle>__local</refentrytitle></citerefentry>
          or <citerefentry><refentrytitle>local</refentrytitle></citerefentry> qualifier from
          other kernel functions is implementation-defined.
        </para>

        <para>
          The <function>__kernel</function> and <function>kernel</function> names are reserved
          for use as functions qualifiers and shall not be used otherwise.
        </para>

        <bridgehead>Optional Attribute Qualifiers</bridgehead>

        <para>
          The <function>__kernel</function> qualifier can be used with the keyword <citerefentry
          href="attribute"><refentrytitle>__attribute__</refentrytitle></citerefentry> to
          declare additional information about the kernel function as described below.
        </para>

        <para>
          The optional <code>__attribute__((vec_type_hint(&lt;type&gt;)))</code>
          is a hint to the compiler and is intended to be a representation of the computational
          <emphasis>width</emphasis> of the <function>__kernel</function>, and should serve as the
          basis for calculating processor bandwidth utilization when the compiler is looking to
          autovectorize the code.  In the <code>__attribute__((vec_type_hint(&lt;type&gt;)))</code>
          qualifier &lt;type&gt; is one of the built-in vector types or the constituent scalar
          element types.  If <code>vec_type_hint (&lt;type&gt;)</code> is not specified, the
          kernel is assumed to have the <code>__attribute__((vec_type_hint(int)))</code> qualifier.
        </para>

        <para>
          Implicit in autovectorization is the assumption that any libraries called from
          the <function>__kernel</function> must be recompilable at run time to handle cases
          where the compiler decides to merge or separate workitems. This probably means that
          such libraries can never be hard coded binaries or that hard coded binaries must be
          accompanied either by source or some retargetable intermediate representation. This
          may be a code security question for some.
        </para>

        <para>
          For example, where the developer specified a width of <type>float4</type>, the compiler
          should assume that the computation usually uses up 4 lanes of a float vector, and would
          decide to merge work-items or possibly even separate one work-item into many threads
          to better match the hardware capabilities. A conforming implementation is not required
          to autovectorize code, but shall support the hint. A compiler may autovectorize, even
          if no hint is provided. If an implementation merges <constant>N</constant> work-items
          into one thread, it is responsible for correctly handling cases where the number of
          global or local work-items in any dimension modulo <constant>N</constant> is not zero.
        </para>

        <para>
          If for example, a <function>__kernel</function> function is declared with
          <code>__attribute__(( vec_type_hint (float4)))</code> (meaning that most
          operations in the <function>__kernel</function> are explicitly vectorized using
          <type>float4</type>) and the kernel is running using Intel&#174; Advanced Vector
          Instructions (Intel&#174; AVX) which implements a 8-float-wide vector unit, the
          autovectorizer might choose to merge two work-items to one thread, running a second
          work-item in the high half of the 256-bit AVX register.
        </para>

        <para>
          As another example, a Power4 machine has two scalar double precision floating-point
          units with an 6-cycle deep pipe. An autovectorizer for the Power4 machine might choose
          to interleave six kernels declared with the <code>__attribute__(( vec_type_hint
          (double2)))</code> qualifier into one hardware thread, to ensure that there is
          always 12-way parallelism available to saturate the FPUs. It might also choose to
          merge 4 or 8 work-items (or some other number) if it concludes that these are better
          choices, due to resource utilization concerns or some preference for divisibility by 2.
        </para>

        <para>
          The optional <code>__attribute__((work_group_size_hint(<emphasis>X</emphasis>,
          <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</code> is a hint to the
          compiler and is intended to specify the work-group size that may be used i.e. value
          most likely to be specified by the <varname>local_work_size</varname> argument to
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>.
          For example the <code>__attribute__((work_group_size_hint(1, 1, 1)))</code> is
          a hint to the compiler that the kernel will most likely be executed with a work-group
          size of 1.
        </para>

        <para>
          The optional <code>__attribute__((reqd_work_group_size(<emphasis>X</emphasis>,
          <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</code> is the work-group
          size that must be used as the <varname>local_work_size</varname> argument to
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>.  This
          allows the compiler to optimize the generated code appropriately for this kernel. The
          optional  <code>__attribute__((reqd_work_group_size(<emphasis>X</emphasis>,
          <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</code>,
          if specified, must be (1, 1, 1) if the kernel is executed via
          <citerefentry><refentrytitle>clEnqueueTask</refentrytitle></citerefentry>.
        </para>

        <para>
          If <varname>Z</varname> is one, the <varname>work_dim</varname> argument to
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>
          can be 2 or 3. If <varname>Y</varname> and <varname>Z</varname>
          are one, the <varname>work_dim</varname> argument to
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>
          can be 1, 2 or 3.
        </para>
    </refsect1>

<!-- ================================ NOTES  -->

    <refsect1 id="notes"><title>Notes</title>
        <para>
          Kernel functions with variables declared inside the function with the
          <citerefentry href="local"><refentrytitle>__local</refentrytitle></citerefentry>
          or <citerefentry><refentrytitle>local</refentrytitle></citerefentry>
          qualifier can be called by the host using appropriate APIs such as
          <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>,
          and <citerefentry><refentrytitle>clEnqueueTask</refentrytitle></citerefentry>.
        </para>

        <para>
          The behavior of calling kernel functions with variables declared inside the function with
          the <citerefentry href="local"><refentrytitle>__local</refentrytitle></citerefentry>
          and <citerefentry><refentrytitle>local</refentrytitle></citerefentry> qualifier from
          other kernel functions is implementation-defined.
        </para>

        <para>
          The <function>__kernel</function> and <function>kernel</function> names are reserved
          for use as functions qualifiers and shall not be used otherwise.
        </para>
    </refsect1>

<!-- ================================ EXAMPLE  -->

    <refsect2 id="example1">
        <title>
            Example
        </title>

        <informaltable frame="none">
            <tgroup cols="1" align="left" colsep="0" rowsep="0">
                <colspec colname="col1" colnum="1" />
                <tbody>
                    <row>
                        <entry>
// autovectorize assuming float4 as the
// basic computation width
__kernel __attribute__((vec_type_hint(float4)))
void foo( __global float4 *p ) { ....

// autovectorize assuming double as the
// basic computation width
__kernel __attribute__((vec_type_hint(double)))
void foo( __global float4 *p ){ ....

// autovectorize assuming int (default)
// as the basic computation width
__kernel
void foo( __global float4 *p ){ ....
                        </entry>
                    </row>
                </tbody>
            </tgroup>
        </informaltable>
    </refsect2>

<!-- ================================ SPECIFICATION  -->
<!-- Set the "uri" attribute in the <olink /> element to the "named destination" for the PDF page
-->
    <refsect1 id="specification"><title>Specification</title>
        <para>
            <imageobject>
                <imagedata fileref="pdficon_small1.gif" format="gif" />
            </imageobject>

            <olink uri="functionQualifiers">OpenCL Specification</olink>
        </para>
    </refsect1>

<!-- ================================ ALSO SEE -->

    <refsect1 id="seealso"><title>Also see</title>
        <para>
            <citerefentry href="attribute"><refentrytitle>__attribute__</refentrytitle></citerefentry>,
            <citerefentry><refentrytitle>clEnqueueNDRangeKernel</refentrytitle></citerefentry>,
            <citerefentry><refentrytitle>clEnqueueTask</refentrytitle></citerefentry>
        </para>
    </refsect1>

<!-- ============================== COPYRIGHT -->
<!-- Content included from copyright.inc.xsl -->

    <refsect3 id="Copyright"><title></title>
        <imageobject>
                <imagedata fileref="KhronosLogo.jpg" format="jpg" />
        </imageobject>
        <para />
    </refsect3>

<!-- 17-Oct-2011 -->
</refentry>