blob: e56b95506fa994c1e50252e7abf259f7abce993c [file] [log] [blame]
<?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>2007-2009</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>2</manvolnum>
</refmeta>
<!-- ================================ SYNOPSIS -->
<refnamediv id="Function Qualifiers">
<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>
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
<constant>__attribute__((vec_type_hint(&lt;type<emphasis>n</emphasis>&gt;)))</constant>
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.
<constant>vec_type_hint (&lt;type<emphasis>n</emphasis>&gt;)</constant>
shall be one of the built-in scalar or vector data type described in
tables 6.1 and 6.2.
If
<constant>vec_type_hint (&lt;type<emphasis>n</emphasis>&gt;)</constant>
is not specified, the default value is <type>int</type>.
</para>
<para>
The
<constant>__attribute__((vec_type_hint(int)))</constant>
is the default type.
</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> is declared with
<constant>__attribute__(( vec_type_hint (float4)))</constant>
(meaning that most operations in the <function>__kernel</function>
are explicitly vectorized using
<type>float4</type>) and the kernel is running using
Intel&reg; Advanced Vector Instructions
(Intel&reg; 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
<constant>__attribute__(( vec_type_hint (double2))) __kernel</constant>s
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
<constant>__attribute__((work_group_size_hint(<emphasis>X</emphasis>, <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</constant>
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
<constant>__attribute__((work_group_size_hint(1, 1, 1)))</constant>
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
<constant>__attribute__((reqd_work_group_size(<emphasis>X</emphasis>, <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</constant>
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
<constant>__attribute__((reqd_work_group_size(<emphasis>X</emphasis>, <emphasis>Y</emphasis>, <emphasis>Z</emphasis>)))</constant>,
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>
Implicit in autovectorization is the assumption that
any libraries called from the
__kernel 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>
</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><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>
</refentry>