=========================================
Intel® SPMD Program Compiler User's Guide
=========================================

``ispc`` is a compiler for writing SPMD (single program multiple data)
programs to run on the CPU.  The SPMD programming approach is widely known
to graphics and GPGPU programmers; it is used for GPU shaders and CUDA\* and
OpenCL\* kernels, for example.  The main idea behind SPMD is that one writes
programs as if they were operating on a single data element (a pixel for a
pixel shader, for example), but then the underlying hardware and runtime
system executes multiple invocations of the program in parallel with
different inputs (the values for different pixels, for example).

The main goals behind ``ispc`` are to:

* Build a small C-like language that can deliver good performance to
  performance-oriented programmers who want to run SPMD programs on
  CPUs.
* Provide a thin abstraction layer between the programmer and the
  hardware--in particular, to follow the lesson from C for serial programs
  of having an execution and data model where the programmer can cleanly
  reason about the mapping of their source program to compiled assembly
  language and the underlying hardware.
* Harness the computational power of the Single Program, Multiple Data (SIMD) vector
  units without the extremely low-programmer-productivity activity of directly
  writing intrinsics.
* Explore opportunities from close-coupling between C/C++ application code
  and SPMD ``ispc`` code running on the same processor--lightweight funcion
  calls betwen the two languages, sharing data directly via pointers without
  copying or reformating, etc.

``ispc`` has already successfully delivered significant speedups for a
number of non-trivial workloads that aren't handled well by other
compilation approaches (e.g. loop auto-vectorization.)

**We are very interested in your feedback and comments about ispc and
in hearing your experiences using the system.  We are especially interested
in hearing if you try using ispc but see results that are not as you
were expecting or hoping for.** We encourage you to send a note with your
experiences or comments to the `ispc-users`_ mailing list or to file bug or
feature requests with the ``ispc`` `bug tracker`_. (Thanks!)

.. _ispc-users: http://groups.google.com/group/ispc-users
.. _bug tracker: https://github.com/ispc/ispc/issues?state=open


Contents:

* `Recent Changes to ISPC`_

* `Getting Started with ISPC`_

  + `Installing ISPC`_
  + `Compiling and Running a Simple ISPC Program`_

* `Using The ISPC Compiler`_

  + `Command-line Options`_

* `The ISPC Language`_

  + `Lexical Structure`_
  + `Basic Types and Type Qualifiers`_
  + `Enumeration Types`_
  + `Short Vector Types`_
  + `Struct and Array Types`_
  + `Declarations and Initializers`_
  + `Function Declarations`_
  + `Expressions`_
  + `Control Flow`_
  + `Functions`_
  + `C Constructs not in ISPC`_

* `Parallel Execution Model in ISPC`_

  + `The SPMD-on-SIMD Execution Model`_
  + `Uniform and Varying Qualifiers`_
  + `Mapping Data to Program Instances`_
  + `"Coherent" Control Flow Statements`_
  + `Program Instance Convergence`_
  + `Data Races`_
  + `Uniform Variables and Varying Control Flow`_
  + `Task Parallelism: Language Syntax`_
  + `Task Parallelism: Runtime Requirements`_

* `The ISPC Standard Library`_

  + `Math Functions`_
  + `Output Functions`_
  + `Cross-Program Instance Operations`_
  + `Packed Load and Store Operations`_
  + `Conversions To and From Half-Precision Floats`_
  + `Atomic Operations and Memory Fences`_
  + `Prefetches`_
  + `System Information`_
  + `Low-Level Bits`_

* `Interoperability with the Application`_

  + `Interoperability Overview`_
  + `Data Layout`_
  + `Data Alignment and Aliasing`_

* `Using ISPC Effectively`_

  + `Restructuring Existing Programs to Use ISPC`_
  + `Understanding How to Interoperate With the Application's Data`_
  + `Communicating Between SPMD Program Instances`_
  + `Gather and Scatter`_
  + `8 and 16-bit Integer Types`_
  + `Low-level Vector Tricks`_
  + `Debugging`_
  + `The "Fast math" Option`_
  + `"Inline" Aggressively`_
  + `Small Performance Tricks`_
  + `Instrumenting Your ISPC Programs`_
  + `Using Scan Operations For Variable Output`_
  + `Application-Supplied Execution Masks`_
  + `Explicit Vector Programming With Uniform Short Vector Types`_

* `Disclaimer and Legal Information`_

* `Optimization Notice`_

Recent Changes to ISPC
======================

See the file ``ReleaseNotes.txt`` in the ``ispc`` distribution for a list
of recent changes to the compiler.

Getting Started with ISPC
=========================

Installing ISPC
---------------

The `ispc downloads web page`_ has prebuilt executables for Windows\*,
Linux\* and Mac OS\* available for download.  Alternatively, you can
download the source code from that page and build it yourself; see see the
`ispc wiki`_ for instructions about building ``ispc`` from source.

.. _ispc downloads web page: downloads.html
.. _ispc wiki: http://github.com/ispc/ispc/wiki

Once you have an executable for your system, copy it into a directory
that's in your ``PATH``.  Congratulations--you've now installed ``ispc``.

Compiling and Running a Simple ISPC Program
-------------------------------------------

The directory ``examples/simple`` in the ``ispc`` distribution includes a
simple example of how to use ``ispc`` with a short C++ program.  See the
file ``simple.ispc`` in that directory (also reproduced here.)

::

    export void simple(uniform float vin[], uniform float vout[], 
                       uniform int count) {
        for (uniform int i = 0; i < count; i += programCount) {
            int index = i + programIndex;
            float v = vin[index];
            if (v < 3.)
                v = v * v;
            else
                v = sqrt(v);
            vout[index] = v;
        }
    }

This program loops over an array of values in ``vin`` and computes an
output value for each one.  For each value in ``vin``, if its value is less
than three, the output is the value squared, otherwise it's the square root
of the value.

The first thing to notice in this program is the presence of the ``export``
keyword in the function definition; this indicates that the function should
be made available to be called from application code.  The ``uniform``
qualifiers on the parameters to ``simple`` as well as for the variable
``i`` indicate that the correpsonding variables are non-vector
quantities--they are discussed in detail in the `Uniform and Varying
Qualifiers`_ section.

Each iteration of the for loop works on a number of input values in
parallel.  The built-in ``programCount`` variable indicates how many
program instances are running in parallel; it is equal to the SIMD width of
the machine.  (For example, the value is four on Intel® SSE, eight on
Intel® AVX, etc.)  Thus, we can see that each execution of the loop will
work on that many output values in parallel.  There is an implicit
assumption that ``programCount`` divides the ``count`` parameter without
remainder; the more general case case can be handled with a small amount of
additional code.

To load the ``programCount``-worth of values, the program computes an index
using the sum of ``i``, which gives the first value to work on in this
iteration, and ``programIndex``, which gives a unique integer identifier
for each running program instance, counting from zero.  Thus, the load from
``vin`` loads the values at offset ``i+0``, ``i+1``, ``i+2``, ..., from the
``vin`` array into the vector variable ``v``.  This general idiom should be
familiar to CUDA\* or OpenCL\* programmers, where thread ids serve a
similar role to ``programIndex`` in ``ispc``.  See the section `Mapping
Data to Program Instances`_ for more detail.

The program can then proceed, doing computation and control flow based on
the values loaded.  The result from the running program instances is
written to the ``vout`` array before the next loop iteration runs.

For a simple program like this one, the performance difference versus a
regular scalar C/C++ implementation are minimal.  For more
complex programs that do more substantial amounts of computation, doing
that computation in parallel across the machine's SIMD lanes can have a
substantial performance benefit.

On Linux\* and Mac OS\*, the makefile in that directory compiles this program.
For Windows\*, open the ``examples/examples.sln`` file in Microsoft Visual
C++ 2010\* to build this (and the other) examples.  In either case,
build it now!  We'll walk through the details of the compilation steps in
the following section, `Using The ISPC Compiler`_.)  In addition to
compiling the ``ispc`` program, in this case the ``ispc`` compiler also
generates a small header file, ``simple.h``.  This header file includes the
declaration for the C-callable function that the above ``ispc`` program is
compiled to.  The relevant parts of this file are:

::

  #ifdef __cplusplus
  extern "C" {
  #endif // __cplusplus
      extern void simple(float vin[], float vout[], int32_t count);
  #ifdef __cplusplus
  }
  #endif // __cplusplus

It's not mandatory to ``#include`` the generated header file in your C/C++
code (you can alternatively use a manually-written ``extern`` declaration
of the ``ispc`` functions you use), but it's a helpful check to ensure that
the function signatures are as expected on both sides.

Here is the main program, ``simple.cpp``, which calls the ``ispc`` function
above.

::

  #include <stdio.h>
  #include "simple.h"

  int main() {
      float vin[16], vout[16];
      for (int i = 0; i < 16; ++i)
          vin[i] = i;

      simple(vin, vout, 16);

      for (int i = 0; i < 16; ++i)
          printf("%d: simple(%f) = %f\n", i, vin[i], vout[i]);
  }

Note that the call to the ``ispc`` function in the middle of ``main()`` is
a regular function call.  (And it has the same overhead as a C/C++ function
call, for that matter.)

When the executable ``simple`` runs, it generates the expected output:
 
::

    0: simple(0.000000) = 0.000000
    1: simple(1.000000) = 1.000000
    2: simple(2.000000) = 4.000000
    3: simple(3.000000) = 1.732051
    ...

There is also a small example of using ``ispc`` to compute the Mandelbrot
set; see the `Mandelbrot set example`_ page on the ``ispc`` website for a
walkthrough of it.

.. _Mandelbrot set example: http://ispc.github.com/example.html

Using The ISPC Compiler
=======================

To go from a ``ispc`` source file to an object file that can be linked
with application code, enter the following command

::

   ispc foo.ispc -o foo.o

``ispc`` automatically runs the C preprocessor on your input program before
compiling it.  (This functionality can be disabled with the ``--nocpp``
command-line argument.)

Command-line Options
--------------------

The ``ispc`` executable can be run with ``--help`` to print a list of
accepted command-line arguments.  By default, the compiler compiles the
provided program (and issues warnings and errors), but doesn't
generate any output.  

If the ``-o`` flag is given, it will generate an output file (a native
object file by default).  To generate a text assembly file, pass
``--emit-asm``:

::

   ispc foo.ispc -o foo.s --emit-asm

To generate LLVM bitcode, use the ``--emit-llvm`` flag.

By default, an optimized x86-64 object file tuned for Intel® Core
processors CPUs is built.  You can use the ``--arch`` command line flag to
specify a 32-bit x86 target:

::

   ispc foo.ispc -o foo.obj --arch=x86

Optimizations can be turned off with ``-O0``:

::

   ispc foo.ispc -o foo.obj -O0

On Mac\* and Linux\*, there is early support for generating debugging
symbols; this is enabled with the ``-g`` command-line flag.

The ``-h`` flag can also be used to direct ``ispc`` to generate a C/C++
header file that includes C/C++ declarations of the C-callable ``ispc``
functions and the types passed to it.

On Linux\* and Mac OS\*, ``-D`` can be used to specify definitions to be
passed along to the C pre-prcessor, which runs over the program input
before it's compiled.  On Windows®, pre-processor definitions should be
provided to the ``cl`` call.

By default, the compiler generates x86-64 Intel® SSE4 code.  To generate
32-bit code, you can use the ``--arch=x86`` command-line flag.  To
select Intel® SSE2, use ``--target=sse2``.

``ispc`` supports an alternative method for generating Intel® SSE4 code,
where the program is "doubled up" and eight instances of it run in
parallel, rather than just four.  For workloads that don't require large
numbers of registers, this method can lead to significantly more efficient
execution thanks to greater instruction level parallelism.  This option is
selected with ``--target=sse4x2``.

The compiler issues a number of performance warnings for code constructs
that compile to relatively inefficient code.  These warnings can be
silenced with the ``--wno-perf`` flag (or by using ``--woff``, which turns
off all warnings.)


The ISPC Language
=================

``ispc``'s syntax is based on C and is designed to be as similar to C
as much as possible.  Between syntactic differences and the fundamentally
parallel execution model (versus C's serial model), C code is not directly
portable to ``ispc``, although starting with working C code and porting it
to ``ispc`` can be an efficient way to write ``ispc`` programs. 

Lexical Structure
-----------------

Tokens in ``ispc`` are delimted by white-space and comments.  The
white-space characters are the usual set of spaces, tabs, and carriage
returns/line feeds.  Comments can be delinated with ``//``, which starts a
comment that continues to the end of the line, or the start of a comment
can be delinated with ``/*`` and the end with ``*/``.  Like in C/C++,
comments can't be nested.

Identifiers in ``ispc`` are sequences of characters that start with an
underscore or an upper-case or lower-case letter, and then followed by
zero or more letters, numbers, or underscores.

Integer numeric constants can be specified in base 10 or in hexidecimal.
Base 10 constants are given by a sequence of one or more digits from 0 to
9.  Hexidecimal constants are denoted by a leading ``0x`` and then one or
more digits from 0-9, a-f, or A-F.

Floating-point constants can be specified in one of three ways.  First,
they may be a sequence of zero or more digits from 0 to 9, followed by a
period, followed by zero or more digits from 0 to 9. (There must be at
least one digit before or after the period).

The second option is scientific notation, where a base value is specified
as the first form of a floating-point constant but is then followed by an
"e" or "E", then a plus sign or a minus sign, and then an exponent.

Finally, floating-point constants may be specified as hexidecimal
constants; this form can ensure a perfectly bit-accurate representation of
a particular floating-point number.  These are specified with an "0x"
prefix, followed by a zero or a one, a period, and then the remainder of
the mantissa in hexidecimal form, with digits from 0-9, a-f, or A-F.  The
start of the exponent is denoted by a "p", which is then followed by an
optional plus or minus sign and then digits from 0 to 9.  For example:

::

  float two = 0x1p+1;  // 2.0
  float pi  = 0x1.921fb54442d18p+1;  // 3.1415926535...
  float neg = -0x1.ffep+11;  // -4095.

Floating-point constants can optionally have a "f" or "F" suffix (``ispc``
currently treats all floating-point constants as having 32-bit precision,
making this suffix unnecessary.)

String constants in ``ispc`` are denoted by an opening double quote ``"``
followed by any character other than a newline, up to a closing double
quote.  Within the string, a number of special escape sequences can be used
to specify special characters.  These sequences all start with an initial
``\`` and are listed below:

.. list-table:: Escape sequences in strings

  * - ``\\``
    - backslash: ``\``
  * - ``\"``
    - double quotation mark: ``"``
  * - ``\'``
    - single quotation mark: ``'``
  * - ``\a`` 
    - bell (alert)
  * - ``\b``
    - backspace character
  * - ``\f``
    - formfeed character
  * - ``\n``
    - newline
  * - ``\r``
    - carriabe return
  * - ``\t``
    - horizontal tab
  * - ``\v``
    - vertical tab
  * - ``\`` followed by one or more digits from 0-8
    - ASCII character in octal notation
  * - ``\x``, followed by one or more digits from 0-9, a-f, A-F 
    - ASCII character in hexidecimal notation

``ispc`` doesn't support a string data type; string constants can be passed
as the first argument to the ``print()`` statement, however.  ``ispc`` also
doesn't support character constants.

The following identifiers are reserved as language keywords: ``bool``,
``break``, ``case``, ``cbreak``, ``ccontinue``, ``cdo``, ``cfor``,
``char``, ``cif``, ``cwhile``, ``const``, ``continue``, ``creturn``,
``default``, ``do``, ``double``, ``else``, ``enum``, ``export``,
``extern``, ``false``, ``float``, ``for``, ``goto``, ``if``, ``inline``, ``int``,
``int8``, ``int16``, ``int32``, ``int64``, ``launch``, ``print``,
``reference``, ``return``,
``signed``, ``sizeof``, ``soa``, ``static``, ``struct``, ``switch``,
``sync``, ``task``, ``true``, ``typedef``, ``uniform``, ``union``,
``unsigned``, ``varying``, ``void``, ``volatile``, ``while``.

``ispc`` defines the following operators and punctuation:

.. list-table:: Operators

  * - Symbols
    - Use
  * - ``=``
    - Assignment
  * - ``+``, ``-``, \*, ``/``, ``%``
    - Arithmetic operators
  * - ``&``, ``|``, ``^``, ``!``, ``~``, ``&&``, ``||``, ``<<``, ``>>``
    - Logical and bitwise operators
  * - ``++``, ``--``
    - Pre/post increment/decrement
  * - ``<``, ``<=``, ``>``, ``>=``, ``==``, ``!=``
    - Relational operators
  * - ``*=``, ``/=``, ``+=``, ``-=``, ``<<=``, ``>>=``, ``&=``, ``|=``
    - Compound assignment operators
  * - ``?``, ``:``
    - Selection operators
  * - ``;``
    - Statement separator
  * - ``,``
    - Expression separator
  * - ``.``
    - Member access

A number of tokens are used for grouping in ``ispc``:

.. list-table:: Grouping Tokens

  * - ``(``, ``)``
    - Parenthesization of expressions, function calls, delimiting specifiers
      for control flow constructs.
  * - ``[``, ``]``
    - Array and short-vector indexing
  * - ``{``, ``}``
    - Compound statements  


Basic Types and Type Qualifiers
-------------------------------

``ispc`` is a statically-typed language.  It supports a variety of basic
types.

* ``void``: "empty" type representing no value.
* ``bool``: boolean value; may be assigned ``true``, ``false``, or the
  value of a boolean expression.
* ``int8``: 8-bit signed integer.
* ``unsigned int8``: 8-bit unsigned integer.
* ``int16``: 16-bit signed integer.
* ``unsigned int16``: 16-bit unsigned integer.
* ``int``: 32-bit signed integer; may also be specified as ``int32``.
* ``unsigned int``: 32-bit unsigned integer; may also be specified as
  ``unsigned int32``.
* ``float``: 32-bit floating point value
* ``int64``: 64-bit signed integer.
* ``unsigned int64``: 64-bit unsigned integer.
* ``double``: 64-bit double-precision floating point value.

Implicit type conversion between values of different types is done
automatically by the ``ispc`` compiler.  Thus, a value of ``float`` type
can be assigned to a variable of ``int`` type directly.  In binary
arithmetic expressions with mixed types, types are promoted to the "more
general" of the two types, with the following precedence:

::

  double > uint64 > int64 > float > uint32 > int32 > 
      uint16 > int16 > uint8 > int8 > bool

In other words, adding an ``int64`` to a ``double`` causes the ``int64`` to
be converted to a ``double``, the addition to be performed, and a
``double`` value to be returned.  If a different conversion behavior is
desired, then explicit type-casts can be used, where the destination type
is provided in parenthesis around the expression:

::

    double foo = 1. / 3.;
    int bar = (float)bar + (float)bar;  // 32-bit float addition

If a ``bool`` is converted to an integer numeric type (``int``, ``int64``,
etc.), then the result is the value one if the ``bool`` has the value
``true`` and has the value zero otherwise.

Variables can be declared with the ``const`` qualifier, which prohibits
their modification.

::

    const float PI = 3.1415926535;

As in C, the ``extern`` qualifier can be used to declare a function or
global variable defined in another source file, and the ``static``
qualifier can be used to define a variable or function that is only visible
in the current scope.  The values of ``static`` variables declared in
functions are preserved across function calls.

The ``typedef`` keyword can be used to name types:

::

  typedef Float3 float[3];

``typedef`` doesn't create a new type: it just provides an alternative name
for an existing type.  Thus, in the above example, it is legal to pass a
value with ``float[3]`` type to a function that has been declared to take a
``Float3`` parameter.

``ispc`` provides a ``reference`` qualifier that can be used for passing
values to functions by reference so that functions can return multiple
results or modify existing variables.

::

    void increment(reference float f) {
        ++f;
    }

``ispc`` doesn't currently support pointer types.


Enumeration Types
-----------------

It is possible to define user-defined enumeration types in ``ispc`` with
the ``enum`` keyword, which is followed by an option enumeration type name
and then a brace-delimited list of enumerators with optional values:

::

    enum Color { RED, GREEN, BLUE };
    enum Flags { 
        UNINITIALIZED = 0,
        INITIALIZED = 2,
        CACHED = 4
    };

Each ``enum`` declaration defines a new type; an attempt to implicitly
convert between enumerations of different types gives a compile-time error,
but enuemrations of different types can be explicitly cast to one other.

::

    Color c = (Color)CACHED;

Enumerators are implicitly converted to integer types, however, so they can
be directly passed to routines that take integer parameters and can be used
in expressions including integers, for example.  However, the integer
result of such an expression must be explicitly cast back to the enumerant
type if it to be assigned to a variable with the enuemrant type.

::

    Color c = RED;
    int nextColor = c+1;
    c = (Color)nextColor;

In this particular case, the explicit cast could be avoided using an
increment operator.

::

    Color c = RED;
    ++c;  // c == GREEN now


Short Vector Types
------------------

``ispc`` supports a parameterized type to define short vectors.  These
short vectors can only be used with basic types like ``float`` and ``int``;
they can't be applied to arrays or structures.  Note: ``ispc`` does *not*
use these short vectors to facilitate program vectorization; they are
purely a syntactic convenience.  Using them or writing the corresponding
code without them shouldn't lead to any noticeable performance differences
between the two approaches.

Syntax similar to C++ templates is used to declare these types:

::

    float<3> foo;   // vector of three floats
    double<6> bar;

The length of these vectors can be arbitrarily long, though the expected
usage model is relatively short vectors.

You can use ``typedef`` to create types that don't carry around
the brackets around the vector length:

::

    typedef float<3> float3;

``ispc`` doesn't support templates in general.  In particular,
not only must the vector length be a compile-time constant, but it's
also not possible to write functions that are parameterized by vector
length.

::

    uniform int i = foo();
    // ERROR: length must be compile-time constant
    float<i> vec; 
    // ERROR: can't write functions parameterized by vector length
    float<N> func(float<N> val); 
    
Arithmetic on these short vector types works as one would expect; the
operation is applied component-wise to the values in the vector.  Here is a
short example:

::

    float<3> func(float<3> a, float<3> b) {
        a += b;    // add individual elements of a and b
        a *= 2.;   // multiply all elements of a by 2
        bool<3> test = a < b;  // component-wise comparison
        return test ? a : b;   // return each minimum component
    }

As shown by the above code, scalar types automatically convert to
corresponding vector types when used in vector expressions.  In this
example, the constant ``2.`` above is converted to a three-vector of 2s for
the multiply in the second line of the function implementation.

Type conversion between other short vector types also works as one would
expect, though the two vector types must have the same length:

::

    float<3> foo = ...;
    int<3> bar = foo;    // ok, cast elements to ints
    int<4> bat = foo;    // ERROR: different vector lengths
    float<4> bing = foo; // ERROR: different vector lengths

For convenience, short vectors can be initialized with a list of individual
element values:

::

    float x = ..., y = ..., z = ...;
    float<3> pos = { x, y, z };


There are two mechanisms to access the individual elements of these short
vector data types.  The first is with the array indexing operator:

::

    float<4> foo;
    for (uniform int i = 0; i < 4; ++i)
        foo[i] = i;

``ispc`` also provides a specialized mechanism for naming and accessing
the first few elements of short vectors based on an overloading of
the structure member access operator.  The syntax is similar to that used
in HLSL, for example.

::

    float<3> position;
    position.x = ...;
    position.y = ...;
    position.z = ...;

More specifically, the first element of any short vector type can be
accessed with ``.x`` or ``.r``, the second with ``.y`` or ``.g``, the third
with ``.z`` or ``.b``, and the fourth with ``.w`` or ``.a``.  Just like
using the array indexing operator with an index that is greater than the
vector size, accessing an element that is beyond the vector's size is
undefined behavior and may cause your program to crash.

It is also possible to construct new short vectors from other short vector
values using this syntax, extended for "swizzling".  For example, 

::

    float<3> position = ...;
    float<3> new_pos = position.zyx;  // reverse order of components
    float<2> pos_2d = position.xy;

Though a single element can be assigned to, as in the examples above, it is
not currently possible to use swizzles on the left-hand side of assignment
expressions:

::

    int8<2> foo = ...;
    int8<2> bar = ...;
    foo.yz = bar;   // Error: can't assign to left-hand side of expression

Struct and Array Types
----------------------

More complex data structures can be built using ``struct`` and arrays.

::

    struct Foo {
        float time;
        int flags[10];
    };

The size of arrays must be a compile-time constant, though functions can be
declared to take "unsized arrays" as parameters so that arrays of any size
may be passed:

::

    void foo(float array[], int length);

As in C++, after a ``struct`` is declared, an instance can be created using
the ``struct``'s name:

::

    Foo f;

Alternatively, ``struct`` can be used before the structure name:

::

    struct Foo f;


Declarations and Initializers
-----------------------------

Variables are declared and assigned just as in C:

::

    float foo = 0, bar[5];
    float bat = func(foo);

If a variable is declared without an initializer expression, then its value
is undefined until a value is assigned to it.  Reading an undefined
variable may lead to unexpected program behavior.

Any variable that is declared at file scope (i.e. outside a function) is a
global variable.  If a global variable is qualified with the ``static``
keyword, then its only visible within the compilation unit in which it was
defined.  As in C/C++, a variable with a ``static`` qualifier inside a
functions maintains its value across function invocations.

Like C++, variables don't need to be declared at the start of a basic
block:

::

    int foo = ...;
    if (foo < 2) { ... }
    int bar = ...;

Variables can also be declared in ``for`` statement initializers:

::

    for (int i = 0; ...)

Arrays can be initialized with individual element values in braces:

::

    int bar[2][4] = { { 1, 2, 3, 4 }, { 5, 6, 7, 8 } };

Structures can also be initialized only with element values in braces:

::

    struct Color { float r, g, b; };
    ....
    Color d = { 0.5, .75, 1.0 }; // r = 0.5, ...


Function Declarations
---------------------

Functions can be declared with a number of qualifiers that affect their
visibility and capabilities.  As in C/C++, functions have global visibility
by default.  If a function is declared with a ``static`` qualifier, then it
is only visible in the file in which it was declared.

Any function that can be launched with the ``launch`` construct in ``ispc``
must have a ``task`` qualifier; see `Task Parallelism: Language Syntax`_
for more discussion of launching tasks in ``ispc``.

Functions that are intended to be called from C/C++ application code must
have the ``export`` qualifier.  This causes them to have regular C linkage
and to have their declarations included in header files, if the ``ispc``
compiler is directed to generated a C/C++ header file for the file it
compiled.

Finally, any function defined with an ``inline`` qualifier will always be
inlined by ``ispc``; ``inline`` is not a hint, but forces inlining.  The
compiler will opportunistically inline short functions depending on their
complexity, but any function that should always be inlined should have the
``inline`` qualifier.


Expressions
-----------

All of the operators from C that you'd expect for writing expressions are
present.  Rather than enumerating all of them, here is a short summary of
the range of them available in action.

::

    unsigned int i = 0x1234feed;
    unsigned int j = (i << 3) ^ ~(i - 3);
    i += j / 6;
    float f = 1.234e+23;
    float g = j * f / (2.f * i);
    double h = (g < 2) ? f : g/5;

Structure member access and array indexing also work as in C.

::

   struct Foo { float f[5]; int i; };
   Foo foo = { { 1,2,3,4,5 }, 2 };
   return foo.f[4] - foo.i;
    

Control Flow
------------

``ispc`` supports most of C's control flow constructs, including ``if``,
``for``, ``while``, ``do``.  You can use ``break`` and ``continue``
statements in ``for``, ``while``, and ``do`` loops.

There are variants of the ``if``, ``do``, ``while``, ``for``, ``break``,
``continue``, and ``return`` statements (``cif``, ``cdo``, ``cwhile``,
``cfor``, ``cbreak``, ``ccontinue``, and ``creturn``, respectively) that
provide the compiler a hint that the control flow is expected to be
coherent at that particular point, thus allowing the compiler to do
additional optimizations for that case.  These are described in the
`"Coherent" Control Flow Statements`_ section.

``ispc`` does not support ``switch`` statements or ``goto``.

Functions
---------

Like C, functions must be declared before they are called, though a forward
declaration can be used before the actual function definition.  Functions
can be overloaded by parameter type.  Given multiple definitions of a
function, ``ispc`` uses the following methods to try to find a match.  If
a single match of a given type is found, it is used; if multiple matches of
a given type are found, an error is issued.

* All parameter types match exactly.
* All parameter types match exactly, where any ``reference``-qualified
  parameters are considered equivalent to their underlying type.
* Parameters match with only promotions from ``uniform`` to ``varying``
  type.
* Parameters match using standard type conversion (``int`` to ``float``,
  ``float`` to ``int``.)

Also like C, arrays are passed to functions by reference.


C Constructs not in ISPC
-------------------------

The following C features are not available in ``ispc``.

* Pointers and function pointers
* ``char`` and ``short`` types
* ``switch`` statements
* bitfield members in structures
* ``union``
* ``goto``


Parallel Execution Model in ISPC
================================

Though ``ispc`` has C-based syntax, it is inherently a language for
parallel computation.  Understanding the details of ``ispc``'s parallel
execution model is critical for writing efficient and correct programs in
``ispc``.

``ispc`` supports both task parallelism to parallelize across multiple
cores and SPMD parallelism to parallelize across the SIMD vector lanes on a
single core.  This section focuses on SPMD parallelism.  See the sections
`Task Parallelism: Language Syntax`_ and `Task Parallelism: Runtime
Requirements`_ for discussion of task parallelism in ``ispc``.

The SPMD-on-SIMD Execution Model
--------------------------------

In the SPMD model as implemented in ``ispc``, you programs that compute a
set of outputs based on a set of inputs.  You must write these
programs so that it is safe to run multiple instances of them in
parallel--i.e. given a program an a set of inputs, the programs shouldn't
have any assumptions about the order in which they will be run over the
inputs, whether one program instances will have completed before another
runs. [#]_

.. [#] This is essentially the same requirement that languages like CUDA\*
   and OpenCL\* place on the programmer.

Given this guarantee, the ``ispc`` compiler can safely execute multiple
program instances in parallel, across the SIMD lanes of a single CPU.  In
many cases, this execution approach can achieve higher overall performance
than if the program instances had been run serially.

Upon entry to a ``ispc`` function, the execution model switches from
the application's serial model to SPMD.  Conceptually, a number of
``ispc`` program instances will start running in parallel.  This
parallelism doesn't involve launching hardware threads.  Rather, one
program instance is mapped to each of the SIMD lanes of the CPU's vector
unit (Intel® SSE or Intel® AVX).

If a ``ispc`` program is written to do a the following computation:

::

    float x = ..., y = ...;
    return x+y;

and if the ``ispc`` program is running four-wide on a CPU that supports the
Intel® SSE instructions, then four program instances are running in
parallel, each adding a pair of scalar values.  However, these four program
instances store their individual scalar values for ``x`` and ``y`` in the
lanes of an Intel® SSE vector register, so the addition operation for all
four program instances can be done in parallel with a single ``addps``
instruction.

Program execution is more complicated in the presence of control flow.  The
details are handled by the ``ispc`` compiler, but you may find it helpful
to understand what is going on in order to be a more effective ``ispc``
programmer.  In particular, the mapping of SPMD to SIMD lanes can lead to
reductions in this SIMD efficiency as different program instances want to
perform different computations.  For example, consider a simple ``if``
statement:

::

   float x = ..., y = ...;
   if (x < y) {
      ... 
   } else {
      ...
   }

In general, the test ``x<y`` has a different result for different running
SPMD program instances.  Some of the currently running program instances
want to execute the statements for the "true" case and some want to execute
the statements for the "false" case.  ``ispc`` processes this case by
generating code that executes for both cases and masking the results, such
that the "true" code doesn't have any side effects for the program
instances that want to run the "false" code, and vice versa.  Thus, the
correct reusult is computed for all of the program instances in the end,
though with some overhead relative to a scalar implementation where code
for only one of the two cases needs to run.

``for``, ``while``, and ``do`` statements are similar.  Their loops must
run until all of the running SPMD program instances are ready to exit the
loop.  Thus in an extreme case of a loop like:

::

    // assume limit has the values (1,1,1,1000) for the
    // current running program instances
    int limit = ...;  
    for (int i = 0; i < limit; ++i) {
        ...
    }

The loop body needs to execute 1000 times, since one of the SPMD
program instances has a value of 1000 for ``limit``.  For the other three
running program instances, the right result will still be computed, as the
code run the additional 999 times won't have any side effects for them.  However,
the result will have poor SIMD utilization as the majority of the loop
iterations don't benefit three out of the four currently running program
instances.  Thus, finding ways to structure the computation
so that the currently running program instances have similar desired
control flow paths leads to better overall efficiency.


Uniform and Varying Qualifiers
------------------------------

To write high-performance code, you need to understand the distinction
between ``uniform`` and ``varying`` data types.

If a variable has a ``uniform`` qualifier, then there is only a single
instance of that variable for all of the currently-executing program
instances.  (As such, it necessarily has the same value across all of the
program instances.)  ``uniform`` variables can be modified as the program
executes, but only in ways that preserve the property that they have the
same value across all of the program instances.  Assigning a
non-``uniform`` (i.e., ``varying``) value to a ``uniform`` variable causes
a compile-time error.

When appropriate, declaring variables as ``uniform`` types can allow the
compiler to produce substantially better code.  Consider for example an
image filtering operation where the program loops over adjacent pixels:

::

    float box3x3(uniform float image[32][32], int x, int y) {
        float sum = 0;
        for (int dy = -1; dy <= 1; ++dy)
            for (int dx = -1; dx <= 1; ++dx)
                sum += image[y+dy][x+dx];
        return sum / 9.;
    }

Under the SPMD execution model, a number of program instances are running
this function in parallel (and in general, we will assume that this
function will end up being called with different values for ``x`` and ``y``
for the running program instances.)  However, all of the program instances
will want to execute the same number of iterations of the ``for`` loops,
with all of them having the same values for ``dx`` and ``dy`` each time
through. [#]_

.. [#] In this case, a sufficiently smart compiler could determine that
   ``dx`` and ``dy`` have the same value for all program instances and thus
   generate more optimized code from the start, though ``ispc`` isn't yet
   this clever.  Put another way, the ``ispc`` approach is generally that
   the programmer shouldn't have to wonder if the compiler was smart or not
   in a particular case, thus avoiding performance surprises.

If these are instead implemented with ``dx`` and ``dy`` declared as
``uniform`` variables, then the ``ispc`` compiler can generate more
efficient code for the loops, taking advantage of the fact that these
values are the same for all program instances.

::

        for (uniform int dy = -1; dy <= 1; ++dy)
            for (uniform int dx = -1; dx <= 1; ++dx)
                sum += image[y+dy][x+dx];

In particular, ``ispc`` can avoid the overhead of checking to see if any
of the running program instances wants to do another loop iteration.
Instead, ``ispc`` can
generate code where all instances always do the same iterations.

A related benefit comes in ``if`` statements--if the test in an ``if``
statement is purely based on ``uniform`` quantities, then the result will
by definition be the same for all of the running program instances. Thus,
the code for only one of the two cases needs to execute.  ``ispc`` can
generate code that jumps to one of the two, avoiding the overhead of
needing to run the code for both cases.

``uniform`` variables will implicitly type-convert to varying types as
required:

::

   uniform int x = ...;
   int y = ...;
   int z = x * y;

Conversely, it is a compile-time error to assign a varying value to a
``uniform`` type:

::

    float f = ....;
    uniform float uf = f;  // ERROR

Arrays themselves aren't uniform or varying, but the elements that they
store are:

::

    float foo[10];
    uniform float bar[10];

Continuing the connection to data types in memory, the first declaration
corresponds to 10 four-wide float values (on Intel® SSE), and the second to
10 single float values.


Mapping Data to Program Instances
---------------------------------

An important part of SPMD programming is how to map the set of running
instances to the set of inputs to the program.  

If the application has created an array of floating-point values on which
the following computation needs to be completed:

::

    // C++ code
    int count = ...;
    float *data = new float[count];
    float *result = new float[count];
    ... initialize data ...
    ispc_func(data, count, result);

And if we have a ``ispc`` function declared as follows, then, given a
number of program instances running in parallel, how do the program
instances determine which elements of the array to work on?

::

    // ispc code
    export void ispc_func(uniform float data[],
                          uniform int count,
                          uniform float result[]) {
        ...
     
``ispc`` provides two built-in variables to help with this data mapping
across the set of running SPMD program instances.  The first,
``programCount`` gives the number of program instances that are executing
in parallel; for example, it may have the value 4 on most targets that
support Intel® and 8 on targets that support Intel® AVX.  The second,
``programIndex``, gives the index of the SIMD-lane being used for the
current program instance.  (In other words, it's a varying integer value
that has value zero for the first program instance, and so forth.)

Given these, ``ispc_func`` might be implemented as:

::

    for (uniform int i = 0; i < count; i += programCount) {
        float d = data[i + programIndex];
        float r = ....
        result[i + programIndex] = r;
    }

This code implicitly assumes that ``programCount`` evenly divides
``count``.  The more general case could be:

::

    for (uniform int i = 0; i < count; i += programCount) {
        if (i + programIndex < count) {
            float d = data[i + programIndex];
            ...

Some performance improvement may come from removing the ``if`` test from
the loop:

::

    uniform int fullCount = count - (count % programCount);
    uniform int i;
    for (i = 0; i < fullCount; i += programCount) {
        float d = data[i + programIndex];
         ...
    }
    if (i + programIndex < count) {
        float d = data[i + programIndex];
        ...
    }

For a more complex example, consider a ray tracer that wants to trace 4
rays per pixel.  To write code that works on one pixel at a time on a
machine that supports Intel® SSE, and 2 pixels at a time on a machine that
supports Intel® AVX, see the following:

::

    // compute sample offsets for the pixel or pixels being processed    
    uniform float xOffsetBase[4] = { 0, 0, 0.5, 0.5 };
    uniform float yOffsetBase[4] = { 0, 0.5, 0, 0.5 };
    float xOffset[programIndex % 4], yOffset[programIndex % 4];

    // compute steps
    uniform int dx, dy;
    if (programCount == 4) { dx = dy = 1; }
    else if (programCount == 8) { 
        dx = 2; dy = 1; 
        xOffset += programIndex / 4;
    }
    else if (programCount == 16) { 
        xOffset += programIndex / 8;
        yOffset += (programIndex / 4) & 0x1;
        dx = dy = 2; 
    }

    for (uniform int y = 0; y < height; y += dy) {
       for (uniform int x = 0; x < width; x += dx) {
           float xSample = x + xOffset, ySample = y + yOffset;
           // process samples in parallel ... 
       }
    }

"Coherent" Control Flow Statements
----------------------------------

``ispc`` provides a few mechanisms for you to supply a hint that control
flow is expected to be coherent at a particular point in the program's
execution.  These mechanisms provide the compiler a hint that it's worth
emitting extra code to check to see if the control flow is in fact coherent
at run-time, in which case it can jump to a simpler code path or otherwise
save work.

The first of these statements is ``cif``, indicating an ``if`` statement
that is expected to be coherent.  Recall from the `The
SPMD-on-SIMD Execution Model`_ section that ``if`` statements with a
``uniform`` test compile to more efficient code than ``if`` tests with
varying tests.  ``cif`` can provide many benefits of ``if`` with a
uniform test in the case where the test is actually varying.

The usage of ``cif`` in code is just the same as ``if``:

::

    cif (x < y) { 
        ...
    } else {
        ...
    }

``cif`` provides a hint to the compiler that you expect that most of the
executing SPMD programs will all have the same result for the ``if``
condition.  In this case, the code the compiler generates for the ``if``
test is along the lines of the following pseudo-code:

::

   bool expr = /* evaluate cif condition */
   if (all(expr)) {
       // run "true" case of if test only
   } else if (!any(expr)) {
       // run "false" case of if test only
   } else {
       // run both true and false cases, updating mask appropriately
   }

(For comparison, see the discussion of how regular ``if`` statements are
executed from the `The SPMD-on-SIMD Execution Model`_
section.)

For ``if`` statements where the different running SPMD program instances
don't have coherent values for the boolean ``if`` test, using ``cif``
introduces some additional overhead from the ``all`` and ``any`` tests as
well as the corresponding branches.  For cases where the program
instances often do compute the same boolean value, this overhead is
worthwhile.  If the control flow is in fact usually incoherent, this
overhead only costs performance.

In a similar fashion, ``ispc`` provides ``cfor``, ``cwhile``, ``cdo``,
``cbreak``, ``ccontinue``, and ``creturn`` statements.  These statements
are semantically the same as the corresponding non-"c"-prefixed functions.

For example, when ``ispc`` encounters a regular ``continue`` statement in
the middle of loop, it disables the mask bits for the program instances
that executed the ``continue`` and then executes the remainder of the loop
body, under the expectation that other executing program instances will
still need to run those instructions.  If you expect that all running
program instances will often execute ``continue`` together, then
``ccontinue`` provides the compiler a hint to do extra work to check if
every running program instance continued, in which case it can jump to the
end of the loop, saving the work of executing the otherwise meaningless
instructions.


Program Instance Convergence
----------------------------

Unlike languages such as OpenCL\* and CUDA\*, these executing program
instances are guaranteed to be maximally converged--if two program
instances follow the same control path, they are guaranteed to execute each
operation at the same time.  In the presence of divergent control flow:

::

   if (test) {
     // true
   }
   else {
     // false
   }

It is guaranteed that all program instances that were running before the
``if`` test will also be running after the end of the ``else`` block.
There is thus no need for a ``syncthreads``--type construct to synchronize
the executing program instances in cases where program instances would like
to share data or commicate with each other.


Data Races
----------

Although the SPMD model assumes that program instances are independent, you
can write code that has data races across the program instances.  For
example, the following code causes all program instances to try to write
different values to the same location:

::

    uniform int array[32] = 0;
    int index = 0;
    array[index] = programIndex;
    
In this case, the behavior of the program is undefined.


Uniform Variables and Varying Control Flow
------------------------------------------

Operations may be executed even if none of the program instances needs to
run them based on their control flow.  Consider an ``if``/``else`` test;
the statements in the ``else`` block may be executed even if the test
evaluates to ``true`` for all of the running program instances.  In
general, the executed statements are masked, such that they have no side
effects for the program instances that don't want to be running them, so
there is no visible side-effect of executing the ``else`` statements.
There is, however, one case where this part of the execution model can
become apparent.

Consider the cast of modifying the value of a ``uniform`` variable under
varying control flow:

::

   extern void foo();
   uniform int a;

   if (test) { // varying test
       ++a;    // modifying uniform under varying control flow
       foo();
   }

When possible, ``ispc`` detects that the control flow is varying and issues
an warning if a uniform variable is modified in this case.  Here, ``a`` may
be modified in the above code even if *none* of the program instances
evaluated a true value for the test, given the ``ispc`` execution model.


Task Parallelism: Language Syntax
---------------------------------

One option for combining task-parallelism with ``ispc`` is to just use
regular task parallelism in the C/C++ application code (be it through
Intel® Cilk(tm), Intel® Thread Building Blocks or another task system), and
for tasks to use ``ispc`` for SPMD parallelism across the vector lanes as
appropriate.  Alternatively, ``ispc`` also has support for launching tasks
from ``ispc`` code.  The approach is similar to Intel® Cilk's task launch
feature.  (See the ``examples/mandelbrot_tasks`` example to see it used in
a small example.)

First, any function that is launched as a task must be declared with the
``task`` qualifier:

::

    task void func(uniform float a[], uniform int index) {
        ...
        a[index] = ....
    }

Tasks must return ``void``; a compile time error is issued if a
non-``void`` task is defined.

Given a task definitions, there are two ways to write code that launches
tasks, using the ``launch`` construct.  First, one task can be launched at
a time, with parameters passed to the task to help it determine what part
of the overall computation it's responsible for:

::

    for (uniform int i = 0; i < 100; ++i)
        launch < func(a, i) >;

Note the ``launch`` keyword and the brackets around the function call.
This code launches 100 tasks, each of which presumably does some
computation that is keyed off of given the value ``i``.  In general, one
should launch many more tasks than there are processors in the system to
ensure good load-balancing, but not so many that the overhead of scheduling
and running tasks dominates the computation.

Alternatively, a number of tasks may be launched from a single ``launch``
statement.  We might instead write the above example with a single
``launch`` like this:

::

    launch[100] < func2(a) >;

Where an integer value (not necessarily a compile-time constant) is
provided to the ``launch`` keyword in square brackets; this number of tasks
will be enqueued to be run asynchronously.  Within each of the tasks, two
special built-in variables are available--``taskIndex``, and ``taskCount``.
The first, ``taskIndex``, ranges from zero to one minus the number of tasks
provided to ``launch``, and ``taskCount`` equals the number of launched
taks.  Thus, we might use ``taskIndex`` in the implementation of ``func2``
to determine which array element to process.

::

    task void func2(uniform float a[]) {
        ...
        a[taskIndex] = ...
    }

Program execution continues asynchronously after a ``launch`` statement;
thus, a function shouldn't access values being generated by the tasks it
has launched within the function without synchronization.  If results are
needed before function return, a function can use a ``sync`` statement to
wait for all launched tasks to finish:

::

    launch[100] < func2(a) >;
    sync;
    // now safe to use computed values in a[]...

Alternatively, any function that launches tasks has an automatically-added
``sync`` statement before it returns, so that functions that call a
function that launches tasks don't have to worry about outstanding
asynchronous computation from that function.

Inside functions with the ``task`` qualifier, two additional built-in
variables are provided in addition to ``taskIndex`` and ``taskCount``:
``threadIndex`` and ``threadCount``.  ``threadCount`` gives the total
number of hardware threads that have been launched by the task system.
``threadIndex`` provides an index between zero and ``threadCount-1`` that
gives a unique index that corresponds to the hardware thread that is
executing the current task.  The ``threadIndex`` can be used for accessing
data that is private to the current thread and thus doesn't require
synchronization to access under parallel execution.

Task Parallelism: Runtime Requirements
--------------------------------------

If you use the task launch feature in ``ispc``, you must provide C/C++
implementations of three specific functions that manage launching and
synchronizing parallel tasks; these functions must be linked into your
executable.  Although these functions may be implemented in any
language, they must have "C" linkage (i.e. their prototypes must be
declared inside an ``extern "C"`` block if they are defined in C++.)

By using user-supplied versions of these functions, ``ispc`` programs can
easily interoperate with software systems that have existing task systems
for managing parallelism.  If you're using ``ispc`` with a system that
isn't otherwise multi-threaded and don't want to write custom
implementations of them, you can use the implementations of these functions
provided in the ``examples/tasksys.cpp`` file in the ``ispc``
distributions.

If you are implementing your own task system, the remainder of this section
discusses the requirements for these calls.  You will also likely want to
review the example task systems in ``examples/tasksys.cpp`` for reference.
If you are not implmenting your own task system, you can skip reading the
remainder of this section.

Here are the declarations of the three functions that must be provided to
manage tasks in ``ispc``:

::

    void *ISPCAlloc(void **handlePtr, int64_t size, int32_t alignment);
    void ISPCLaunch(void **handlePtr, void *f, void *data, int count);
    void ISPCSync(void *handle);

All three of these functions take an opaque handle (or a pointer to an
opaque handle) as their first parameter.  This handle allows the task
system runtime to distinguish between calls to these functions from
different functions in ``ispc`` code.  In this way, the task system
implementation can efficiently wait for completion on just the tasks
launched from a single function.

The first time one of ``ISPCLaunch()`` or ``ISPCAlloc()`` is called in an
``ispc`` functon, the ``void *`` pointed to by the ``handlePtr`` parameter
will be ``NULL``.  The implementations of these function should then
initialize ``*handlePtr`` to a unique handle value of some sort.  (For
example, it might allocate a small structure to record which tasks were
launched by the current function.)  In subsequent calls to these functions
in the emitted ``ispc`` code, the same value for ``handlePtr`` will be
passed in, such that loading from ``*handlePtr`` will retrieve the value
stored in the first call.

At function exit (or at an explicit ``sync`` statement), a call to
``ISPCSync()`` will be generated if ``*handlePtr`` is non-``NULL``.
Therefore, the handle value is passed directly to ``ISPCSync()``, rather
than a pointer to it, as in the other functions.

The ``ISPCAlloc()`` function is used to allocate small blocks of memory to
store parameters passed to tasks.  It should return a pointer to memory
with the given aize and alignment.  Note that there is no explicit
``ISPCFree()`` call; instead, all memory allocated within an ``ispc``
function should be freed when ``ISPCSync()`` is called.

``ISPCLaunch()`` is called to launch to launch one or more asynchronous
tasks.  Each ``launch`` statement in ``ispc`` code causes a call to
``ISPCLaunch()`` to be emitted in the generated code.  The three parameters
after the handle pointer to thie function are relatively straightforward;
the ``void *f`` parameter holds a pointer to a function to call to run the
work for this task, ``data`` holds a pointer to data to pass to this
function, and ``count`` is the number of instances of this function to
enqueue for asynchronous execution.  (In other words, ``count`` corresponds
to the value ``n`` in a multiple-task launch statement like ``launch[n]``.)

The signature of the provided function pointer ``f`` is

::

    void (*TaskFuncPtr)(void *data, int threadIndex, int threadCount,
                        int taskIndex, int taskCount)

When this function pointer is called by one of the hardware threads managed
bythe task system, the ``data`` pointer passed to ``ISPCLaunch()`` should
be passed to it for its first parameter; ``threadCount`` gives the total
number of hardware threads that have been spawned to run tasks and
``threadIndex`` should be an integer index between zero and ``threadCount``
uniquely identifying the hardware thread that is running the task.  (These
values can be used to index into thread-local storage.)

The value of ``taskCount`` should be the number of tasks launched in the
``launch`` statement that caused the call to ``ISPCLaunch()`` and each of
the calls to this function should be given a unique value of ``taskIndex``
between zero and ``taskCount``, to distinguish which of the instances
of the set of launched tasks is running.

The ISPC Standard Library
=========================

``ispc`` has a standard library that is automatically available when
compiling ``ispc`` programs.  (To disable the standard library, pass the
``--nostdlib`` command-line flag to the compiler.)

Math Functions
--------------

The math functions in the standard library provide a relatively standard
range of mathematical functionality.

A number of different implementations of the transcendental math functions
are available; the math library to use can be selected with the
``--math-lib=`` command line argument.  The following values can be provided
for this argument. 

* ``default``: ``ispc``'s default built-in math functions.  These have
  reasonably high precision. (e.g. ``sin`` has a maximum absolute error of
  approximately 1.45e-6 over the range -10pi to 10pi.)
* ``fast``: more efficient but lower accuracy versions of the default ``ispc``
  implementations.
* ``svml``: use Intel "Short Vector Math Library".  Use
  ``icc`` to link your final executable so that the appropriate libraries
  are linked.
* ``system``: use the system's math library.  On many systems, these
  functions are more accurate than both of ``ispc``'s implementations.
  Using these functions may be quite
  inefficient; the system math functions only compute one result at a time
  (i.e. they aren't vectorized), so ``ispc`` has to call them once per
  active program instance.  (This is not the case for the other three
  options.)

In addition to an absolute value call, ``abs()``, ``signbits()`` extracts
the sign bit of the given value, returning ``0x80000000`` if the sign bit
is on (i.e. the value is negative) and zero if it is off.

::

    float abs(float a)
    uniform float abs(uniform float a)
    unsigned int signbits(float x)

Standard rounding functions are provided.  (On machines that support Intel®
SSE or Intel® AVX, these functions all map to variants of the ``roundss`` and
``roundps`` instructions, respectively.)

::

    float round(float x)
    uniform float round(uniform float x)
    float floor(float x)
    uniform float floor(uniform float x)
    float ceil(float x)
    uniform float ceil(uniform float x)

``rcp()`` computes an approximation to ``1/v``.  The amount of error is
different on different architectures.

::

    float rcp(float v)
    uniform float rcp(uniform float v)

The square root of a given value can be computed with ``sqrt()``, which
maps to hardware square root intrinsics when available.  An approximate
reciprocal square root, ``1/sqrt(v)`` is computed by ``rsqrt()``.  Like
``rcp()``, the error from this call is different on different
architectures.

::

    float sqrt(float v)
    uniform float sqrt(uniform float v)
    float rsqrt(float v)
    uniform float rsqrt(uniform float v)

A standard set of minimum and maximum functions is available.  These
functions also map to corresponding intrinsic functions.

::

    float min(float a, float b)
    uniform float min(uniform float a, uniform float b)
    float max(float a, float b)
    uniform float max(uniform float a, uniform float b)
    unsigned int min(unsigned int a, unsigned int b)
    uniform unsigned int min(uniform unsigned int a,
                             uniform unsigned int b)
    unsigned int max(unsigned int a, unsigned int b)
    uniform unsigned int max(uniform unsigned int a,
                             uniform unsigned int b)

The ``clamp()`` functions clamp the provided value to the given range.
(Their implementations are based on ``min()`` and ``max()`` and are thus
quite efficient.)

::

    float clamp(float v, float low, float high)
    uniform float clamp(uniform float v, uniform float low,
                        uniform float high)
    unsigned int clamp(unsigned int v, unsigned int low,
                       unsigned int high)
    uniform unsigned int clamp(uniform unsigned int v,
                               uniform unsigned int low,
                               uniform unsigned int high)

``ispc`` provides a standard variety of calls for trigonometric functions:

::

    float sin(float x)
    uniform float sin(uniform float x)
    float cos(float x)
    uniform float cos(uniform float x)
    float tan(float x)
    uniform float tan(uniform float x)

Arctangent functions are also available:

::

   float atan(float x)
   float atan2(float x, float y)
   uniform float atan(uniform float x)
   uniform float atan2(uniform float x, uniform float y)

If both sine and cosine are needed, then the ``sincos()`` call computes
both more efficiently than two calls to the respective individual
functions:

::

    void sincos(float x, reference float s, reference float c)
    void sincos(uniform float x, uniform reference float s,
                uniform reference float c)


The usual exponential and logarithmic functions are provided.

::

    float exp(float x)
    uniform float exp(uniform float x)
    float log(float x)
    uniform float log(uniform float x)
    float pow(float a, float b)
    uniform float pow(uniform float a, uniform float b)

Some functions that end up doing low-level manipulation of the
floating-point representation in memory are available.  As in the standard
math library, ``ldexp()`` multiplies the value ``x`` by 2^n, and
``frexp()`` directly returns the normalized mantissa and returns the
normalized exponent as a power of two in the ``pw2`` parameter.

::

    float ldexp(float x, int n)
    uniform float ldexp(uniform float x, uniform int n)
    float frexp(float x, reference int pw2)
    niform float frexp(uniform float x,
                       reference uniform int pw2)


A simple random number generator is provided.  State for the RNG
is maintained in an instance of the ``RNGState`` structure, which is seeded
with ``seed_rng()``.

::

    struct RNGState;
    unsigned int random(reference uniform RNGState state)
    float frandom(reference uniform RNGState state)
    void seed_rng(reference uniform RNGState state,
                  uniform int seed)

Output Functions
----------------

``ispc`` has a simple ``print`` statement for printing values during
program execution.  In the following short ``ispc`` program, there are
three uses of the ``print`` statement: 

::

    export void foo(uniform float f[4], uniform int i) {
        float x = f[programIndex];
        print("i = %, x = %\n", i, x);
        if (x < 2) {
            ++x;
            print("added to x = %\n", x);
        }
        print("last print of x = %\n", x);
    }

There are a few things to note.  First, the function is called ``print``,
not ``printf`` (unlike C).  Second, the formatting string passed to this
function only uses a single percent sign to denote where the corresponding
value should be printed.  You don't need to match the types of formatting
operators with the types being passed.  However, you can't currently use
the rich data formatting options that ``printf`` provides (e.g. constructs
like ``%.10f``.).

If this function is called with the array of floats (0,1,2,3) passed in for
the ``f`` parameter and the value ``10`` for the ``i`` parameter, it
generates the following output on a four-wide compilation target:

::

    i = 10, x = [0.000000,1.000000,2.000000,3.000000]
    added to x = [1.000000,2.000000,_________,_________]
    last print of x = [1.000000,2.000000,2.000000,3.000000]

All values of "varying" variables for each executing program instance is
printed when a "varying" variable is printed.  The result from the second
print statement, which was called under control flow in the function
``foo()`` above, and given the input array (0,1,2,3), only includes the
first two program instances entered the ``if`` block.  Therefore, the
values for the inactive program instances aren't printed.  (In other cases,
they may have garbage values or be otherwise undefined.)


Cross-Program Instance Operations
---------------------------------

Usually, ``ispc`` code expresses independent programs performing
computation on separate data elements.  There are, however, a number of
cases where it's useful for the program instances to be able to cooperate
in computing results.  The cross-lane operations described in this section
provide primitives for communication between the running program instances.
 
A few routines that evaluate conditions across the running program
instances.  For example, ``any()`` returns ``true`` if the given value
``v`` is ``true`` for any of the SPMD program instances currently running,
and ``all()`` returns ``true`` if it true for all of them.

::

    uniform bool any(bool v)
    uniform bool all(bool v)

To broadcast a value from one program instance to all of the others, a
``broadcast()`` function is available.  It broadcasts the value of the
``value`` parameter for the program instance given by ``index`` to all of
the running program instances.

::

    int8 broadcast(int8 value, uniform int index)
    int16 broadcast(int16 value, uniform int index)
    int32 broadcast(int32 value, uniform int index)
    int64 broadcast(int64 value, uniform int index)
    float broadcast(float value, uniform int index)
    double broadcast(double value, uniform int index)

The ``rotate()`` function allows each program instance to find the value of
the given value that their neighbor ``offset`` steps away has.  For
example, on an 8-wide target, if ``offset`` has the value (1, 2, 3, 4, 5,
6, 7, 8) in each of the running program instances, then ``rotate(value,
-1)`` causes the first program instance to get the value 8, the second
program instance to get the value 1, the third 2, and so forth.  The
provided offset value can be positive or negative, and may be greater than
``programCount`` (it is masked to ensure valid offsets).

::

    int8 rotate(int8 value, uniform int offset)
    int16 rotate(int16 value, uniform int offset)
    int32 rotate(int32 value, uniform int offset)
    int64 rotate(int64 value, uniform int offset)
    float rotate(float value, uniform int offset)
    double rotate(double value, uniform int offset)


Finally, the ``shuffle()`` functions allow two variants of fully general
shuffling of values among the program instances.  For the first version,
each program instance's value of permutation gives the program instance
from which to get the value of ``value``.  The provided values for
``permutation`` must all be between 0 and ``programCount-1``.

::

    int8 shuffle(int8 value, int permutation)
    int16 shuffle(int16 value, int permutation)
    int32 shuffle(int32 value, int permutation)
    int64 shuffle(int64 value, int permutation)
    float shuffle(float value, int permutation)
    double shuffle(double value, int permutation)


The second variant of ``shuffle()`` permutes over the extended vector that
is the concatenation of the two provided values.  In other words, a value
of 0 in an element of ``permutation`` corresponds to the first element of
``value0``, the value ``2*programCount-1`` corresponds to the last element
of ``value1``, etc.)

::

    int8 shuffle(int8 value0, int8 value1, int permutation)
    int16 shuffle(int16 value0, int16 value1, int permutation)
    int32 shuffle(int32 value0, int32 value1, int permutation)
    int64 shuffle(int64 value0, int64 value1, int permutation)
    float shuffle(float value0, float value1, int permutation)
    double shuffle(double value0, double value1, int permutation)

The various variants of ``popcnt()`` return the population count--the
number of bits set in the given value.

::

    uniform int popcnt(uniform int v)
    int popcnt(int v)
    uniform int popcnt(bool v)

The ``lanemask()`` function returns an integer that encodes which of the
current SPMD program instances are currently executing.  The i'th bit is
set if the i'th SIMD lane is currently active.

::

    uniform int lanemask()

You can compute reductions across the program instances.  For example, the
values in each of the SIMD lanes ``x`` are added together by
``reduce_add()``.  If this function is called under control flow, it only
adds the values for the currently active program instances.

::

    uniform float reduce_add(float x)
    uniform int reduce_add(int x)
    uniform unsigned int reduce_add(unsigned int x)

You can also use functions to compute the minimum and maximum value of the
given value across all of the currently-executing vector lanes.

::

    uniform float reduce_min(float a, float b)
    uniform int reduce_min(int a, int b)
    uniform unsigned int reduce_min(unsigned int a, unsigned int b)
    uniform float reduce_max(float a, float b)
    uniform int reduce_max(int a, int b)
    uniform unsigned int reduce_max(unsigned int a, unsigned int b)

Finally, you can check to see if a particular value has the same value in
all of the currently-running program instances:

::

    uniform bool reduce_equal(int32 v)
    uniform bool reduce_equal(unsigned int32 v)
    uniform bool reduce_equal(float v)
    uniform bool reduce_equal(int64 v)
    uniform bool reduce_equal(unsigned int64 v)
    uniform bool reduce_equal(double)

There are also variants of these functions that return the value as a
``uniform`` in the case where the values are all the same.

::

    uniform bool reduce_equal(int32 v, reference uniform int32 sameval)
    uniform bool reduce_equal(unsigned int32 v,
                              reference uniform unsigned int32 sameval)
    uniform bool reduce_equal(float v, reference uniform float sameval)
    uniform bool reduce_equal(int64 v, reference uniform int64 sameval)
    uniform bool reduce_equal(unsigned int64 v,
                              reference uniform unsigned int64 sameval)
    uniform bool reduce_equal(double, reference uniform double sameval)

If called when none of the program instances are running,
``reduce_equal()`` will return ``false``.

There are also a number of functions to compute "scan"s of values across
the program instances.  For example, the ``exclusive_scan_and()`` function
computes, for each program instance, the sum of the given value over all of
the preceeding program instances.  (The scans currently available in
``ispc`` are all so-called "exclusive" scans, meaning that the value
computed for a given element does not include the value provided for that
element.)  In C code, an exclusive add scan over an array might be
implemented as:

::

    void scan_add(int *in_array, int *result_array, int count) {
        result_array[0] = 0;
        for (int i = 0; i < count; ++i)
            result_array[i] = result_array[i-1] + in_array[i-1];
    }

``ispc`` provides the following scan functions--addition, bitwise-and, and
bitwise-or are available:

::

    int32 exclusive_scan_add(int32 v) 
    unsigned int32 exclusive_scan_add(unsigned int32 v) 
    float exclusive_scan_add(float v) 
    int64 exclusive_scan_add(int64 v) 
    unsigned int64 exclusive_scan_add(unsigned int64 v) 
    double exclusive_scan_add(double v) 
    int32 exclusive_scan_and(int32 v) 
    unsigned int32 exclusive_scan_and(unsigned int32 v) 
    int64 exclusive_scan_and(int64 v) 
    unsigned int64 exclusive_scan_and(unsigned int64 v) 
    int32 exclusive_scan_or(int32 v) 
    unsigned int32 exclusive_scan_or(unsigned int32 v) 
    int64 exclusive_scan_or(int64 v) 
    unsigned int64 exclusive_scan_or(unsigned int64 v) 


Packed Load and Store Operations
--------------------------------

The standard library also offers routines for writing out and reading in
values from linear memory locations for the active program instances.  The
``packed_load_active()`` functions load consecutive values from the given
array, starting at ``a[offset]``, loading one value for each
currently-executing program instance and storing it into that program
instance's ``val`` variable.  They return the total number of values
loaded.  Similarly, the ``packed_store_active()`` functions store the
``val`` values for each program instances that executed the
``packed_store_active()`` call, storing the results into the given array
starting at the given offset.  They return the total number of values
stored.

::

    uniform int packed_load_active(uniform int a[],
                                   uniform int offset,
                                   reference int val)
    uniform int packed_load_active(uniform unsigned int a[],
                                   uniform int offset,
                                   reference unsigned int val)
    uniform int packed_store_active(uniform int a[],
                                    uniform int offset,
                                    int val)
    uniform int packed_store_active(uniform unsigned int a[],
                                    uniform int offset,
                                    unsigned int val)


As an example of how these functions can be used, the following code shows
the use of ``packed_store_active()``.  The program instances that are
executing each compute some value ``x``; we'd like to record the program
index values of the program instances for which ``x`` is less than zero, if
any.  In following the code, the ``programIndex`` value for each program
instance is written into the ``ids`` array only if ``x < 0`` for that
program instance.  The total number of values written into ``ids`` is
returned from ``packed_store_active()``.

::

    uniform int ids[100];
    uniform int offset = 0;
    float x = ...;
    if (x < 0)
        offset += packed_store_active(ids, offset, programIndex);


Finally, there are primitive operations that extract and set values in the
SIMD lanes.  You can implement all of the operations described
above in this section from these routines, though in general, not as
efficiently.  These routines are useful for implementing other reductions
and cross-lane communication that isn't included in the above, though.
Given a ``varying`` value, ``extract()`` returns the i'th element of it as
a single ``uniform`` value.  Similarly, ``insert`` returns a new value
where the ``i`` th element of ``x`` has been replaced with the value ``v``
.

::

    uniform int8 extract(int8 x, uniform int i)
    uniform int16 extract(int16 x, uniform int i)
    uniform int32 extract(int32 x, uniform int i)
    uniform int64 extract(int64 x, uniform int i)
    uniform float extract(float x, uniform int i)

::

    int8 insert(int8 x, uniform int i, uniform int8 v)
    int16 insert(int16 x, uniform int i, uniform int16 v)
    int32 insert(int32 x, uniform int i, uniform int32 v)
    int64 insert(int64 x, uniform int i, uniform int64 v)
    float insert(float x, uniform int i, uniform float v)


Conversions To and From Half-Precision Floats
---------------------------------------------

There are functions to convert to and from the IEEE 16-bit floating-point
format.  Note that there is no ``half`` data-type, and it isn't possible
to do floating-point math directly with ``half`` types in ``ispc``; these
functions facilitate converting to and from half-format data in memory.

To use them, half-format data should be loaded into an ``int16`` and the
``half_to_float()`` function used to convert it the a 32-bit floating point
value.  To store a value to memory in half format, the ``float_to_half()``
function returns the 16 bits that are the closest match to the given
``float``, in half format.

::

    float half_to_float(unsigned int16 h)
    uniform float half_to_float(uniform unsigned int16 h)
    int16 float_to_half(float f)
    uniform int16 float_to_half(uniform float f)

There are also faster versions of these functions that don't worry about
handling floating point infinity, "not a number" and denormalized numbers
correctly.  These are faster than the above functions, but are less
precise.

::

    float half_to_float_fast(unsigned int16 h)
    uniform float half_to_float_fast(uniform unsigned int16 h)
    int16 float_to_half_fast(float f)
    uniform int16 float_to_half_fast(uniform float f)


Atomic Operations and Memory Fences
-----------------------------------

The usual range of atomic memory operations are provided in ``ispc``.  As an
example, consider the 32-bit integer atomic add routine:

::

  int32 atomic_add_global(reference uniform int32 val, int32 delta)

The semantics are the expected ones for an atomic add function: the value
"val" has the value "delta" added to it atomically, and the old value of
"val" is returned from the function.  (Thus, if multiple processors 
simultaneously issue atomic adds to the same memory location, the adds will
be serialized by the hardware so that the correct result is computed in the
end.)

One thing to note is that that the value being added to here is a
``uniform`` integer, while the increment amount and the return value are
``varying``.  In other words, the semantics of this call are that each
running program instance individually issues the atomic operation with its
own ``delta`` value and gets the previous value of ``val`` back in return.
The atomics for the running program instances may be issued in arbitrary
order; it's not guaranteed that they will be issued in ``programIndex``
order, for example.

Here are the declarations of the ``int32`` variants of these functions.
There are also ``int64`` equivalents as well as variants that take
``unsigned`` ``int32`` and ``int64`` values.  (The ``atomic_swap_global()``
function can be used with ``float`` and ``double`` types as well.)

::

  int32 atomic_add_global(reference uniform int32 val, int32 value)
  int32 atomic_subtract_global(reference uniform int32 val, int32 value)
  int32 atomic_min_global(reference uniform int32 val, int32 value)
  int32 atomic_max_global(reference uniform int32 val, int32 value)
  int32 atomic_and_global(reference uniform int32 val, int32 value)
  int32 atomic_or_global(reference uniform int32 val, int32 value)
  int32 atomic_xor_global(reference uniform int32 val, int32 value)
  int32 atomic_swap_global(reference uniform int32 val, int32 newval)

There are also variants of these functions that take ``uniform`` values for
the operand and return a ``uniform`` result:

::

  uniform int32 atomic_add_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_subtract_global(reference uniform int32 val,
                                       uniform int32 value)
  uniform int32 atomic_min_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_max_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_and_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_or_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_xor_global(reference uniform int32 val,
                                  uniform int32 value)
  uniform int32 atomic_swap_global(reference uniform int32 val,
                                   uniform int32 newval)

There are also an atomic swap and "compare and exchange" functions.
Compare and exchange atomically compares the value in "val" to
"compare"--if they match, it assigns "newval" to "val".  In either case,
the old value of "val" is returned.  (As with the other atomic operations,
there are also ``unsigned`` and 64-bit variants of this function.
Furthermore, there are ``float`` and ``double`` variants as well.)

::

  int32 atomic_swap_global(reference uniform int32 val, int32 new)
  uniform int32 atomic_swap_global(reference uniform int32 val,
                                   uniform int32 new)
  int32 atomic_compare_exchange_global(reference uniform int32 val,
                                       int32 compare, int32 newval)
  uniform int32 atomic_compare_exchange_global(reference uniform int32 val,
                                  uniform int32 compare, uniform int32 newval)

``ispc`` also has a standard library routine that inserts a memory barrier
into the code; it ensures that all memory reads and writes prior to be
barrier complete before any reads or writes after the barrier are issued.
See the `Linux kernel documentation on memory barriers`_ for an excellent
writeup on the need for and the use of memory barriers in multi-threaded
code.

.. _Linux kernel documentation on memory barriers: http://www.kernel.org/doc/Documentation/memory-barriers.txt

::

    void memory_barrier();


Prefetches
----------

The standard library has a variety of functions to prefetch data into the
processor's cache.  While modern CPUs have automatic prefetchers that do a
reasonable job of prefetching data to the cache before its needed, high
performance applications may find it helpful to prefetch data before it's
needed.

For example, this code shows how to prefetch data to the processor's L1
cache while iterating over the items in an array.  

::

   uniform int32 array[...];
   for (uniform int i = 0; i < count; ++i) {
       // do computation with array[i]
       prefetch_l1(array[i+32]);
   }

The standard library has routines to prefetch to the L1, L2, and L3
caches.  It also has a variant, ``prefetch_nt()``, that indicates that the
value being prefetched isn't expected to be used more than once (so should
be high priority to be evicted from the cache).

::

    void prefetch_{l1,l2,l3,nt}(reference TYPE)

These functions are available for all of the basic types in the
language--``int8``, ``int16``, ``int32``, ``float``, and so forth.


System Information
------------------

A routine is available to find the number of CPU cores available in the
system:

::

    int num_cores()

This value can be useful for adapting the granularity of parallel task
decomposition depending on the number of processors in the system.


Low-Level Bits
--------------

Sometimes it's useful to convert a ``bool`` value to an integer using sign
extension so that the integer's bits are all on if the ``bool`` has the
value ``true`` (rather than just having the value one).  The
``sign_extend()`` functions provide this functionality:

::

    int sign_extend(bool value) 
    uniform int sign_extend(uniform bool value) 

The ``intbits()`` and ``floatbits()`` functions can be used to implement
low-level floating-point bit twiddling.  For example, ``intbits()`` returns
an ``unsigned int`` that is a bit-for-bit copy of the given ``float``
value.  (Note: it is **not** the same as ``(int)a``, but corresponds to
something like ``*((int *)&a)`` in C.

::

    float floatbits(unsigned int a);
    uniform float floatbits(uniform unsigned int a);
    unsigned int intbits(float a);
    uniform unsigned int intbits(uniform float a);


The ``intbits()`` and ``floatbits()`` functions have no cost at runtime;
they just let the compiler know how to interpret the bits of the given
value.  They make it possible to efficiently write functions that take
advantage of the low-level bit representation of floating-point values.

For example, the ``abs()`` function in the standard library is implemented
as follows:

::

    float abs(float a) {
        unsigned int i = intbits(a);
        i &= 0x7fffffff;
        return floatbits(i);
    }

It, it clears the high order bit, to ensure that the given floating-point
value is positive.  This compiles down to a single ``andps`` instruction
when used with an Intel® SSE target, for example.

Interoperability with the Application
=====================================

One of ``ispc``'s key goals is to make it easy to interoperate between the
C/C++ application code and parallel code written in ``ispc``.  This
section describes the details of how this works and describes a number of
the pitfalls.

Interoperability Overview
-------------------------

As described in `Compiling and Running a Simple ISPC Program`_ it's
relatively straightforward to call ``ispc`` code from C/C++.  First, any
``ispc`` functions to be called should be defined with the ``export``
keyword:

::

    export void foo(uniform float a[]) { 
        ...
    }


This function corresponds to the following C-callable function:

::

   void foo(float a[]);


(Recall from the `Uniform and Varying Qualifiers`_ section 
that ``uniform`` types correspond to a single instances of the
corresponding type in C/C++.) 

In addition to variables passed from the application to ``ispc`` in the
function call, you can also share global variables between the application
and ``ispc``.  To do so, just declare the global variable as usual (in
either ``ispc`` or application code), and add an ``extern`` declaration on
the other side.

For example, given this ``ispc`` code:

::

    // ispc code
    uniform float foo;
    extern uniform float bar[10];

And this C++ code:

::

   // C++ code
   extern float foo;
   float bar[10];

Both the ``foo`` and ``bar`` global variables can be accessed on each
side.

``ispc`` code can also call back to C/C++.  On the ``ispc`` side, any
application functions to be called must be declared with the ``extern "C"``
qualifier.

::

   extern "C" void foo(uniform float f, uniform float g);

Unlike in C++, ``extern "C"`` doesn't take braces to delineate
multiple functions to be declared; thus, multiple C functions to be called
from ``ispc`` must be declared as follows:

::

   extern "C" void foo(uniform float f, uniform float g);
   extern "C" uniform int bar(uniform int a);

It is illegal to overload functions declared with ``extern "C"`` linkage;
``ispc`` issues an error in this case.

Function calls back to C/C++ are not made if none of the program instances
want to make the call.  For example, given code like:

::

    uniform float foo = ...;
    float x = ...;
    if (x != 0)
        foo = appFunc(foo);


``appFunc()`` will only be called if one or more of the running program
instances evaluates ``true`` for ``x != 0``.  If the application code would
like to determine which of the running program instances want to make the
call, a mask representing the active SIMD lanes can be passed to the
function.

::

   extern "C" float appFunc(uniform float x,
                            uniform int activeLanes);

If the function is then called as:
   
::

   ...
   x = appFunc(x, lanemask());

The ``activeLanes`` parameter will have the value one in the 0th bit if the
first program instance is running at this point in the code, one in the
first bit for the second instance, and so forth.  (The ``lanemask()``
function is documented in `Low-Level Bits`_.)  Application code can thus be
written as:

::

    float appFunc(float x, int activeLanes) {
        for (int i = 0; i < programCount; ++i)
            if ((activeLanes & (1 << i)) != 0) {
                // do computation for i'th SIMD lane
            }
    }


Data Layout
-----------

In general, ``ispc`` tries to ensure that ``struct`` s and other complex
datatypes are laid out in the same way in memory as they are in C/C++.
Matching structure layout is important for easy interoperability between C/C++
code and ``ispc`` code.

The main complexity in sharing data between ``ispc`` and C/C++ often comes
from reconciling data structures between ``ispc`` code and application
code; it can be useful to declare the shared structures in ``ispc`` code
and then examine the generated header file (which will have the C/C++
equivalents of them.)  For example, given a structure in ``ispc``:

::

  // ispc code
  struct Node {
     uniform int count;
     uniform float pos[3];
  };

If the ``Node`` structure is used in the parameters to an ``export`` ed
function, then the header file generated by the ``ispc`` compiler will
have a declaration like:

::

  // C/C++ code
  struct Node {
     int count;
     float pos[3];
  };

Because ``varying`` types have different sizes on different processor
architectures, ``ispc`` prohibits any varying types from being used in
parameters to functions with the ``export`` qualifier.  (``ispc`` also
prohibits passing structures that themselves have varying types as members,
etc.)  Thus, all datatypes that is shared with the application must have
the ``uniform`` qualifier applied to them.  (See `Understanding How to
Interoperate With the Application's Data`_ for more discussion of how to
load vectors of SoA or AoSoA data from the application.)

While ``ispc`` doesn't support pointers, there are two mechanisms to work
with pointers to arrays from the application.  First, ``ispc`` passes
arrays by reference (like C), if the application has allocated an array by:

::

   // C++ code
   float *array = new float[count];

It can pass ``array`` to a ``ispc`` function defined as:

::

   export void foo(uniform float array[], uniform int count)

Similarly, ``struct`` s from the application can have embedded pointers.
This is handled with similar ``[]`` syntax:

::

  // C code
  struct Foo {
      float *foo, *bar;
  };

On the ``ispc`` side, the corresponding ``struct`` declaration is:

::

  // ispc
  struct Foo {
      uniform float foo[], bar[];
  };

There are two subtleties related to data layout to be aware of.  First, the
C++ specification doesn't define the size or memory layout of ``bool`` s.
Therefore, it's dangerous to share ``bool`` values in memory between
``ispc`` code and C/C++ code.

Second, ``ispc`` stores ``uniform`` short-vector types in memory with their
first element at the machine's natural vector alignment (i.e. 16 bytes for
a target that is using Intel® SSE, and so forth.)  This implies that these
types will have different layout on different compilation targets.  As
such, applications should in general avoid accessing ``uniform`` short
vector types from C/C++ application code if possible.

Data Alignment and Aliasing
---------------------------

There are are two important constraints that must be adhered to when
passing pointers from the application to ``ispc`` programs.

The first is that it is required that it be valid to read memory at the
first element of any array that is passed to ``ispc``.  In practice, this
should just happen naturally, but it does mean that it is illegal to pass a
``NULL`` pointer as a parameter to a ``ispc`` function called from the
application.

The second constraint is that pointers and references in ``ispc`` programs
must not alias.  The ``ispc`` compiler assumes that different pointers
can't end up pointing to the same memory location, either due to having the
same initial value, or through array indexing in the program as it
executed.

This aliasing constraint also applies to ``reference`` parameters to
functions.  Given a function like:

::

    void func(reference int a, reference int b) {
        a = 0;
        if (b == 0) { ... }
    }

Then if the same variable must not be passed to ``func()``.  This is
another case of aliasing, and if the caller calls the function as ``func(x,
x)``, it's not guaranteed that the ``if`` test will evaluate to true, due
to the compiler's requirement of no aliasing.

(In the future, ``ispc`` will have a mechanism to indicate that pointers
may alias.)

Using ISPC Effectively
======================

Restructuring Existing Programs to Use ISPC
-------------------------------------------

``ispc`` is designed to enable you to incorporate
SPMD parallelism into existing code with minimal modification; features
like the ability to share memory and data structures betwen C/C++ and
``ispc`` code and the ability to directly call back and forth between
``ispc`` and C/C++ are motivated by this.  These features also make it
easy to incrementally transform a program to use ``ispc``; the most
computationally-intensive localized parts of the computation can be
transformed into ``ispc`` code while the remainder of the system is left
as is.

For a given section of code to be transitioned to run in ``ispc``, the
next question is how to parallelize the computation.  Generally, there will
be obvious loops inside which a large amount of computation is done ("for
each ray", "for each pixel", etc.)  Mapping these to the SPMD computational
style is often effective.

Carefully choose how to do the exact mapping of computation to SPMD program
instances.  This choice can impact the mix of gather/scatter memory access
versus coherent memory access, for example.  (See more on this in the
section `Gather and Scatter`_ below.)  This decision can also impact the
coherence of control flow across the running SPMD program instances, which
can also have a significant effect on performance; in general, creating
groups of work that will tend to do similar computation across the SPMD
program instances improves performance.

Understanding How to Interoperate With the Application's Data
-------------------------------------------------------------

One of ``ispc``'s key goals is to be able to interoperate with the
application's data, in whatever layout it is stored in.  You don't need to
worry about reformatting of data or the overhead of a driver model that
abstracts the data layout.  This section illustrates some of the
alternatives with a simple example of computing the length of a large
number of vectors.

Consider for starters a ``Vector`` data-type, defined in C as:

::

   struct Vector { float x, y, z; };

We might have (still in C) an array of ``Vector`` s defined like this:

::

   Vector vectors[1024];

This is called an "array of structures" (AoS) layout.  To compute the
lengths of these vectors in parallel, you can write ``ispc`` code like
this:

::

  export void length(Vector vectors[1024], uniform float len[]) {
      for (uniform int i = 0; i < 1024; i += programCount) {
          int index = i+programIndex;
          float x = vectors[index].x;
          float y = vectors[index].y;
          float z = vectors[index].z;
          float l = sqrt(x*x + y*y + z*z);
          len[index] = l;
      }
  }

The ``vectors`` array has been indexed using ``programIndex`` in
order to "peel off" ``programCount`` worth of values to compute the length
of each time through the loop.

The problem with this implementation is that the indexing into the array of
structures, ``vectors[index].x`` is relatively expensive.  On a target
machine that supports four-wide Intel® SSE, this turns into four loads of
single ``float`` values from non-contiguous memory locations, which are
then packed into a four-wide register corresponding to ``float x``.  Once the
values are loaded into the local ``x``, ``y``, and ``z`` variables,
SIMD-efficient computation can proceed; getting to that point is
relatively inefficient.

An alternative would be the "structure of arrays" (SoA) layout.  In C, the
data would be declared as:

::

    float x[1024], y[1024], z[1024];

The ``ispc`` code might be:

::

  export void length(uniform float x[1024], uniform float y[1024],
                     uniform float z[1024], uniform float len[]) {
      for (uniform int i = 0; i < 1024; i += programCount) {
          int index = i+programIndex;
          float xx = x[index];
          float yy = y[index];
          float zz = z[index];
          float l = sqrt(xx*xx + yy*yy + zz*zz);
          len[index] = l;
      }
  }

In this example, the loads into ``xx``, ``yy``, and ``zz`` are single
vector loads of ``programCount`` values into the corresponding registers.
This processing is more efficient than the multiple scalar loads that are
required with the AoS layout above.

A final alternative is "array of structures of arrays" (AoSoA), a hybrid
between these two.  A structure is declared that stores a small number of
``x``, ``y``, and ``z`` values in contiguous memory locations:

::

  struct Vector16 {
      float x[16], y[16], z[16];
  };


The ``ispc`` code has an outer loop over ``Vector16`` elements and
then an inner loop that peels off values from the element members:

::

  #define N_VEC (1024/16)
  export void length(Vector16 v[N_VEC], uniform float len[]) {
      for (uniform int i = 0; i < N_VEC; ++i) {
          for (uniform int j = 0; j < 16; j += programCount) {
              int index = j + programIndex;
              float x = v[i].x[index];
              float y = v[i].y[index];
              float z = v[i].z[index];
              float l = sqrt(x*x + y*y + z*z);
              len[index] = l;
          }
      }
  }

(This code assumes that ``programCount`` divides 16 equally.  See below for
discussion of the more general case.)  One advantage of the AoSoA layout is
that the memory accesses to load values are to nearby memory locations,
where as with SoA, each of the three loads above is to locations separated
by a few thousand bytes.  Thus, AoSoA can be more cache friendly.  For
structures with many members, this difference can lead to a substantial
improvement.

``ispc`` can also efficiently process data in AoSoA layout where the inner
array length is less than the machine vector width.  For example, consider
doing computation with this AoSoA structure definition on a machine with an
8-wide vector unit (for example, an Intel® AVX target):

::

  struct Vector4 {
      float x[4], y[4], z[4];
  };


The ``ispc`` code to process this loads elements four at a time from
``Vector4`` instances until it has a full ``programCount`` number of
elements to work with and then proceeds with the computation.

::

  #define N_VEC (1024/4)
  export void length(Vector4 v[N_VEC], uniform float len[]) {
      for (uniform int i = 0; i < N_VEC; i += programCount / 4) {
          float x, y, z;
          for (uniform int j = 0; j < programCount / 4; ++j) {
              if (programIndex >= 4 * j &&
                  programIndex <  4 * (j+1)) {
                  int index = (programIndex & 0x3);
                  x = v[i+j].x[index];
                  y = v[i+j].y[index];
                  z = v[i+j].z[index];
              }
          }
          float l = sqrt(x*x + y*y + z*z);
          len[4*i + programIndex] = l;
      }
  }


Communicating Between SPMD Program Instances
--------------------------------------------

The ``broadcast()``, ``rotate()``, and ``shuffle()`` standard library
routines provide a variety of mechanisms for the running program instances
to communicate values to each other during execution.  See the section
`Cross-Program Instance Operations`_ for more information about their
operation.


Gather and Scatter
------------------

The CPU is a poor fit for SPMD execution in some ways, the worst of which
is handling of general memory reads and writes from SPMD program instances.
For example, in a "simple" array index:

::

    int i = ....;
    uniform float x[10] = { ... };
    float f = x[i];

Since the index ``i`` is a varying value, the various SPMD program
instances will in general be reading different locations in the array
``x``.  Because the CPU doesn't have a gather instruction, the ``ispc``
compiler has to serialize these memory reads, performing a separate memory
load for each running program instance, packing the result into ``f``.
(And the analogous case would happen for a write into ``x[i]``.)

In many cases, gathers like these are unavoidable; the running program
instances just need to access incoherent memory locations.  However, if the
array index ``i`` could actually be declared and used as a ``uniform``
variable, the resulting array index is substantially more
efficient.  This is another case where using ``uniform`` whenever applicable
is of benefit.

In some cases, the ``ispc`` compiler is able to deduce that the memory
locations accessed are either all the same or are uniform.  For example,
given:

::

  uniform int x = ...;
  int y = x;
  return array[y];

The compiler is able to determine that all of the program instances are
loading from the same location, even though ``y`` is not a ``uniform``
variable.  In this case, the compiler will transform this load to a regular vector
load, rather than a general gather.

Sometimes the running program instances will access a
linear sequence of memory locations; this happens most frequently when
array indexing is done based on the built-in ``programIndex`` variable.  In
many of these cases, the compiler is also able to detect this case and then
do a vector load.  For example, given:

::

    uniform int x = ...;
    return array[2*x + programIndex];

A regular vector load is done from array, starting at offset ``2*x``.


8 and 16-bit Integer Types
--------------------------

The code generated for 8 and 16-bit integer types is generally not as
efficient as the code generated for 32-bit integer types.  It is generally
worthwhile to use 32-bit integer types for intermediate computations, even
if the final result will be stored in a smaller integer type.

Low-level Vector Tricks
-----------------------

Many low-level Intel® SSE coding constructs can be implemented in ``ispc``
code.  For example, the following code efficiently reverses the sign of the
given values.

::

  float flipsign(float a) {
      unsigned int i = intbits(a);
      i ^= 0x80000000;
      return floatbits(i);
  }

This code compiles down to a single XOR instruction.

Debugging
---------

Support for debugging in ``ispc`` is in progress.  On Linux\* and Mac
OS\*, the ``-g`` command-line flag can be supplied to the compiler,
which causes it to generate debugging symbols.  Running ``ispc`` programs
in the debugger, setting breakpoints, printing out variables and the like
all generally works, though there is occasional unexpected behavior.

Another option for debugging (the only current option on Windows\*) is
to use the ``print`` statement for ``printf()``
style debugging.  You can also use the ability to call back to
application code at particular points in the program, passing a set of
variable values to be logged or otherwise analyzed from there.

The "Fast math" Option
----------------------

``ispc`` has a ``--fast-math`` command-line flag that enables a number of
optimizations that may be undesirable in code where numerical preceision is
critically important.  For many graphics applications, the
approximations may be acceptable.  The following two optimizations are
performed when ``--fast-math`` is used.  By default, the ``--fast-math``
flag is off.

* Expressions like ``x / y``, where ``y`` is a compile-time constant, are
  transformed to ``x * (1./y)``, where the inverse value of ``y`` is
  precomputed at compile time.

* Expressions like ``x / y``, where ``y`` is not a compile-time constant,
  are transformed to ``x * rcp(y)``, where ``rcp()`` maps to the
  approximate reciprocal instruction from the standard library.


"Inline" Aggressively
---------------------

Inlining functions aggressively is generally beneficial for performance
with ``ispc``.  Definitely use the ``inline`` qualifier for any short
functions (a few lines long), and experiment with it for longer functions.

Small Performance Tricks
------------------------

Performance is slightly improved by declaring variables at the same block
scope where they are first used.  For example, in code like the
following, if the lifetime of ``foo`` is only within the scope of the
``if`` clause, write the code like this:  

::

    float func() {
        ....
        if (x < y) {
            float foo;
            ... use foo ...
        }
    }

Try not to write code as:

::

    float func() {
        float foo;
        ....
        if (x < y) {
            ... use foo ...
        }
    }

Doing so can reduce the amount of masked store instructions that the
compiler needs to generate.

Instrumenting Your ISPC Programs
--------------------------------

``ispc`` has an optional instrumentation feature that can help you
understand performance issues.  If a program is compiled using the
``--instrument`` flag, the compiler emits calls to a function with the
following signature at various points in the program (for
example, at interesting points in the control flow, when scatters or
gathers happen.)

::

    extern "C" {
        void ISPCInstrument(const char *fn, const char *note, 
                            int line, int mask);
    }

This function is passed the file name of the ``ispc`` file running, a short
note indicating what is happening, the line number in the source file, and
the current mask of active SPMD program lanes.  You must provide an
implementation of this function and link it in with your application.

For example, when the ``ispc`` program runs, this function might be called
as follows:

::

   ISPCInstrument("foo.ispc", "function entry", 55, 0xf);

This call indicates that at the currently executing program has just
entered the function defined at line 55 of the file ``foo.ispc``, with a
mask of all lanes currently executing (assuming a four-wide Intel® SSE
target machine).

For a fuller example of the utility of this functionality, see
``examples/aobench_instrumented`` in the ``ispc`` distribution.  Ths
example includes an implementation of the ``ISPCInstrument`` function that
collects aggregate data about the program's execution behavior.

When running this example, you will want to direct to the ``ao`` executable
to generate a low resolution image, because the instrumentation adds
substantial execution overhead.  For example:

::

    % ./ao 1 32 32

After the ``ao`` program exits, a summary report along the following lines
will be printed.  In the first few lines, you can see how many times a few
functions were called, and the average percentage of SIMD lanes that were
active upon function entry.

:: 

    ao.ispc(0067) - function entry: 342424 calls (0 / 0.00% all off!), 95.86% active lanes
    ao.ispc(0067) - return: uniform control flow: 342424 calls (0 / 0.00% all off!), 95.86% active lanes
    ao.ispc(0071) - function entry: 1122 calls (0 / 0.00% all off!), 97.33% active lanes
    ao.ispc(0075) - return: uniform control flow: 1122 calls (0 / 0.00% all off!), 97.33% active lanes
    ao.ispc(0079) - function entry: 10072 calls (0 / 0.00% all off!), 45.09% active lanes
    ao.ispc(0088) - function entry: 36928 calls (0 / 0.00% all off!), 97.40% active lanes
    ...


Using Scan Operations For Variable Output
-----------------------------------------

One important application of the ``exclusive_scan_add()`` function in the
standard library is when program instances want to generate a variable amount
of output and when one would like that output to be densely packed in a
single array.  For example, consider the code fragment below:

::

    uniform int func(uniform float outArray[], ...) {
       int numOut = ...;  // figure out how many to be output
       float outLocal[MAX_OUT]; // staging area
       // put results in outLocal[0], ..., outLocal[numOut-1]
       int startOffset = exclusive_scan_add(numOut);
       for (int i = 0; i < numOut; ++i)
           outArray[startOffset + i] = outLocal[i];
       return reduce_add(numOut);
    }

Here, each program instance has computed a number, ``numOut``, of values to
output, and has stored them in the ``outLocal`` array.  Assume that four
program instances are running and that the first one wants to output one
value, the second two values, and the third and fourth three values each.
In this case, ``exclusive_scan_add()`` will return the values (0, 1, 3, 6)
to the four program instances, respectively.  The first program instance
will write its one result to ``outArray[0]``, the second will write its two
values to ``outArray[1]`` and ``outArray[2]``, and so forth.  The
``reduce_add`` call at the end returns the total number of values that the
program instances have written to the array.

Application-Supplied Execution Masks
------------------------------------

Recall that when execution transitions from the application code to an
``ispc`` function, all of the program instances are initially executing.
In some cases, it may desired that only some of them are running, based on
a data-dependent condition computed in the application program.  This
situation can easily be handled via an additional parameter from the
application.

As a simple example, consider a case where the application code has an
array of ``float`` values and we'd like the ``ispc`` code to update
just specific values in that array, where which of those values to be
updated has been determined by the application.  In C++ code, we might
have:

::

    int count = ...;
    float *array = new float[count];
    bool *shouldUpdate = new bool[count];
    // initialize array and shouldUpdate
    ispc_func(array, shouldUpdate, count);

Then, the ``ispc`` code could process this update as:

::

    export void ispc_func(uniform float array[], uniform bool update[],
                          uniform int count) {
        for (uniform int i = 0; i < count; i += programCount) {
            cif (update[i+programIndex] == true)
                // update array[i+programIndex]...
        }
    }

(In this case a "coherent" if statement is likely to be worthwhile if the
``update`` array will tend to have sections that are either all-true or
all-false.)

Explicit Vector Programming With Uniform Short Vector Types
-----------------------------------------------------------

The typical model for programming in ``ispc`` is an *implicit* parallel
model, where one writes a program that is apparently doing scalar
computation on values and the program is then vectorized to run in parallel
across the SIMD lanes of a processor.  However, ``ispc`` also has some
support for explicit vector unit programming, where the vectorization is
explicit.  Some computations may be more effectively described in the
explicit model rather than the implicit model.

This support is provided via ``uniform`` instances of short vectors 
(as were introduced in the `Short Vector Types`_ section).  Specifically, 
if this short program

::

    export uniform float<8> madd(uniform float<8> a, 
                                 uniform float<8> b, uniform float<8> c) {
        return a + b * c;
    }

is compiled with the AVX target, ``ispc`` generates the following assembly:

::
    _madd:
	vmulps	%ymm2, %ymm1, %ymm1
	vaddps	%ymm0, %ymm1, %ymm0
	ret

(And similarly, if compiled with a 4-wide SSE target, two ``mulps`` and two
``addps`` instructions are generated, and so forth.)

Note that ``ispc`` doesn't currently support control-flow based on
``uniform`` short vector types; it is thus not possible to write code like:

::

    export uniform int<8> count(uniform float<8> a, uniform float<8> b) {
        uniform int<8> sum = 0;
        while (a++ < b)
            ++sum;
    }


Disclaimer and Legal Information
================================

INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL(R) PRODUCTS.
NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL
PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS
AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER,
AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE
OF INTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A
PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT
OR OTHER INTELLECTUAL PROPERTY RIGHT.

UNLESS OTHERWISE AGREED IN WRITING BY INTEL, THE INTEL PRODUCTS ARE NOT DESIGNED
NOR INTENDED FOR ANY APPLICATION IN WHICH THE FAILURE OF THE INTEL PRODUCT COULD
CREATE A SITUATION WHERE PERSONAL INJURY OR DEATH MAY OCCUR.

Intel may make changes to specifications and product descriptions at any time,
without notice. Designers must not rely on the absence or characteristics of any
features or instructions marked "reserved" or "undefined." Intel reserves these
for future definition and shall have no responsibility whatsoever for conflicts
or incompatibilities arising from future changes to them. The information here
is subject to change without notice. Do not finalize a design with this
information.

The products described in this document may contain design defects or errors
known as errata which may cause the product to deviate from published
specifications. Current characterized errata are available on request.

Contact your local Intel sales office or your distributor to obtain the latest
specifications and before placing your product order.

Copies of documents which have an order number and are referenced in this
document, or other Intel literature, may be obtained by calling 1-800-548-4725,
or by visiting Intel's Web Site.

Intel processor numbers are not a measure of performance. Processor numbers
differentiate features within each processor family, not across different
processor families. See http://www.intel.com/products/processor_number for
details.

BunnyPeople, Celeron, Celeron Inside, Centrino, Centrino Atom,
Centrino Atom Inside, Centrino Inside, Centrino logo, Core Inside, FlashFile,
i960, InstantIP, Intel, Intel logo, Intel386, Intel486, IntelDX2, IntelDX4,
IntelSX2, Intel Atom, Intel Atom Inside, Intel Core, Intel Inside,
Intel Inside logo, Intel. Leap ahead., Intel. Leap ahead. logo, Intel NetBurst,
Intel NetMerge, Intel NetStructure, Intel SingleDriver, Intel SpeedStep,
Intel StrataFlash, Intel Viiv, Intel vPro, Intel XScale, Itanium,
Itanium Inside, MCS, MMX, Oplus, OverDrive, PDCharm, Pentium, Pentium Inside,
skoool, Sound Mark, The Journey Inside, Viiv Inside, vPro Inside, VTune, Xeon,
and Xeon Inside are trademarks of Intel Corporation in the U.S. and other
countries.

* Other names and brands may be claimed as the property of others.

Copyright(C) 2011, Intel Corporation. All rights reserved.


Optimization Notice
===================

Intel compilers, associated libraries and associated development tools may
include or utilize options that optimize for instruction sets that are
available in both Intel and non-Intel microprocessors (for example SIMD
instruction sets), but do not optimize equally for non-Intel
microprocessors.  In addition, certain compiler options for Intel
compilers, including some that are not specific to Intel
micro-architecture, are reserved for Intel microprocessors.  For a detailed
description of Intel compiler options, including the instruction sets and
specific microprocessors they implicate, please refer to the "Intel
Compiler User and Reference Guides" under "Compiler Options."  Many library
routines that are part of Intel compiler products are more highly optimized
for Intel microprocessors than for other microprocessors.  While the
compilers and libraries in Intel compiler products offer optimizations for
both Intel and Intel-compatible microprocessors, depending on the options
you select, your code and other factors, you likely will get extra
performance on Intel microprocessors.

Intel compilers, associated libraries and associated development tools may
or may not optimize to the same degree for non-Intel microprocessors for
optimizations that are not unique to Intel microprocessors.  These
optimizations include Intel® Streaming SIMD Extensions 2 (Intel® SSE2),
Intel® Streaming SIMD Extensions 3 (Intel® SSE3), and Supplemental
Streaming SIMD Extensions 3 (Intel SSSE3) instruction sets and other
optimizations.  Intel does not guarantee the availability, functionality,
or effectiveness of any optimization on microprocessors not manufactured by
Intel.  Microprocessor-dependent optimizations in this product are intended
for use with Intel microprocessors.

While Intel believes our compilers and libraries are excellent choices to
assist in obtaining the best performance on Intel and non-Intel
microprocessors, Intel recommends that you evaluate other compilers and
libraries to determine which best meet your requirements.  We hope to win
your business by striving to offer the best performance of any compiler or
library; please let us know if you find we do not.