ispc/docs/ReleaseNotes.txt

=== v1.9.1 === (8 July 2016)

An ISPC update with new native AVX512 target for future Xeon CPUs and
improvements for debugging, including new switch --dwarf-version to support
debugging on old systems.

The release is based on patched LLVM 3.8.

=== v1.9.0 === (12 Feb 2016)

An ISPC release with AVX512 (KNL flavor) support and a number of bug fixes,
based on fresh LLVM 3.8 backend.

For AVX512 two modes are supported - generic and native. For instructions on how
to use them, please refer to the wiki. Going forward we assume that native mode
is the primary way to get AVX512 support and that generic mode will be deprecated.
If you observe significantly better performance in generic mode, please report
it via github issues.

Starting this release we are shipping two versions on Windows:
(1) for VS2013 and earlier releases
(2) for VS2015 and newer releases
The reason for doing this is the redesigned C run-time library in VS.
An implementation of "print" ISPC standard library function relies on C runtime
library, which has changed. If you are not using "print" function in your code,
you are safe to use either version.

A new options was introduced to improve debugging: --no-omit-frame-pointer.

=== v1.8.2 === (29 May 2015)

An ISPC update with several important stability fixes and an experimental
AVX512 support.

Current level of AVX512 support is targeting the new generation of Xeon Phi
codename Knights Landing. It's implemented in two different ways: as generic and
native target. Generic target is similar to KNC support and requires Intel C/C++
Compiler (15.0 and newer) and is available in regular ISPC build, which is
based on LLVM 3.6.1. For the native AVX512 target, we have a separate ISPC
build, which is based on LLVM trunk (3.7). This build is less stable and has
several known issues. Nevertheless, if you are interested in AVX512 support for
your code, we encourage you to try it and report the bugs. We actively working
with LLVM maintainers to fix all AVX512 bugs, so your feedback is important for
us and will ensure that bugs affecting your code are fixed by LLVM 3.7 release.

Other notable changes and fixes include:

* Broadwell support via --cpu=broadwell.

* Changed cpu naming to accept cpu codenames. Check help for more details.

* --cpu switch disallowed in multi-target mode.

* Alignment of structure fields (in generated header files) is changed to be
  more consistent regardless used C/C++ compiler.

* --dllexport switch is added on Windows to make non-static functions DLL
  export.

* --print-target switch is added to dump details of LLVM target machine.
  This may help you to debug issues with code generation for incorrect target
  (or more likely to ensure that code generation is done right).

* A bug was fixed, which triggered uniform statements to be executed with
  all-off mask under some circumstances.

* The restriction for using some uniform types as return type in multi-target
  mode with targets of different width was relaxed.

Also, if you are using ISPC for code generation for current generation of
Xeon Phi (Knights Corner), the following changes are for you:

* A bunch of stability fixes for KNC.

* A bug, which affects projects with multiple ISPC source files compiled with generic
  target is fixed. As side effect, you may see multiple warnings about unused static
  functions - you need to add "-wd177" switch to ICC compiling generic output files.

The release includes LLVM 3.6.1 binaries for Linux, MacOS, Windows and Windows based
cross-compiler for Sony PlayStation4. LLVM 3.5 based experimental Linux binary with
NVPTX support (now supporting also K80).

Native AVX512 support is available in the set of less stable LLVM 3.7 based binaries
for Linux, MacOS and Windows.

=== v1.8.1 === (31 December 2014)

A minor update of ``ispc`` with several important stability fixes, namely:

* Auto-dispatch mechanism is fixed in pre-built Linux binaries (it used to
  select too conservative target).

* Compile crash with "-O2 -g" is fixed.

Also KNC (Xeon Phi) support is further improved.

The release includes experimental build for Sony PlayStation4 target (Windows
cross compiler), as well NVPTX experimental support (64 bit Linux binaries
only). Note that there might be NVPTX compilation fails with CUDA 7.0.

Similar to 1.8.0 all binaries are based on LLVM 3.5. MacOS binaries are built
for MacOS 10.9 Mavericks. Linux binaries are compatible with kernel 2.6.32
(ok for RHEL6) and later.

=== v1.8.0 === (16 October 2014)

A major new version of ISPC, which introduces experimental support for NVPTX
target, brings numerous improvements to our KNC (Xeon Phi) support, introduces
debugging support on Windows and fixes several bugs. We also ship experimental
build for Sony PlayStation4 target in this release. Binaries for all platforms
are based on LLVM 3.5.

Note that MacOS binaries are build for MacOS 10.9 Mavericks. Linux binaries are
compatible with kernel 2.6.32 (ok for RHEL6) and later.

More details:

* Experimental NVPTX support is available for users of our binary distribution
  on Linux only at the moment. MacOS and Windows users willing to experiment
  with this target are welcome to build it from source. Note that GPU imposes
  some limitation on ISPC language, which are discussed in corresponding section
  of ISPC User's Guide. Implementation of NVPTX support was done by our
  contributor Evghenii Gaburov.

* KNC support was greatly extended in knc.h header file. Beyond new features
  there are stability fixes and changes for icc 15.0 compatibility. Stdlib
  prefetch functions were improved to map to KNC vector prefetches.

* PS4 experimental build is Windows to PS4 cross compiler, which disables arch
  and cpu selection (which are preset to PS4 hardware).

* Debug info support on Windows (compatible with VS2010, VS2012 and VS2013).

* Critical bug fix, which caused code generation for incorrect target, despite
  explicit target switches, under some conditions.

* Stability fix of the bug, which caused print() function to execute under
  all-off mask under some conditions.

=== v1.7.0 === (18 April 2014)

A major new version of ISPC with several language and library extensions and
fixes in debug info support. Binaries for all platforms are based on patched
version on LLVM 3.4. There also performance improvements beyond switchover to
LLVM 3.4.

The list of language and library changes:

* Support for varying types in exported functions was added. See documentation
  for more details.

* get_programCount() function was moved from stdlib.ispc to
  examples/util/util.isph, which needs to be included somewhere in your
  project, if you want to use it.

* Library functions for saturated arithmetic were added. add/sub/mul/div
  operations are supported for signed and unsigned 8/16/32/64 integer types
  (both uniform and varying).

* The algorithm for selecting overloaded function was extended to cover more
  types of overloading. Handling of reference types in overloaded functions was
  fixed. The rules for selecting the best match were changed to match C++,
  which requires the function to be the best match for all parameters. In
  ambiguous cases, a warning is issued, but it will be converted to an error
  in the next release.

* Explicit typecasts between any two reference types were allowed.

* Implicit cast of pointer to const type to void* was disallowed.

The list of other notable changes is:

* Number of fixes for better debug info support.

* Memory corruption bug was fixed, which caused rare but not reproducible
  compile time fails.

* Alias analysis was enabled (more aggressive optimizations are expected).

* A bug involving inaccurate handling of "const" qualifier was fixed. As a
  result, more "const" qualifiers may appear in .h files, which may cause
  compilation errors.

=== v1.6.0 === (19 December 2013)

A major new version of ISPC with major improvements in performance and
stability. Linux and MacOS binaries are based on patched version of LLVM 3.3,
while Windows version is based on LLVM 3.4rc3. LLVM 3.4 significantly improves
stability on Win32 platform, so we've decided not to wait for official LLVM 3.4
release.

The list of the most significant changes is:

* New avx1-i32x4 target was added. It may play well for you, if you are focused
  on integer computations or FP unit in your hardware is 128 bit wide.

* Support for calculations in double precision was extended with two new
  targets avx1.1-i64x4 and avx2-i64x4.

* Language support for overloaded operators was added.

* New library shift() function was added, which is similar to rotate(), but is
  non-circular.

* The language was extended to accept 3 dimensional tasking - a syntactic sugar,
  which may facilitate programming of some tasks.

* Regression, which broke --opt=force-aligned-memory is fixed.

If you are not using pre-built binaries, you may notice the following changes:

* VS2012/VS2013 are supported.

* alloy.py (with -b switch) can build LLVM for you on any platform now
  (except MacOS 10.9, but we know about the problem and working on it).
  This is a preferred way to build LLVM for ISPC, as all required patches for
  better performance and stability will automatically apply.

* LLVM 3.5 (current trunk) is supported.

There are also multiple fixes for better performance and stability, most
notable are:

* Fixed performance problem for x2 targets.

* Fixed a problem with incorrect vzeroupper insertion on AVX target on Win32.

=== v1.5.0 === (27 September 2013)

A major new version of ISPC with several new targets and important bug fixes.
Here's a list of the most important changes, if you are using pre-built
binaries (which are based on patched version of LLVM 3.3):

* The naming of targets was changed to explicitly include data type width and
  a number of threads in the gang. For example, avx2-i32x8 is avx2 target,
  which uses 32 bit types as a base and has 8 threads in a gang. Old naming
  scheme is still supported, but depricated.

* New SSE4 targets for calculations based on 8 bit and 16 bit data types:
  sse4-i8x16 and sse4-i16x8.

* New AVX1 target for calculations based on 64 bit data types: avx1-i64x4.

* SVML support was extended and improved.

* Behavior of -g switch was changed to not affect optimization level.

* ISPC debug infrastructure was redesigned. See --help-dev for more info and
  enjoy capabilities of new --debug-phase=<value> and --off-phase=<value>
  switches.

* Fixed an auto-dispatch bug, which caused AVX code execution when OS doesn't
  support AVX (but hardware does).

* Fixed a bug, which discarded uniform/varying keyword in typedefs.

* Several performance regressions were fixed.

If you are building ISPC yourself, then following changes are also available
to you:

* --cpu=slm for targeting Intel Atom codename Silvermont (if LLVM 3.4 is used).

* ARM NEON targets are available (if enabled in build system).

* --debug-ir=<value> is available to generate debug information based on LLVM
  IR (if LLVM 3.4 is used). In debugger you'll see LLVM IR instead of source
  code.

* A redesigned and improved test and configuration management system is
  available to facilitate the process of building LLVM and testing ISPC
  compiler.

Standard library changes/fixes:

* __pause() function was removed from standard library.

* Fixed reduce_[min|max]_[float|double] intrinsics, which were producing
  incorrect code under some conditions.

Language changes:

* By default a floating point constant without a suffix is a single precision
  constant (32 bit). A new suffix "d" was introduced to allow double precision
  constant (64 bit). Please refer to tests/double-consts.ispc for syntax
  examples.

=== v1.4.4 === (19 July 2013)

A minor version update with several stability fixes requested by the customers.

=== v1.4.3 === (25 June 2013)

A minor version update with several stability improvements:

* Two bugs were fixed (including a bug in LLVM) to improve stability on 32 bit
  platforms.

* A bug affecting several examples was fixed.

* --instrument switch is fixed.

All tests and examples now properly compile and execute on native targets on
Unix platforms (Linux and MacOS).

=== v1.4.2 === (11 June 2013)

A minor version update with a few important changes:

* Stability fix for AVX2 target (Haswell) - problem with gather instructions was
  released in LLVM 3.4, if you build with LLVM 3.2 or 3.3, it's available in our
  repository (llvm_patches/r183327-AVX2-GATHER.patch) and needs to be applied
  manually.

* Stability fix for widespread issue on Win32 platform (#503).

* Performance improvements for Xeon Phi related to mask representation.

Also LLVM 3.3 has been released and now it's the recommended version for building ISPC.
Precompiled binaries are also built with LLVM 3.3. 

=== v1.4.1 === (28 May 2013)

A major new version of ispc has been released with stability and performance
improvements on all supported platforms (Windows, Linux and MacOS). 
This version supports LLVM 3.1, 3.2, 3.3 and 3.4. The released binaries are built with 3.2.

New compiler features:

* ISPC memory allocation returns aligned memory with platform natural alignment
  of vector registers by default. Alignment can also be managed via
  --force-alignment=<value>.

Important bug fixes/changes:

* ISPC was fixed to be fully functional when built by GCC 4.7.

* Major cleanup of build and test scripts on Windows.

* Gather/scatter performance improvements on Xeon Phi.

* FMA instructions are enabled for AVX2 instruction set.

* Support of RDRAND instruction when available via library function rdrand (Ivy Bridge).

Release also contains numerous bug fixes and minor improvements.

=== v1.3.0 === (29 June 2012)

This is a major new release of ispc, with support for more compilation
targets and a number of additions to the language.  As usual, the quality
of generated code has also been improved in a number of cases and a number
of small bugs have been fixed.

New targets:

* This release provides "beta" support for compiling to Intel® Xeon
  Phi™ processor, code named Knights Corner, the first processor in
  the Intel® Many Integrated Core Architecture.  See
  http://ispc.github.com/ispc.html#compiling-for-the-intel-xeon-phi-architecture
  for more details on this support.

* This release also has an "avx1.1" target, which provides support for the
  new instructions in the Intel Ivy Bridge microarchitecutre. 

New language features:

* The foreach_active statement allows iteration over the active program
  instances in a gang.  (See
  http://ispc.github.com/ispc.html#iteration-over-active-program-instances-foreach-active)

* foreach_unique allows iterating over subsets of program instances in a
  gang that share the same value of a variable.  (See
  http://ispc.github.com/ispc.html#iteration-over-unique-elements-foreach-unique)

* An "unmasked" function qualifier and statement in the language allow
  re-activating execution of all program instances in a gang.  (See
  http://ispc.github.com/ispc.html#re-establishing-the-execution-mask

Standard library updates:

* The seed_rng() function has been modified to take a "varying" seed value
  when a varying RNGState is being initialized.

* An isnan() function has been added, to check for floating-point "not a
  number" values.

* The float_to_srgb8() routine does high performance conversion of
  floating-point color values to SRGB8 format.

Other changes:

* A number of bugfixes have been made for compiler crashes with malformed
  programs.

* Floating-point comparisons are now "unordered", so that any comparison
  where one of the operands is a "not a number" value returns false.  (This
  matches standard IEEE floating-point behavior.)

* The code generated for 'break' statements in "varying" loops has been
  improved for some common cases. 

* Compile time and compiler memory use have both been improved,
  particularly for large input programs.

* A nubmer of bugs have been fixed in the debugging information generated
  by the compiler when the "-g" command-line flag is used.

=== v1.2.2 === (20 April 2012)

This release includes a number of small additions to functionality and a
number of bugfixes.  New functionality includes:

* It's now possible to forward declare structures as in C/C++: "struct
  Foo;".  After such a declaration, structs with pointers to "Foo" and
  functions that take pointers or references to Foo structs can be declared
  without the entire definition of Foo being available.

* New built-in types size_t, ptrdiff_t, and [u]intptr_t are now available,
  corresponding to the equivalent types in C.

* The standard library now provides atomic_swap*() and
  atomic_compare_exchange*() functions for void * types.

* The C++ backend has seen a number of improvements to the quality and
  readability of generated code.

A number of bugs have been fixed in this release as well.  The most
significant are:

* Fixed a bug where nested loops could cause a compiler crash in some
  circumstances (issues #240, and #229)

* Gathers could access invlaid mamory (and cause the program to crash) in
  some circumstances (#235)

* References to temporary values are now handled properly when passed to a
  function that takes a reference typed parameter.

* A case where incorrect code could be generated for compile-time-constant
  initializers has been fixed (#234).

=== v1.2.1 === (6 April 2012)

This release contains only minor new functionality and is mostly for many
small bugfixes and improvements to error handling and error reporting.
The new functionality that is present is:

* Significantly more efficient versions of the float / half conversion
  routines are now available in the standard library, thanks to Fabian
  Giesen.

* The last member of a struct can now be a zero-length array; this allows
  the trick of dynamically allocating enough storage for the struct and
  some number of array elements at the end of it.

Significant bugs fixed include:

* Issue #205: When a target ISA isn't specified, use the host system's
  capabilities to choose a target for which it will be able to run the
  generated code.

* Issues #215 and #217: Don't allocate storage for global variables that
  are declared "extern".

* Issue #197: Allow NULL as a default argument value in a function
  declaration.

* Issue #223: Fix bugs where taking the address of a function wouldn't work
  as expected.

* Issue #224: When there are overloaded variants of a function that take
  both reference and const reference parameters, give the non-const
  reference preference when matching values of that underlying type.

* Issue #225: An error is issed when a varying lvalue is assigned to a
  reference type (rather than crashing).

* Issue #193: Permit conversions from array types to void *, not just the
  pointer type of the underlying array element.

* Issue #199: Still evaluate expressions that are cast to (void).

The documentation has also been improved, with FAQs added to clarify some
aspects of the ispc pointer model.

=== v1.2.0 === (20 March 2012)

This is a major new release of ispc, with a number of significant
improvements to functionality, performance, and compiler robustness.  It
does, however, include three small changes to language syntax and semantics
that may require changes to existing programs:

* Syntax for the "launch" keyword has been cleaned up; it's now no longer
  necessary to bracket the launched function call with angle brackets.
  (In other words, now use "launch foo();", rather than "launch < foo() >;".

* When using pointers, the pointed-to data type is now "uniform" by
  default.  Use the varying keyword to specify varying pointed-to types when
  needed.  (i.e. "float *ptr" is a varying pointer to uniform float data,
  whereas previously it was a varying pointer to varying float values.)
  Use "varying float *" to specify a varying pointer to varying float data,
  and so forth.

* The details of "uniform" and "varying" and how they interact with struct
  types have been cleaned up.  Now, when a struct type is declared, if the
  struct elements don't have explicit "uniform" or "varying" qualifiers,
  they are said to have "unbound" variability.  When a struct type is
  instantiated, any unbound variability elements inherit the variability of
  the parent struct type. See http://ispc.github.com/ispc.html#struct-types
  for more details.

ispc has a new language feature that makes it much easier to use the
efficient "(array of) structure of arrays" (AoSoA, or SoA) memory layout of
data.  A new "soa<n>" qualifier can be applied to structure types to
specify an n-wide SoA version of the corresponding type.  Array indexing
and pointer operations with arrays SoA types automatically handles the
two-stage indexing calculation to access the data.  See
http://ispc.github.com/ispc.html#structure-of-array-types for more details.

For more efficient access of data that is still in "array of structures"
(AoS) format, ispc has a new "memory coalescing" optimization that
automatically detects series of strided loads and/or gathers that can be
transformed into a more efficient set of vector loads and shuffles.  A
diagnostic is emitted when this optimization is successfully applied. 

Smaller changes in this release:

* The standard library now provides memcpy(), memmove() and memset()
  functions, as well as single-precision asin() and acos() functions.

* -I can now be specified on the command-line to specify a search path for
  #include files.

* A number of improvements have been made to error reporting from the
  parser, and a number of cases where malformed programs could cause the
  compiler to crash have been fixed.

* A number of small improvements to the quality and performance of generated
  code have been made, including finding more cases where 32-bit addressing
  calculations can be safely done on 64-bit systems and generating better
  code for initializer expressions.

=== v1.1.4 === (4 February 2012)

There are two major bugfixes for Windows in this release.  First, a number
of failures in AVX code generation on Windows have been fixed; AVX on
Windows now has no known issues.  Second, a longstanding bug in parsing 64-bit
integer constants on Windows has been fixed.

This release features a new experimental scalar target, contributed by Gabe
Weisz <gweisz@cs.cmu.edu>.  This target ("--target=generic-1") compiles
gangs of single program instances (i.e. programCount == 1); it can be
useful for debugging ispc programs.

The compiler now supports dynamic memory allocation in ispc programs (with
"new" and "delete" operators based on C++).  See
http://ispc.github.com/ispc.html#dynamic-memory-allocation in the
documentation for more information.

ispc now performs "short circuit" evaluation of the || and && logical
operators and the ? : selection operator.  (This represents the correction
of a major incompatibility with C.)  Code like "(index < arraySize &&
array[index] == 1)" thus now executes as in C, where "array[index]" won't
be evaluated unless "index" is less than "arraySize".

The standard library now provides "local" atomic operations, which are
atomic across the gang of program instances (but not across other gangs or
other hardware threads.  See the updated documentation on atomics for more
information:
http://ispc.github.com/ispc.html#atomic-operations-and-memory-fences.

The standard library now offers a clock() function, which returns a uniform
int64 value that counts processor cycles; it can be used for
fine-resolution timing measurements.

Finally (of limited interest now): ispc now supports the forthcoming AVX2
instruction set, due with Haswell-generation CPUs.  All tests and examples
compile and execute correctly with AVX2.  (Thanks specifically to Craig
Topper and Nadav Rotem for work on AVX2 support in LLVM, which made this
possible.)
 
=== v1.1.3 === (20 January 2012)

With this release, the language now supports "switch" statements, with the
same semantics and syntax as in C.

This release includes fixes for two important performance related issues:
the quality of code generated for "foreach" statements has been
substantially improved (https://github.com/ispc/ispc/issues/151), and a
performance regression with code for "gathers" that was introduced in
v1.1.2 has been fixed in this release. 

A number of other small bugs were fixed in this release as well, including
one where invalid memory would sometimes be incorrectly accessed
(https://github.com/ispc/ispc/issues/160).

Thanks to Jean-Luc Duprat for a number of patches that improve support for
building on various platforms, and to Pierre-Antoine Lacaze for patches so
that ispc builds under MinGW.

=== v1.1.2 === (9 January 2012)

The major new feature in this release is support for "generic" C++
vectorized output; in other words, ispc can emit C++ code that corresponds
to the vectorized computation that the ispc program represents.  See the
examples/intrinsics directory in the ispc distribution for two example
implementations of the set of functions that must be provided map the
vector calls generated by ispc to target specific functions.

ispc now has partial support for 'goto' statements; specifically, goto is
allowed if any enclosing control flow statements (if/for/while/do) have
'uniform' test expressions, but not if they have 'varying' tests.

A number of improvements have been made to the code generated for gathers
and scatters--one of them (better matching x86's "free" scale by 2/4/8 for
addressing calculations) improved the performance of the noise example by
14%.

Many small bugs have been fixed in this release as well, including issue
numbers 138, 129, 135, 127, 149, and 142.

=== v1.1.1 === (15 December 2011)

This release doesn't include any significant new functionality, but does
include a small improvements in generated code and a number of bug fixes.

The one user-visible language change is that integer constants may be
specified with 'u' and 'l' suffixes, like in C.  For example, "1024llu"
defines the constant with unsigned 64-bit type.

More informative and useful error messages are printed when function
overload resolution fails.

Masking is avoided in additional cases when the mask can be
statically-determined to be all on. 

A number of small bugs have been fixed:
- Under some circumstances, incorrect masks were used when assigning a
  value to a reference and when doing gathers/scatters.
- Incorrect code could be generated in some cases when some instances
  returned part way through a function but others contineud executing.
- Type checking wasn't being performed for calls through function pointers;
  now an error is issued if the arguments don't match up, etc.
- Incorrect code was being generated for gather/scatter to structs that had
  elements with varying short-vector types.
- Typechecking wasn't being performed for "foreach" statements; this led to
  problems like function overload resolution not being performed if an
  overloaded function call was used to determine the iteration range..
- A number of symbols would be multiply-defined when compiling to multiple
  targets and using the sse2-x2 target as one of them (issue #131).

=== v1.1.0 === (5 December 2011)

This is a major new release of the compiler, with significant additions to
language functionality and capabilities.  It includes a number of small
language syntax changes that will require modification of existing
programs.  These changes should generally be straightforward and all are
steps toward eliminating parts of ispc syntax that are incompatible with
C/C++.  See
http://ispc.github.com/ispc.html#updating-ispc-programs-for-changes-in-ispc-1-1
for more information about these changes.

ispc now fully supports pointers, including pointer arithmetic, implicit
conversions of arrays to pointers, and all of the other capabilities of
pointers in C.  See http://ispc.github.com/ispc.html#pointer-types for more
information about pointers in ispc and
http://ispc.github.com/ispc.html#function-pointer-types for information
about function pointers in ispc.

Reference types are now declared with C++ syntax (e.g. "const float &foo").

ispc now supports 64-bit addressing.  For performance reasons, this
capability is disabled by default (even on 64-bit targets), but can be
enabled with a command-line flag:
http://ispc.github.com/ispc.html#selecting-32-or-64-bit-addressing.

This release features new parallel "foreach" statements, which make it
easier in many instances to map program instances to data for data-parallel
computation than the programIndex/programCount mechanism:
http://ispc.github.com/ispc.html#parallel-iteration-statements-foreach-and-foreach-tiled.

Finally, all of the system's documentation has been significantly revised.
The documentation of ispc's parallel execution model has been rewritten:
http://ispc.github.com/ispc.html#the-ispc-parallel-execution-model, and
there is now a more specific discussion of similarities and differences
between ispc and C/C++:
http://ispc.github.com/ispc.html#relationship-to-the-c-programming-language.
There is now a separate FAQ (http://ispc.github.com/faq.html), and a
Performance Guide (http://ispc.github.com/perfguide.html).
 
=== v1.0.12 === (20 October 2011)

This release includes a new "double-pumped" 8-wide target for SSE2,
"sse2-x2".  Like the sse4-x2 and avx-x2 targets, this target may deliver
higher performance for some workloads than the regular sse2 target.  (For
other workloads, it may be slower.)

The ispc language now includes an "assert()" statement.  See
http://ispc.github.com/ispc.html#assertions for more information.

The compiler now sets a preprocessor #define based on the target ISA; for
example, ISPC_TARGET_SSE4 is defined for the sse4 targets, and so forth.

The standard library now provides high-performance routines for converting
between some "array of structures" and "structure of arrays" formats.
See
http://ispc.github.com/ispc.html#converting-between-array-of-structures-and-structure-of-arrays-layout
for more information.

Inline functions now have static linkage.

A number of improvements have been made to the optimization passes that
detect when gathers and scatters can be transformed into vector stores and
loads, respectively.  In particular, these passes now handle variables that
are used as loop induction variables much better.

=== v1.0.11 === (6 October 2011)

The main new feature in this release is support for generating code for
multiple targets (e.g., SSE2, SSE4, and AVX) and having the compiled code
select the best variant at execution time.  For more information, see
http://ispc.github.com/ispc.html#compiling-with-support-for-multiple-instruction-sets.

All of the examples now take advantage of the support for multiple
compilation targets; thus, if one has an AVX system, it's not necessary to
recompile the examples to use the AVX target.

Performance of the built-in task system that is used in the examples has
been improved.

Finally, the print() statement now works on OSX; it had been broken for the
last few releases.

=== v1.0.10 === (30 September 2011)

This release features an extensive new example showing the application of
ispc to a deferred shading algorithm for scenes with thousands of lights
(examples/deferred).  This is an implementation of the algorithm that Johan
Andersson described at SIGGRAPH 2009 and was implemented by Andrew
Lauritzen and Jefferson Montgomery.  The basic idea is that a pre-rendered
G-buffer is partitioned into tiles, and in each tile, the set of lights
that contribute to the tile is computed.  Then, the pixels in the tile are
then shaded using those light sources. (See slides 19-29 of
http://s09.idav.ucdavis.edu/talks/04-JAndersson-ParallelFrostbite-Siggraph09.pdf
for more details on the algorithm.)

The mechanism for launching tasks from ispc code has been generalized to
allow multiple tasks to be launched with a single launch call (see
http://ispc.github.com/ispc.html#task-parallelism-language-syntax for more
information.)

A few new functions have been added to the standard library: num_cores()
returns the number of cores in the system's CPU, and variants of all of the
atomic operators that take 'uniform' values as parameters have been added.

=== v1.0.9 === (26 September 2011)

The binary release of v1.0.9 is the first that supports AVX code
generation.  Two targets are provided: "avx", which runs with a
programCount of 8, and "avx-x2" which runs 16 program instances
simultaneously.  (This binary is also built using the in-progress LLVM 3.0
development libraries, while previous ones have been built with the
released 2.9 version of LLVM.)

This release has no other significant changes beyond a number of small
bugfixes (https://github.com/ispc/ispc/issues/100,
https://github.com/ispc/ispc/issues/101, https://github.com/ispc/ispc/issues/103.)
 
=== v1.0.8 === (19 September 2011)

A number of improvements have been made to handling of 'if' statements in
the language:
  - A bug was fixed where invalid memory could be incorrectly accessed even
    if none of the running program instances wanted to execute the
    corresponding instructions (https://github.com/ispc/ispc/issues/74).
  - The code generated for 'if' statements is a bit simpler and thus more
    efficient.

There is now '--pic' command-line argument that causes position-independent
code to be generated (Linux and OSX only).

A number of additional performance improvements:
  - Loops are now unrolled by default; the --opt=disable-loop-unroll
    command-line argument can be used to disable this behavior.
    (https://github.com/ispc/ispc/issues/78)
  - A few more cases where gathers/scatters could be determined at compile
    time to actually access contiguous locations have been added.
    (https://github.com/ispc/ispc/issues/79)

Finally, warnings are now issued (if possible) when it can be determined
at compile-time that an out-of-bounds array index is being used.
(https://github.com/ispc/ispc/issues/98).


=== v1.0.7 === (3 September 2011)

The various atomic_*_global() standard library functions are generally
substantially more efficient.  They all previously issued one hardware
atomic instruction for each running program instance but now locally
compute a reduction over the operands and issue a single hardware atomic,
giving the same effect and results in the end (issue #57).

CPU/ISA target handling has been substantially improved.  If no CPU is
specified, the host CPU type is used, not just a default of "nehalem".  A
number of bugs were fixed that ensure that LLVM doesn't generate SSE>2
instructions when using the SSE2 target (fixes issue #82).

Shift rights of unsigned integer types use a logical shift right
instruction now, not an arithmetic shift right (fixed issue #88).

When emitting header files, 'extern' declarations of globals used in ispc
code are now outside of the ispc namespace.  Fixes issue #64.

The stencil example has been modified to do runs with and without
parallelism.

Many other small bugfixes and improvements.

=== v1.0.6 === (17 August 2011)

Some additional cross-program instance operations have been added to the
standard library.  reduce_equal() checks to see if the given value is the
same across all running program instances, and exclusive_scan_{and,or,and}()
computes a scan over the given value in the running program instances.
See the documentation of these new routines for more information:
http://ispc.github.com/ispc.html#cross-program-instance-operations.

The simple task system implementations used in the examples have been
improved.  The Windows version no nlonger has a hard limit on the number of
tasks that can be launched, and all versions have less dynamic memory
allocation and less locking.  More of the examples now have paths that also
measure performance using tasks along with SPMD vectorization.

Two new examples have been added: one that shows the implementation of a
ray-marching volume rendering algorithm, and one that shows a 3D stencil
computation, as might be done for PDE solutions.

Standard library routines to issue prefetches have been added.  See the
documentation for more details: http://ispc.github.com/ispc.html#prefetches.

Fast versions of the float to half-precision float conversion routines have
been added.  For more details, see:
http://ispc.github.com/ispc.html#conversions-to-and-from-half-precision-floats.

There is the usual set of small bug fixes.  Notably, a number of details
related to handling 32 versus 64 bit targets have been fixed, which in turn
has fixed a bug related to tasks having incorrect values for pointers
passed to them.

=== v1.0.5 === (1 August 2011)

Multi-element vector swizzles are supported; for example, given a 3-wide
vector "foo", then expressions like "foo.zyx" and "foo.yz" can be used to
construct other short vectors.  See
http://ispc.github.com/ispc.html#short-vector-types
for more details.  (Thanks to Pete Couperus for implementing this code!).

int8 and int16 datatypes are now supported.  It is still generally more
efficient to use int32 for intermediate computations, even if the in-memory
format is int8 or int16.

There are now standard library routines to convert to and from 'half'-format
floating-point values (half_to_float() and float_to_half()).

There is a new example with an implementation of Perlin's Noise function
(examples/noise).  It shows a speedup of approximately 4.2x versus a C
implementation on OSX and a 2.9x speedup versus C on Windows.

=== v1.0.4 === (18 July 2011)

enums are now supported in ispc; see the section on enumeration types in
the documentation (http://ispc.github.com/ispc.html#enumeration-types) for
more informaiton.

bools are converted to integers with zero extension, not sign extension as
before (i.e. a 'true' bool converts to the value one, not 'all bits on'.)
For cases where sign extension is still desired, there is a
sign_extend(bool) function in the standard library.

Support for 64-bit types in the standard library is much more complete than
before.

64-bit integer constants are now supported by the parser.

Storage for parameters to tasks is now allocated dynamically on Windows,
rather than on the stack; with this fix, all tests now run correctly on
Windows.

There is now support for atomic swap and compare/exchange with float and
double types.

A number of additional small bugs have been fixed and a number of cases
where the compiler would crash given a malformed program have been fixed.

=== v1.0.3 === (4 July 2011)

ispc now has a bulit-in pre-processor (from LLVM's clang compiler).
(Thanks to Pete Couperus for this patch!)  It is therefore no longer
necessary to use cl.exe for preprocessing on Windows; the MSVC proejct
files for the examples have been updated accordingly.

There is another variant of the shuffle() function int the standard
library: "<type> shuffle(<type> v0, <type> v1, int permute)", where the
permutation vector indexes over the concatenation of the two vectors
(e.g. the value 0 corresponds to the first element of v0, the value
2*programCount-1 corresponds to the last element of v1, etc.)

ispc now supports the usual range of atomic operations (add, subtract, min,
max, and, or, and xor) as well as atomic swap and atomic compare and
exchange.  There is also a facility for inserting memory fences.  See the
"Atomic Operations and Memory Fences" section of the user's guide
(http://ispc.github.com/ispc.html#atomic-operations-and-memory-fences) for
more information.
 
There are now both 'signed' and 'unsigned' variants of the standard library
functions like packed_load_active() that take references to arrays of
signed int32s and unsigned int32s respectively.  (The
{load_from,store_to}_{int8,int16}() functions have similarly been augmented
to have both 'signed' and 'unsigned' variants.)

In initializer expressions with variable declarations, it is no longer
legal to initialize arrays and structs with single scalar values that then
initialize their members; they now must be initialized with initializer
lists in braces (or initialized after of the initializer with a loop over
array elements, etc.)

=== v1.0.2 === (1 July 2011)

Floating-point hexidecimal constants are now parsed correctly on Windows
(fixes issue #16).

SSE2 is now the default target if --cpu=atom is given in the command line
arguments and another target isn't explicitly specified.

The standard library now provides broadcast(), rotate(), and shuffle()
routines for efficient communication between program instances.

The MSVC solution files to build the examples on Windows now use
/fpmath:fast when building.

=== v1.0.1 === (24 June 2011)

ispc no longer requires that pointers to memory that are passed in to ispc
have alignment equal to the targets vector width; now alignment just has to
be the regular element alignment (e.g. 4 bytes for floats, etc.)  This
change also fixed a number of cases where it previously incorrectly
generated aligned load/store instructions in cases where the address wasn't
actually aligned (even if the base address passed into ispc code was).

=== v1.0 === (21 June 2011)

Initial Release