redgpu/ispc: Intel® Implicit SPMD Program Compiler

docs/ReleaseNotes.txt
1 715 строк75 KB
Arina Neshlyaeva
Update Release Notes for v1.21.0
18 авг 2023, 00:03
18 авг 2023, 00:03f99deb1
=== v1.21.0 === (18 August 2023)

ISPC release with template function specializations support; changed rules for signed integer overflow, which match
C/C++ behavior and lead to more aggressive optimizations; an enhanced ISPC Runtime; multiple stability and performance
fixes and more. The release is based on patched LLVM 15.0.7.

Language changes:

 - Added support for function template specializations with explicit template arguments.
   For more details please refer to Function Templates section of documentation.

 - Modified behavior for signed integer overflow.
   Now, in case of signed integer overflow, `ispc` will assume undefined behavior similar to C and C++. This change may
   cause compatibility issues. You can manage this behavior by using the `--[no-]wrap-signed-int` compiler switch. The default
   behavior (before version 1.21.0) can be preserved by using `--wrap-signed-int`, which maintains defined wraparound
   behavior for signed integers, though it may limit some compiler optimizations.

New hardware support:

Added support of Intel Meteor Lake Xe-LPG graphics:

 - Added two new ISPC targets: `xelpg-x16` and `xelpg-x8`
 - Added two new device names: `mtl-m` and `mtl-p`

Infrastructure changes:

 - ISPC now uses LLVM's new pass manager. Optimization pipeline was modified by introducing early LoopFullUnrollPass
   which matches ISPC unrolled loops with manually unrolled loops in many cases.
 - Introduced ISPC superbuild, which facilitates building ISPC with Xe dependencies (LLVM, L0, vc-intrinsics,
   SPIRV-Translator). It can generate an archive with dependencies or consume a pre-built archive to build ISPC only.
   It also enables generating LTO or LTO+PGO enabled builds of LLVM and ISPC.
 - Supported building ISPC with LLVM 16.

New compiler switches:

 - `--mcmodel` switch, which accepts `small` and `large` values. The definition is similar to gcc/clang. When `large`
   model is used, it enables programs larger than 2Gb.
 - `--opt=disable-gathers` and `--opt=disable-scatters` options, which disable generation of gathers and scatters
   instructions on platforms that support them (for performance experiments).
 - `--[no-]wrap-signed-int` switches, which [does not] preserve(s) wrap-around behavior on signed integer overflow.

ISPC Runtime improvements:

 - Added `ispcrtSetTaskingCallbacks` to the ISPCRT API, allowing the override of default implementations of
   `ISPCLaunch`, `ISPCAlloc`, and `ISPCSync`.
 - Removed compile-time Level Zero dependency from ISPCRT, no longer necessary after the ISPCRT split into CPU and GPU
   parts.


Recommended versions of Runtime Dependencies when targeting GPU:

Linux:

 - Intel(R) Graphics Compute Runtime
   https://github.com/intel/compute-runtime/releases/tag/23.22.26516.18
 - Level Zero Loader
   https://github.com/oneapi-src/level-zero/releases/tag/v1.13.5
 - Threading Building Blocks (TBB)

Alternatively, you can use a validated gfx driver stack supporting Intel® Arc™
available at https://dgpu-docs.intel.com/driver/installation.html

Windows:

 - Intel(R) Graphics Windows(R) DCH Drivers 31.0.101.4644
   https://www.intel.com/content/www/us/en/download/726609/intel-arc-iris-xe-graphics-whql-windows.html
 - Level Zero Loader
   https://github.com/oneapi-src/level-zero/releases/tag/v1.13.5
 - OpenCL™ Offline Compiler (OCLOC)
   https://www.intel.com/content/www/us/en/developer/articles/tool/oneapi-standalone-components.html
   (this is needed for AoT compilation on Windows only)
 - Supported GPU platforms: Intel(R) Arc Graphics, 11th-13th Gen Intel(R) Core
   processor graphics

Components revisions used in GPU-enabled build:
 - https://github.com/KhronosGroup/SPIRV-LLVM-Translator/commit/e82ecc2
 - https://github.com/intel/vc-intrinsics/commit/910db48
 - https://github.com/oneapi-src/level-zero/commit/e1f09b4 (v1.13.5)
 - https://github.com/llvm/llvm-project/commit/8dfdcc7 (llvmorg-15.0.7) +
   patches from llvm_patches folder

=== v1.20.0 === (5 May 2023)

ISPC release with compile time improvements, enhancements in the ISPC Runtime,
and a number of code generation fixes. The release is based on patched LLVM
15.0.7.

ISPC distribution changes.

ISPC binaries got faster and smaller. ISPC binaries got smaller approximately
by 1/3 and a few percent faster. The distribution macOS now includes x86_64,
arm64 and Universal Binaries. On Linux a snap package with the latest ISPC is
available.

ISPC Runtime.

 - ispcrt was split under the hood into GPU and CPU parts, which are loaded
   dynamically. This means you don't need GPU dependencies when running CPU-only
   code using ispcrt.
 - ispcrt got support for fences to enable CPU/GPU asynchronous computations.
 - ispcrt does not depend on OpenMP runtime anymore, but requires TBB.

New targets.

For better fine-tuning when targeting old platforms, sse4 targets were split
into sse4.1 and sse4.2 targets. All changes are backward compatible - sse4 are
aliased to sse4.2 and multi-target compilation allows only one of sse4 target,
so build systems are not confused.

Improvements for contributors

We got a brand new Github Codespaces config, so you are welcome to start
hacking on ISPC in browser. Give it a try!

Recommended versions of Runtime Dependencies when targeting GPU.

Linux:

 - Intel(R) Graphics Compute Runtime
   https://github.com/intel/compute-runtime/releases/tag/23.09.25812.14
 - Level Zero Loader
   https://github.com/oneapi-src/level-zero/releases/tag/v1.10.0
 - Threading Building Blocks (TBB)

Alternatively, you can use a validated gfx driver stack supporting Intel® Arc™
available at https://dgpu-docs.intel.com/releases/stable_602_20230323.html

Windows:

 - Intel(R) Graphics Windows(R) DCH Drivers 31.0.101.4146
   https://www.intel.com/content/www/us/en/download/726609/772016/intel-arc-iris-xe-graphics-whql-windows.html
 - Level Zero Loader
   https://github.com/oneapi-src/level-zero/releases/tag/v1.10.0
 - OpenCL™ Offline Compiler (OCLOC)
   https://www.intel.com/content/www/us/en/developer/articles/tool/oneapi-standalone-components.html
   (this is needed for AoT compilation on Windows only)
 - Supported platforms: Intel(R) Arc Graphics, 11th-13th Gen Intel(R) Core
   processor graphics

Components revisions used in GPU-enabled build:
 - https://github.com/KhronosGroup/SPIRV-LLVM-Translator/commit/855eb27
 - https://github.com/intel/vc-intrinsics/commit/29fe787
 - https://github.com/oneapi-src/level-zero/commit/0d56d8e (v1.10.0)
 - https://github.com/llvm/llvm-project/commit/8dfdcc7 (llvmorg-15.0.7) +
   patches from llvm_patches folder


=== v1.19.0 === (28 February 2022)

ISPC release with long-awaited function templates technical preview; new
hardware support for 4th generation Intel® Xeon® Scalable (codename Sapphire
Rapids) CPUs, Intel® Data Center GPU Max (codename Ponte Vecchio), and updated
support for Intel® Arc™ GPUs; improved performance and compile time; an enhanced
ISPC Runtime; a bunch of stability fixes and more. The release is based on
patched LLVM 14.0.6.

Language changes.

Function templates support was introduced in ISPC and it's currently in
technical preview, meaning that current language definition might change in
future versions. For more details please refer to Function Templates
section of documentation.

ISPC has got several other language changes needed for ISPC/SYCL
interoperability (an experimental feature):
1. Support of `__regcall` attribute.
2. A new language construct `invoke_sycl` which is used to call SYCL function
   from ISPC. The function must be declared on ISPC side with `extern "SYCL"
   __regcall` qualifiers.
3. Support of `extern "C"` functions definitions.

New hardware support.

1. Targets for 4th generation Intel® Xeon® Scalable (codename Sapphire Rapids)
   CPUs were introduced: `avx512spr-x4`, `avx512spr-x8`,`avx512spr-x16`,
   `avx512spr-x32`, `avx512spr-x64`. The key difference with other AVX512 targets
   is native support for FP16.
2. New `xehpc-x16`/`xehpc-x32` targets were added for Intel® Data Center GPU Max
   (codename Ponte Vecchio). A new `pvc` device name was introduced.
3. New device names `acm-g10`, `acm-g11`, and `acm-g12` were added for Intel®
   Arc™ Graphics. The `dg2` device name has been removed.
4. Support for Aarch64 targets was enabled on Windows.

ISPC Runtime.

1. A chunking allocator was introduced that can be enabled with `ISPCRT_MEM_POOL`.
2. An API was added to link input modules through `ispcrtStaticLinkModules`
   (using linking on vISA level under the hood) and `ispcrtDynamicLinkModules`
   (using binary linking under the hood).
3. Support for creating multiple devices within a single context was added, and
   an API was added to get a function pointer from a module. It's also possible to
   construct ISPC RT objects from native handlers now.
4. ISPC RT verbose mode was added that can be enabled through `ISPCRT_VERBOSE`.

Performance.

There's a significant performance boost on Xe targets caused by updates in the
ISPC optimization pipeline and the usage of the new spill-cost IGC finalizer
function, which dramatically reduces spill size.

Utilities.

1. ISPC `link` mode has been introduced, allowing to link several LLVM bitcode
   or SPIR-V files and output the result as LLVM bitcode or SPIR-V. For example:
     ispc link test_a.bc test_b.bc --emit-spirv -o test.spv

2. CMake utilities was improved, and support was added for building an ISPC GPU
   target from multiple ISPC files, linking them with `ispc --link`. An
   application's ISPC CMakeLists would look like this:
     add_ispc_library(my_ispc_lib filea.ispc fileb.ispc)
     ispc_target_include_directories(my_ispc_lib <some directory path>)
     ispc_target_compile_definitions(my_ispc_lib -DMY_DEFINE=1)

     add_ispc_library(my_ispc_kernel filec.ispc)
     ispc_target_link_libraries(my_ispc_kernel my_ispc_lib)

Runtime Dependencies when targeting GPU.

Linux:

- Intel(R) Graphics Compute Runtime
  https://github.com/intel/compute-runtime/releases/tag/22.49.25018.24
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.9.4
- OpenMP Runtime. Consult your Linux distribution documentation for the
  installation of OpenMP runtime instructions. No specific version is required.

Windows:

- Intel(R) Graphics Windows(R) DCH Drivers 30.0.101.4091
  https://www.intel.com/content/www/us/en/download/726609/intel-arc-iris-xe-graphics-whql-windows.html
- Level Zero Loader https://github.com/oneapi-src/level-zero/releases/tag/v1.9.4
- OpenCL™ Offline Compiler (OCLOC)
  https://www.intel.com/content/www/us/en/developer/articles/tool/oneapi-standalone-components.html
  (this is needed for AoT compilation on Windows only)

Components revisions used in GPU-enabled build:

- https://github.com/KhronosGroup/SPIRV-LLVM-Translator/commit/c469fa8
- https://github.com/intel/vc-intrinsics/commit/3ac855c
- https://github.com/oneapi-src/level-zero/commit/4ed13f3 (v1.9.4)
- https://github.com/llvm/llvm-project/commit/f28c006 (llvmorg-14.0.6) +
  patches from llvm_patches folder

=== v1.18.0 === (5 May 2022)

An ISPC release with a bunch of stability and performance fixes, improvements
for ISPC Runtime, and complete stdlib support for `float16` type. This release
is based on patched LLVM 13.0.1.

`-E` switch was introduced to run preprocessor only. An old bug preventing the
compiler to crash in case of preprocessor error was fixed and now the compiler
will properly crash. As some users considered an old behavior convenient in
some cases, `--ignore-preprocessor-errors` switch was introduced to maintain
the old behavior.

Targets naming was changed for the targets with native masking support to drop
"base type" from the naming scheme, the old naming is accepted for
compatibility. This affected AVX512 target names, the new names are
`avx512skx-x4`, `avx512skx-x8`, `avx512skx-x16`, `avx512skx-x32`,
`avx512skx-x64`, and `avx512knl-x16`.

For debugging and for those, who are interested in understanding compiler
internals, `--ast-dump` switch was introduced. The produced dump of AST
(Abstract Syntax Tree) is intentionally made to look like clang AST dump for
convenience.

Standard library gained full support for `float16` type. Note that it is fully
supported only on the targets with native hardware support. On the other
targets emulation is still not guaranteed but may work in some cases.

Among other fixes, it is worth mentioning the following:
- fixed a bug #1308 affecting multi-target compilation
- a bunch of fixes to make it easier to build ISPC on FreeBSD, even though
  FreeBSD is not officially supported

Improvements for the ISPC Runtime in this release:
- flexible task system selection during build
- support of ISPCRT build separate from ISPC
- support of ISPCRT build for CPU only
- version check in CMake
- new API to get the type of allocated memory (`ispcrtGetMemoryViewAllocType`
  and `ispcrtGetMemoryAllocType`)
- new API for memory copy on device (`ispcrtCopyMemoryView`)
- support of device-only memory without corresponding application memory.

Performance on Xe targets was significantly improved in this release due to
optimizations in ISPC and Vector Backend.

Runtime Dependencies when targeting GPU:

Linux:

- Intel(R) Graphics Compute Runtime
  https://github.com/intel/compute-runtime/releases/tag/22.17.23034
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.7.15
- OpenMP Runtime. Consult your Linux distribution documentation for the
  installation of OpenMP runtime instructions. No specific version is required.

Windows:

- Intel(R) Graphics Windows(R) DCH Drivers 30.0.101.1660
  https://www.intel.com/content/www/us/en/download/19344/intel-graphics-windows-dch-drivers.html
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.7.15
- OpenCL™ Offline Compiler (OCLOC)
  https://registrationcenter-download.intel.com/akdlm/irc_nas/18653/ocloc_win_101.1660.zip
  (this is needed for AoT compilation on Windows only)

Components revisions used in GPU-enabled build:

- https://github.com/KhronosGroup/SPIRV-LLVM-Translator/commit/d7a0304
- https://github.com/intel/vc-intrinsics/commit/1e2562d
- https://github.com/oneapi-src/level-zero/commit/bb7fff0 (v1.7.15)
- https://github.com/llvm/llvm-project/commit/75e33f7 (llvmorg-13.0.1) +
  patches from llvm_patches folder

=== v1.17.0 === (14 January 2022)

An ISPC release with massive update of Xe targets, including support for
forthcoming XeHPG GPUs, improvements for `double` type on AVX512 targets, and
multiple standard library improvements. Windows and Linux binaries in this
release support both CPU and GPU targets, while macOS binary supports only CPU.
This release is based on patched LLVM 12.0.1.

Improvements for CPU targets:
- Performance improvements for `double` type on AVX512 targets - better use of
  gather/scatter instructions, 2-5x improvements for `rsqrt()` and `rcp()`
  standard library functions.
- New `avx512skx-i32x4` target.
- `aos_to_soa` and `soa_to_aos` performance improvements for `-x8` and `-x16`
  targets on CPU.
- `--math-lib=svml` mode was fixed and extended - it requires Intel® C++
  Compiler (`icc` or `icx`) to link the binary.
- `zen1`, `zen2`, and `zen3` CPU definitions were added.
- Added experimental support for PS5 platform.

ISPC language got experimental support for IEEE 754 half-precision data type -
`float16`. Not all library functions are supported yet with this type. The key
focus in this release was on hardware natively supporting this type.

This update includes breaking changes in compiler switches for Xe targets:
- Graphics targets `genx-x8` and `genx-x16` were renamed to `gen9-x8` and
  `gen9-x16`.
- Compiler architectures for graphics target were renamed from `genx32` and
  `genx64` to `xe32` and `xe64`.
- Xe targets were renamed from uppercase to lowercase (so instead of SKL/TGLLP
  it is now skl/tgllp).
- A new `--device` switch (which is an alias for the existing `--cpu` switch)
  was introduced. Now the recommended way to specify the required platform for
  CPU and GPU is: `--device=<platform>`

Also this release changes `export` and `task` functions definition on GPU.  Now
GPU kernel is ISPC `task` function only, `export` functions cannot be invoked
from host (i.e. called from ISPC Runtime/L0 Runtime) anymore. `export`
functions are ready to be linked with and called from other GPU modules.
Currently, ISPC experimentally supports such interoperability with Explicit
SIMD SYCL* Extension (ESIMD).

New Xe targets were added:
- `xelp-x8` and `xelp-x16`. XeLP refers to XeLP generation of hardware
  (TigerLake chips and alike).
- `xehpg-x8` and `xehpg-x16`. XeHPG is the architecture name for the
  forthcoming Intel® Arc™ GPUs codename Alchemist..

GPU part has a bunch of stability, performance, and usability improvements
including but not limited to `alloca()` with constant parameter support,
`assume()` support, improved performance for double math functions and integer
division.

`ISPC Runtime` performance was improved several times by fixing the setting of
local group size for kernels, using events as a synchronization mechanism, and
utilizing HW compute and copy engines. There is also a new structure
`ISPCRTModuleOptions` to pass additional options to VC backend if needed.
Currently, `ISPCRTModuleOptions` allows setting of stack size for VC backend
which is used to compile SPIR-V.

Runtime Dependencies when targeting GPU:

Linux:

- Intel(R) Graphics Compute Runtime
  https://github.com/intel/compute-runtime/releases/tag/22.02.22151
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.7.4
- OpenMP Runtime. Consult your Linux distribution documentation for the
  installation of OpenMP runtime instructions. No specific version is required.

Windows:

- Intel(R) Graphics Windows(R) DCH Drivers 30.0.101.1191
  https://www.intel.com/content/www/us/en/download/19344/intel-graphics-windows-dch-drivers.html
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.7.4
- OpenCL™ Offline Compiler (OCLOC)
  https://software.intel.com/sites/downloads/ocloc/ocloc_win_101.1191.zip (this
  is needed for AoT compilation on Windows only)

Components revisions used in GPU-enabled build:

KhronosGroup/SPIRV-LLVM-Translator@ed25f1b intel/vc-intrinsics@3a5f4b4
oneapi-src/level-zero@2824c1f (v1.7.4) llvm/llvm-project@fed4134 (llvmorg-12.0.1) + patches from llvm_patches folder

=== v1.16.1 === (15 July 2021)

A minor ISPC update, which has a bug fix for [issue
#2111](https://github.com/ispc/ispc/issues/2111) and is based on patched
version of LLVM 12.0.1.

The bug fix affects x86 targets only and shows up as incorrect code generation
for the sequence of `shuffle()` and `reduce_add()` stdlib functions.

If you are building `ispc` from the sources, note that the fix is implemented
as a patch for LLVM backend and LLVM must be built with this patch applied in
order for this fix to take an effect. Stock build of LLVM 12.0.1 will not
contain this bug fix.

=== v1.16.0 === (11 June 2021)

An ISPC release with language extensions for performance fine tuning, cpu
definitions for `AlderLake` and `SapphireRapids` targets, support for macOS ARM
targets, and massive update of Intel GPUs support. Windows and Linux binaries
in this release support both CPU and GPU targets, while macOS binary supports
only CPU. This release is based on patched LLVM 12.0.0.

The language changes include the following:
- The ability to directly call LLVM intrinsics from ISPC source. This should be
  handy for performance fine tuning and reaching the hardware instructions not
  yet covered by the standard library. Note that it is an experimental feature
  and is enabled only with `--enable-llvm-intrinsics` switch. Please refer to
  `LLVM Intrinsic Functions` section of the user manual for more details.
- `assume()` optimization hint, which can be used for communicating assumptions
  to the optimizer. It will not lead to runtime check, unlike `assert()` calls.
  This is intended for optimizations like removing null pointer checks, removing
  loop reminders, communicating alignment information to the optimizer, and etc.
  Please refer to `Compiler Optimization Hints` section of the user manual for
  more details.
- Support for stack memory allocations through `alloca()` calls.
- `trunc()` standard library functions.

Changes for CPU targets:
- CPU definitions for `AlderLake` and `SapphireRapids` were added: `alderlake`
  and `sapphirerapids` respectively.
- CPU definition for Apple ARM chips were added: `apple-a7`, `apple-a10`,
  `apple-a11`, `apple-a12`, `apple-a13`, `apple-a14`.
- Support for macOS ARM targets was added.

Using GPU-enabled binaries you can build ISPC programs and run them on
Intel(R) Core(tm) Processors with Gen9 graphics (formerly `Skylake`,
`Kaby Lake`, `Coffee Lake`) and Gen12 graphics (TigerLake mobile CPU) using
`--target` options (`genx-x8` and `genx-x16`) and `--cpu` option for
specifying particular platform (e.g. `--cpu=TGLLP`).

The main GPU feature of the current release is Windows support.
There are also a bunch of stability and performance improvements.
Here are some of them:
- ISPC Runtime got support of unified shared memory and multi GPU. Also, there is
  a new `TaskQueue::submit()` method which allows to start executing, but don't
  wait for the completion.
- Thread private memory was mapped to SVM in VC backend. It greatly improves
  stability of the current release. It may affect performance on Gen9 graphics
  but we do not expect any significant changes on Gen12.
- L0 binary generation was reworked through libocloc. Supported on Linux only.

More details about the current state of GPU support are available here:
https://ispc.github.io/ispc_for_xe.html

For build instructions check our docker recipe:
https://github.com/ispc/ispc/blob/main/docker/ubuntu/xpu_ispc_build/Dockerfile

GPU support is still in Beta stage so you may experience some issues but we
strongly encourage you to try it out and give us feedback! You can reach us
through Github discussions and issues, or on Twitter (@ispc_updates).

Runtime Dependencies when targeting GPU:

Linux:
- Intel(R) Graphics Compute Runtime https://github.com/intel/compute-runtime/releases/tag/21.21.19914
- Level Zero Loader https://github.com/oneapi-src/level-zero/releases/tag/v1.2.3
- OpenMP Runtime. Consult your Linux distribution documentation for the
  installation of OpenMP runtime instructions. No specific version is required.

Windows:
- Intel(R) Graphics - BETA Windows(R) 10 DCH Drivers 30.0.100.9667
  https://downloadcenter.intel.com/download/30522/Intel-Graphics-BETA-Windows-10-DCH-Drivers
- Level Zero Loader https://github.com/oneapi-src/level-zero/releases/tag/v1.2.3

Components revisions used in GPU-enabled build:

KhronosGroup/SPIRV-LLVM-Translator@0592c4f
intel/vc-intrinsics@2d0795c
oneapi-src/level-zero@0d30b1f (v1.2.3)
llvm/llvm-project@d28af7c (llvmorg-12.0.0) + patches from llvm_patches folder

=== v1.15.0 === (18 December 2020)

An ISPC release with several improvements for CPU and Beta support of Intel
graphics hardware architectures. The binaries in this release include CPU versions
for Windows, Linux, and macOS, and a GPU-enabled Linux binary, which supports
both CPU and GPU.
CPU binaries are based on patched LLVM 11.0.0, GPU binary is based on patched
LLVM 10.0.1.

CPU changes include:

- New loop unroll pragmas: '#pragma unroll' and '#pragma nounroll' directives
  provide loop unrolling optimization hints to the compiler. This pragma may be used
  immediately before a loop statement. Currently, this functionality is limited to
  uniform for and do-while.
- More efficient 'packed_[load|store]_active()' stdlib functions implementation
  (up to 2.5x faster), which now supports 64 bit types.
- New cpus: 'icelake-server', 'tigerlake' , 'alderlake', 'sapphirerapids'.
- Several stability fixes related to SOA types, bool varying type initialization,
  broken alignment information, type scoping.
- Compile time improvements.
ISPC support was added to CMake 3.19 so you can use now the standard CMake approach
to find ISPC on the system and use it in your build.
https://cmake.org/cmake/help/latest/release/3.19.html#languages

Using GPU-enabled Linux binary you can build ISPC programs and run them on Intel(R)
Core(tm) Processors with Gen9 graphics (formerly Skylake, Kaby Lake, Coffee Lake) and
Gen12 graphics (TigerLake mobile CPU) using '--target' options ('genx-x8' and
'genx-x16') and '--cpu' option for specifying particular platform (e.g. '--cpu=TGLLP').

Stability and performance were significantly improved in this release. Here is the list
of new features:

- Initial support of ahead of time compilation to oneAPI Level Zero binary format using
  '--emit-zebin' switch. You can use this binary from ISPC Runtime by setting
  ISPCRT_USE_ZEBIN env variable to 1. Please note that SPIR-V format is still a
  recommended and default way.
- Initial function pointers implementation.
- Global atomics support.
- Double math functions support.
- Memory functions support.
- Reworked masking approach. We disabled genx hardware mask by default and use
  a software mask by default.
- Improved address spaces differentiation.
- Initial debug support.
- TGLLP (TigerLake mobile CPU) support ('--cpu=TGLLP').
We also added examples to demonstrate interoperability with oneAPI DPC++ Compiler.
More details about current state of GPU support are available here:
https://ispc.github.io/ispc_for_xe.html

For build instructions check our docker recipe:
https://github.com/ispc/ispc/blob/main/docker/ubuntu/gen/Dockerfile

GPU support is in Beta stage so you may experience some issues but we
strongly encourage to try it out and give us feedback! You can reach us through
Github discussions and issues, ISPC mailing list (ispc-users@googlegroups.com),
or on Twitter (@ispc_updates).

Runtime Dependencies:

Intel(R) Graphics Compute Runtime
https://github.com/intel/compute-runtime/releases/tag/20.50.18716
Level Zero Loader
https://github.com/oneapi-src/level-zero/releases/tag/v1.0.22
OpenMP Runtime. Consult your Linux distribution documentation for the
installation of OpenMP runtime instructions. No specific version is required.

Components revisions used in GPU-enabled build:

KhronosGroup/SPIRV-LLVM-Translator@ab5e12a
intel/vc-intrinsics@2de2dd4
oneapi-src/level-zero@c6fa2cd (v1.0.22)
llvm/llvm-project@ef32c61 (llvmorg-10.0.1) + patches from llvm_patches folder

=== v1.14.1 === (28 August 2020)

A minor ISPC update with a bug fix for AVX512 detection problem on macOS
(for more details see issue #1854) and update of GPU version to use Level0 v1.0.
CPU binaries are based on patched LLVM 10.0.1.

Runtime Dependencies for GPU-enabled build:
- Intel(R) Graphics Compute Runtime
  https://github.com/intel/compute-runtime/releases/tag/20.33.17675
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v1.0
- OpenMP Runtime
  Consult your Linux distribution documentation for installation of OpenMP runtime
  instructions.

Components revisions used in GPU-enabled build:
KhronosGroup/SPIRV-LLVM-Translator@1a5c52f
intel/vc-intrinsics@f39ff1e
oneapi-src/level-zero@fcc7b7a (v1.0)
llvm/llvm-project@ef32c61 (llvmorg-10.0.1) + patches from llvm_patches folder

=== v1.14.0 === (30 July 2020)

An ISPC release with several improvements for CPU and initial support of Intel
graphics hardware architectures. The binaries in this release include CPU versions
for Windows, Linux, and macOS, as previous releases, plus a GPU-enabled Linux binary,
which supports both CPU and GPU. CPU binaries are based on patched LLVM 10.0.1.

CPU changes include:
- new avx2-i8x32, avx2-i16x16, avx512skx-i8x64, avx512skx-i16x32 targets.
- "generic" targets were removed.
- several stability fixes, including bugs discovered during fuzzing ISPC by YARPGen.
- integer division performance improvements.
- support for __vectorcall calling convention on Windows x64 (enabled by
  '--vectorcall')

Using GPU-enabled Linux binary you can build ISPC programs and run them on Intel(R)
Core(tm) Processors with Gen9 graphics (formerly Skylake, Kaby Lake, Coffee Lake)
using new '--target' options: 'genx-x8' and 'genx-x16'. For code generation ISPC
uses Vector Compute backend which is the part of 'Intel(R) Graphics Compute Runtime'
through SPIR-V interface. This release also includes ISPC Runtime based on
'oneAPI Level Zero' for GPU and 'OpenMP Runtime' for CPU, which creates unified
abstraction for executing ISPC code on CPU and GPU.

More details are available here: https://ispc.github.io/ispc_for_xe.html

For build instructions check our docker recipe:
https://github.com/ispc/ispc/blob/main/docker/ubuntu/gen/Dockerfile

The stability and performance of GPU part of this release is not mature yet but we
strongly encourage to try it out and give us feedback! You can reach us through
Github issues, ISPC mailing list (ispc-users@googlegroups.com), or on Twitter
(@ispc_updates).

Runtime Dependencies
- Intel(R) Graphics Compute Runtime
  https://github.com/intel/compute-runtime/releases/tag/20.29.17408
- Level Zero Loader
  https://github.com/oneapi-src/level-zero/releases/tag/v0.91.21
- OpenMP Runtime
  Consult your Linux distribution documentation for installation of OpenMP runtime
  instructions.

Components revisions used in this build:
KhronosGroup/SPIRV-LLVM-Translator@1e661b2
intel/vc-intrinsics.git@a0b66f2
oneapi-src/level-zero@v0.91.21
llvm/llvm-project@llvmorg-10.0.0

=== v1.13.0 === (23 April 2020)

An ISPC update, which graduates cross-compilation support to production and
has multiple code generation improvements and bug fixes. AVX512 targets may
get the biggest performance boost due to changed internal representation of
masks (we observed up to 5% speedups), and new switch '--opt=disable-zmm',
which disables using zmm registers in favour of ymm for avx512skx-i32x16 target.
All targets will definitely benefit from LLVM 10.0 backend used in this release.

Here is the list of other changes:

- new switch '--support-matrix' was added to display information about supported
  cross-compilation targets, which are managed by '--target-os=<os>',
  '--target=<ispc-target>', and '--arch=<arch>' switches.
- representation of 'bool' type in storage was changed to match C/C++ (i.e. one
  'bool' occupies one byte) for better interoperability.
- type aliases for unsigned types were added: 'uint8', 'uint16', 'uint32',
  'uint64', and 'uint'. To detect if these types are supported you can check if
  ISPC_UINT_IS_DEFINED macro is defined.
- 'extract()'/'insert()' for boolean arguments, and 'abs()' for all integer and
  FP types were added to standard library.
- FreeBSD was added to the list of supported target OSes, but it's not well
  tested.

Supported platforms in this release are below. Rows are hosts, columns are
targets. x86 and arm are both 32 and 64 bits, where appropriate.

          Windows | Linux    | macOS | Android  | iOS | PS4 | FreeBSD
Windows | x86     | x86, arm | x86   | x86, arm |     | x86 | x86, arm
Linux   |         | x86, arm | x86   | x86, arm |     |     | x86, arm
macOS   |         | x86, arm | x86   | x86, arm | arm |     | x86, arm

=== v1.12.0 === (15 August 2019)

This ISPC update includes experimental cross OS compilation support, ARM and AARCH64
support and a bunch of language features and stability fixes.

Here are the details:

- ISPC is now a cross OS compiler - You can build ISPC programs for Windows, Linux,
  macOS, iOS, Android and PS4 targets from Windows, Linux and macOS hosts.
- ARM and AARCH64 support has been enabled for ISPC. ARM support currently exists for
  neon-i32x4, neon-i8x16 and neon-i16x8 targets. AARCH64 is supported for neon-i32x4
  as well as for a new "double-pumped" 8-wide target: neon-i32x8.
- A new 128-bit AVX2 target (avx2-i32x4) was added.
- Added a CPU definition for Ice Lake client CPUs (--cpu=icl). Note that there is no
  special target for new instructions in Ice Lake flavor of AVX512 yet. For now, You
  can use SKX targets (avx512skx-i32x8 and avx512skx-i32x16) with --cpu=icl.
- Removed the generic targets for KNC and KNL, so ISPC does not have KNC support anymore.
  KNL is still supported through native target (avx512knl-i32x16).
- Removed AVX1.1 (IvyBridge) targets (use AVX1 targets instead).
- Introduced new language features:
    - 'noinline' function qualifier.
    - 'rsqrt_fast()' and 'rcp_fast()' functions.
    - Static initialization for varying.
- A new command line option '--emit-llvm-text' was added to dump LLVM IR in text format.

An ISPC top-of-trunk build is now available in the Compiler Explorer (https://godbolt.org)

The release is based on a patched LLVM 8.0.0 backend.

=== v1.11.0 === (19 April 2019)

An ISPC update with a bunch of new features and stability bug fixes based on a
patched LLVM 8.0.0 backend.

Notable new features are:
- A new 256-bit AVX512 target (avx512skx-i32x8).
- Modified -O1 switch to optimize for size.
- “#pragma once” in auto-generated headers.
- Better debugging support with -O0.

Also we resumed support for PS4 build.

To efficiently write ISPC programs you can now use the ISPC plug-in for VSCode:
https://marketplace.visualstudio.com/items?itemName=intel-corporation.ispc

=== v1.10.0 === (18 January 2019)

An ISPC update, which brings several new features, has a bunch of stability and
performance bug fixes, and infrastructure improvements for those who are
interested in participating in hacking on the ISPC trunk. We also are also
deprecating KNC support and the KNL-generic target (in favor of the native KNL
target, i.e. avx512knl-i32x16).

We've added:
- a streaming store and load implementation (see "Streaming Load and Store
  Operations" section in documentation)
- support for 64 bit wide types in aos_to_soa/soa_to_aos intrinsics
- an option to specify assembler style (see --x86-asm-syntax switch
  documentation is help message)
- a pragma to disable warnings locally (search for "#pragma ignore" in
  documentation)

Our examples include a new SGEMM example which demonstrates different versions
of matrix multiply with various level of optimality. It is useful for learning
how to start from a naive implementation and then add various optimizations
afterwards. Also, our build system is now based on CMake, as are the examples.
So you can use it as a reference for integrating ISPC to your CMake-based
project.

For those who are interested in hacking ISPC or trying a bleeding edge
development version, we have CI on Linux (Travis-CI) and Windows (Appveyor),
including automatic package builds on Windows. We also have Dockerfiles, which
demonstrate bringing up your environment for ISPC development.

The release is based on a patched LLVM 5.0.2 backend.

=== v1.9.2 === (10 November 2017)

An ISPC update, which brings out-of-the-box debug support on Windows,
better performance of most of the targets and a bunch of stability
and performance bug fixes.

The release is based on patched LLVM 5.0 backend.

Windows build is now supports only VS2015 and newer. If you are using earlier
versions, the only known problem that you may encounter is a problem with
"print" ISPC library function.

AVX512 targets are the main beneficiaries of a newer LLVM backend and
demonstrate the biggest performance improvements. SVML support is also
now available on these targets (requires linking by ICC compiler).

=== 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 deprecated.

* 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.io/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.io/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.io/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.io/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.io/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.io/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.io/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.io/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.io/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.io/ispc.html#pointer-types for more
information about pointers in ispc and
http://ispc.github.io/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.io/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.io/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.io/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.io/ispc.html#relationship-to-the-c-programming-language.
There is now a separate FAQ (http://ispc.github.io/faq.html), and a
Performance Guide (http://ispc.github.io/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.io/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.io/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.io/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.io/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.io/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.io/ispc.html#prefetches.

Fast versions of the float to half-precision float conversion routines have
been added.  For more details, see:
http://ispc.github.io/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.io/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.io/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.io/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