As of this release, a C++14 compiler is required.
New Features:
- AX now generates all target machine (host) optimizations by default. This
includes any supported SIMD (sse, avx etc.) instructions for SLP
vectorization
- A full restructure of the developer Function framework with the switch to
IR functions implementations over C bindings (where available). Up until
now, natively supported functions in AX have used their C bindings by
default. This release defaults to their IR implementations which, along
with full vectorization support, allows for the full potential of all LLVM
optimizations. This is most noticeable when using function inside of loops
or in larger AX code snippets
- New functions added: acosh, asinh, atanh, diag
- New options and output to the vdb_ax binary. --try-compile provides a way
to test if a given AX code snippet will compile for a particular VDB
target. Verbose mode with -v now provides better formatted output with
more detailed timing reports.
- Added the ability to provide external string data to AX which can be
accessed with string$ (or chs() in Houdini)
- Added support for vec4 and string VDB volumes. Note that vec4 volumes are
not registered by default
Improvements:
- Significant performance improvements to AX execution with the introduction of:
- Proper SLP vectorization with supported host features
- Defaulting IR functions to ON
- Loop allocation fixes (see bug fixes)
- Constant Folding for scalar C bindings
- Introduced initial constant folding for scalar C bindings. Functions which
are called with known constants are evaluated at compile time. Note that
CF for IR Functions is left to LLVM's optimization passes
- Added support for vec2 and vec4 arguments for length() and lengthsq()
- Added a new floating point signature for abs() (which calls fabs) to avoid
implicit casting of integer arguments
Bug Fixes:
- Fixed a bug where assignments from functions which return void could cause
a crash
- Fixed a bug where functions which modified reference arguments (such as
transform, posttransform) would cast the modifiable argument, resulting in
the result not being set. These functions will now error with invalid
signatures and require the correct argument types to be provided
- Fixed an issue which could cause stack overflows when using scalar or
array (vec/mat) allocations within loops
- Fixed an issue which could cause a crash when using uninitialized strings
in compound expressions
- Fixed a bug where binary string concatenations could cause spurious
characters to appear with print()
- Fixed a bug where vec4 types were being interpreted as vec3
Code Generation:
- Assignment and binary string operators now always copy the null terminator
- Added new IR function logic for lerp and for normalization of integer
vectors
- The ast::FunctionCall visit now always pushes back the function result
- Strings are always initialized to the empty string
- All allocations, aside from string allocs, have been converted to function
prologue static allocas
- New codegen/ConstantFolding.h which provides the generic framework for
constant folding C bindings
- Renamed the FunctionBase class to FunctionGroup
- Renamed the FunctionSignatureBase class to Function and re-structured its
member implementations
- New classes for derived Functions, CFunctionBase and IRFunctionBase, which
can be further subclassed for the respective function type.
- A new interface class SRetFunction for functions which require structural
returns
- New framework for building and creating a function. The FunctionBuilder
provides a way to interface with the function types in FunctionTypes.h
without having to manually subclass.
- Exposed the ability to customize llvm function attributes.
- Removed the unnecessary use of std::function on templated classes and
methods
- Re-structured all native functions to use the new FunctionBuilder
- A variety of IR fixes to some native functions
- Re-structured all IR functions which required structural returns.
Previously, these function would produce invalid results unless they were
forcefully inlined. SRET IR methods now behave in the same way as SRET C
bindings
- Improvements to the LLVMType trait structures. These have been mostly
consolidated to take advantage of new LLVM methods
- Added a new AliasTypeMap trait structure in Types.h to allow for safer
and restricted front-end to back-end type mapping
- New insertStaticAlloca utility method for creating function prologue
static allocations. New isValidCast utility method to check if a cast is
valid between two types
- Removed a number of unused utility methods
- Removed the function arguments vector from the ComputeGenerator in favor
of new utility methods to extract arguments (extractArgument)
- Renamed the ComputeGenerator functions to prefix with ax and suffix
with the generator type
Compiler:
- Initial implementation of LLVM's new PassManager. Currently disabled by
default. This has not been fully configured to work with target machine
optimizations.
- Introduced an internal compiler function, initializeTargetMachine, to
create the target machine representing the available feature set of the
host
- Renamed the mPrioritiseFunctionIR Function Option to mPrioritiseIR and
defaulted it to ON
- Introduced a mConstantFoldCBindings Function Option which defaults to ON
- Added support for AXStringMetadata during global mapping for external
values
Math:
- Exposed the curlnoise functions from StandardFunctions to the
OpenSimplexNoise headers
CMake / Build / Testing:
- New test suites for various back-end utilities
- Added CMake SIMD options. Note that these optional apply to the compiled
CXX source files of OpenVDB AX, not the JIT LLVM IR. The IR is always
optimized to the host specification.
- Set the required LANGUAGES for OpenVDB AX projects to CXX
- Added significantly more function tests with profiling options
- BISON 3.0 is now marked as required if the grammar is to be regenerated
- New developer options, USE_NEW_PASS_MANAGER and OPENVDB_AX_TEST_PROFILE
- Travis fixes to build against VDB 6.2.1
- Removed the deprecated Makefiles
