CHANGELOG.md

ChangeLog FastR-GPU Compiler and Execution Environment

fastr-gpu-0.13

* Marawacc Multi-Device support imported

fastr-gpu-0.12

* Rsequence improved for OpenCL data transfer optimization (flag and compass sequences)
* Refactor and documentation improved
* OpenCL deoptimization technique improved (resume from last element in the AST interpreter)

fastr-gpu-0.11

* Option for node rewriting for scope detection
* New Parray Strategy for primitives arrays such as int[] or double[], marshal is zero overhead.
* RSequence with the new PArray primitives
* Deoptimisation technique improved (maximum number of deopts and node rewriting supported)	
* Buffer reallocation when needed for the same input function
* MApply parallel version supported 

fastr-gpu-0.10

* Deoptimizations supported with a flag in OpenCL
* New scope phases
* Truffle annotation system for OpenCL node filtering and deopt
* Javadoc
* Refactor of the main OpenCL builtin

fastr-gpu-0.9

* Garbage collector builtin provided
* R root node for OpenCL execution equivalent to normal FastR
* Profiling information improved
* Benchmarks updated 

fastr-gpu-0.8

* Mandelbrot supported
* Lexical scope variables improved
* Cache included for all the parameters related to the RFunction
* Profiling info added 

fastr-gpu-0.7

* NBody, SpectralNorm, DFT and Euler benchmarks supported
* Scope by pattern in the IR detected and use it in the code generation
* PArrays introduced 
* RSequence optimisation 

fastr-gpu-0.6

* Graph (GraalIR) phases with node simplification.
* KMeans suported
* Basic scope supported for KMeans and general pattern detection in the IR

fastr-gpu-0.5

* Benchmarks saxpy, montecarlo and blackscholes provided. There is an assumption in montecarlo for random numbers (still working on it).
* Nodes ZeroExtendNode, XorNode and OrNode supported.
* ForeignCallNode fixed.
* AMD64MathIntrinsicNode supported.
* Builtin for random numbers in Java
* Fix semantic Tuples and RLists

fastr-gpu-0.4

* Tuples support from RList in the R lambda expression
* Graph simplified within the interpreter level to the GraalIR
* Graal extension to get the information whithin the data types from Partial Escape Analysis
* Mararawcc code generator improved for tuples support from Virtual Nodes and CommitAllocationNode

fastr-gpu-0.3

* Basic support for GPU from AST -> Graal IR with Truffle information -> OpenCL code.
  This is just one example workin (examples/gpu/testGPU.R) 
* NA assumptions. There is no GPU code support for NA numbers. 
* Depth assumption is 0 -> It is removed from the code generator.
* Graph simplification phases. It removes FrameState nodes from the original graph.
* Boxing and UnBoxing elimination. 

fastr-gpu-0.2a

* Basic support for kernel generation (simple expressions) and non-generic.
  This is a proof of concept.
* Interception from the Partial Evaluator to the GPU backend
* Mechanism to identify a potencial R callTarget for GPU compilation

fastr-gpu-0.2

* Support async for map operations : get((map(map(map))))
* Support async for reduction operations
* Get node added
* Futures added for async support
* Change number of threads per expression

fastr-gpu-0.1

* Support for map/reduce with Marawacc threads backend. 
* Support for Integer, Boolean and Double arrays
* Function composition from R.