################### ### v1.3 ### ################### General: - Implemented kernel fusion in A-Darwin/Bones - Added command-line options to test individual optimisations - Enabled thread-merging (coarsening) for additional skeletons Skeletons: - Optimised the CUDA skeletons for threadblock sizing - Fixed various bugs in the CPU-targets Bug fixes: - Synchronisation statements are now correctly inserted as pragma's - Re-enabled the memory optimisations (-rfbl) - Minor bug fixes in A-Darwin ################### ### v1.2 ### ################### General: - Integrated A-Darwin (a tool to automatically extract algorithmic species) into the Bones source-tree. - Added memory transfer optimisations (e.g. CPU-GPU copy optimisations). A-Darwin: - Initial test release of A-Darwin (with many limitations) - Added a 'golden' reference set to test the extraction of species - Added dependence-test examples Miscellaneous: - Substituted the simplify method by a symbolic math library gem ('symbolic'). - Updated the readme file - Updated the examples ################### ### v1.1 ### ################### General: - Initial release to RubyGems and initial commit to github. - Added support for a compiler optimisation pass: array substitution with a local register copy in the case of chunk to element species. - Added support for a compiler optimisation pass: thread-merging, potentially improving re-use through locality at the cost of parallelism. - Updated and added examples. - Tuned the skeletons. Skeletons performance/readability/usability tuning: - Added a special GPU-CUDA 'chunk(1,N)' skeleton with pre-shuffling (e.g. atax, bicg, mvt, syrk). Added support for a corresponding transformation in Bones. - Added a special GPU-CUDA 2x'chunk(1,N)' skeleton with pre-shuffling if there are two chunk inputs (e.g. gesummv,syr2k). - Updated the various OpenCL skeletons to create only once the 'context', 'command queue' and 'program'. This saves significant time for programs executing a single kernel multiple times or multiple kernels subsequently. - Tune GPU-CUDA skeletons to prefer L1 cache above scratchpad memory. Examples: - Updated several benchmarks with additional species. - Added the 'saxpy' benchmark. - Added 'example12' to demonstrate the use of species in a separate function. Bug fixes: - Removed the generation of global code in case no species are detected. Miscellaneous: - Renamed the original main function and created a new 'main' in 'common/globals.c' with initialisation and clean-up. The use of '#pragma species initialize' is now deprecated, but a main function is required in the original code. - Added the identification of a species' function, making variable definition detection local. - Improved the compiler run-time for algorithms for which a skeleton is unavailable. - Added a command-line option to be able to generate code for a single species only (-only_alg_number). - Added a command-line option to set a fixed thread-merge factor (-merge_factor). - Added a gemspec file. - Updated the documentation and readme. ################### ### v1.0 ### ################### General: - Added a benchmark set to the examples directory based on the PolyBench/C benchmark set. - Performed major code refactoring, improving maintainability and performance of Bones. - Updated all examples to work with the upcoming 'automatic species extraction tool' (ASET). Skeletons performance/readability/usability tuning: - Updated the 'GPU-CUDA' reduction skeleton to support initial values, loops not starting at zero or not a power of 2, and more (see the new 'shared/example5.c'). - Improved the way skeletons are matched with species. Introduced a 'default' skeleton and modified others to accomodate these changes. Improved code/species support: - Added atomic support for OpenCL based targets - Added support for loop iterator variables inside the 'algorithmic species' with selective memory copy for the 'GPU-CUDA' target, see 'element/example11.c'. Bug fixes: - Fixed a compatibility problem with CAST version 0.2.0. Miscellaneous: - Added proper error messaging to catch various exceptions - Reversed the order of loop-flattening to obtain coalesced memory accesses (in particular for the 'GPU-CUDA' target). - Renamed 'tile' into 'chunk', since 'tile' might imply using a 2D chunk of data. - Changed '#pragma bones' into '#pragma species' to match the algorithmic species naming. - Added a warning message for negative or zero range dimensions. - Organized the search-and-replace parameters found in skeletons, renamed most of them and added a few new. - Added headers to the code examples and cleaned them up. - Added a warning message if an 'endkernel' pragma is missing. - Improved the way error messages are thrown. ################### ### v0.9 (Beta) ### ################### General: - Implemented the new 'algorithmic species' using ranges, thus creating support for for-loops with affine functions as loop-bounds. - Updated the examples Supported targets: - Changed the names of the targets 'GPU-OPENCL' and 'CPU-OPENCL' into 'GPU-OPENCL-AMD' and 'CPU-OPENCL-INTEL' respectively. - Added a new target 'CPU-C' which implements a C-to-C pass-through. - Added a new target 'CPU-OPENMP', which uses OpenMP to create 4 CPU threads. - Added a new target 'CPU-OPENCL-AMD'. This target is similar to 'CPU-OPENCL-INTEL', but targets the AMD APP. Skeletons performance/readability/usability tuning: - Added a basic prefetching technique in the local memory for the neighbourhood skeleton for the 'GPU-CUDA' target. - Removed the first entry in the transformation settings, which was used in previous versions to set the dimensions (now automatically detected). - Changed the 'GPU-CUDA' skeletons such that host files can be compiled with a C99 compiler. - Tuned 'CPU-OPENCL-INTEL' performance for Intel's OpenCL SDK. - Created aligned memory allocation functions to enable zero-copy possibilities for the 'CPU-OPENCL-INTEL' target. - Completed the addition of 'bones_' for every variable in the skeletons. Verification code: - Create a verification function specific to each output. - Moved the verification code (including the original code) to a separate file, which is only generated if '-c' is provided as a flag to Bones. Bug fixes: - Fixed a bug which would create function names starting with a digit. - Adjusted the use of directory structures in the code for Windows-compatibility. - Fixed a bug where variables for verification would have duplicate names. - Fixed a bug where verification code would not compile for 'unsigned int' types. - Fixed a memory leak which would occur when verification is enabled. - Fixed a bug where statements of the form 'a[i]++' would not be recognized as input nor as output. They are now rewritten as 'a[i]=a[i]+1'. Miscellaneous: - Added support for selective copying-out (based on array access ranges) - Added support for defines found in header functions (pre-processor now also pre-processes the header files) - Added the possibility to specify the order of inputs/outputs in the classification by giving their names (if not given, the default ordering is assumed). - Writing to a specific location in an array followed by a read no longer considers the array as input and output, it is now output only. - Added a check to see if for-loops start and end as expected (as provided by the ranges given through the 'algorithmic species'). - Create a 'simplify' function, which simplifies math expressions to a certain extend. A test is included to give a few examples of what it can do. - Clean-up of the Rakefile, addition of stub tasks to compile and execute example code, and the addition of an 'add new target' task. - Changed the code such that the core components of Bones (the 'lib' folder) do not have to be adjusted to add a new target. - Added performance measurement for original code in case verification is enabled. - Renamed 'Tribe' into 'Species' and 'Primitive' into 'Algorithm'. ################### ### v0.8 (Beta) ### ################### Initial release. ###################