Memory model. Memory architectures have a significant effect on the programming model. Uniform memory Some programming models provide a uniform view of memory so that parallel code has exactly the same view of memory as the serial code. Memory allocated by the serial code can be used directly by the parallel code, and in some cases memory can be allocated directly by the parallel code. While this is the most convenient model for the programmer it excludes the use of parallel processors with segmented memory architectures, which includes most of the massively parallel processors. Segmented memory Processors with segmented memory architectures require data to be explicitly copied between the different areas of memory. Therefore programming models need a mechanism to describe how input data is written to the parallel processor and how results are read back. There is often a limited amount of memory available so there also needs to be a mechanism for allocating memory on the parallel processor that is separate from memory allocation on the serial processor. Programming models will typically allow the program to define blocks of shared memory that have matching allocations on both the serial processor memory and parallel processor memory. Some programming models require the programmer to explicitly copy data between the different memory segments at the right time. Other models will automatically copy the data to where it is needed, in which case there must be a mechanism for describing whether memory is an input to the parallel code, output from the parallel code or both. Private memory Each thread has its own stack but some models also provide access to private memory which is additional local memory for each individual thread. This memory is uninitialized when the thread starts and is discarded when the thread ends and can therefore be used only as temporary storage. In many architectures this memory is significantly faster than other memory and can be used to store intermediate results that are too large to be held in registers. Tile memory When tiles are supported a form of local memory is usually available that can be accessed by all threads in a tile. This can be use to share results between threads efficiently, but some form of tile synchronization is required to ensure that reading and writing this memory happens in the right order. Some programming models allow efficient copying of data between the different types of memory in a segmented memory system and, in parti...
Memory model. Renderscript uses a split memory model. Any memory to be used in parallel code must be explicitly allocated from serial code and initialized if necessary. If a memory object is used by parallel Renderscript code, subsequent attempts to use it from serial code trigger a synchronization event. Compilation Model. Renderscript code is compiled to an intermediate representation at compile time. At run time it translated into machine code for the specific machine it is running on. Minimum Steps Needed to Run Parallel Code. Renderscript has some explicit context and memory management requirements:
