Modern GPU outperforms contemporary CPU by order of magnitude both in memory bandwidth and arithmetic performance. To be able to unleash the power of GPU it is necessary to carry out enough arithmetical operations compared to the amount of data transfered between GPU on-chip and global memory. It is generally not possible to achieve good ratio of arithmetical operations to memory transfers in the case of mapping functions on a set of small elements in which case the data are not reused enough once is loaded from the global memory. If several consequent functions are being mapped it is possible to fuse these functions and store the intermediate results in faster on-chip memory. However, higher consumption of this memory can reduce the degree of parallelism achieved on the GPU and consequently decrease performance.

Implementation of source-to-source compiler seems to be a suitable solution, enabling to fuse particular functions in automated manner based on a simple description of the computation defining the functions call sequence. The compiler will be able to find the effective balance between reducing the memory transfers by fusing the particular functions and lowering the consumption of the on-chip resources limiting attainable parallelism.

The subject of this thesis is to design simple language allowing programmer to define the sequence of functions call and to generate code of fused kernels and kernels calls according to defined call sequence.

The language for functions call will allow to allocate particular data elements and apply functions to them. Moreover, it will allow to use loops of fixed length and index the data elements. The parser transforming the code in this language to direct acyclic graph representing function calls and parameters passing will be developed.

The DAG is processed by optimizing part of the compiler, which is subject of another thesis. On the output of the optimizer, the subsets of functions which is to be fused is defined as well as the implementation of particular fusions (functions call linearization, memory allocation and reuse etc.). The student will implement the code generator, transforming output of the optimizer to CUDA source code. The code generator will be able to generate GPU kernels, CPU functions calling them and helper code allowing to check GPU results again CPU and benchmark the performance of GPU implementation.