1. What are the contributions in "Optimization space pruning without regrets" ?
The authors present a novel approach to automatically discover the best performing code from a given set of possible implementations.. While the space considered in this paper focuses on GPUs, the approach is generic enough to be applied to other architectures.. The authors implemented their algorithm in a tool called Telamon and demonstrate its effectiveness on a huge, architecture-specific and input-sensitive optimization space.. The information provided by the performance model also helps to identify ways to enrich the search space to consider better candidates, or to highlight architectural bottlenecks.
read more
2. What future works have the authors mentioned in the paper "Optimization space pruning without regrets" ?
The intermediate representation would need to be extended to capture that domain ’ s computational structure and to adapt the performance model to consider the bottlenecks of the target platform.. One interesting possibility would be to generate domain-specific intermediate representations and the associated exploration scheme from a high level description of the optimization requirements.
read more
3. What is the way to compute the bounds at the level of a thread?
When the authors compute the bounds at the level of a thread or block, the authors assume it runs in isolationas contentions between threads or between blocks are taken into account by the bounds at a coarser level of parallelism.
read more
4. What is the way to solve the problem of parallelism?
It requires a deep knowledge of the architecture and many trials and errors for the programmer to explore implementation alternatives at each level of parallelism, from thread-local optimizations to the mapping of computations across the entire processor.
read more
