《高性能处理器体系结构》TLP-推测多线程和事务存储.ppt

* * Other thread spawning schemes Based on Cache misses Based on more sophisticated analysis of the code in order to quantify some relevant metrics * * Microarchitectural support for Speculative Architectural Threads Spawning process Looking for an idle context, if there are no free contexts, different solutions can be taken: Not to spawn the thread Wait for a free context Squash one thread and spawn the new one at the recently freed context Computing the logical position of the thread among all the active threads Since it is possible that values flow from one speculative thread to another, the logical order among threads must be known by the processor. The approach to identify the order may depend on the spawning scheme. * * Microarchitectural support for Speculative Architectural Threads Spawning process Initializing the values for the spawned thread In order to perform useful work, speculative threads have to be executed with the correct input values. So, it is necessary to initialize the register and memory values that are to be used by the spawned thread. If such values are available at the spawning point, they can be directly copied from the spawner thread to the new spawned thread. If the values have to be produced by a preceding thread, they may be predicted as described in later subsections or a synchronization may be required. * * Microarchitectural support for Speculative Architectural Threads Committing process When a thread finds a special instruction or when it reaches the TSP of any other active thread, this thread finishes and stops fetching instructions and stalls until it becomes the non-speculative thread. When the non-speculative thread reaches the TSP of any other active thread, it verifies thread that corresponds to the next thread in sequential program order. All input values that consumed but not produced by a thread have to be validated. The verification of the input values can be done either when the non-speculative thread has reached t