摘要:
In one embodiment, the present invention includes a processor having a core to execute instructions. This core can include various structures and logic that enable instructions of different atomic regions to be executed in an overlapping manner. To this end, the core can include a register file having registers to store data for use in execution of the instructions, and multiple shadow register files each to store a register checkpoint on initiation of a given atomic region. In this way, overlapping execution of atomic regions identified by a programmer or compiler can occur. Other embodiments are described and claimed.
摘要:
Technologies for performing flexible code acceleration on a computing device includes initializing an accelerator virtual device on the computing device. The computing device allocates memory-mapped input and output (I/O) for the accelerator virtual device and also allocates an accelerator virtual device context for a code to be accelerated. The computing device accesses a bytecode of the code to be accelerated and determines whether the bytecode is an operating system-dependent bytecode. If not, the computing device performs hardware acceleration of the bytecode via the memory-mapped I/O using an internal binary translation module. However, if the bytecode is operating system-dependent, the computing device performs software acceleration of the bytecode.
摘要:
Methods and apparatus relating to compact trace trees for dynamic binary parallelization are described. In one embodiment, a compact trace tree (CTT) is generated to improve the effectiveness of dynamic binary parallelization. CTT may be used to determine which traces are to be duplicated and specialized for execution on separate processing elements. Other embodiments are also described and claimed.
摘要:
Disclosed are methods, machine readable medium and systems that dynamically translate binary programs. The dynamic binary translation may include identifying a hot code trace of a program. The translation may further include determining a completion ratio for the hot code trace. The translation may also include packaging the hot code trace into a transactional memory region in response to the completion ratio having a predetermined relationship to a threshold ratio.
摘要:
An apparatus, method and system is described herein for enabling intelligent recirculation of hot code sections. A hot code section is determined and marked with a begin and end instruction. When the begin instruction is decoded, recirculation logic in a back-end of a processor enters a detection mode and loads decoded loop instructions. When the end instruction is decoded, the recirculation logic enters a recirculation mode. And during the recirculation mode, the loop instructions are dispatched directly from the recirculation logic to execution stages for execution. Since the loop is being directly serviced out of the back-end, the front-end may be powered down into a standby state to save power and increase energy efficiency. Upon finishing the loop, the front-end is powered back on and continues normal operation, which potentially includes propagating next instructions after the loop that were prefetched before the front-end entered the standby mode.
摘要:
A method and apparatus for ensuring integrity of transaction exit functions is herein described. Dead local data in a transaction is prevented from overwriting local variables associated with a transaction exit function. In a write-buffering Software Transactional Memory (STM) system, a commit function is associated with a private stack to store local variables to ensure write-back of local dead data in a write-buffer does not corrupt the commit function. Similarly, in a roll-back STM, an abort function is associated with a private stack to store local variables to ensure the roll-back of a program stack with local dead data from a write log does not corrupt the abort function. Alternatively, one stack may be used for the transaction including a first function and an exit function. Here, local dead variables are detected and prevented from overwriting local variables of the exit function.
摘要:
Systems, apparatuses, and methods for a hardware and software system to automatically decompose a program into multiple parallel threads are described. In some embodiments, the systems and apparatuses execute a method of original code decomposition and/or generated thread execution.
摘要:
Dynamic optimization of application code is performed by selecting a portion of the application code as a possible transaction. A transaction has a property that when it is executed, it is either atomically committed or atomically aborted. Determining whether to convert the selected portion of the application code to a transaction includes determining whether to apply at least one of a group of code optimizations to the portion of the application code. If it is determined to apply at least one of the code optimizations of the group of optimizations to the portion of application code, then the optimization is applied to the portion of the code and the portion of the code is converted to a transaction.
摘要:
An apparatus and method for redundant transient fault detection. In one embodiment, the method includes the replication of an application into two communicating threads, a leading thread and a trailing thread. The trailing thread may repeat computations performed by the leading thread to detect transient faults, referred to herein as “soft errors.” A first in, first out (FIFO) buffer of shared memory is reserved for passing data between the leading thread and the trailing thread. The FIFO buffer may include a buffer head variable to write data to the FIFO buffer and a buffer tail variable to read data from the FIFO buffer. In one embodiment, data passing between the leading thread data buffering is restricted according to a data unit size and thread synchronization between a leading thread and the trailing thread is limited to buffer overflow/underflow detection. Other embodiments are described and claimed.
摘要:
A method and apparatus for dynamic binary translator to support precise exceptions with minimal optimization constraints. In one embodiment, the method includes the translation of a source binary application generated for a source instruction set architecture (ISA) into a sequential, intermediate representation (IR) of the source binary application. In one embodiment, the sequential IR is modified to incorporate exception recovery information for each of the exception instructions identified from the source binary application to enable a dynamic binary translator (DBT) to represent exception recovery values as regular values used by IR instructions. In one embodiment, the sequential IR may be optimized with a constraint on movement of an exception instruction downward past an irreversible instruction to form a non-sequential IR. In one embodiment, the non-sequential IR is optimized to form a translated binary application for a target ISA. Other embodiments are described and claimed.