International Business Machines Corporation

Huawei Technologies Co., Limited

Guansong Zhang

Raul Esteban Silvera

Ettore Tiotto

Tyler Bryce Nowicki

Weiwei Li

Zhixing Ren

Alexandre E Eichenberger

Kevin Alexander Stoodley

Yaoqing Gao

Robert J Blainey

Roch Georges Archambault

Shimin Cui

Peng Zhao

Priya Unnikrishnan

Ahmed Mohammed ElShafiey Mohammed Eltantawy

Yan Luo

Liangxiao Hu

Amy K Wang

Erik P Charlebois

Andrew Siu Doug Lee

Abraham Davidson Fai Chung Chan

Kevin Lin

Yue Zhao

The disclosed systems, structures, and methods are directed to optimizing address calculations in a computer. This is achieved in a compiler that identifies an address calculation in code that is being compiled and transforms the code by splitting the address calculation into a first portion in which an offset is determined and a second portion, in which the offset is combined with a base pointer to generate an address. The address and the base pointer have a first bit-length, and the offset has a second bit-length shorter than the first bit-length. The offset is determined using an operation performed at the second bit-length. In some implementations the first bit-length is 64 bits and the second bit-length is 32 bits.

Method and system for optimizing address calculations

Methods, systems and apparatus are provided for handling control flow structures in data-parallel architectures. A method includes receiving, by a processing unit (PU), a program for execution. The method further includes applying, by the PU, a branching solution to the program to obtain data on control flow structures of the program. The method further includes determining, by the PU and based at least in part on the obtained data, one or more control flow structures of the program to predicate. The method further includes applying, by the PU, predication to the one or more control flow structures of the program.

Methods and systems for handling control flow structures in data-parallel architectures

There is described a method of shading a group of pixels in a fragment shader in a raster graphics pipeline. At least one first pilot pixel of the group of pixels is shaded under a first precision. At least one second pilot pixel of the group of pixels is shaded under a second precision. An error value representing a difference between the first and second pilot pixels is calculated. At least one other pixel of the group of pixels is shaded under the first precision if the error value is greater than an error threshold. The at least one other pixel is shaded under the second precision if the error value is smaller than the error threshold.

Apparatus, method, and computer-readable medium for image processing using variable-precision shading

Systems and methods are described for extending a live range for a virtual scalar register during compiling of a program, comprising: receiving an intermediate representation (IR) of a source code configured for implementing single-instruction-multiple-thread (SIMT) execution, the IR representing the source code as control flow graph including a plurality of basic blocks (BB); and when a virtual scalar register defined in a first BB of the IR is last used in a second BB of the IR that is a divergent BB, modifying the IR to extend the live range of the virtual scalar register.

Systems and methods for extending a live range of a virtual scalar register

The disclosed systems, structures, and methods are directed to optimizing memory access to constants in heterogeneous parallel computers, including systems that support OpenCL. This is achieved in an optimizing compiler that transforms program scope constants and constants at the outermost scope of kernels into implicit constant pointer arguments. The optimizing compiler also attempts to determine access patterns for constants at compile-time and places the constants in a variety of memory types available in a compute device architecture based on these access patterns.

Method and system for optimizing access to constant memory

A compiler method for exploiting data value locality for computation reuse. When a code region having single entry and exit points and in which a potential computation reuse opportunity exists is identified during runtime, a helper thread is created separate from the master thread. One of the helper thread and master thread performs a computation specified in the code region, and the other of the helper thread and master thread looks up a value of the computation previously executed and stored in a lookup table. If the value of the computation previously executed is located in the lookup table, the other thread retrieves the value from the table, and ignores the computation performed by the thread. If the value of the computation is not located, the other thread obtains a result of the computation performed by the thread and stores the result in the lookup table for future computation reuse.

Compiler method of exploiting data value locality for computation reuse

A computer-implemented method for creating a threaded package of computer executable instructions from software compiler generated code includes allocating, through a computer processor, the computer executable instructions into a plurality of stacks, differentiating between different types of computer executable instructions for each computer executable instruction allocated to each stack of the plurality of stacks, creating switch points for each stack of the plurality of stacks based upon the differentiating, and inserting the switch points within each stack of the plurality of stacks.

Software compiler generated threaded environment

A method for promotion of a child procedure in a software application for a heterogeneous architecture, wherein the heterogeneous architecture comprises a first architecture type and a second architecture type, comprises inserting a parameter representing a parallel frame pointer to a parent procedure of the child procedure into the child procedure; and modifying a reference in the child procedure to a stack variable of the parent procedure to include an indirect access to the parent procedure via the parallel frame pointer.

Promotion of a child procedure in heterogeneous architecture software

A method and system of auto parallelization of zero-trip loops that substitutes a nested basic linear induction variable by exploiting a parallelizing compiler is provided. Provided is a use of a max{0,N} variable for loop iterations in case of no information is known about the value of N, for a typical loop iterating from 1 to N, in which N is the loop invariant. For the nested basic induction variables, an induction variable substitution process is applied to the nested loops starting from the innermost loop to the outermost one. Then a removal of the max operator afterwards through a copy propagation pass of the IBM compiler is provided. In doing so, the loop dependency on the induction variable is eliminated and an opportunity for a parallelizing compiler to parallel the outermost loop is provided.

Auto parallelization of zero-trip loops through the induction variable substitution

Embodiments of the present invention address deficiencies of the art in respect to loop parallelization for a target architecture implementing a shared memory model and provide a novel and non-obvious method, system and computer program product for SIMD code generation for parallel loops using versioning and scheduling. In an embodiment of the invention, within a code compilation data processing system a parallel SIMD loop code generation method can include identifying a loop in a representation of source code as a parallel loop candidate, either through a user directive or through auto-parallelization. The method also can include selecting a trip count condition responsive to a scheduling policy set for the code compilation data processing system and also on a minimal simdizable threshold, determining a trip count and an alignment constraint for the selected loop, and generating a version of a parallel loop in the source code according to the alignment constraint and a comparison of the trip count to the trip count condition.

Single instruction multiple data (SIMD) code generation for parallel loops using versioning and scheduling

A computer implemented method, computer usable program code, and a system for parallelizing a loop. A parameter that will be used to limit parallelization of the loop is identified to limit parallelization of the loop. The parameter specifies a minimum number of loop iterations that a thread should execute. The parameter can be adjusted based on a parallel performance factor. A parallel performance factor is a factor that influences the performance of parallel code. A number of threads from a plurality of threads is selected for processing iterations of the loop based on the parameter. The number of threads is selected prior to execution of the first iteration of the loop.

Mechanism to restrict parallelization of loops

A computer-implemented method for determining whether an array within a loop can be privatized for that loop is presented. The method calculates the array sections that require first or last privatization and copies only those sections, reducing the privatization overhead of the known solutions.

Method of partially copying first and last private arrays for parallelized loops based on array data flow

The present disclosure is directed to a method for providing an OpenMP reduction implementation. The method may comprise creating an aggregate of at least one reduction variable in a parallel region or a work-sharing construct; defining a pointer variable, the pointer variable pointing to a dynamic array of the aggregate; creating an initialization routine, an outlined routine and a reduction accumulation routine; replacing the parallel region or the work-sharing construct with a runtime routine, the runtime routine taking a plurality of arguments including an address of the initialization routine, an address of the outlined routine, an address of the reduction accumulation routine, an address of the pointer variable, and a size of the aggregate; and executing the runtime routine when the at least one reduction variable is in the parallel region or the work-sharing construct.

Open multi-processing reduction implementation in cell broadband engine (CBE) single source compiler

The present invention provides an approach for barrier synchronization. The barrier has a first array of elements with each element of the first array having an associated process, and a second array of elements with each element of the second array having an associated process. Prior to use, the values or states of the elements in each array may be initialized. As each process finishes its phase and arrives at the barrier, it may update the value or state of its associated element in the first array. Each process may then proceed to spin at its associated element in the second array, waiting for that element to switch. When the values or states of the elements of the first array reach a predetermined value or state, an instruction is sent to all of the elements in the second array to switch their values or states, allowing all processes to leave.

Software barrier synchronization

A method, system and apparatus for barrier synchronization using distributed counters and a centralized sensor. The system can include multiple distributed counters coupled to corresponding application processes in a computing application. The barrier synchronization system further can include a centralized sensor coupled for observation by the application processes. Preferably, the application processes can be separate threads of execution in the computing application. The barrier synchronization centralized sensor yet further can be managed by a designated master one of the application processes. Moreover, preferably the system further can include a backup sensor coupled for observation by the application processes and managed by the designated master one of the application processes.

Growing community of inventors

Markham, Canada

Guansong Zhang