Advanced Micro Devices Corporation

Qualcomm Incorporated

Lee W Howes

Benedict R Gaster

Michael C Houston

Michael Mantor

Vinod Tipparaju

Thomas Scogland

Mark Douglas Hummel

Brian D Emberling

Maxim V Kazakov

Mark M Leather

Norman Rubin

Dongping Zhang

Dominik Behr

With the success of programming models such as OpenCL and CUDA, heterogeneous computing platforms are becoming mainstream. However, these heterogeneous systems are low-level, not composable, and their behavior is often implementation defined even for standardized programming models. In contrast, the method and system embodiments for the heterogeneous parallel primitives (HPP) programming model disclosed herein provide a flexible and composable programming platform that guarantees behavior even in the case of developing high-performance code.

Heterogeneous parallel primitives programming model

A method, system, and computer program product synchronize a group of workitems executing an instruction stream on a processor. The processor is yielded by a first workitem responsive to a synchronization instruction in the instruction stream. A first one of a plurality of program counters is updated to point to a next instruction following the synchronization instruction in the instruction stream to be executed by the first workitem. A second workitem is run on the processor after the yielding.

Method and system for yield operation supporting thread-like behavior

A system and methods embodying some aspects of the present embodiments for efficient load balancing using predication flags are provided. The load balancing system includes a first processing unit, a second processing unit, and a shared queue. The first processing unit is in communication with a first queue. The second processing unit is in communication with a second queue. The first and second queues are each configured to hold a packet. The shared queue is configured to maintain a work assignment, wherein the work assignment is to be processed by either the first or second processing unit.

Efficient processor load balancing using predication

A system, method, and computer program product are provided for improving resource utilization of multithreaded applications. Rather than requiring threads to block while waiting for data from a channel or requiring context switching to minimize blocking, the techniques disclosed herein provide an event-driven approach to launch kernels only when needed to perform operations on channel data, and then terminate in order to free resources. These operations are handled efficiently in hardware, but are flexible enough to be implemented in all manner of programming models.

Multithreaded computing

Examples are described for a device to receive intermediate code that was generated from compiling source code of an application. The intermediate code includes information generated from the compiling that identifies a hierarchical structure of lower level sub-routines in higher level sub-routines, and the lower level sub-routines are defined in the source code of the application to execute more frequently than the higher level sub-routines that identify the lower level sub-routines. The device is configured to compile the intermediate code to generate object code based on the information that identifies lower level sub-routines in higher level sub-routines, and store the object code.

Generating object code from intermediate code that includes hierarchical sub-routine information

Techniques are described for copying data only from a subset of memory locations allocated to a set of instructions to free memory locations for higher priority instructions to execute. Data from a dynamic portion of one or more general purpose registers (GPRs) allocated to the set of instructions may be copied and stored to another memory unit while data from a static portion of the one or more GPRs allocated to the set of instructions may not be copied and stored to another memory unit.

Register spill management for general purpose registers (GPRs)

In an example, a method for speculative scalarization may include receiving, by a first processor, vector code. The method may include determining, during compilation of the vector code, whether at least one instruction of the plurality of instructions is a speculatively uniform instruction. The method may include generating, during complication of the vector code, uniformity detection code for the at least one speculatively uniform instruction. The uniformity detection code, when executed, may be configured to determine whether the at least one speculatively uniform instruction is uniform during runtime. The method may include generating, during complication of the vector code, scalar code by scalarizing the at least one speculatively uniform instruction. The scalar code may be configured to be compiled for execution by the first processor, a scalar processor, a scalar processing unit of the vector processor, or a vector pipeline of the vector processor.

Speculative scalarization in vector processing

Efficient processor load balancing using predication flags

Disclosed methods, systems, and computer program products embodiments include synchronizing a group of workitems on a processor by storing a respective program counter associated with each of the workitems, selecting at least one first workitem from the group for execution, and executing the selected at least one first workitem on the processor. The selecting is based upon the respective stored program counter associated with the at least one first workitem.

Method and system for synchronization of workitems with divergent control flow

A method, a system, and a non-transitory computer readable medium for parallelizing computer program code including a loop are presented. An intermediate language version of the computer program code is generated based on a parallel type of the loop, wherein the intermediate language version includes information about parallelism in the computer program code. The intermediate language version is optimized at runtime based on the device characteristics where the computer program code is to be executed. The parallel type may include a thread parallel type, wherein the loop is dispatched to multiple threads for execution, or a general parallel type, wherein the loop is dispatched to a single thread and may be vectorized for execution. The intermediate language version may be saved separate from the computer program code.

Layered programming for heterogeneous devices

In a computing system, memory may be managed by using a distributed array, which is a global set of local memory regions. A segment in the distributed array is allocated and is bound to a physical memory region. The segment is used by a workgroup in a dispatched data parallel kernel, wherein a workgroup includes one or more work items. When the distributed array is declared, parameters of the distributed array may be defined. The parameters may include an indication whether the distributed array is persistent (data written to the distributed array during one parallel dispatch is accessible by work items in a subsequent dispatch) or an indication whether the distributed array is shared (nested kernels may access the distributed array). The segment may be deallocated after it has been used.

Allocating memory and using the allocated memory in a workgroup in a dispatched data parallel kernel

A medium, method, and apparatus are disclosed for eliding superfluous function invocations in a vector-processing environment. A compiler receives program code comprising a width-contingent invocation of a function. The compiler creates a width-specific executable version of the program code by determining a vector width of a target computer system and omitting the function from the width-specific executable if the vector width meets one or more criteria. For example, the compiler may omit the function call if the vector width is greater than a minimum size.

Growing community of inventors

San Jose, CA, United States of America

Lee W Howes