WO2012088508A2 - Extensible data parallel semantics - Google Patents
Extensible data parallel semantics Download PDFInfo
- Publication number
- WO2012088508A2 WO2012088508A2 PCT/US2011/067185 US2011067185W WO2012088508A2 WO 2012088508 A2 WO2012088508 A2 WO 2012088508A2 US 2011067185 W US2011067185 W US 2011067185W WO 2012088508 A2 WO2012088508 A2 WO 2012088508A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data parallel
- compute
- level
- resource
- code
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/40—Transformation of program code
- G06F8/41—Compilation
- G06F8/43—Checking; Contextual analysis
- G06F8/436—Semantic checking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline, look ahead
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/30—Creation or generation of source code
- G06F8/31—Programming languages or programming paradigms
- G06F8/314—Parallel programming languages
Definitions
- Computer systems often include one or more general purpose processors (e.g., central processing units (CPUs)) and one or more specialized data parallel compute nodes (e.g., graphics processing units (GPUs) or single instruction, multiple data (SIMD) execution units in CPUs).
- general purpose processors generally perform general purpose processing on computer systems
- data parallel compute nodes generally perform data parallel processing (e.g., graphics processing) on computer systems.
- General purpose processors often have the ability to implement data parallel algorithms but do so without the optimized hardware resources found in data parallel compute nodes. As a result, general purpose processors may be far less efficient in executing data parallel algorithms than data parallel compute nodes.
- Data parallel compute nodes have traditionally played a supporting role to general purpose processors in executing programs on computer systems. As the role of hardware optimized for data parallel algorithms increases due to enhancements in data parallel compute node processing capabilities, it would be desirable to enhance the ability of programmers to program data parallel compute nodes and make the programming of data parallel compute nodes easier. Data parallel algorithms, however, are typically
- a high level programming language provides extensible data parallel semantics.
- User code specifies hardware and software resources for executing data parallel code using a compute device object and a resource view object.
- the user code uses the objects and semantic metadata to allow execution by new and / or updated types of compute nodes and new and / or updated types of runtime libraries.
- the extensible data parallel semantics allow the user code to be executed by the new and / or updated types of compute nodes and runtime libraries.
- Figure 1 is a block diagram illustrating an embodiment of a runtime environment with extensible data parallel semantics.
- Figure 2 is a computer code diagram illustrating an embodiment of code that implements extensible data parallel semantics.
- Figures 3A-3C is are block diagrams illustrating embodiments of runtime libraries in a runtime environment that implements extensible data parallel semantics.
- Figures 4A-4B are block diagrams illustrating embodiments of data structures that support extensible data parallel semantics.
- Figure 5 is a block diagram illustrating an embodiment of a computer system configured to compile and execute data parallel code with extensible data parallel semantics.
- FIG 1 is a block diagram illustrating an embodiment of a runtime environment 2 with extensible data parallel semantics in a computer system such as computer system 100 shown in Figure 5.
- Runtime environment 2 represents a runtime mode of operation in the computer system where the computer system is executing instructions from user code 10 and a set of one or more runtime libraries 20 on one or more compute nodes 121 (also shown in Figure 5 and described in additional detail below).
- Code 10 includes a sequence of instructions from a high level general purpose or data parallel programming language that may be compiled into one or more executables (e.g., DP executable 138 shown in Figure 5) for execution by one or more compute nodes 121.
- Code 10 executes in conjunction with one or more runtime libraries 20 where runtime libraries 20 include data parallel application programming interfaces (APIs) that provide data parallel functions.
- APIs application programming interfaces
- Code 10 causes a compute device object 12 to be generated from a runtime library 20 to specify a compute node 121 for executing at least a portion of code 10 and causes a resource view object 14 to be generated from a runtime library 20 to specify a runtime library 20 to be used in executing code 10.
- Compute device object 12 forms an abstraction of hardware that specifies a device level (i.e., a type of compute node 121).
- Resource view object 14 specifies a resource level that describes how to use the hardware specified by compute device object 12.
- Resource view object 14, for example, may describe different DirectX implementations (e,g., DirectX 11, DirectX 11.1, DirectX 12, and
- DirectX 13 or SSE/AVX implementations with native code generation or with WARP (a DirectX software emulator).
- Resource view object 14 may also include memory management and kernel execution services.
- the use of compute device object 12 and resource view object 14 along with associated semantic metadata provide extensible data parallel semantics for handling semantic changes of the underlying programming language of code 10.
- the extensible data parallel semantics allow code 10 to be executed with new and / or updated types of compute nodes 121 and new and / or updated types of runtime libraries 20.
- constructs of code 10 that were designed for use with specific types of compute nodes 121 may be executed by new and / or updated types of compute nodes 121.
- code 10 includes a sequence of instructions from a high level general purpose programming language with data parallel extensions (hereafter GP language) that form a program stored in a set of one or more modules.
- GP language may allow the program to be written in different parts (i.e., modules) such that each module may be stored in separate files or locations accessible by the computer system.
- the GP language provides a single language for programming a computing environment that includes one or more general purpose processors and one or more special purpose, DP optimal compute nodes.
- DP optimal compute nodes are typically graphic processing units (GPUs) or SIMD units of general purpose processors but may also include the scalar or vector execution units of general purpose processors, field programmable gate arrays (FPGAs), or other suitable devices in some computing environments.
- a programmer may include both general purpose processor and DP source code in code 10 for execution by general purpose processors and DP compute nodes, respectively, and coordinate the execution of the general purpose processor and DP source code.
- Code 10 may represent any suitable type of code in this embodiment, such as an application, a library function, or an operating system service.
- the GP language may be formed by extending a widely adapted, high level, and general purpose programming language such as C or C++ to include data parallel features.
- general purpose languages in which DP features may appear include JavaTM, PHP, Visual Basic, Perl, PythonTM, C#, Ruby, Delphi, Fortran, VB, F#, OCaml, Haskell, Erlang, NESL, Chapel, and JavaScriptTM.
- the GP language implementation may include rich linking capabilities that allow different parts of a program to be included in different modules.
- the data parallel features provide programming tools that take advantage of the special purpose architecture of DP optimal compute nodes to allow data parallel operations to be executed faster or more efficiently than with general purpose processors (i.e., non-DP optimal compute nodes).
- the GP language may also be another suitable high level general purpose programming language that allows a programmer to program for both general purpose processors and DP optimal compute nodes.
- code 10 includes a sequence of instructions from a high level data parallel programming language (hereafter DP language) that form a program.
- DP language provides a specialized language for programming a DP optimal compute node in a computing environment with one or more DP optimal compute nodes.
- a programmer uses the DP language, a programmer generates DP source code in code 10 that is intended for execution on DP optimal compute nodes.
- the DP language provides programming tools that take advantage of the special purpose architecture of DP optimal compute nodes to allow data parallel operations to be executed faster or more efficiently than with general purpose processors.
- the DP language may be an existing DP programming language such as HLSL, GLSL, Cg, C, C++, NESL, Chapel, CUD A, OpenCL, Accelerator, Ct, PGI GPGPU Accelerator, CAPS GPGPU Accelerator, Brook+, CAL, APL, Fortran 90 (and higher), Data Parallel C, DAPPLE, or APL.
- Code 10 may represent any suitable type of DP source code in this embodiment, such as an application, a library function, or an operating system service.
- Code 10 includes code portions designated for execution on a DP optimal compute node 121.
- a DP optimal compute node 121 has one or more computational resources with a hardware architecture that is optimized for data parallel computing (i.e., the execution of DP programs or algorithms).
- the GP language allows a programmer to designate DP source code using an annotation 29 (e.g., rl annote) when defining a vector function.
- the annotation 29 is associated with a function name 27 (e.g., vector func) of the vector function that is intended for execution on a DP optimal compute node.
- Code 10 may also include one or more invocations 28 of a vector function (e.g., forall ...
- a vector function corresponding to a call site is referred to as a kernel function.
- a kernel function may call other vector functions in code 10 (i.e., other DP source code) and may be viewed as the root of a vector function call graph.
- a kernel function may also use types (e.g., classes or structs) defined by code 10. The types may or may not be annotated as DP source code.
- other suitable programming language constructs may be used to designate portions of code 10 as DP source code and / or general purpose processor code.
- annotations 29 may be omitted in embodiments where code 10 is written in a DP language.
- Annotation 29 designates resource level semantics for the vector function.
- Annotation 29 allows a compiler to ensure that the semantic state of the vector function is compatible with the semantics and other characteristics of the target compute node 121 as reflected in compute device object 12 and resource view object 14. As the semantic restrictions ease over time, newer compute nodes 121 with fewer semantic restrictions may execute vector functions with older annotations 29 that indicate a higher level of semantic restrictions.
- Runtime libraries 20 include any suitable type and / or number of libraries that provide task parallel and / or data parallel (DP) execution capabilities.
- runtime libraries 20 may include DirectX runtime libraries and Concurrency Runtime libraries with Parallel Patterns Library (PPL) in one embodiment.
- Runtime libraries 20 provide application programming interfaces (APIs) or other suitable programming constructs that offer functions with task parallel and / or data parallel capabilities.
- APIs application programming interfaces
- Figures 3A-3C illustrate embodiments of 20(l)-20(3), respectively, of runtime libraries 20 for use in runtime environment 2 shown in Figure 1.
- runtime library 20(1) includes a compute device function 42, a resource view function 43, data parallel (DP) functions 44, and a call-site abstraction function 46.
- Compute device function 42 creates compute device object 12 (shown in Figure 1) according to parameters supplied by user code 10.
- user code 10 passes a device level (e.g, device level) that specifies a type of compute node for executing at least a portion of code 10 to create compute device object 12.
- compute device function 42 recognizes the device levels shown in the data structure of Figure 4A (i.e., none, custom, GPU, WARP, REF, NATIVE, SSE, AVX, and LRB). Accordingly, user code 10 may pass one of these device levels in this embodiment.
- Resource view function 43 creates resource view object 14 (shown in Figure 1) according to parameters supplied by user code 10.
- user code 10 passes a resource level (e.g, resource_level) that describes how to use the hardware specified by compute device object 12.
- Resource view function 43 recognizes the resource levels shown in the data structure of Figure 4B (i.e., none, NATIVE, CUSTOM, DX11, DX11.1, and DX12) in one embodiment. Accordingly, user code 10 may pass one of these resource levels in this embodiment.
- the device and resource levels of user code 10 indicate a level of semantic restrictions that user code 10 is required to meet. These semantic restrictions are assumed to decrease over time as new and updated compute nodes 121 and runtime libraries 20 are used to execute user code 10. Accordingly, compute nodes 121 and runtime libraries 20 are able to execute user code 10 if the compute nodes 121 and runtime libraries 20 have the same or fewer semantic restrictions than the device and resource levels of user code 10.
- user code 10 creates compute device object 12 with a device level of GPU and a resource view object 14 with a resource level of DX11.
- User code 10 also includes at least one vector function with an annotation 29 that indicates DX11.
- the device level of GPU indicates that at least a portion of user code 10 is written for execution on a GPU
- the resource level of DX11 indicates that at least a portion of user code 10 is written for execution with a runtime library 20 that includes or otherwise supports DirectX 11. Accordingly, the vector function or functions conform to the semantic restrictions of DirectX 11 and may be executed with DirectX 11 on a GPU.
- resource view object 14 has a resource level of DX12 rather than DX11 (i.e., user code 10 is written for execution with a runtime library 20 that includes or otherwise supports DirectX 12), then annotation(s) 29 of the vector function(s) may be DX11, DX 11.1, or DX12 because DirectX 12 includes fewer semantic restrictions than DirectX 11 and DirectX 11.1.
- user code 10 creates compute device object 12 with a device level of WARP and a resource view object 14 with a resource level of DX11 or DX12.
- User code 10 also includes at least one vector function with an annotation 29 that indicates DX11 if the resource level is DX11 or DX11 , DX 11.1 , or DX12 if the resource level is DX12.
- the device level of WARP indicates that at least a portion of user code 10 is written for execution by an SSE, AVX, or LRBni enabled DirectX simulator, and the resource level of DX11 or DX12 indicates that at least a portion of user code 10 is written for execution with a runtime library 20 that includes or otherwise supports DirectX 11 or DirectX 12.
- vector functions that conform to the semantic restrictions of DirectX 11 may be executed with DirectX 11 or higher on a WARP simulator and vector functions that conform to the semantic restrictions of DirectX 12 may be executed with DirectX 12 on a WARP simulator.
- user code 10 creates compute device object 12 with a device level of REF and a resource view object 14 with a resource level of DX11 or DX12.
- User code 10 also includes at least one vector function with an annotation 29 that indicates DX11 if the resource level is DX11 or DX11 , DX 11.1 , or DX12 if the resource level is DX12.
- the device level of REF indicates that at least a portion of user code 10 is written for execution by a single threaded CPU based DirectX simulator
- the resource level of DX11 or DX12 indicates that at least a portion of user code 10 is written for execution with a runtime library 20 that includes or otherwise supports DirectX 11 or DX12, respectively.
- vector functions that conform to the semantic restrictions of DirectX 11 may be executed with DirectX 11 or higher on a REF simulator and vector functions that conform to the semantic restrictions of DirectX 12 may be executed with DirectX 12 on a REF simulator.
- runtime libraries 20 that support DirectX.
- other hardware implementations may be supported by indicating the hardware type as a device level and native as the resource level.
- user code 10 may creates compute device object 12 with a device level of SSE to enable SSE vector units on CPUs from Intel and AMD, a device level of AVX to enable the Sandy Bridge CPU from Intel, or a device level of LRB to enable the Knights Ferry specialized data parallel optimized CPU from Intel.
- user code 10 may create a resource view object 14 with a native level (i.e., NATIVE) where vector functions of user code 10 are unrestricted with regard to data parallel semantics but conform to the semantics of the underlying general purpose language (e.g., C++).
- a native level i.e., NATIVE
- C++ general purpose language
- DP functions 44 provide data parallel functions to implement call-sites such as forall, scan, reduce, and sort for selected types of compute nodes 121 such as GPUs or those with WARP or REF software simulators.
- runtime library 20(1) provides a call-site abstraction function 46 that abstracts call-sites.
- Call-site abstraction function 46 may be used to implement call-site functionality for new and / or updated types of compute nodes 121.
- a user may also provide one or more runtime libraries 20(2) that include a custom compute device function 48, a custom resource view function, and / or custom DP functions 50.
- Custom compute device function 48 may be
- Custom resource view function 49 may also be implemented as an abstract base class and allows the user to provide new and / or updated resource levels that describe how to use the hardware for executing user code 10.
- Custom DP functions 50 allow the user to provide custom call-site implementations such as forall, scan, reduce, and sort that may be executed on the new and / or updated types of compute nodes 121.
- user code 10 creates compute device object 12 with a device level of CUSTOM and a resource view object 14 with a resource level of NATIVE.
- User code 10 also invokes custom compute device function 48 to provide the semantic metadata for the new and / or updated type of compute node 121.
- user code 10 may either use call-site abstraction function 46 of runtime library 20(1) to implement suitable call-site functionality for the compute node 121 or provide custom DP functions 50 that implement the call-sites.
- user code 10 creates compute device object 12 with a device level of NATIVE and a resource view object 14 with a resource level of CUSTOM.
- User code 10 also invokes custom resource view function 49 to provide a new and / or updated resource level that describes how to use the compute node 121 specified by compute device object 12 for executing user code 10.
- the resource level may be a custom version of Intel® Thread Building Blocks (TBB).
- TLB Thread Building Blocks
- user code 10 may either use call-site abstraction function 46 of runtime library 20(1) to implement suitable call-site functionality for the compute node 121 or provide custom DP functions 50 that implement the call-sites.
- user code 10 creates compute device object 12 with a device level of CUSTOM and a resource view object 14 with a resource level of CUSTOM.
- User code 10 invokes custom compute device function 48 to provide the semantic metadata for the new and / or updated type of compute node 121.
- User code 10 also invokes custom resource view function 49 to provide a new and / or updated resource level that describes how to use the compute node 121 specified by compute device object 12 for executing user code 10.
- user code 10 may either use call-site abstraction function 46 of runtime library 20(1) to implement suitable call-site functionality for the compute node 121 or provide custom DP functions 50 that implement the call-sites.
- user code 10 To execute vector functions on a host (i.e., a non- vector CPU), user code 10 creates compute device object 12 with a device level of NATIVE and a resource view object 14 with a resource level of NATIVE.
- the device level of NATIVE indicates that at least a portion of user code 10 may be executed on a host.
- the resource level of NATIVE allows vector functions of user code 10 to be unrestricted with regard to data parallel semantics but conform to the semantics of the underlying general purpose language (e.g., C++).
- user code 10 may either use call-site abstraction function 46 of runtime library 20(1) to implement suitable call-site functionality for the compute node 121 or provide custom DP functions 50 that implement the call-sites.
- a runtime library 20(3) with task parallel functions 52 may be used in conjunction with DP functions 44 ( Figure 3 A).
- task parallel functions 52 represents Concurrency Runtime libraries (ConcRT) with Parallel Patterns Library (PPL)
- Concurrency Runtime libraries ConcRT
- PPL Parallel Patterns Library
- an implementation of forall using task parallel functions 52 may be generated for execution on a host using virtual processors.
- implementations of foreach and transform from PPL and the C++ Standard Template Library (STL) may include range based signatures instead of linear iterators.
- the options for range patterns are: enumerator (basically the classical STL iterator pattern with
- the 'range' is the analogue of the compute domain in forall.
- a naive implementation may call forall inside of foreach when presented with a random access 'range' with data parallel sub-trait.
- the limited number of kernel arguments in foreach or transform may be mitigated by utilizing lambda closures.
- the built-in 'range'-based foreach and transform implementations activate the 'range '-traits to decide which implementation pattern.
- the new 'range' type may be overloaded.
- integration of data parallel functions 44 with task-based parallel programming runtime libraries 20 e.g., runtime library 20(3) with task parallel functions 52
- runtime library 20(3) with task parallel functions 52 may be achieved with user extensibility using existing language mechanisms.
- Figure 5 is a block diagram illustrating an embodiment of a computer system 100 configured to compile and execute data parallel code 10 with extensible data parallel semantics.
- Computer system 100 includes a host 101 with one or more processing elements (PEs) 102 housed in one or more processor packages (not shown) and a memory system 104. Computer system 100 also includes zero or more input / output devices 106, zero or more display devices 108, zero or more peripheral devices 110, and zero or more network devices 112. Computer system 100 further includes a compute engine 120 with one or more DP optimal or other types of compute nodes 121 where each DP optimal compute node 121 includes a set of one or more processing elements (PEs) 122 and a memory 124 that stores DP executable 138. [0041] Host 101, input / output devices 106, display devices 108, peripheral devices 110, network devices 112, and compute engine 120 communicate using a set of
- interconnections 114 that includes any suitable type, number, and configuration of controllers, buses, interfaces, and / or other wired or wireless connections.
- Computer system 100 represents any suitable processing device configured for a general purpose or a specific purpose. Examples of computer system 100 include a server, a personal computer, a laptop computer, a tablet computer, a smart phone, a personal digital assistant (PDA), a mobile telephone, and an audio/video device.
- the components of computer system 100 i.e., host 101, input / output devices 106, display devices 108, peripheral devices 110, network devices 112, interconnections 114, and compute engine 120
- Processing elements 102 each form execution hardware configured to execute instructions (i.e., software) stored in memory system 104.
- the processing elements 102 in each processor package may have the same or different architectures and / or instruction sets.
- the processing elements 102 may include any combination of in-order execution elements, superscalar execution elements, and data parallel execution elements (e.g., GPU execution elements).
- Each processing element 102 is configured to access and execute instructions stored in memory system 104.
- the instructions may include a basic input output system (BIOS) or firmware (not shown), an operating system (OS) 132, code 10, compiler 134, GP executable 136, and DP executable 138.
- BIOS basic input output system
- OS operating system
- GP executable 136 GP executable 136
- DP executable 138 DP executable 138.
- Each processing element 102 may execute the instructions in conjunction with or in response to information received from input / output devices 106, display devices 108, peripheral devices 110, network devices 112, and / or
- OS 132 includes instructions executable by the processing elements to manage the components of computer system 100 and provide a set of functions that allow programs to access and use the components. In one
- OS 132 is the Windows operating system. In other embodiments, OS 132 is another operating system suitable for use with computer system 100.
- compiler 134 When computer system executes compiler 134 to compile code 10, compiler 134 generates one or more executables - e.g., one or more GP executables 136 and one or more DP executables 138. In other embodiments, compiler 134 may generate one or more GP executables 136 to each include one or more DP executables 138 or may generate one or more DP executables 138 without generating any GP executables 136. GP executables 136 and / or DP executables 138 are generated in response to an invocation of compiler 134 with data parallel extensions to compile all or selected portions of code 10. The invocation may be generated by a programmer or other user of computer system 100, other code in computer system 100, or other code in another computer system (not shown), for example.
- GP executable 136 represents a program intended for execution on one or more general purpose processing elements 102 (e.g., central processing units (CPUs)). GP executable 136 includes low level instructions from an instruction set of one or more general purpose processing elements 102.
- general purpose processing elements 102 e.g., central processing units (CPUs)
- CPUs central processing units
- DP executable 138 represents a data parallel program or algorithm (e.g., a shader) that is intended and optimized for execution on one or more data parallel (DP) optimal compute nodes 121.
- DP executable 138 includes DP byte code or some other intermediate representation (IL) that is converted to low level instructions from an instruction set of a DP optimal compute node 121 using a device driver (not shown) prior to being executed on the DP optimal compute node 121.
- DP executable 138 includes low level instructions from an instruction set of one or more DP optimal compute nodes 121 where the low level instructions were inserted by compiler 134.
- GP executable 136 is directly executable by one or more general purpose processors (e.g., CPUs), and DP executable 138 is either directly executable by one or more DP optimal compute nodes 121 or executable by one or more DP optimal compute nodes 121 subsequent to being converted to the low level instructions of the DP optimal compute node 121.
- general purpose processors e.g., CPUs
- DP executable 138 is either directly executable by one or more DP optimal compute nodes 121 or executable by one or more DP optimal compute nodes 121 subsequent to being converted to the low level instructions of the DP optimal compute node 121.
- Computer system 100 may execute GP executable 136 using one or more processing elements 102, and computer system 100 may execute DP executable 138 using one or more PEs 122 as described in additional detail below.
- Memory system 104 includes any suitable type, number, and configuration of volatile or non-volatile storage devices configured to store instructions and data.
- the storage devices of memory system 104 represent computer readable storage media that store computer-executable instructions (i.e., software) including OS 132, code 10, compiler 134, GP executable 136, and DP executable 138.
- the instructions are executable by computer system 100 to perform the functions and methods of OS 132, code 10, compiler 134, GP executable 136, and DP executable 138 as described herein.
- Memory system 104 stores instructions and data received from processing elements 102, input / output devices 106, display devices 108, peripheral devices 110, network devices 112, and compute engine 120.
- Memory system 104 provides stored instructions and data to processing elements 102, input / output devices 106, display devices 108, peripheral devices 110, network devices 112, and compute engine 120.
- Examples of storage devices in memory system 104 include hard disk drives, random access memory (RAM), read only memory (ROM), flash memory drives and cards, and magnetic and optical disks such as CDs and DVDs.
- Input / output devices 106 include any suitable type, number, and configuration of input / output devices configured to input instructions or data from a user to computer system 100 and output instructions or data from computer system 100 to the user.
- Examples of input / output devices 106 include a keyboard, a mouse, a touchpad, a touchscreen, buttons, dials, knobs, and switches.
- Display devices 108 include any suitable type, number, and configuration of display devices configured to output textual and / or graphical information to a user of computer system 100. Examples of display devices 108 include a monitor, a display screen, and a projector.
- Peripheral devices 110 include any suitable type, number, and configuration of peripheral devices configured to operate with one or more other components in computer system 100 to perform general or specific processing functions.
- Network devices 112 include any suitable type, number, and configuration of network devices configured to allow computer system 100 to communicate across one or more networks (not shown).
- Network devices 112 may operate according to any suitable networking protocol and / or configuration to allow information to be transmitted by computer system 100 to a network or received by computer system 100 from a network.
- Compute engine 120 is configured to execute DP executable 138.
- Compute engine 120 includes one or more compute nodes 121.
- Each compute node 121 is a collection of computational resources that share a memory hierarchy.
- Each compute node 121 includes a set of one or more PEs 122 and a memory 124 that stores DP executable 138.
- PEs 122 execute DP executable 138 and store the results generated by DP executable 138 in memory 124.
- a compute node 121 that has one or more computational resources with a hardware architecture that is optimized for data parallel computing (i.e., the execution of DP programs or algorithms) is referred to as a DP optimal compute node 121.
- Examples of a DP optimal compute node 121 include a node 121 where the set of PEs 122 includes one or more GPUs and a node 121 where the set of PEs 122 includes the set of SIMD units in a general purpose processor package.
- a compute node 121 that does not have any computational resources with a hardware architecture that is optimized for data parallel computing e.g., processor packages with only general purpose processing elements 102
- memory 124 may be separate from memory system 104 (e.g., GPU memory used by a GPU) or a part of memory system 104 (e.g., memory used by SIMD units in a general purpose processor package).
- Host 101 forms a host compute node that is configured to provide DP executable 138 to a compute node 121 for execution and receive results generated by DP executable 138 using interconnections 114.
- the host compute node includes is a collection of general purpose computational resources (i.e., general purpose processing elements 102) that share a memory hierarchy (i.e., memory system 104).
- the host compute node may be configured with a symmetric multiprocessing architecture (SMP) and may also be configured to maximize memory locality of memory system 104 using a non-uniform memory access (NUMA) architecture, for example.
- SMP symmetric multiprocessing architecture
- NUMA non-uniform memory access
- OS 132 of the host compute node is configured to execute a DP call site to cause a DP executable 138 to be executed by a DP optimal or non-DP optimal compute node 121.
- the host compute node causes DP executable 138 and one or more indexable types 14 to be copied from memory system 104 to memory 124.
- the host compute node may designate a copy of DP executable 138 and / or one or more indexable types 14 in memory system 104 as memory 124 and / or may copy DP executable 138 and / or one or more indexable types 14 from one part of memory system 104 into another part of memory system 104 that forms memory 124.
- the copying process between compute node 121 and the host compute node may be a synchronization point unless designated as asynchronous.
- the host compute node and each compute node 121 may concurrently execute code independently of one another.
- the host compute node and each compute node 121 may interact at synchronization points to coordinate node computations.
- compute engine 120 represents a graphics card where one or more graphics processing units (GPUs) include PEs 122 and a memory 124 that is separate from memory system 104.
- a driver of the graphics card (not shown) may convert byte code or some other intermediate representation (IL) of DP executable 138 into the instruction set of the GPUs for execution by the PEs 122 of the GPUs.
- compute engine 120 is formed from the combination of one or more GPUs (i.e. PEs 122) that are included in processor packages with one or more general purpose processing elements 102 and a portion of memory system 104 that includes memory 124.
- additional software may be provided on computer system 100 to convert byte code or some other intermediate representation (IL) of DP executable 138 into the instruction set of the GPUs in the processor packages.
- compute engine 120 is formed from the combination of one or more SIMD units in one or more of the processor packages that include processing elements 102 and a portion of memory system 104 that includes memory 124.
- additional software may be provided on computer system 100 to convert the byte code or some other intermediate representation (IL) of DP executable 138 into the instruction set of the SIMD units in the processor packages.
- IL intermediate representation
- compute engine 120 is formed from the combination of one or more scalar or vector processing pipelines in one or more of the processor packages that include processing elements 102 and a portion of memory system 104 that includes memory 124.
- additional software may be provided on computer system 100 to convert the byte code or some other intermediate representation (IL)of DP executable 138 into the instruction set of the scalar processing pipelines in the processor packages.
- IL intermediate representation
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP11850729.2A EP2656203A4 (en) | 2010-12-23 | 2011-12-23 | Extensible data parallel semantics |
JP2013546453A JP5957006B2 (en) | 2010-12-23 | 2011-12-23 | Extensible data parallel semantics |
CA2822100A CA2822100A1 (en) | 2010-12-23 | 2011-12-23 | Extensible data parallel semantics |
KR1020137016150A KR101962484B1 (en) | 2010-12-23 | 2011-12-23 | Extensible data parallel semantics |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US12/977,207 US9841958B2 (en) | 2010-12-23 | 2010-12-23 | Extensible data parallel semantics |
US12/977,207 | 2010-12-23 |
Publications (2)
Publication Number | Publication Date |
---|---|
WO2012088508A2 true WO2012088508A2 (en) | 2012-06-28 |
WO2012088508A3 WO2012088508A3 (en) | 2012-11-29 |
Family
ID=46314971
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2011/067185 WO2012088508A2 (en) | 2010-12-23 | 2011-12-23 | Extensible data parallel semantics |
Country Status (8)
Country | Link |
---|---|
US (1) | US9841958B2 (en) |
EP (1) | EP2656203A4 (en) |
JP (1) | JP5957006B2 (en) |
KR (1) | KR101962484B1 (en) |
CN (1) | CN102566980B (en) |
CA (1) | CA2822100A1 (en) |
HK (1) | HK1172968A1 (en) |
WO (1) | WO2012088508A2 (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9841958B2 (en) * | 2010-12-23 | 2017-12-12 | Microsoft Technology Licensing, Llc. | Extensible data parallel semantics |
US9632771B2 (en) * | 2012-12-13 | 2017-04-25 | Microsoft Technology Licensing, Llc | Association of metadata with source code and applications and services premised thereon |
US10101982B2 (en) * | 2013-01-31 | 2018-10-16 | Htc Corporation | Methods for application management in an electronic device supporting hardware acceleration |
KR101632027B1 (en) * | 2014-11-11 | 2016-06-20 | 포트리스이노베이션 주식회사 | Method for converting program using pseudo code based comment and computer-readable recording media storing the program performing the said mehtod |
US20170046168A1 (en) * | 2015-08-14 | 2017-02-16 | Qualcomm Incorporated | Scalable single-instruction-multiple-data instructions |
US10409560B1 (en) * | 2015-11-18 | 2019-09-10 | Amazon Technologies, Inc. | Acceleration techniques for graph analysis programs |
JP6563363B2 (en) * | 2016-05-13 | 2019-08-21 | 日本電信電話株式会社 | Setting server, setting method and setting program |
WO2018094087A1 (en) * | 2016-11-17 | 2018-05-24 | The Mathworks, Inc. | Systems and methods for generating code for parallel processing units |
US11372633B2 (en) * | 2018-01-04 | 2022-06-28 | Shenzhen Tinysoft Co., Ltd. | Method, device and terminal apparatus for code execution and computer readable storage medium |
CN109918133A (en) * | 2019-01-24 | 2019-06-21 | 董栋挺 | A kind of electrical power transmission system multi-core task processing method |
KR20200093105A (en) | 2019-01-25 | 2020-08-05 | 삼성전자주식회사 | Method implemented by processor of electronic device and processor to operate electronic device for heterogeneous processors |
CN111832736B (en) * | 2019-04-19 | 2024-04-12 | 伊姆西Ip控股有限责任公司 | Method, apparatus and computer readable storage medium for processing machine learning model |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080109795A1 (en) | 2006-11-02 | 2008-05-08 | Nvidia Corporation | C/c++ language extensions for general-purpose graphics processing unit |
WO2009138812A1 (en) | 2008-05-16 | 2009-11-19 | Ateji | Process and system for development of parallel programs |
US20100205580A1 (en) | 2009-02-10 | 2010-08-12 | International Business Machines Corporation | Generating parallel simd code for an arbitrary target architecture |
Family Cites Families (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH11316688A (en) | 1998-05-06 | 1999-11-16 | Mitsubishi Electric Corp | Device and method for processing programming language and recording medium |
US6675189B2 (en) | 1998-05-28 | 2004-01-06 | Hewlett-Packard Development Company, L.P. | System for learning and applying integrated task and data parallel strategies in dynamic applications |
US6828975B2 (en) | 2001-03-01 | 2004-12-07 | Microsoft Corporation | Method and system for managing graphics objects in a graphics display system |
US7733347B2 (en) * | 2004-11-05 | 2010-06-08 | Microsoft Corporation | Automated construction of shader programs |
US8136104B2 (en) * | 2006-06-20 | 2012-03-13 | Google Inc. | Systems and methods for determining compute kernels for an application in a parallel-processing computer system |
US7822594B2 (en) * | 2006-08-07 | 2010-10-26 | Voltaire Ltd. | Service-oriented infrastructure management |
US8341611B2 (en) * | 2007-04-11 | 2012-12-25 | Apple Inc. | Application interface on multiple processors |
EP2135163B1 (en) | 2007-04-11 | 2018-08-08 | Apple Inc. | Data parallel computing on multiple processors |
US8069190B2 (en) | 2007-12-27 | 2011-11-29 | Cloudscale, Inc. | System and methodology for parallel stream processing |
US8566830B2 (en) | 2008-05-16 | 2013-10-22 | Microsoft Corporation | Local collections of tasks in a scheduler |
US8214814B2 (en) * | 2008-06-24 | 2012-07-03 | International Business Machines Corporation | Sharing compiler optimizations in a multi-node system |
US8645933B2 (en) * | 2008-08-01 | 2014-02-04 | Leon Schwartz | Method and apparatus for detection and optimization of presumably parallel program regions |
US8495603B2 (en) * | 2008-08-11 | 2013-07-23 | International Business Machines Corporation | Generating an executable version of an application using a distributed compiler operating on a plurality of compute nodes |
US8549496B2 (en) * | 2009-02-27 | 2013-10-01 | Texas Tech University System | Method, apparatus and computer program product for automatically generating a computer program using consume, simplify and produce semantics with normalize, transpose and distribute operations |
US8239847B2 (en) * | 2009-03-18 | 2012-08-07 | Microsoft Corporation | General distributed reduction for data parallel computing |
US8209664B2 (en) * | 2009-03-18 | 2012-06-26 | Microsoft Corporation | High level programming extensions for distributed data parallel processing |
US9349154B2 (en) * | 2010-04-05 | 2016-05-24 | Nvidia Corporation | Bindless texture and image API |
US9424087B2 (en) * | 2010-04-29 | 2016-08-23 | International Business Machines Corporation | Optimizing collective operations |
US20110314256A1 (en) * | 2010-06-18 | 2011-12-22 | Microsoft Corporation | Data Parallel Programming Model |
US8589867B2 (en) * | 2010-06-18 | 2013-11-19 | Microsoft Corporation | Compiler-generated invocation stubs for data parallel programming model |
US8756590B2 (en) * | 2010-06-22 | 2014-06-17 | Microsoft Corporation | Binding data parallel device source code |
US9395957B2 (en) * | 2010-12-22 | 2016-07-19 | Microsoft Technology Licensing, Llc | Agile communication operator |
US9841958B2 (en) * | 2010-12-23 | 2017-12-12 | Microsoft Technology Licensing, Llc. | Extensible data parallel semantics |
-
2010
- 2010-12-23 US US12/977,207 patent/US9841958B2/en active Active
-
2011
- 2011-12-22 CN CN201110436176.8A patent/CN102566980B/en active Active
- 2011-12-23 EP EP11850729.2A patent/EP2656203A4/en not_active Ceased
- 2011-12-23 WO PCT/US2011/067185 patent/WO2012088508A2/en active Application Filing
- 2011-12-23 CA CA2822100A patent/CA2822100A1/en not_active Abandoned
- 2011-12-23 JP JP2013546453A patent/JP5957006B2/en active Active
- 2011-12-23 KR KR1020137016150A patent/KR101962484B1/en active IP Right Grant
-
2013
- 2013-01-02 HK HK13100019.2A patent/HK1172968A1/en not_active IP Right Cessation
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080109795A1 (en) | 2006-11-02 | 2008-05-08 | Nvidia Corporation | C/c++ language extensions for general-purpose graphics processing unit |
WO2009138812A1 (en) | 2008-05-16 | 2009-11-19 | Ateji | Process and system for development of parallel programs |
US20100205580A1 (en) | 2009-02-10 | 2010-08-12 | International Business Machines Corporation | Generating parallel simd code for an arbitrary target architecture |
Non-Patent Citations (1)
Title |
---|
See also references of EP2656203A4 |
Also Published As
Publication number | Publication date |
---|---|
CN102566980B (en) | 2015-03-18 |
EP2656203A2 (en) | 2013-10-30 |
JP2014501412A (en) | 2014-01-20 |
EP2656203A4 (en) | 2016-03-30 |
JP5957006B2 (en) | 2016-07-27 |
WO2012088508A3 (en) | 2012-11-29 |
KR20130137652A (en) | 2013-12-17 |
KR101962484B1 (en) | 2019-03-26 |
CN102566980A (en) | 2012-07-11 |
CA2822100A1 (en) | 2012-06-28 |
HK1172968A1 (en) | 2013-05-03 |
US20120166772A1 (en) | 2012-06-28 |
US9841958B2 (en) | 2017-12-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US9841958B2 (en) | Extensible data parallel semantics | |
US8756590B2 (en) | Binding data parallel device source code | |
US9489183B2 (en) | Tile communication operator | |
JP6027021B2 (en) | Agile communication operator | |
US10282179B2 (en) | Nested communication operator | |
KR20140097548A (en) | Software libraries for heterogeneous parallel processing platforms | |
US8402450B2 (en) | Map transformation in data parallel code | |
US10620916B2 (en) | Read-only communication operator | |
US8713039B2 (en) | Co-map communication operator | |
US10496433B2 (en) | Modification of context saving functions | |
KR101632027B1 (en) | Method for converting program using pseudo code based comment and computer-readable recording media storing the program performing the said mehtod | |
US8984473B2 (en) | Methods for type analysis in systems for code generation | |
Jain | Programming Languages | |
Linderman | A programming model and processor architecture for heterogeneous multicore computers | |
Swaine | Electrical Engineering and Computer Science Department |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 11850729 Country of ref document: EP Kind code of ref document: A2 |
|
ENP | Entry into the national phase |
Ref document number: 2822100 Country of ref document: CA |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2011850729 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 20137016150 Country of ref document: KR Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 2013546453 Country of ref document: JP Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |