computerized method of managing a plurality of subbuffers, non-transitory machine readable means and

专利摘要:
SUBBUFFER OBJECTS. The present invention relates to a method and apparatus for a parallel computer program using subbuffers to perform a data processing task in parallel between heterogeneous computing units. Computing units can include a heterogeneous mix of central processing units (CPUs) and graphics processing units (GPUs). A system creates a subbuffer of a master buffer for each of a plurality of heterogeneous computing units. If a subbuffer is not associated with the same computing unit as the main buffer, the system copies data from the subbuffer to the memory of that computing unit. The system even tracks updates to the data and transfers those updates back to the subbuffer.
公开号:BR112012027950B1
申请号:R112012027950-5
申请日:2011-04-20
公开日:2020-11-10
发明作者:Aaftab A. Munshi；Ian R. Ollman
申请人:Aplle Inc；
IPC主号:

专利说明:

RELATED REQUESTS
[001] The applicant claims the benefit of the priority of the previous copending provisional application Serial No. 61 / 346,866, deposited on May 20, 2010, all of which is incorporated by reference. FIELD OF THE INVENTION
[002] The present invention relates in general to parallel data computing. More particularly, this invention relates to the management of subbuffer objects associated with a buffer in a heterogeneous multicomputing unit environment. BACKGROUND
[003] As GPUs continue to evolve into high-performance parallel computing devices, more and more applications are written to perform parallel computing of data on GPUs similar to general-purpose computing devices. Today, these applications are designed to run on specific GPUs using vendor-specific interfaces. Thus, these applications are not able to leverage CPU processing resources even when both GPUs and CPUs are available in the data processing system. Nor can processing resources be leveraged through GPUs from different vendors where such an application is running.
[004] However, as more and more CPUs adopt multiple cores to perform parallel data computations, more and more processing tasks can be supported by both CPUs and / or GPUs, no matter which one is available. Traditionally, GPUs and CPUs are configured through separate programming environments that are not compatible with each other. Most GPUs require dedicated programs that are specific to the vendor. As a result, it is very difficult for an application to leverage the processing capabilities of both CPUs and GPUs, for example, by leveraging the processing capabilities of GPUs with parallel data computing capabilities along with multicore CPUs.
[005] In addition, CPUs and GPUs use separate memory address areas. The memory buffer needs to be allocated and copied into the GPU's memory for the GPU to process the data. If an application wants the CPU and one or more GPUs to operate in regions of a data buffer, the application needs to manage the allocation and copying of data from appropriate buffer regions that must be shared between CPU and GPU or through GPUs. Consequently, there is a need for modern data processing systems to have a heterogeneous mix of CPUs and GPUs sharing a buffer. DESCRIPTION SUMMARY
[006] A method and apparatus for a parallel computing program using subbuffers to perform a data processing task in parallel between heterogeneous computing units are described. Computing units can include a heterogeneous mix of central processing units (CPUs) and graphics processing units (GPUs). A system creates a subbuffer from a master buffer for each of a plurality of heterogeneous computing units. If a subbuffer is not associated with the same computing unit as the main buffer, the system copies data from the subbuffer to the memory of that computing unit. The system additionally tracks updates to the data and transfers those updates back to the subbuffer.
[007] Other characteristics of the present invention will be evident from the accompanying drawings and from the detailed description that follows. BRIEF DESCRIPTION OF THE DRAWINGS
[008] The present invention is illustrated by way of example and not of limitation in the figures in the accompanying drawings, in which similar references indicate similar elements and in which:
[009] figure 1 is a block diagram illustrating a modality of a system for configuring computing devices including CPUs and / or GPUs to perform parallel data computing for applications;
[0010] figure 2 is a block diagram illustrating an example of a computing device with multiple computing processors operating in parallel to run multiple lines at the same time;
[0011] figure 3 is a block diagram illustrating a modality of a plurality of physical computing devices configured as a logical computing device using a computing device identifier;
[0012] figure 4 is a block diagram illustrating a modality of a buffer subdivided into multiple subbuffers;
[0013] figure 5 is a block diagram illustrating a modality of multiple subbuffers in a one-dimensional buffer;
[0014] figure 6 is a block diagram illustrating a modality of a two-dimensional image subdivided into multiple subbuffers;
[0015] figure 7 is a block diagram illustrating a modality of a three-dimensional image subdivided into multiple subbuffers;
[0016] figure 8 is a flow diagram illustrating an embodiment of a process for configuring a plurality of physical computing devices with a computing device identifier by combining a capacity requirement received from an application;
[0017] figure 9 is a flow diagram illustrating a modality of a process for executing a computable executable on a logical computing device;
[0018] figure 10 is a flow diagram illustrating a modality of a runtime process to create and use subbuffers with multiple computing units;
[0019] figure 11 is a flow diagram illustrating a modality of a process to perform returns associated with events that have external and internal dependencies;
[0020] figure 12 is a block diagram illustrating a modality of a chain of events with external and internal dependencies;
[0021] figure 13 is an exemplary source code illustrating an example of a computing kernel source for a computing kernel executable to be run on a plurality of physical computing devices;
[0022] Figures 14A to 14C include an exemplary source code illustrating an example for configuring a logical computing device to run one of a plurality of executables on a plurality of physical computing devices calling APIs;
[0023] figure 15 illustrates an example of a typical computer system with a plurality of CPUs and GPUs (graphics processing unit) that can be used in conjunction with the modalities described here. DETAILED DESCRIPTION
[0024] A method and apparatus for parallel computing of data on multiple processors using subbuffers created from a main buffer are described here. In the following description, numerous specific details are presented to provide a complete explanation of the modalities of the present invention. It will be apparent, however, to a person skilled in the art, that the modalities of the present invention can be realized without these specific details. In other cases, well-known techniques, structures and components were not shown in detail in order not to impair the understanding of this description.
[0025] Reference in the specification to "a modality" or "a modality" means that a particular characteristic, structure or attribute described in connection with the modality can be included in at least one embodiment of the invention. The appearance of the phrase "in one modality" in various parts of the specification does not necessarily refer to the same modality.
[0026] The processes described in the figures that follow are carried out through the processing logic that comprises hardware (for example, circuit, dedicated logic, etc.), software (as it is executed in a general purpose computer system or a machine dedicated), or a combination of both. Although the processes are described below in terms of some sequential operations, it must be taken into account that some of the described operations can be performed in a different order. In addition, some operations can be performed in parallel instead of sequentially.
[0027] A Graphics Processing Unit (GPU) can be a dedicated graphics processor implementing highly efficient graphics operations, such as 2D, 3D graphics operation and / or functions related to digital video. A GPU can include special (programmable) hardware to perform graphics operations, for example, blitter operations, texture mapping, polygon rendering, pixel shading and vertex shading. GPUs are known to fetch data from a frame buffer and to mix pixels together to render an image back into the frame buffer for display. GPUs can also control the frame buffer and allow the frame buffer to be used to restore a screen, for example, a CRT or LCD screen. Both a CRT or LCD screen is a small persistent screen that requires restoration at a rate of at least 20Hz (for example, every 1/30 of a second, the screen is restored with data from a frame buffer). Usually, GPUs can take graphics processing tasks from CPUs coupled with GPUs to produce images of output raster graphics to display devices via screen controllers. References in the specification to "GPU" can be a graphics processor or a programmable graphics processor as described in "Method and Apparatus for Processing Multi-Transaction Data in a Programmable Graphics Processor", Lindholdm et al., Pa-try US No. 7015913, and "Method for Deinterlacing Video Interlaced by a Graphics Processor", Swan et al., US Patent No. 6970206, which are incorporated herein by reference.
[0028] In one embodiment, a plurality of different types of processors, such as CPUs or GPUs, can perform parallel data processing tasks for one or more applications at the same time to increase the effectiveness of using processing resources available in one data processing system. Processing capabilities of a data processing system can be based on a plurality of physical computing devices, such as CPUs or GPUs. A physical computing device can include one or more computing units. In one embodiment, parallel data processing tasks (or parallel data tasks) can be assigned to a plurality of types of processors, for example, CPUs or GPUs capable of carrying out the tasks. A parallel data task may require certain processor-specific processing capabilities. Processing capabilities can be, for example, hardware support for dedicated texturing, double precision floating point arithmetic, dedicated local memory, flow data cache or simple synchronization elements. Separate types of processors can provide different but overlapping groups of processing capabilities. For example, both CPU and GPU may be able to perform double precision floating point computing. In one embodiment, an application is able to leverage both a CPU and a GPU, no matter what is available, to perform a task of parallel data processing.
[0029] In another mode, the system can allocate a main buffer and additionally subdivide that main buffer into multiple subbuffers. If the computing unit for the subbuffer is the same computing unit as the one associated with the main buffer, that computing unit accesses the data from the subbuffer using pointers. If the computing unit for the subbuffer is different from the computing unit for the main buffer, the system copies the data from the subbuffer. In addition, the system tracks updates for the copied data and transfers the updated data back to the subfuffer.
[0030] Figure 1 is a block diagram illustrating a modality of a system 100 for configuring computing devices including CPUs and / or GPUs to perform parallel data computing for applications. System 100 can implement a parallel computing architecture. In one embodiment, system 100 may be a graphics system including one or more main processors coupled to one or more central processors 117 and one or more other processors such as media processors 115 over a data bus 113. A plurality of main processors can be interconnected in hosting systems 101. The plurality of central processors 117 may include multi-core CPUs from different vendors. A computing processor or computing unit, such as a CPU or GPU, can be associated with a group of capabilities. For example, a media processor can be a GPU with dedicated texture rendering hardware. Another media processor may be a GPU supporting both dedicated texture rendering hardware and double precision floating point arithmetic. Multiple GPUs can be connected together for Scalable Link Interface (SLI) or CrossFire configurations.
[0031] In one embodiment, the hosting systems 101 can support a stack of software. The software stack may include software stack components such as applications 103, a computing platform layer 141, for example, an OpenCL platform (Open Computing Language), a computing runtime layer 109, a computer compiler. computing 107 and computing application libraries 105. An application 103 can interact with other components stacked through API calls. One or more lines can be running at the same time for application 103 on hosting systems 101. The computing platform layer 141 can maintain a data structure, or a computing device data structure, by storing processing capabilities for each connected physical computing device. In one embodiment, an application can retrieve information about processing resources available from hosting systems 101 through the computing platform layer 141. An application can select and specify capacity requirements to perform a processing task through the computing platform layer 141. Consequently, computing platform layer 141 can determine a configuration for physical computing devices to allocate and initialize processing resources from connected 117 CPUs and / or GPUs 115 for the processing task. In one embodiment, the computing platform layer 141 can generate one or more logical computing devices for the application corresponding to one or more actual physical computing devices configured.
[0032] The computing runtime layer 109 can manage the execution of a processing task according to the processing resources configured for an application 103, for example, based on one or more logical computing devices. In one embodiment, executing a processing task may include creating a computer object program representing the processing task and allocating memory resources, for example, waiting executables, input / output data, etc. An executable loaded into a computer object program can be a computer program executable. A computer program executable can be included in a computer object program to run on a computer processor or computer unit, such as a CPU or GPU. The computation runtime layer 109 can interact with the physical devices allocated to perform the actual execution of the processing task. In one embodiment, the computing runtime layer 109 can coordinate multiple processing tasks running from different applications according to the runtime states of each processor, such as CPU or GPU configured for processing tasks . The computing runtime layer 109 can select, based on the runtime states, one or more processors from the physical computing devices configured to perform the processing tasks. Performing a processing task can include running multiple lines of one or more executables on a plurality of physical computing devices at the same time. In one embodiment, the computation runtime layer 109 can track the status of each processing task performed by monitoring the execution state of the execution time of each processor.
[0033] The runtime layer can load one or more executables as computer program executables corresponding to a processing task from an application 103. In one embodiment, the computing runtime layer 109 automatically loads executables additional data required to perform a processing task from the computing application library 105. The computing runtime layer 109 can load both an executable and its corresponding source program into an application 103 or object computing program. computer application library 105. A source program for a computer object program can be a computer program source. A plurality of executables based on a single computer program source can be loaded according to a logical computing device configured to include multiple different types and / or versions of physical computing devices. In one embodiment, the computing runtime layer 109 can activate computing compiler 107 to compile online a source program loaded into an executable optimized for a target processor, for example, a CPU or a GPU configured to run the executable .
[0034] An executable compiled online can be stored to be called in the future in addition to the existing executables according to a corresponding source program. Additionally, executables can be compiled offline and loaded into computation runtime 109 using API calls. Compute application library 105 and / or application 103 can load an associated executable in response to API library requests from an application. Recently compiled executables can be dynamically updated to compute application library 105 or to application 103. In one embodiment, compute runtime 109 can replace an existing compute program executable in an application with a new compiled executable online computing compiler 107 for a recently updated version of the computing device. Compute runtime 109 can insert a new compiled executable online to update compute application library 105. In one embodiment, compute runtime 109 can call compiler 107 when loading an executable for a task. processing. In another embodiment, computer compiler 107 can be called offline to establish executables for computer application library 105. Computer compiler 107 can compile and link a computer kernel program to generate a computer program executable. In one embodiment, the computing application library 105 may include a plurality of functions to support, for example, development tool kits and / or image processing. Each function in the library can correspond to a computer program source and one or more computer program executables stored in computer application library 105 for a plurality of physical computing devices.
[0035] Figure 2 is a block diagram illustrating an example of a computing device with multiple computing processors (for example, computing units) operating in parallel to run multiple lines at the same time. Each computing processor can execute a plurality of lines in parallel (or at the same time). Lines that can be run in parallel on a computing processor or computing unit can be referred to as a line group. A computing device could have multiple line groups that can be run in parallel. For example, M lines are shown to run as a line group on the computing device 205. Multiple groups of lines, for example, line 1 of the computer processor_1 205 and line N of the computer processor_L 203, can be run in parallel through separate computing processors on a computing device or across multiple computing devices. A plurality of line groups across multiple computing processors can run a computer program executable in parallel. More than one computing processor can be based on a single chip, such as an ASIC (Application Specific Integrated Circuits) device. In one embodiment, multiple lines of an application can be run at the same time on more than one computing processor across multiple chips.
[0036] A computing device can include one or more computing processors or computing units such as a Processor_1 205 and Processor_L 203. A local memory can be coupled to a computing processor. Local memory, shared between lines in a single line group running on a computing processor, can be supported by the local memory attached to the computing processor. Multiple lines from different line groups, such as line 1 213 and line N 209, can share a computing memory object, such as a stream, stored in a computing device memory 217 coupled to computing device 201. A computing device memory 217 can include a global memory and a constant memory. Global memory can be used to allocate computing memory objects, such as streams. A computer memory object can include a collection of data elements that can be operated by a computer program executable. A computing memory object can represent an image, a texture, a buffer frame, an arrangement of a scalar data type, an arrangement of a user-defined structure, buffer, subbuffer, or a variable, etc. A constant memory can be read-only memory by storing constant variables frequently used by a computer program executable.
[0037] In one embodiment, a local memory for a computing processor or computing unit can be used to allocate shared variables across the entire line in a line group or a line group. Local memory can be implemented as dedicated local storage, such as local shared memory 219 for Processor_1 and local shared memory 211 for Processor_L. In another embodiment, a local memory for a computing processor can be implemented as a read / write cache for a computing device memory for one or more computing processors on a computing device, such as data cache 215 for processes. -computer servers 205, 203 on computing device 201. Dedicated local storage may not be shared by lines across different line groups. If the local memory of a computing processor, such as a Processor_1 205, is implemented as a read / write cache, for example, data cache 215, a variable declared to be in local memory can be allocated from the memory of the device. computation 217 and cached in the read / write cache, for example, data cache 215 that implements local memory. Lines within a line group can share local variables allocated in computing device memory 217 when, for example, neither the read / write cache nor dedicated local storage is available for the corresponding computing device. In one embodiment, each line is associated with a private memory to store private line variables that are used by functions called on the line. For example, private memory 1 211 may not be seen by lines other than line 1 213.
[0038] In addition, in one embodiment, computing device memory 217 includes a buffer 223 that is used to store data used by processor_1 205 - processor_L 203. Temporary storage 223 can be a one-dimensional buffer, two-dimensional image, three-dimensional image, or another type of buffer as known in the art. In one embodiment, computing device 201 stores data to be operated by processors (for example, processor_1 205 - processor_L 203) in buffer 223. For example, and in one embodiment, the buffer can store an array of data, a two-dimensional image. , a three-dimensional image, etc., and / or other data as known in the art. In one embodiment, data between buffer 223 and other memory in system 201 (private memory 211, 207, local shared memory 219, 221, data cache 215, etc.) can be transferred using any method known in the art for data transfer between memory (direct PCIe transfer, direct asynchronous memory access, etc.)
[0039] Figure 3 is a block diagram illustrating a modality of a plurality of physical computing devices configured as a logical computing device using a computing device identifier. In one embodiment, an application 303 and a platform layer 305 may be running on a main CPU 301. Application 303 may be one of the applications 103 in Figure 1. Hosting systems 101 may include a main CPU 301. Each one of the physical computing devices Physical_Compute_Device-1 305 through Physi- cal_Compute_Device-N 311 can be one of the 117 CPUs or GPUs 115 of figure 1. In one embodiment, the computing platform layer 141 can generate a computing device identifier 307 in response to API requests from the 303 application to configure parallel data processing capabilities according to a list of capacity requirements included in API requests. The computing device identifier 307 can refer to a selection of actual physical computing devices Physi- cal_Compute_Device-1 305 via Physical_Compute_Device-N 311 according to the configuration by the computing platform layer 141. In one embodiment, a device logic computing 309 can represent the group of selected real physical computing devices separate from the main CPU 301.
[0040] Figure 4 is a block diagram illustrating a modality of a buffer subdivided into multiple subbuffers. In one embodiment, buffer 400 is buffer 223 as shown in figure 2 above. In figure 4, buffer 408 is allocated memory that is used to store data that is used by the 402A-D computing units. Temporary storage 408 can be a one-dimensional arrangement, two-dimensional image, three-dimensional image, or other type of buffer known in the art. Temporary storage 408 is further subdivided into multiple 410A-D subbuffers. In one embodiment, each subbuffer 410A-D is referred to by a pointer 412A-D inside the buffer. For example and in one embodiment, subbuffer 410A is referred to by pointer 412A, subbuffer 410B is referred to by pointer 412B, subbuffer 410C is referred to by pointer 412C, and subbuffer 410D is referred to by pointer 412D. In one embodiment, these 412A-D hands indicate the start of each buffer. In this modality, to access the data in subbuffers 410A-D, computing units 402A-D would provide the corresponding pointer 412A-D and an offset to the desired region of subbuffer 410-D.
[0041] In one embodiment, each 402A-D computing unit is associated with one of the 410A-D subbuffers in buffer 408. In one embodiment, each of these 402A-D computing units uses the data for the computing task assigned to each computing unit. Each of the computing units can read and / or write data to the corresponding subbuffer 410A-D. For example and in one embodiment, computing unit 402A uses subbuffer 410A, computing unit 402B uses subbuffer 410B, computing unit 402C uses subbuffer 410C, and computing unit 402D uses subbuffer 410D. In this mode, to access the data in the subbuffers 410A-D, the 402A-D computing units would provide the corresponding pointer 412A-D and an offset to the desired region of the subbuffer 410-D. The offsets can be an index arrangement, two-dimensional reference, three-dimensional reference, etc. The structure of buffer 408 is further described in figures 5 to 7 below.
[0042] In one embodiment, each subbuffer is created by a function call and providing a buffer pointer and subbuffer size value. The creation of a subbuffer is additionally described in figure 10 below.
[0043] In one embodiment, a 402A-D computing unit transfers data from the corresponding 402A-D subbuffer to the private memory 404A-D of that 402A-D computing unit. In one embodiment, private memory 404A-D is memory that is local to the computing unit (for example, private memory 1-M 211, private memory 1-N 207, local shared memory 219 and 221 and / or data cache 215, as illustrated in figure 2). In one embodiment, the 402A-D computing unit transfers data over a bus by coupling the 402A-D computing units and the memory that contains the 408 buffer. For example, and in one embodiment, the docking bus. it is a Peripheral Component Interface (PCI, PCI-Express (PCIe), etc.) type bus and the transfer mechanism is a direct PCI memory transfer.
[0044] Figure 5 is a block diagram illustrating a modality of multiple subbuffers 502A-D in a one-dimensional buffer 500. In figure 5, while buffer 500 is illustrated with four subbuffers 502A-D, in alternative modalities, buffer 500 can have more or less subbuffers and / or subbuffers of varying size. In one embodiment, buffer 500 is a one-dimensional arrangement of a data type (ints, floats, strings, user-defined structs, user-defined objects, etc.). For reference data from one of the 502A-D subbuffers, an offset from a 504A-D start pointer from the 502A-D subbuffer can be used. For example and in one embodiment, buffer 500 is two arrangements of one billion floats each. In this example, the computing units will add the contents of the arrangement together, and each 502A-D subbuffer contains parts of the two arrangements (for example, each 502A-D subbuffer has half a billion floats for each of the two arrangements, billion floats in total). The computing units in this example transfer the data from the subbuffer corresponding to the computing unit, add the floats and store the resulting value within the subbuffer.
[0045] Figure 6 is a block diagram illustrating a modality of a two-dimensional image buffer 600 subdivided into multiple 602A-D subbuffers. In figure 6, while buffer 600 is illustrated with four subbuffers 602A-D, in alternative modalities, buffer 600 may have more or less subbuffers and / or subbuffers of varying size. In figure 6, two-dimensional image buffer 600 is a two-dimensional buffer that contains data that is referred to by an x-offset and a y-offset. This buffer can store data of different types (ints, floats, strings, user-defined structs, user-defined objects, etc.). For example and in one mode, buffer 600 can store a two-dimensional image of pixels in the x and y direction. For example, in one embodiment, buffer 600 stores a two-dimensional image in order to compute a color histogram of the stored image. In this example, the image is subdivided into four 602A-D subbuffers and each 602A-D subbuffer is used by a computing unit to maintain the portion of the image that the computing unit is processing. In addition, each 602A-D computing unit copies a relevant portion of the corresponding sub-buffer image into the computing unit's private memory. The computing unit computes the histogram information using that image data and returns the histogram information.
[0046] Figure 7 is a block diagram illustrating a modality of a three-dimensional image buffer 700 subdivided into multiple subbuffers 702A-D. In figure 7, while buffer 700 is illustrated with four subbuffers 702A-D, in alternative embodiments, buffer 700 can have more or less subbuffers and / or subbuffers of varying size. In figure 7, three-dimensional image buffer 700 is a three-dimensional buffer that contains data that is referred to as an x, y and z offset or another suitable system for referencing a location in a three-dimensional space. Like buffers 500 and 600, this buffer 700 can store data of various types (ints, floats, strings, user-defined structs, user-defined objects, etc.). For example and in one embodiment, buffer 700 can store a three-dimensional pixel image in the x, y and z direction.
[0047] Figure 8 is a flow diagram illustrating an embodiment of a process 800 for configuring a plurality of physical computing devices with a computing device identifier by combining a resource request received from an application. Exemplary process 800 can be performed by a processing logic that can comprise hardware (circuit, dedicated logic, etc.), software (as it runs on a dedicated machine), or a combination of both. For example, process 800 can be performed according to system 100 of figure 1 on a data processing system hosted by hosting systems 101. The data processing system can include a main processor hosting a platform layer, such as a computing platform layer 141 of figure 1, and multiple physical computing devices connected to the main processor, such as CPUs 117 and GPUs 115 of figure 1.
[0048] In a block 801, in one embodiment, the processing logic of process 800 may establish a data structure (or a computing device data structure) representing multiple physical computing devices associated with one or more corresponding capacities. Each physical computing device can be connected to the processing system by performing process processing logic 800. The computing capabilities or capabilities of a physical computing device, such as a CPU or GPU, can include whether the computing device physical supports a processing feature, a memory access mechanism, a named extension or associated limitations. A processing feature can be related to hardware support for dedicated texturing, double precision floating point arithmetic or synchronization support (for example, mutual exclusion).
[0049] Capabilities of a computing device may include a type indicating processing characteristics or limitations associated with a computing device. An application can specify a type of computing device required or query the type of a specific computing device using APIs. Examples of different types of computing devices are shown in the following table: Table 1


[0050] Additionally, the capabilities of a computing device may include, for example, configuration values as shown in the following table: Table 2






1 The platform profile returns the profile that is implemented by the OpenCL framework. If the platform profile returned is FULL_PROFILE, the OpenCL structure will support devices that are FULL_PROFILE and can also support devices that are EMBEDDED_PROFILE. The compiler must be available for all devices, that is,

[0051] A memory access mechanism for a physical processing device can be related to a variable cache type (for example, unsupported, read-only, write-only), a type of computing memory object cache, size support cache, dedicated local memory support or associated limitations. Memory access limitations can include a maximum number of computing memory objects that can be read or written simultaneously by a computer program executable, a maximum number of computing memory objects that can be allocated, or a maximum size together. with a dimension of a multidimensional computing memory object, for example, a maximum width of a computing memory object for a 2D (two-dimensional) image. A data processing system system application can update the data structure in response to the connection of a new physical computing device to a data processing system. In one embodiment, the capabilities of a physical computing device can be predetermined. In another embodiment, a data processing system system application can discover a physical processing device recently connected during run time. The system application can restore the capabilities of the recently discovered physical computing device to update the data structure representing the connected physical computing devices and their corresponding capabilities.
[0052] According to a modality, the processing logic of process 800 may receive a requirement for computing capacity of an application in an 803 block. The application can send the request for computing capacity to a system application by calling APIs. The system application can correspond to a platform layer of a software stack on a hosting system for the application. In one embodiment, a computing capacity requirement can identify a list of capacities required for request processing capabilities to perform a task for the application. In one embodiment, the application may require the processing resources required to perform the task on multiple lines at the same time. In response, the processing logic of process 800 can select a group of physical computing devices from physical computing devices connected in block 805. Selection can be determined based on a combination of computing capacity requirements versus computing capabilities. computations stored in the capacity data structure. In one embodiment, the processing logic of process 800 can perform the combination according to a warning provided by the capacity requirement.
[0053] The processing logic of process 800 can determine a combination score according to the number of computing capacities combined between a physical computing device and the computing capacity requirement. In one embodiment, the processing logic of process 800 can select multiple physical computing devices with the highest combination scores. In another embodiment, the processing logic of process 800 can select a physical computing device if each capacity in the capacity requirement is combined. The processing logic of process 800 can determine multiple groups of physical computing devices matching in block 805. In one embodiment, each group of physical computing devices matching is selected according to the load-balancing capacity of each device. In block 807, in one embodiment, the processing logic of process 800 can generate a computing device identifier for each group of physical computing devices selected in block 805. The processing logic of process 800 can return one or more of the computing device identifiers generated back to the application through API calls. An application can choose which processing resource to use to perform a task according to the computing device identifiers. In one embodiment, the processing logic of process 800 can generate at most one computing device identifier in block 807 for each received capacity requirement.
[0054] In block 809, in one embodiment, the processing logic of process 800 can allocate resources to initialize a logical computing device for a group of physical computing devices selected in block 805 according to a corresponding computing device identifier . A logical computing device can be a group of computing devices including one or more physical computing devices. The processing logic of process 800 can initialize a logical computing device in response to API requests from an application that received one or more computing device identifiers according to the selection in block 805.
[0055] The processing logic of process 800 can create a context object in the logical computing device for an application in block 811. Commands that operate on the computing memory object, computer program objects and / or computer program executables computation for a context object can be performed in order (for example, synchronously) or out of order (for example, asynchronously) according to the parameters specified in API requests when creating the context object. Profiling commands that operate on computing memory objects, computing programs, or computing kernels can be allowed for a context object using API requests. In one embodiment, a context object is associated with an application line on a hosting system running the application. Multiple lines performing processing tasks on one logical computing device or across different logical computing devices at the same time can be based on separate context objects.
[0056] In one embodiment, the processing logic of process 800 can be based on multiple APIs including clCreateContext, cIRetainContext and cIReleaseContext. The cICreateContext API creates a computing context. A computing context can correspond to a computing context object. The cIRetainContext API increases the number of instances using a particular computing context identified by a context as an input argument for cIRetainContext. The cICreateContext API does an implicit hold. This is useful for third party libraries, which typically receive context passed to them by the application. However, it is possible that the application can delete the context without informing the library. Allowing multiple instances to connect to a context and free themselves from a context solves the problem of a computing context being used by a library that is no longer valid. If an input argument for cIRetainContext does not match a valid computing context object, cIRetainContext returns CU_INVALID_CONTEXT. The cIReleaseContext API releases an instance of a valid computing context. If an input argument for cIReleaseContext does not match a valid computing context object, cIReleaseContext returns CUJNVALID-CONTEXT.
[0057] Figure 9 is a flow diagram illustrating a modality of an exemplary process 900 for executing a computing executable on a logical computing device. In one embodiment, process 900 can be performed by a runtime layer in a data processing system such as the computing runtime layer 109 of figure 1. In block 901, the processing logic of process 900 it can allocate one or more computing memory objects (for example, streams) in a logical computing device to run a computing executable. A computing memory object can include one or more data elements to represent, for example, an image memory object or an array memory object. An array memory object can be a one-dimensional collection of the data element. An image memory object can be a collection for storing two-dimensional, three-dimensional or other multidimensional data, such as texture, frame buffer or an image. A processing task can be performed by a computer program executable operating on computing memory objects or streams using computing memory APIs including reading from incoming computing memory objects and writing to outgoing computing memory objects. In one embodiment, a computing memory object can be connected to a data object, such as a buffer object, texture object, or a rendering buffer object, to update the data object using memory memory APIs. computing. A data object can be associated with APIs that enable graphics processing operations, such as rendering text, on a data object. In one embodiment, a memory object is a buffer with multiple subbuffers as described in figure 2 above.
[0058] When allocating a computing memory object, the processing logic of process 900 can determine where the allocation should reside according to the specifications of an API. For example, a computing memory object can be allocated outside a main memory, such as a main memory for hosting systems 101 of Figure 1 and / or a computing device memory, such as a global memory or a constant memory. 217 of figure 2. A computing memory object allocated in main memory may need to be cached in a computing device memory. The processing logic of process 900 can load data asynchronously into computing memory objects allocated using non-blocking API interfaces, for example, based on generated event objects that include synchronization data indicating whether the data has been loaded into a data object. computing memory. In one embodiment, the processing logic of process 900 can program memory access operations when reading from or writing to computing memory objects. The processing logic of process 900 can map a stream memory allocated to form an application's logical address. In one embodiment, the processing logic of process 900 can perform operations on block 901 based on API requests from an application running on a main processor, such as applications 103 in Figure 1.
[0059] In block 903, according to a modality, the processing logic of process 900 can create a computer program object for the logical computer device (for example, a computer device group). A computer program object can include a group of computer kernels representing exported functions or entry points for a parallel data program. A computing kernel can include a pointer to a computer program executable that can be run on a computer unit to perform a parallel data task (for example, a function). Each computing kernel can be associated with a group of function arguments including computing memory objects or flows allocated to function inputs or outputs, such as flows allocated in block 901.
[0060] The processing logic of process 900 can load a computer program binary and / or a computer program source into the computer program object in block 909. A computer program binary can include bits that describe a computer program executable that will run on a computing device. A computer program binary can be a computer program executable and / or an intermediate representation of a computer program source to be converted into a computer program executable. In one embodiment, a computer program executable may include description data associated with, for example, the type of target physical computing devices (for example, a GPU or CPU), versions and / or build options or flags, such as line group sizes and / or line group dimensions. A computer program source can be the source code from which a computer program executable is compiled. The processing logic of process 900 can load multiple computer program executables corresponding to a computer program source in block 909. In one embodiment, the processing logic of process 900 can load a computer program executable from an application or through a computer library such as a computer application library 105 of figure 1. A computer program executable can be loaded with the corresponding computer program source. Process processing logic 900 can configure function arguments for a computer program object in block 905. In one embodiment, process processing logic 900 can perform operations on blocks 903, 905 and 909 according to requests of an application's API.
[0061] In block 911, the processing logic of process 900 can update a execution list to execute the computing kernel object with a logical computing device. The processing logic of process 900 can run a computer kernel in response to API calls with appropriate arguments for a computing runtime, for example, computing runtime 109 in figure 1, from an application or a library of computing application, such as applications 103 or computing application library 105 in figure 1. Running a computing kernel can include running a computer program executable associated with the computing kernel. In one embodiment, the processing logic of process 900 can generate an instance of computing kernel execution to execute a computing kernel. API calls to a computing runtime, such as a computing runtime 109 of figure 1, to run a computing kernel can be asynchronous in nature. A run instance can be identified by a compute event object that can be returned by a compute run time, such as a compute run time 109 in figure 1. A compute kernel run instance can be added to a playlist to run a compute kernel instance.
[0062] In one embodiment, API calls for a computing runtime to run a computing kernel can include the number of lines that are executed simultaneously in parallel on a computing processor as a group of lines. An API call can include the number of computing processors to be used. A compute kernel execution instance can include a priority value indicating a desired priority to run the corresponding compute program executable. A computing kernel execution instance may also include an event object identifying an instance of previous execution and / or expected total number of lines and number of line groups to perform the execution. The number of line groups and the total number of lines can be specified in API calls. In one embodiment, an event object can indicate an order of execution relationship between the execution instance that includes the event object and another execution instance identified by the event object. An execution instance including an event object may be required to be executed after another execution instance identified by the end event object execution. An event object can be referred to as a queue_after_event_object. Events and event dependencies are further described in figures 11 and 12 below. In one instance, a run list can include multiple compute kernel run instances to run corresponding compute program executables. One or more compute kernel execution instances for a compilation program executable can be programmed to run in a playlist. In one embodiment, the 900 processing logic can update the execution list in response to API requests for an application. The playlist can be hosted by the hosting data systems where the application is running.
[0063] In block 913, the processing logic of process 900 can select a computing kernel execution instance from the execution list for execution. In one embodiment, the processing logic of process 900 can select more than one computing kernel execution instance to be executed at the same time according to the corresponding logical computing devices. The processing logic of process 900 can determine whether a computing kernel execution instance is selected from the execution list based on its dependency and priority relationships associated with other execution instances in the execution list. A compute kernel execution instance can be executed by executing its corresponding compute kernel object according to an executable loaded to the compute kernel object.
[0064] In block 917, in one embodiment, the processing logic of process 900 can select one of the plurality of executables loaded to the computing kernel object corresponding to the computing kernel instance selected to run on a physical computing device associated with the logical computing device for the computing kernel object. The processing logic of process 900 can select more than one executable to run on more than one physical computing device in parallel for an computing kernel execution instance. The selection can be based on the current execution states of the physical computing devices corresponding to the logical computing device associated with the selected computing kernel execution instance. A running state of a physical computing device can include the number of lines being run, the level of local memory usage and the level of processor usage (for example, maximum number of operations per unit of time), etc. . In one embodiment, the selection can be based on predetermined levels of use. In another embodiment, the selection can be based on the number of lines and number of line groups associated with the computing kernel execution instance. The processing logic of process 900 can retrieve a running state from a physical computing device. In one embodiment, the processing logic of process 900 can perform operations to select an instance of computing kernel execution from the execution list to execute in blocks 913 917 asynchronously for applications that are running on hosting systems.
[0065] In block 919, the processing logic of process 900 can check the execution status of a computation kernel execution instance programmed for execution in the execution list. Each execution instance can be identified by a single compute event object. An event object can be returned to an application or a computing application library, such as an application 103 or computing application library 105 of figure 9, which calls APIs to run the execution instance, when the execution instance The corresponding computing kernel has been listed according to a computation runtime, such as runtime 109 in figure 1. In one embodiment, the processing logic of process 900 can perform the execution state by checking in response to requests of an application's API. The processing logic of process 900 can determine the end of execution of a compute kernel execution instance by asking for a state of the compute event object by identifying the compute kernel execution instance. The processing logic of process 900 can wait until the execution of a compute kernel execution instance is complete to return an application's API calls. The processing logic of process 900 can control instances of processing execution by reading and / or writing from various streams based on the computation event objects.
[0066] In block 921, according to one modality, the processing logic of process 900 can retrieve execution results from a computation kernel execution instance. Subsequently, the processing logic of process 900 can clear processing resources allocated to run the compute kernel execution instance. In one embodiment, the processing logic of the 900 process can copy a stream memory while maintaining results from running a computing kernel executable in local memory. The processing logic of process 900 can delete variable streams or image flows allocated in block 901. The processing logic of process 900 can delete a kernel event object to detect when a computation kernel run is complete. If each computing kernel execution instance associated with a specific computing kernel object has been completely executed, the processing logic of process 900 can delete the specific computing kernel object. In one embodiment, the processing logic of process 900 can perform operations on block 921 based on API requests initiated by an application.
[0067] Figure 10 is a flow diagram illustrating a modality of a 1000 runtime process for creating and using subbuffers with multiple computing units. Exemplary 1000 processes can be performed by processing logic that can comprise hardware (circuit, dedicated logic, etc.), software (as it runs on a dedicated machine) or a combination of both. For example, process 1000 can be performed according to system 100 of figure 1 on a data processing system hosted by hosting systems 101. The data processing system can include a main processor hosting a platform layer, such as such as the computing platform layer 141 of figure 1, and multiple physical computing devices connected to the main processor, such as CPUs 117 and GPUs 115 of figure 1.
[0068] In figure 10, process 1000 creates a subbuffer for a computing unit, where the subbuffer is associated with a buffer. In one embodiment, process 1000 creates a subbuffer of a buffer currently allocated. For example and in one embodiment, process 1000 creates a subbuffer from an allocated buffer using a function call: cl_memclCreatesubbuffer (cl_mem buffer, cl_mem_flags flags, cl_buffer_create_type buffer_create_type, const void * buffer_create_info, cljnt * errcode_ret) (flags) is a bit field that is used to specify allocation and usage information about the image memory object being created and is described in table 3, without size,) is the size in bytes of the subbuffer memory object to be allocated, buffer_create_type and buffer_create_info describe the type of buffer object to be created. The list of supported values for buffer_create_type and corresponding descriptor that buffer_create_info points to is described in table 4. Table 3. Signs for creating subbuffers memory.

Table 4. CL TEMPORARY STORAGE CREATE TYPE values.


[0069] In block 1004, process 1000 determines whether the computing unit for the subbuffer is the same computing unit as the main buffer. For example, in one embodiment, process 1000 determines that the subbuffer is created for a CPU. If the computing unit is different, process 1000 copies the data to the private memory of the computing unit associated with the subbuffer. For example, and in one embodiment, if the computing unit is a GPU and the computing unit associated with the buffer is a CPU, process 1000 would copy the data associated with the subbuffer into the GPU's memory. Referring again to figure 4, process 1000 would copy data from one of the subbuffers (for example, the subbuffer 410A) into the memory of the GPU (for example, private memory 404A of computing unit 402A). If the computing units are the same for the subbuffer and the buffer, process 1000 uses pointers to access data in the subbuffer in block 1006. For example, and in one embodiment, process 1000 would use pointer 412A to access data in subbuffer 410A as described in figure 4 above. Because process 1000 is using pointers to access data and does not need to update data that is changed, process 1000 ends at 1006.
[0070] On the other hand, if process 1000 copied the data into the private memory of the computing unit associated with the subbuffer, process 1000 tracks updates to the data in the private memory of that computing unit. For example, and in a modality in block 1010. Based on the tracked updates, process 1000 sends the updates to the main buffer in block 1012. While in a modality, process 1000 sends the updates at once, in an alternate mode, process 1000 sends updates in a different way (for example, sends updates periodically, sends updates automatically, etc.).
[0071] In addition to creating, using and / or managing subbuffers for computing units, system 100 can use events to synchronize operations in a context as described above with reference to figures 8 and 9. In one embodiment, an event object summarizes the status of an operation such as a command. In this mode, these objects can be used to synchronize operations in a context. Additionally, system 100 can use event waiting lists to control when a particular command starts executing. An event waiting list is a list of event objects. Figure 11 is a flow diagram illustrating a modality of a 1100 process to perform returns associated with events that have internal and external dependencies. In one embodiment, a return is used to report events (for example, errors, etc.) that occur within a context. As described above with reference to figure 8, a context is created with one or more computing units and is used to manage objects such as command lists, memory, programs, kernel objects and to run kernels on one or more computing units. specified in the context.
[0072] Exemplary process 1100 can be performed by processing logic that can comprise hardware (circuit, dedicated logic, etc.), software (as it is executed on a dedicated machine) or a combination of both. For example, process 1100 can be performed according to system 100 of figure 1 on the data processing system hosted by hosting systems 101. The data processing system can include a main processor hosting a platform layer, such as a computing platform layer 141 of figure 1, and multiple physical computing devices connected to the main processor, such as CPUs 117 and GPUs 115 of figure 1.
[0073] Process 1100 records an event to perform a return with a context, where the event has external dependencies in block 1102. In one mode, an event can have internal, external dependencies and / or not have dependencies. An event with an internal dependency means that before the return associated with the event can be executed, the internal dependency must be resolved. In one embodiment, internal dependency is a recognized system event, such as a kernel execution command or management commands (for example, reading, writing, mapping, copy commands on memory objects). An external dependency is a user-defined event and that external dependency must be resolved before the return is performed. For example, and in one embodiment, a user-defined event may allow applications for queued commands that wait for the user's event to be terminated before the queued command is executed by the corresponding computing unit. In another embodiment, a user event object can be used to report an application-specific error condition. In one embodiment, event dependencies can be stored in an event waiting list.
[0074] In block 1104, process 1100 receives notification that the registered event has occurred. In one embodiment, process 1100 receives notification of the event by calling a function that waits for events. In block 1106, process 1100 determines whether the recorded event has any unresolved internal events. For example and in one embodiment, process 1100 determines whether an event waiting list associated with the logged event has any internal dependencies. If there are any internal dependencies, process 1100 delays the execution of the return in block 1112. In one embodiment, process 1100 delays execution until the internal dependencies are resolved. For example, and in one mode, resolving a dependency may include waiting for a command associated with a dependent event to be complete.
[0075] If there is no internal dependency for the registered event, process 1100 determines whether the registered event has any external dependencies in block 1108. For example, and in one mode, process 1100 determines whether an event waiting list associated with a registered event has any external dependencies. If there are any external dependencies, process 1100 delays the execution of the return in block 1112. In one embodiment, process 1100 delays execution until the external dependencies are resolved. For example and in one mode, resolving a dependency may include waiting for a command associated with a dependent event to be complete.
[0076] Figure 12 is a block diagram illustrating a modality of a chain of events 1202A-D with external and internal dependencies. In figure 12, event 1202A has a dependency chain including three internal events 1202B-D and an external event, user event 1204. For example, and in one embodiment, event 1202A is dependent on event 1202B, which in turn is dependent on event 1202C, which in turn is dependent on event 1202D, which in turn is dependent on user event 1204. In this mode, event 1202D waits for user event 1204 to be resolved, the event 1202C waits for events 1202D and 1204 to be resolved, event 1202B waits for events 1202C-D and 1204 to be resolved, and event 1202B waits for events 1202B-D and 1204 to be resolved.
[0077] Figure 13 is an exemplary source code illustrating an example of a computer program source code for a computer program executable to be run on multiple physical computing devices. Example 1300 can represent an API function with arguments including variables 1301 and flows (or computing memory objects) 1303. Example 1300 can be based on a programming language for a parallel computing environment such as system 131 in figure 1. In a fashion, the parallel programming language can be specified according to the ANSI (American National Standards Institute) C standard with additional extensions and restrictions designed to implement one or more of the modalities described here. Extensions can include a function qualifier, such as qualifier 1305, to specify a computing kernel function to be performed on a computing device. A compute kernel function may not be called by the other compute kernel functions. In one embodiment, a computing kernel function can be called by a main function in the parallel program language. A main function can be a regular ANSI C function. A main function can be performed on a main processor separate from the computing device by executing a computing kernel function. In one embodiment, extensions can include a local qualifier to describe variables that need to be allocated to a local memory associated with a computing device to be shared across all lines in a group of lines. The local qualifier can be declared within a compute kernel function. The constraints of the parallel programming language can be enforced during compiler time or runtime to generate error conditions, such as exit error messages or exiting an execution, when restrictions are violated.
[0078] Figures 14A to 14C include an exemplary source code illustrating an example for configuring a local computing device to run one of multiple executables on multiple physical computing devices by calling APIs. Examples 1400A to 1400C can be executed by an application being run on a main system coupled to multiple physical computing devices, such as hosting systems 101 of figure 1. Examples 1400A to 1400C can specify a main function of a parallel programming language. Processing operations in examples 1400A to 1400C can be performed as API calls by a process such as process 800 in figure 8 and / or process 900 in figure 9. Processing operations to create a context object from a computing device , a computing device group or a logical computing device 1401, can be performed by process processing logic 800 in block 811 of figure 8. Processing operations to allocate input / output image memory objects (for example, computing memory objects) can be performed by the processing logic of process 900 in block 901 of figure 9.
[0079] Now returning to figure 14B, the processing operations to allocate and load memory objects of arrangement 1403b can be performed by the processing logic of process 900 in block 901 of figure 9. The processing operation to create an object of computer program 1405 can be performed by the processing logic of process 900 in block 903 of figure 9. Processing operation 1407 can load a computer program source, such as example 900 of figure 9, into the computer program object. computation created. Processing operation 1409 can explicitly establish a computer program executable from the loaded computer program source. In one embodiment, processing operation 1409 may load a computer program executable already established for the computer program object created. Subsequently, processing operation 1411 can create a compute kernel object pointing to the compute program executable established to schedule a run on a compute device.
[0080] Now returning to figure 14C, in a modality, processing operation 1413 can connect variables and computing memory objects as function arguments for the created computing kernel object. Processing operation 1413 can be performed by the processing logic of process 900 in block 905 of figure 9. Processing operation 1415 can execute the computation kernel object created. In one embodiment, processing operation 1415 can be performed by the processing logic of process 900 in block 911 of figure 9. Processing operation 1415 can cause a playlist to be updated with a computing kernel execution instance. corresponding to the computation kernel object created. Processing operation 1417 can wait synchronously for the execution of the computation kernel object created to finish. In one embodiment, processing operation 1419 can restore a result from the execution of a compute kernel object. Subsequently, 1191 processing operations can clear resources allocated to run the compute kernel object, such as an event object, the compute kernel object created and the allocated memories. In one embodiment, the processing operation 1417 can be performed asynchronously based on whether the kernel event object is set or not. Processing operation 1417 can be performed by process 900 in block 919 of figure 9.
[0081] Figure 15 shows an example of a computer system 1500 that can be used with an embodiment of the present invention. For example, system 1500 may be implemented as a part of the systems shown in figure 1. Note that while figure 15 illustrates various components of a computer system, it is not intended to represent any particular architecture or way of interconnecting the components, since such details are not relevant to the present invention. It will also be appreciated that network computers and other data processing systems (for example, handheld computers, personal digital assistants (PDAs), cell phones, entertainment systems, consumer electronic devices, etc.) that have fewer components or perhaps more components can also be used to implement one or more embodiments of the present invention.
[0082] As shown in figure 15, the computer system 1500, which is a form of a data processing system, includes a bus 1503 which is coupled to microprocessor (s) 1505, such as CPUs and / or GPUs, a ROM (Read Only Memory) 1507, volatile RAM 1509 and a non-volatile memory 1911. The microprocessor 1505 can restore the instructions from memories 1507, 1509, 1911 and execute the instructions using Cache 1521 to perform the operations described above. The 1503 bus interconnects these various components and also interconnects these 1505, 1507, 1509 and 1911 components to a 1913 display controller and display device and to peripheral devices such as input / output (I / O) devices that can be mouse, keyboards, modems, network interfaces, printers and other devices that are well known in the art. Typically, 915 input / output devices are coupled to the system via 1917 input / output controllers. Volatile RAM (Random Access Memory) 1509 is typically implemented as dynamic RAM (DRAM) that requires power continuously in order to regenerate or keep the data in memory. The display controller attached to the 1913 display device can optionally include one or more GPUs for process display data. Optionally, 1919 GPU memory can be provided to support GPUs included in the 1913 display device.
[0083] 1911 mass storage is typically a magnetic hard disk drive or optical drive or a DVD RAM or fast memory or other types of memory systems that hold data (for example, large amounts of data) even after the power is removed from the system. Typically, 1911 mass storage will also be a random access memory, although this is not required. Although figure 15 shows that mass storage 1911 is a local device coupled directly to the rest of the components in the data processing system, it will be appreciated that the present invention can use a non-volatile memory that is distant from the system, such as a device network storage device that is coupled to the data processing system via a network interface such as a modem or Ethernet interface or wireless network interface. The bus 1503 can include one or more buses connected to each other via various bridges, controllers and / or adapters as is well known in the art.
[0084] Portions of what has been described above can be implemented with logic circuit such as a dedicated logic circuit or with a microcontroller or other form of processing core that executes program code instructions. Thus, the processes taught by the above description can be carried out with program code such as machine executable instructions that cause the machine to execute those instructions to perform certain functions. In this context, a "machine" can be a machine that converts instructions in an intermediate (or "abstract") manner into specific processor instructions (for example, an abstract execution environment such as a "virtual machine" (for example, a Machine Virtual Java, an interpreter, a Common Language Runtime, a high-level language virtual machine, etc.), and / or electronic circuit arranged on a semiconductor chip (for example, "logic circuit" implemented with transistors) designed to execute instructions such as a general purpose processor and / or a specific use processor. The processes taught by the description above can also be performed by (as an alternative to a machine or in combination with a machine) electronic circuit designed to carry out the processes (or a portion thereof) without executing program code.
[0085] A manufactured product can be used to store program code, for example, including multiple cards. A manufactured product that stores program code can be incorporated as, but is not limited to, one or more memories (for example, one or more quick memories, random access memories (static, dynamic or other)), optical discs , CD-ROMs, DVD ROMs, EPROMs, EEPROMs, optical or magnetic cards or other types of machine-readable media suitable for storing electronic instructions. Program code can also be loaded from a remote computer (for example, a server) to a requesting computer (for example, a client) via data signals embedded in a propagation medium (for example, using a communication link (for example, a network connection)).
[0086] The preceding detailed descriptions are presented in terms of algorithms and symbolic representations of operations on bits of data within a computer memory. These algorithmic descriptions and representations are tools used by those versed in data processing techniques to communicate the content of their work more effectively to others versed in the technique. An algorithm is here, and generally, understood to be a self-faithful sequence of operations that lead to a desired result. Operations are those that require physical manipulation of physical quantities. Usually, although not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared or otherwise manipulated. It has proven to be convenient at times, mainly for reasons of common use, to refer to these signs as bits, values, elements, symbols, characters, terms, numbers or the like.
[0087] It should be kept in mind, however, that all of these terms and the like must be associated with the appropriate physical quantities and are merely convenient labels applied to those quantities. Unless specifically stated otherwise, as evident from the above description, it is appreciated that during the description, descriptions using terms such as "processing" or "computing" or "calculating" or "determining" or "displaying" or " copying "or" tracking "or" sending "or similar, refer to the actions and processes of a computer system, or similar electronic computing device, which manipulates and transforms data re-presented as physical (electronic) quantities within of the computer system's records and memories into other data similarly represented as physical quantities within the computer system's memories or records or other such information storage, transmission or display devices.
[0088] The present invention also relates to an apparatus for carrying out the operations described here. This device can be built especially for the required purpose, or it can comprise a general purpose computer selectively activated or reconfigured by a computer program stored on the computer. Such a computer program may be stored on a computer-readable storage medium, such as, but not limited to, any type of disk including diskettes, optical disks, CD-ROMs, and optical-magnetic disks, read-only memories (ROMs) ), RAMs, EPROMs, EE-PROMs, magnetic or optical cards, or any type of suitable means for storing electronic instructions, and each coupled to a computer system bus.
[0089] The processes and displays presented here are not inherently related to any particular computer or other device. Various general-purpose systems can be used with programs according to the teachings here, or they can prove to be convenient for building a more specialized device to perform the described operations. The structure required for a variety of these systems will be evident from the description below. In addition, the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages can be used to implement the teachings of the invention as described here.
[0090] The foregoing description merely describes some exemplary embodiments of the present invention. A person skilled in the art will readily recognize from such a description, the accompanying drawings and the embodiments that various modifications can be made without departing from the spirit and scope of the invention.

权利要求:
Claims (14)
[0001]
1. Computerized method of managing a plurality of subbuffers associated with a main buffer in a heterogeneous computing environment, the method characterized by the fact that it comprises the steps of: allocating the main buffer to a process, in which the process uses a plurality of heterogeneous computing units, plurality of heterogeneous computing units includes a central processing unit and a graphics processing unit, and the plurality of heterogeneous computing units resides in a single device; for each subbuffer in the plurality of subbuffers, create that subbuffer for one of a plurality of heterogeneous computing units from the main buffer, with a different subbuffer for each of the plurality of heterogeneous computing units, and store data from subbuffer in that subbuffer; for each subbuffer buffer that corresponds to one of the plurality of heterogeneous computing units not associated with the main buffer, manage updates to the subbuffer data in a private memory of a corresponding computing unit.
[0002]
2. Computerized method, according to claim 1, characterized by the fact that managing updates to the subbuffer data in private memory comprises: copying the data in that subbuffer to private memory to the computing unit corresponding to that subbuffer; track updates to data on that subbuffer; and send updates to the main buffer.
[0003]
3. Computerized method, according to claim 1, characterized by the fact that it additionally comprises: if the computing unit associated with that subbuffer is the same computing unit associated with the main buffer, create a pointer to that subbuffer in the main buffer to access the subbuffer data using the pointer.
[0004]
4. Computerized method, according to claim 3, characterized by the fact that the pointer is an offset within the main buffer.
[0005]
5. Computerized method, according to claim 1, characterized by the fact that the main buffer is selected from a group consisting of a one-dimensional buffer, a two-dimensional image, and a three-dimensional image.
[0006]
6. Computerized method, according to claim 1, characterized by the fact that the main buffer is formed from the system memory.
[0007]
7. Non-transitory machine-readable medium characterized by the fact that it is to make one or more processing units perform a method of managing a plurality of subbuffers associated with a main buffer in a heterogeneous computing environment, the method comprising the steps of: allocating the main buffer to a process, where the process uses a plurality of heterogeneous computing units, the plurality of heterogeneous computing units includes a central processing unit and a graphics processing unit, and the plurality of heterogeneous computing units resides on a single device; for each subbuffer in the plurality of subbuffers, create that subbuffer for one of a plurality of heterogeneous computing units from the main buffer, with a different subbuffer for each of the plurality of heterogeneous computing units, and store data from subbuffer in that subbuffer; for each subbuffer buffer that corresponds to one of the plurality of heterogeneous computing units not associated with the main buffer, manage updates to the subbuffer data in a private memory of a corresponding computing unit.
[0008]
8. Non-transitory machine-readable medium, according to claim 7, characterized by the fact that managing updates to the subbuffer data in private memory comprises: copying the data on that subbuffer from private memory to the computing unit corresponding to that subbuffer; track updates to data on that subbuffer; and send updates to the main buffer.
[0009]
9. Non-transitory machine-readable medium, according to claim 7, characterized by the fact that the method still comprises: if the computing unit associated with that subbuffer is the same computing unit associated with the main buffer, create a pointer to that subbuffer in the main buffer to access the subbuffer data using the pointer.
[0010]
10. Non-transitory machine-readable medium, according to claim 7, characterized by the fact that the main buffer is selected from a group consisting of a one-dimensional buffer, a two-dimensional image and a three-dimensional image.
[0011]
11. Apparatus for managing a plurality of subbuffers associated with a main buffer for managing a plurality of subbuffers associated with a main buffer in a heterogeneous computing environment, the apparatus characterized by the fact that it comprises: means for allocating the main buffer for a process, where the process uses a plurality of heterogeneous computing units, the plurality of heterogeneous computing units includes a central processing unit and a graphics processing unit, and the plurality of heterogeneous computing units resides on a single device ; for each subbuffer in the plurality of subbuffers, means for creating that subbuffer for one of a plurality of heterogeneous computing units from the main buffer, with a different subbuffer for each of the plurality of heterogeneous computing units, and means for storing subbuffer data in that subbuffer; for each subbuffer buffer that corresponds to one of the plurality of heterogeneous computing units not associated with the main buffer, means for managing updates to the subbuffer data in a private memory of a corresponding computing unit.
[0012]
12. Apparatus according to claim 11, characterized by the fact that the means for managing updates to the subbuffer data in private memory comprise: means for copying the data in that subbuffer from the private memory to the computing unit corresponding to that subbuffer; means for tracking updates to data in that sub-buffer; and means for sending updates to the main buffer.
[0013]
13. Apparatus according to claim 11, characterized by the fact that it still comprises: if the computing unit associated with that subbuffer is the same computing unit associated with the main buffer, means for creating a pointer to that subbufer in the main buffer to access the subbuffer data using the pointer.
[0014]
14. Apparatus according to claim 11, characterized by the fact that the plurality of heterogeneous computing units includes a central processing unit and a graphics processing unit.

类似技术:

---

公开号 | 公开日 | 专利标题

---

BR112012027950B1|2020-11-10|computerized method of managing a plurality of subbuffers, non-transitory machine readable means and apparatus for managing a plurality of subbuffers

---

US20200250005A1|2020-08-06|Data parallel computing on multiple processors

---

US9766938B2|2017-09-19|Application interface on multiple processors

---

US9471401B2|2016-10-18|Parallel runtime execution on multiple processors

---

US20180203737A1|2018-07-19|Data parallel computing on multiple processors

---

US8108633B2|2012-01-31|Shared stream memory on multiple processors

---

AU2011253819B2|2014-05-22|Parallel runtime execution on multiple processors

同族专利:

---

公开号 | 公开日

---

GB2480536A|2011-11-23|

---

CN102870096A|2013-01-09|

---

CA2795365C|2015-12-08|

---

US8723877B2|2014-05-13|

---

GB201108084D0|2011-06-29|

---

WO2011146197A1|2011-11-24|

---

DE112011101725T5|2013-04-25|

---

CN102870096B|2016-01-13|

---

JP5583844B2|2014-09-03|

---

KR20140117689A|2014-10-07|

---

JP2015007982A|2015-01-15|

---

MX2012012534A|2012-12-17|

---

JP5939524B2|2016-06-22|

---

CA2795365A1|2011-11-24|

---

JP2013528861A|2013-07-11|

---

US20150187322A1|2015-07-02|

---

US20140313214A1|2014-10-23|

---

EP2572276B1|2018-07-11|

---

EP2572276A1|2013-03-27|

---

AU2011256745B2|2014-01-23|

---

KR101477882B1|2014-12-30|

---

US20110285729A1|2011-11-24|

---

US9691346B2|2017-06-27|

---

KR101558831B1|2015-10-08|

---

GB2480536B|2013-01-02|

---

AU2011256745A1|2012-11-01|

---

KR20130004351A|2013-01-09|

---

US8957906B2|2015-02-17|

引用文献:

---

公开号 | 申请日 | 公开日 | 申请人 | 专利标题

---

---

CA2122139C|1991-10-24|2000-08-29|David Leroy Sprague|Data processing system|

---

US6115761A|1997-05-30|2000-09-05|Lsi Logic Corporation|First-In-First-Out memories having dual descriptors and credit passing for efficient access in a multi-processor system environment|

---

US6754887B1|1999-10-22|2004-06-22|International Business Machines Corporation|Methods for implementing virtual bases with fixed offsets in object oriented applications|

---

AU2603901A|1999-12-29|2001-07-09|Johns Hopkins University, The|System, method, and computer program product for high speed backplane messaging|

---

US6970206B1|2000-04-20|2005-11-29|Ati International Srl|Method for deinterlacing interlaced video by a graphics processor|

---

US7673304B2|2003-02-18|2010-03-02|Microsoft Corporation|Multithreaded kernel for graphics processing unit|

---

US7015913B1|2003-06-27|2006-03-21|Nvidia Corporation|Method and apparatus for multithreaded processing of data in a programmable graphics processor|

---

US20050071578A1|2003-09-25|2005-03-31|International Business Machines Corporation|System and method for manipulating data with a plurality of processors|

---

US7523157B2|2003-09-25|2009-04-21|International Business Machines Corporation|Managing a plurality of processors as devices|

---

US7634776B2|2004-05-13|2009-12-15|Ittiam Systems Ltd.|Multi-threaded processing design in architecture with multiple co-processors|

---

US7818507B2|2005-04-04|2010-10-19|Sony Computer Entertainment Inc.|Methods and apparatus for facilitating coherency management in distributed multi-processor system|

---

JP4082706B2|2005-04-12|2008-04-30|学校法人早稲田大学|Multiprocessor system and multigrain parallelizing compiler|

---

JP4836491B2|2005-05-20|2011-12-14|株式会社ソニー・コンピュータエンタテインメント|Information processing apparatus, system, method and processor|

---

JP2007066759A|2005-08-31|2007-03-15|Toyota Motor Corp|Fuel cell|

---

JP2007172456A|2005-12-26|2007-07-05|Toshiba Corp|Plotting device|

---

US7865898B2|2006-01-27|2011-01-04|Oracle America, Inc.|Repartitioning parallel SVM computations using dynamic timeout|

---

TWI355645B|2006-04-24|2012-01-01|Sony Corp|

---

US8468532B2|2006-06-21|2013-06-18|International Business Machines Corporation|Adjusting CPU time allocated to next thread based on gathered data in heterogeneous processor system having plurality of different instruction set architectures|

---

US8453132B2|2006-07-28|2013-05-28|Hewlett-Packard Development Company, L.P.|System and method for recompiling code based on locality domain and thread affinity in NUMA computer systems|

---

US20080109795A1|2006-11-02|2008-05-08|Nvidia Corporation|C/c++ language extensions for general-purpose graphics processing unit|

---

US7701459B1|2006-11-03|2010-04-20|Nvidia Corporation|Primitive oriented assembly for parallel vertex/geometry processing|

---

US8269782B2|2006-11-10|2012-09-18|Sony Computer Entertainment Inc.|Graphics processing apparatus|

---

AT529808T|2007-02-07|2011-11-15|Bosch Gmbh Robert|ADMINISTRATIVE MODULE, MANUFACTURER AND CONSUMER CALCULATOR, ARRANGEMENT THEREOF, AND METHOD FOR COMMUNICATING BETWEEN CALCULATORS OVER A MEMORY USED IN COMMUNITY|

---

US8108845B2|2007-02-14|2012-01-31|The Mathworks, Inc.|Parallel programming computing system to dynamically allocate program portions|

---

JP5224498B2|2007-02-28|2013-07-03|学校法人早稲田大学|MEMORY MANAGEMENT METHOD, INFORMATION PROCESSING DEVICE, PROGRAM CREATION METHOD, AND PROGRAM|

---

EP2135163B1|2007-04-11|2018-08-08|Apple Inc.|Data parallel computing on multiple processors|

---

US8108633B2|2007-04-11|2012-01-31|Apple Inc.|Shared stream memory on multiple processors|

---

US8806426B2|2008-06-04|2014-08-12|Microsoft Corporation|Configurable partitioning of parallel data for parallel processing|

---

JP5112386B2|2008-06-05|2013-01-09|株式会社東芝|Image processing apparatus and image processing method|

---

JP5161305B2|2008-06-05|2013-03-13|北陸電気工業株式会社|Display device provided with touch panel and piezoelectric actuator|

---

JP2009295032A|2008-06-06|2009-12-17|Canon Inc|Image processing device and its control method|

---

US8225325B2|2008-06-06|2012-07-17|Apple Inc.|Multi-dimensional thread grouping for multiple processors|

---

US8095735B2|2008-08-05|2012-01-10|Convey Computer|Memory interleave for heterogeneous computing|

---

US8531471B2|2008-11-13|2013-09-10|Intel Corporation|Shared virtual memory|

---

CN101551761A|2009-04-30|2009-10-07|浪潮电子信息产业股份有限公司|Method for sharing stream memory of heterogeneous multi-processor|

---

US9354944B2|2009-07-27|2016-05-31|Advanced Micro Devices, Inc.|Mapping processing logic having data-parallel threads across processors|

---

US8669990B2|2009-12-31|2014-03-11|Intel Corporation|Sharing resources between a CPU and GPU|

---

US8723877B2|2010-05-20|2014-05-13|Apple Inc.|Subbuffer objects|US4842803A|1988-03-25|1989-06-27|Union Camp Corporation|Method and apparatus for extruding blown thermoplastic film tubes|

---

JPH0760833A|1993-08-23|1995-03-07|Mitsubishi Chem Corp|Molding method for inflation resin film|

---

US8669990B2|2009-12-31|2014-03-11|Intel Corporation|Sharing resources between a CPU and GPU|

---

US8723877B2|2010-05-20|2014-05-13|Apple Inc.|Subbuffer objects|

---

US9552299B2|2010-06-11|2017-01-24|California Institute Of Technology|Systems and methods for rapid processing and storage of data|

---

US8752018B2|2011-06-21|2014-06-10|Nvidia Corporation|Emitting coherent output from multiple threads for printf|

---

US20130141443A1|2011-12-01|2013-06-06|Michael L. Schmit|Software libraries for heterogeneous parallel processing platforms|

---

WO2013103571A1|2012-01-06|2013-07-11|Intel Corporation|Reducing the number of read/write operations performed by a cpu to duplicate source data to enable parallel processing on the source data|

---

US9996394B2|2012-03-01|2018-06-12|Microsoft Technology Licensing, Llc|Scheduling accelerator tasks on accelerators using graphs|

---

US9864638B2|2012-06-22|2018-01-09|Intel Corporation|Techniques for accessing a graphical processing unit memory by an application|

---

US20140101761A1|2012-10-09|2014-04-10|James Harlacher|Systems and methods for capturing, replaying, or analyzing time-series data|

---

US9424079B2|2013-06-27|2016-08-23|Microsoft Technology Licensing, Llc|Iteration support in a heterogeneous dataflow engine|

---

US9286328B2|2013-07-19|2016-03-15|International Business Machines Corporation|Producing an image copy of a database object based on information within database buffer pools|

---

US9703696B1|2013-09-11|2017-07-11|Altera Corporation|Guided memory buffer allocation|

---

US10564949B2|2013-09-20|2020-02-18|Reservoir Labs, Inc.|System and method for generation of event driven, tuple-space based programs|

---

US20150091912A1|2013-09-27|2015-04-02|Nvidia Corporation|Independent memory heaps for scalable link interface technology|

---

US9411652B2|2014-08-22|2016-08-09|Advanced Micro Devices, Inc.|Runtime for automatically load-balancing and synchronizing heterogeneous computer systems with scoped synchronization|

---

CN104714850B|2015-03-02|2016-03-30|心医国际数字医疗系统有限公司|A kind of isomery based on OPENCL calculates equalization methods jointly|

---

US9916634B2|2015-06-12|2018-03-13|Intel Corporation|Facilitating efficient graphics command generation and execution for improved graphics performance at computing devices|

---

GB2540940B|2015-07-31|2018-01-03|Advanced Risc Mach Ltd|An apparatus and method for transferring a plurality of data structures between memory and one or more vectors of data elements stored in a register bank|

---

EP3188013A1|2015-12-29|2017-07-05|Dassault Systèmes|Management of a plurality of graphic cards|

---

WO2017126924A1|2016-01-19|2017-07-27|서울대학교 산학협력단|Data distribution technique in heterogeneous system|

---

US10152243B2|2016-09-15|2018-12-11|Qualcomm Incorporated|Managing data flow in heterogeneous computing|

---

WO2018119778A1|2016-12-28|2018-07-05|Intel Corporation|System and method for vector communication|

---

KR102066212B1|2017-01-19|2020-01-14|서울대학교산학협력단|Method for transferring data in parallel system, and parallel system for performing the same|

法律状态:
2018-12-26| B06F| Objections, documents and/or translations needed after an examination request according [chapter 6.6 patent gazette]|

---

2019-11-12| B06U| Preliminary requirement: requests with searches performed by other patent offices: procedure suspended [chapter 6.21 patent gazette]|

---

2020-05-26| B09A| Decision: intention to grant [chapter 9.1 patent gazette]|

---

2020-11-10| B16A| Patent or certificate of addition of invention granted [chapter 16.1 patent gazette]|Free format text: PRAZO DE VALIDADE: 20 (VINTE) ANOS CONTADOS A PARTIR DE 20/04/2011, OBSERVADAS AS CONDICOES LEGAIS. |

优先权:

---

申请号 | 申请日 | 专利标题

---

US34686610P| true| 2010-05-20|2010-05-20|

---

US61/346,866|2010-05-20|

---

US12/892,834|2010-09-28|

---

US12/892,834|US8723877B2|2010-05-20|2010-09-28|Subbuffer objects|

---

PCT/US2011/033282|WO2011146197A1|2010-05-20|2011-04-20|Subbuffer objects|

---