Graphics processor and graphics processing unit having hybrid floating-point format dot product accumulate instruction - Patents.com

Abstract

To provide graphic processors and graphic processing units having dot product accumulate instructions for a hybrid floating point format.SOLUTION: A graphic multiprocessor, a parallel processor, comprises an instruction unit for dispatching instructions, and a processing resource, comprising a functional unit and FPU, coupled to the instruction unit. The processing resource is configured to receive a dot product accumulate instruction from the instruction unit and process the dot product accumulate instruction using a Bfloat16 number (BF16) format.SELECTED DRAWING: Figure 2D

Description

[Technical field]

[0001] [CROSS REFERENCE TO RELATED APPLICATIONS] This application is related to and claims the benefit of and priority under 35 U.S.C. §119(e) to the following provisional applications: U.S. Provisional Application No. 62 / 819,337, entitled "GRAPHICS PROCESSING," filed March 15, 2019 by Abhishek Appu et al. (Attorney Docket No. AC0271-Z); U.S. Provisional Application No. 62 / 819,435, entitled "GRAPHICS DATA PROCESSING," filed March 15, 2019 by Lakshminarayanan Striramassarma et al. (Attorney Docket No. AC0285-Z); and U.S. Provisional Application No. 62 / 819,361, entitled "SYSTEM AND METHOD FOR REDUCING CACHE ACCESS LATENCY AND CACHE PARTITIONING METHOD," filed March 15, 2019 by Subramaniam Maiyuran et al. (Attorney Docket No. AC0286-Z), all of which are incorporated herein by reference in their entirety.

[0002] The present disclosure relates generally to data processing, and more particularly to data processing by general purpose graphics processing units. [Background technology]

[0003] Modern parallel graphics data processing includes systems and methods developed to perform specific operations on graphics data, such as linear interpolation, tessellation, rasterization, texture mapping, depth testing, etc. Historically, graphics processors processed graphics data using fixed function computation units, but more recently, some graphics processors have become programmable and such processors can support a wide range of operations that process vertex and fragment data.

[0004] Additionally, to increase performance, graphics processors typically implement processing techniques such as pipelining, which attempts to process as much graphics data as possible in parallel across various parts of the graphics pipeline. Parallel graphics processors with single instruction multiple threads (SIMT) architectures are designed to maximize the amount of parallelism in the graphics pipeline. In a SIMT architecture, a group of parallel threads attempts to synchronously execute program instructions together as often as possible to increase processing efficiency. A software and hardware overview of the SIMT architecture can be found in "CUDA Programming" (Shane Cook, 2013), Chapter 3, pages 37-51. [Brief description of the drawings]

[0005] In a manner that allows the above-mentioned features of the present embodiment to be understood in detail, a more particular description of the embodiments briefly summarized above may be made by reference to these embodiments, some of which are illustrated in the accompanying drawings, however, it should be noted that these drawings only illustrate representative embodiments and therefore should not be considered as limiting the scope thereof.

[0006] [Figure 1] FIG. 1 is a block diagram illustrating a computer system configured to implement one or more aspects of the embodiments described herein.

[0007] [Figure 2A] 1 shows the components of a parallel processor. [Figure 2B] 1 shows the components of a parallel processor. [Figure 2C] 1 shows the components of a parallel processor. [Figure 2D] 1 shows the components of a parallel processor.

[0008] [Figure 3A] FIG. 1 is a block diagram of a graphics multiprocessor and a multiprocessor-based GPU. [Figure 3B] FIG. 1 is a block diagram of a graphics multiprocessor and a multiprocessor-based GPU. [Figure 3C] FIG. 1 is a block diagram of a graphics multiprocessor and a multiprocessor-based GPU.

[0009] [Figure 4A] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors. [Figure 4B] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors. [Figure 4C] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors. [Figure 4D] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors. [Figure 4E] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors. [Figure 4F] 1 illustrates an exemplary architecture in which multiple GPUs are communicatively coupled to multiple multi-core processors.

[0010] [Diagram 5] 1 illustrates a graphics processing pipeline.

[0011] [Figure 6] Shows the machine learning software stack.

[0012] [Figure 7] 1 illustrates a general-purpose graphics processing unit.

[0013] [Figure 8] 1 illustrates a multi-GPU computer processing system.

[0014] [Figure 9A] 1 illustrates layers of an exemplary deep neural network. [Figure 9B] 1 illustrates layers of an exemplary deep neural network.

[0015] [Figure 10] 1 illustrates an exemplary recurrent neural network.

[0016] [Figure 11] 1 illustrates training and deploying deep neural networks.

[0017] [Figure 12] FIG. 2 is a block diagram illustrating distributed learning.

[0018] [Figure 13] 1 illustrates an exemplary inference system-on-chip (SoC) suitable for performing inference using a trained model.

[0019] [Figure 14] FIG. 1 is a block diagram of a processing system.

[0020] [Figure 15A] 1 illustrates a computer processing system and a graphics processor. [Figure 15B] 1 illustrates a computer processing system and a graphics processor. [Figure 15C] 1 illustrates a computer processing system and a graphics processor.

[0021] [Figure 16A] 1 shows a block diagram of another graphics processor architecture. [Figure 16B] 1 shows a block diagram of another graphics processor architecture. [Figure 16C] 1 illustrates a block diagram of another compute accelerator architecture.

[0022] [Figure 17] FIG. 2 is a block diagram of a graphics processing engine of a graphics processor.

[0023] [Figure 18A] 1 illustrates thread execution logic including an array of processing elements used in a graphics processor core. [Figure 18B] 1 illustrates thread execution logic including an array of processing elements used in a graphics processor core.

[0024] [Figure 19] 1 illustrates another execution unit.

[0025] [Figure 20] FIG. 2 is a block diagram showing an instruction format of a graphics processor.

[0026] [Figure 21] FIG. 2 is a block diagram of another graphics processor architecture.

[0027] [Figure 22A] 2 shows a graphics processor command format. [Figure 22B] 2 shows a graphics processor command sequence.

[0028] [Figure 23] 1 illustrates an exemplary graphics software architecture for a data processing system.

[0029] [Figure 24A]FIG. 1 is a block diagram showing an IP core development system.

[0030] [Figure 24B] 1 shows a vertical cross-sectional view of an integrated circuit package assembly.

[0031] [Figure 24C] 1 illustrates a package assembly that includes multiple hardware logic chiplet units connected to a substrate (eg, a base die).

[0032] [Figure 24D] 1 shows a package assembly containing replaceable chiplets.

[0033] [Diagram 25] FIG. 1 is a block diagram illustrating an exemplary system-on-chip integrated circuit.

[0034] [Figure 26A] FIG. 1 is a block diagram illustrating an exemplary graphics processor for use within a SoC. [Figure 26B] FIG. 1 is a block diagram illustrating an exemplary graphics processor for use within a SoC.

[0035] [Figure 27] 27 illustrates a block diagram of a circuit 2700 for executing DP2A instructions in BF16 format on a processing resource (eg, a functional unit, FPU) according to one embodiment.

[0036] [Figure 28] 28 shows a block diagram of a circuit 2800 for executing a DP2A instruction in BF16 format in the FPU according to another embodiment.

[0037] [Figure 29] 29 shows a block diagram of a circuit 2900 for executing a DP2A instruction in BF16 format in the FPU according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0038] A graphics processing unit (GPU) is communicatively coupled to a host / processor core to, for example, accelerate graphics operations, machine learning operations, pattern analysis operations, and / or various general-purpose GPU (GPGPU) functions. The GPU may be communicatively coupled to the host processor / core by a bus or another interconnect (e.g., a high-speed interconnect such as PCIe or NVLink). Alternatively, the GPU may be integrated in the same package or chip as the core and communicatively coupled to the core (i.e., internal to the package or chip) by an internal processor bus / interconnect. Regardless of the manner in which the GPU is connected, the processor core may assign work to the GPU in the form of a sequence of commands / instructions contained in a work descriptor. The GPU then uses dedicated circuitry / logic to efficiently process these commands / instructions.

[0039] In the following description, numerous specific details are set forth to provide a more thorough understanding. However, it will be apparent to one of ordinary skill in the art that the embodiments described herein may be practiced without one or more of these specific details. In other instances, well-known features have not been described so as not to obscure the details of the embodiments. [System Overview]

[0040] 1 is a block diagram illustrating a computer processing system 100 configured to implement one or more aspects of the embodiments described herein. The computer processing system 100 includes a processing subsystem 101 having one or more processors 102 and a system memory 104 that communicate through an interconnect path that may include a memory hub 105. The memory hub 105 may be a separate component in a chipset component or may be integrated into the one or more processors 102. The memory hub 105 is coupled to an I / O subsystem 111 via a communication link 106. The I / O subsystem 111 includes an I / O hub 107 that may enable the computer processing system 100 to receive input from one or more input devices 108. Additionally, the I / O hub 107 may enable a display controller that may be included in the one or more processors 102 to provide output to one or more display devices 110A. In one embodiment, one or more display devices 110A coupled to I / O hub 107 may include local, built-in, or embedded display devices.

[0041] The processing subsystem 101 includes, for example, one or more parallel processors 112 coupled to a memory hub 105 via a bus or other communication link 113. The communication link 113 may be one of any number of standards-based communication link technologies or protocols (such as, but not limited to, PCI Express) or may be a vendor-specific communication interface or fabric. The one or more parallel processors 112 may form a computationally intensive parallel or vector processing system (such as a many integrated core (MIC) processor) that may include multiple processing cores and / or processing clusters. For example, the one or more parallel processors 112 form a graphics processing subsystem that may output pixels to one of one or more display devices 110A coupled via an I / O hub 107. The one or more parallel processors 112 may also include a display controller and display interface (not shown) that allows for direct connection to one or more display devices 110B.

[0042] In the I / O subsystem 111, a system storage unit 114 may connect to the I / O hub 107 to provide a storage mechanism for the computer processing system 100. An I / O switch 116 may be used to provide an interface mechanism to allow connection between the I / O hub 107 and other components such as a network adapter 118 and / or a wireless network adapter 119 that may be integrated into the platform, and various other devices that may be added via one or more add-in devices 120. The add-in devices 120 may also include, for example, one or more external graphics processor devices and / or compute accelerators. The network adapter 118 may be an Ethernet adapter or another wired network adapter. The wireless network adapter 119 may include one or more of a Wi-Fi network device, a Bluetooth network device, a Near Field Communication (NFC) network device, or other network device including one or more radios.

[0043] Computer processing system 100 may include other components not expressly shown, including USB or other port connections, optical storage drives, video capture devices, and the like, and may also be connected to an I / O hub 107. Communication paths interconnecting the various components of Figure 1 may be implemented using any suitable protocol, such as a Peripheral Component Interconnect (PCI) based protocol (e.g., PCI-Express), or any other bus or point-to-point communication interface and / or protocol, such as an NV-Link high speed interconnect, or interconnect protocol known in the art.

[0044] The one or more parallel processors 112 may incorporate circuitry optimized for graphics and video processing (e.g., including video output circuitry) to form a graphics processing unit (GPU). Alternatively or additionally, the one or more parallel processors 112 may incorporate circuitry optimized for general-purpose processing while retaining the underlying computational architecture described in more detail herein. Each component of the computer processing system 100 may be integrated with one or more other system elements on a single integrated circuit. For example, the one or more parallel processors 112, the memory hub 105, the processor 102, and the I / O hub 107 may be integrated into a system-on-chip (SoC) type integrated circuit. Alternatively, each component of the computer processing system 100 may be integrated into a single package to form a system-in-package (SIP) configuration. In one embodiment, at least some of the components of the computer processing system 100 may be integrated into a multi-chip module (MCM) that may be interconnected with other multi-chip modules to form a modular computer processing system.

[0045] It will be understood that the computer processing system 100 shown herein is exemplary and that variations and modifications are possible. The connection topology, including the number and arrangement of bridges, the number of processors 102, and the number of parallel processors 112, may be modified as desired. For example, the system memory 104 may be directly connected to the processors 102 rather than through a bridge, and other devices communicate with the system memory 104 via the memory hub 105 and the processors 102. In other alternative topologies, the parallel processors 112 are connected to the I / O hub 107 rather than the memory hub 105, or directly to one of the one or more processors 102. In other embodiments, the I / O hub 107 and the memory hub 105 may be integrated into a single chip. Two or more sets of processors 102 may also be installed via multiple sockets, which may be coupled with two or more instances of the parallel processors 112.

[0046] Some of the specific components illustrated herein are optional and may not be included in all implementations of computer processing system 100. For example, any number of add-in cards or peripherals may be supported, and some components may be omitted. Additionally, some architectures may use different terminology for components similar to those illustrated in FIG. 1. For example, memory hub 105 may be referred to as a northbridge in some architectures, and I / O hub 107 may be referred to as a southbridge.

[0047] 2A illustrates a parallel processor 200. The parallel processor 200 may be a GPU or GPGPU as described herein, or the like. Various components of the parallel processor 200 may be implemented using one or more integrated circuit devices, such as a programmable processor, an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA). The illustrated parallel processor 200 may be the parallel processor 112, or one of the parallel processors 112, illustrated in FIG. 1.

[0048] Parallel processor 200 includes parallel processing units 202. Parallel processing units include I / O units 204 that enable communication with other devices, including other instances of parallel processing units 202. I / O units 204 may be directly connected to other devices. For example, I / O units 204 connect to other devices through the use of a hub, such as memory hub 105, or a switch interface. The connection between memory hub 105 and I / O units 204 forms communication link 113. In parallel processing units 202, I / O units 204 connect to host interface 206 and memory crossbar 216. Host interface 206 receives commands directed to the execution of processing operations, and memory crossbar 216 receives commands directed to the execution of memory operations.

[0049] Once the host interface 206 receives command buffers via the I / O unit 204, the host interface 206 can send work operations to the front end 208 to execute these commands. In one embodiment, the front end 208 couples to a scheduler 210 configured to distribute commands or other work items to the processing cluster array 212. The scheduler 210 ensures that the processing cluster array 212 is properly configured and enabled before tasks are distributed to the processing clusters of the processing cluster array 212. The scheduler 210 may be implemented by firmware logic executing on a microcontroller. The microcontroller-implemented scheduler 210 can be configured to perform complex scheduling and work distribution operations at both coarse and fine grains, allowing for fast preemption and context switching of threads executing on the processing array 212. The host software is preferably able to ascertain the workload to schedule to the processing array 212 via one of multiple graphics processing doorbells. The workload may then be automatically distributed across the processing array 212 by scheduler 210 logic within the scheduler microcontroller.

[0050] The processing cluster array 212 may include up to “N” processing clusters (e.g., cluster 214A, cluster 214B, ..., cluster 214N). Each cluster 214A-214N of the processing cluster array 212 may execute multiple concurrent threads. The scheduler 210 may assign work to the clusters 214A-214N of the processing cluster array 212 using various scheduling and / or work distribution algorithms that may vary depending on the workload generated by each type of program or computation. Scheduling may be handled dynamically by the scheduler 210 or may be assisted in part by compiler logic during compilation of program logic configured for execution by the processing cluster array 212. As desired, individual clusters 214A-214N of the processing cluster array 212 may be assigned to process different types of programs or perform different types of computations.

[0051] The processing cluster array 212 may be configured to perform various types of parallel processing operations. For example, the cluster array 212 may be configured to perform general-purpose parallel compute operations. For example, the processing cluster array 212 may include logic to perform processing tasks including filtering video and / or audio data, performing modeling operations including physics operations, and performing data transformations.

[0052] The processing cluster array 212 is configured to perform parallel graphics processing operations. In such embodiments in which the parallel processor 200 is configured to perform graphics processing operations, the processing cluster array 212 may include other logic to support the execution of such graphics processing operations, including, but not limited to, texture sampling logic to perform texture operations, as well as tessellation logic and other vertex processing logic. Additionally, the processing cluster array 212 may be configured to execute graphics processing related shader programs, including, but not limited to, vertex shaders, tessellation shaders, geometry shaders, and pixel shaders. The parallel processing units 202 may transfer data from the system memory for processing via the I / O units 204. During processing, the transferred data may be stored in an on-chip memory (e.g., the parallel processor memory 222) during processing and then written back to the system memory.

[0053] In embodiments in which graphics processing is performed using parallel processing units 202, scheduler 210 may be configured to divide the processing workload into multiple tasks of approximately equal size to more effectively distribute graphics processing operations among multiple clusters 214A-214N of processing cluster array 212. In some of these embodiments, portions of processing cluster array 212 may be configured to perform different types of processing. For example, to generate an image for rendered display, a first portion may be configured to perform vertex shading and topology generation, a second portion may be configured to perform tessellation and geometry shading, and a third portion may be configured to perform pixel shading or other screen space operations. Intermediate data generated by one or more of clusters 214A-214N may be stored in a buffer to allow the intermediate data to be sent to clusters 214A-214N for further processing.

[0054] In operation, the processing cluster array 212 may receive processing tasks to be performed via the scheduler 210, which receives commands from the front end 208 defining the processing tasks. In a graphics processing operation, a processing task may include an index of the data to be processed (e.g., surface (patch) data, primitive data, vertex data, and / or pixel data), as well as state parameters and commands that define how the data is processed (e.g., which program is to be executed). The scheduler 210 may be configured to fetch an index corresponding to the task, and may receive the index from the front end 208. The front end 208 may be configured to configure the processing cluster array 212 to be in a valid state before a workload specified by a received command buffer (e.g., a batch buffer, a push buffer, etc.) is initiated.

[0055] Each of the one or more instances of the parallel processing unit 202 can be coupled to a parallel processor memory 222. The parallel processor memory 222 can be accessed via a memory crossbar 216, which can receive memory requests from the processing cluster array 212 and the I / O unit 204. The memory crossbar 216 can access the parallel processor memory 222 via a memory interface 218. The memory interface 218 can include a number of partition units (e.g., partition unit 220A, partition unit 220B, ..., partition unit 220N), each of which can be coupled to a portion (e.g., a memory unit) of the parallel processor memory 222. The number of partition units 220A-220N can be configured to be equal to the number of memory units, such that the first partition unit 220A has a corresponding first memory unit 224A, the second partition unit 220B has a corresponding memory unit 224B, and the Nth partition unit 220N has a corresponding Nth memory unit 224N. In other embodiments, the number of partition units 220A-220N may not be equal to the number of memory devices.

[0056] The memory units 224A-224N may include various types of memory devices including dynamic random access memory (DRAM) or graphics random access memory such as synchronous graphics random access memory (SGRAM) including graphics double data rate (GDDR) memory. If desired, the memory units 224A-224N may also include 3D stacked memory including but not limited to high bandwidth memory (HBM). Those skilled in the art will appreciate that the specific implementation of the memory units 224A-224N may vary and may be selected from one of various conventional schemes. Render targets such as frame buffers or texture maps may be stored entirely in the memory units 224A-224N, and the partition units 220A-220N may write portions of each render target in parallel, allowing efficient use of the available bandwidth of the parallel processor memory 222. In some embodiments, local instances of the parallel processor memory 222 may be omitted in favor of a unified memory scheme that utilizes system memory in conjunction with a local cache memory.

[0057] As needed, any one of the clusters 214A-214N of the processing cluster array 212 is capable of processing data that is written to any one of the memory units 224A-224N in the parallel processor memory 222. The memory crossbar 216 may be configured to forward the output of each cluster 214A-214N to any of the partition units 220A-220N or to another cluster 214A-214N, which may perform other processing operations on the output. Each cluster 214A-214N may communicate with a memory interface 218 through the memory crossbar 216 to read from or write to various external memory devices. In one embodiment having a memory crossbar 216, the memory crossbar 216 has a connection to a memory interface 218 for communicating with the I / O units 204 and a connection to a local instance of a parallel processor memory 222, allowing processing units in different processing clusters 214A-214N to communicate with system memory or other memory that is not local to the parallel processing units 202. In general, the memory crossbar 216 may be capable of using virtual channels to separate traffic streams between the clusters 214A-214N and the partition units 220A-220N, for example.

[0058] Although one instance of parallel processing unit 202 is shown in parallel processor 200, any number of instances of parallel processing unit 202 may be included. For example, multiple instances of parallel processing unit 202 may be provided on a single add-in card, and multiple add-in cards may be interconnected. Individual instances of parallel processing unit 202 may be configured to work with each other, even though the individual instances may have different numbers of processing cores, different amounts of local parallel processor memory, and / or other configuration differences. If desired, some instances of parallel processing unit 202 may include a higher precision floating point unit 264 than other instances. Systems incorporating one or more instances of parallel processing unit 202 or parallel processor 200 may be implemented in a variety of configurations and form factors, including, but not limited to, desktops, laptops, handheld personal computers, servers, workstations, game consoles, and / or embedded systems.

[0059] FIG. 2B is a block diagram of partition unit 220. Partition unit 220 may be an instance of one of partition units 220A-220N shown in FIG. 2A. As shown, partition unit 220 includes L2 cache 221, metadata cache 224 for storing metadata (e.g., control surface updates, compressed values), frame buffer interface 225, and ROP 226 (raster operations unit). L2 cache 221 is a read / write cache configured to perform load and store operations received from memory crossbar 216 and ROP 226. Read misses and urgent writeback requests are output by L2 cache 221 to frame buffer interface 225 for processing. These updates may also be sent to a frame buffer via frame buffer interface 225 for processing. In one embodiment, frame buffer interface 225 interfaces with one of multiple memory units in a parallel processor memory (e.g., in parallel processor memory 222), such as memory units 224A-224N shown in FIG. 2A. The partition unit 220 may additionally or alternatively interface with one of multiple memory units in the parallel processor memory via a memory controller (not shown).

[0060] In graphics applications, ROP 226 is a processing unit that performs raster operations such as stencil, z-test, and blending. ROP 226 then outputs the processed graphics data that is stored in the graphics memory. In some embodiments, ROP 226 includes compression logic that compresses depth and color data that is written to memory and restores depth or color data that is read from memory. The compression logic may be lossless compression logic that utilizes one or more of a number of compression algorithms. The type of compression performed by ROP 226 may vary based on statistical characteristics of the data being compressed. For example, in one embodiment, delta color compression is performed on the depth and color data on a tile-by-tile basis.

[0061] ROP 226 may be included in each processing cluster (e.g., clusters 214A-214N of FIG. 2A) instead of being included in partition unit 220. In such an embodiment, read and write requests for pixel data are routed by memory crossbar 216 instead of pixel fragment data. Processed graphics data may be displayed on a display device, such as one of one of one or more display devices 110 shown in FIG. 1, routed for further processing by processor 102, or routed for further processing by one of multiple processing entities in parallel processor 200 shown in FIG. 2A.

[0062] FIG. 2C is a block diagram of a processing cluster 214 in a parallel processing unit. For example, the processing cluster is an instance of one of the processing clusters 214A-214N shown in FIG. 2A. The processing cluster 214 may be configured to execute many threads in parallel, with the term "thread" meaning an instance of a particular program executed on a particular set of input data. If desired, rather than providing multiple independent instruction units, a single instruction multiple data (SIMD) type instruction issue technique may be used to support the parallel execution of multiple threads. Alternatively, a single instruction multiple thread (SIMT) technique may be used to support the parallel execution of multiple threads, typically synchronized, with a common instruction unit configured to issue instructions to a set of processing engines in each of multiple processing clusters. Unlike a SIMD execution scheme, where all processing engines typically execute the exact same instructions, SIMT execution allows individual threads to more easily follow divergent execution paths with a given thread program. Those skilled in the art will appreciate that a SIMD processing scheme represents a functional subset of a SIMT processing scheme.

[0063] The operation of the processing cluster 214 may be controlled through a pipeline manager 232 that distributes processing tasks to the SIMT parallel processors. The pipeline manager 232 receives instructions from the scheduler 210 of FIG. 2A and manages the execution of these instructions through the graphics multiprocessor 234 and / or the texture unit 236. The illustrated graphics multiprocessor 234 is an exemplary instance of a SIMT parallel processor. However, various types of SIMT parallel processors of different architectures may be included in the processing cluster 214. One or more instances of the graphics multiprocessor 234 may be included in the processing cluster 214. The graphics multiprocessor 234 may process data, and a data crossbar 240 may be used to distribute the processed data to one of several possible destinations, including other shader units. The pipeline manager 232 may facilitate the distribution of the processed data by specifying the destination of the processed data to be distributed via the data crossbar 240.

[0064] Each graphics multiprocessor 234 in a processing cluster 214 may include the exact same set of function execution logic (e.g., arithmetic logic units, load store units, etc.). The function execution logic may be configured in a pipelined manner that allows new instructions to issue before previous instructions complete. The function execution logic supports a variety of operations including integer and floating point arithmetic, comparison operations, Boolean operations, bit shifts, and calculation of various algebraic functions. The same functional unit hardware may be utilized to perform a variety of operations, and any combination of functional units may exist.

[0065] An instruction sent to a processing cluster 214 constitutes a thread. A set of threads that execute across a set of parallel processing engines is a thread group. A thread group executes the same program on different input data. Each thread in a thread group may be assigned to a different processing engine in the graphics multiprocessor 234. A thread group may include fewer threads than the number of processing engines in the graphics multiprocessor 234. If a thread group includes fewer threads than the number of processing engines, one or more of these processing engines may be idle during the processing cycle of the thread group. A thread group may also include more threads than the number of processing engines in the graphics multiprocessor 234. If a thread group includes more threads than the number of processing engines in the graphics multiprocessor 234, processing may be performed over consecutive clock cycles. Multiple thread groups may be executed simultaneously in the graphics multiprocessor 234 if desired.

[0066] The graphics multiprocessor 234 may include an internal cache memory for performing load and store operations. If desired, the graphics multiprocessor 234 may do without an internal cache and use a cache memory (e.g., L1 cache 248) in the processing cluster 214. Each graphics multiprocessor 234 may also access an L2 cache in a partition unit (e.g., partition units 220A-220N of FIG. 2A) that is shared by all processing clusters 214 and may be used to transfer data between threads. The graphics multiprocessor 234 may also access off-chip global memory, which may include one or both of local parallel processor memory and / or system memory. Any memory external to the parallel processing unit 202 may be used as global memory. In an embodiment in which the processing cluster 214 includes multiple instances of the graphics multiprocessor 234, they may share common instructions and data that may be stored in the L1 cache 248.

[0067] Each processing cluster 214 may include a memory management unit (MMU) 245 configured to map virtual addresses to physical addresses. In other embodiments, one or more instances of MMU 245 may be present in memory interface 218 of FIG. 2A. MMU 245 includes a set of page table entries (PTEs) used to map virtual addresses to physical addresses of tiles and, if necessary, cache line indexes. MMU 245 may include an address translation lookaside buffer (TLB) or cache that may be present in graphics multiprocessor 234 or an L1 cache or processing cluster 214. The physical addresses are processed to distribute locality of surface data access and enable efficient request interleaving among partition units. The cache line index may be used to determine whether a request for a cache line is a hit or a miss.

[0068] In graphics and computer processing applications, processing clusters 214 may be configured such that each graphics multiprocessor 234 is coupled to a texture unit 236 for performing texture mapping operations, such as determining texture sample locations, retrieving texture data, and filtering the texture data. Texture data is retrieved from an internal texture L1 cache (not shown), or in some embodiments, from an L1 cache within graphics multiprocessor 234, and fetched as needed from an L2 cache, local parallel processor memory, or system memory. Each graphics multiprocessor 234 outputs processed tasks to data crossbar 240, provides processed tasks to another processing cluster 214 for further processing, or stores processed tasks in an L2 cache, local parallel processor memory, or system memory via memory crossbar 216. A pre-ROP 242 (pre-raster operations unit) is configured to receive data from the graphics multiprocessor 234 and send the data to the ROP unit, which may be co-located with the partition units described herein (e.g., partition units 220A-220N of FIG. 2A). The pre-ROP 242 unit may perform color blending optimizations, organize pixel color data, and perform address conversion.

[0069] It will be understood that the core architecture described herein is exemplary and that variations and modifications are possible. Any number of processing units may be included in the processing clusters 214, e.g., graphics multiprocessor 234, texture unit 236, pre-ROP 242, etc. Additionally, while only one processing cluster 214 is shown, the parallel processing units described herein may include any number of instances of processing clusters 214. If desired, each processing cluster 214 may be configured to operate independently from the other processing clusters 214, using separate and distinct processing units, L1 caches, etc.

[0070] 2D illustrates an example of a graphics multiprocessor 234, where the graphics multiprocessor 234 couples with a pipeline manager 232 of a processing cluster 214. The graphics multiprocessor 234 has an execution pipeline including, but not limited to, an instruction cache 252, an instruction unit 254, an address mapping unit 256, a register file 258, one or more general purpose graphics processing unit (GPGPU) cores 262, and one or more load / store units 266. The GPGPU cores 262 and the load / store units 266 are coupled to a cache memory 272 and a shared memory 270 via a memory and cache interconnect 268. The graphics multiprocessor 234 may further include a tensor and / or ray tracing core 263 that includes hardware logic to accelerate matrix operations and / or ray tracing operations.

[0071] The instruction cache 252 may receive a sequence of instructions to execute from the pipeline manager 232. These instructions are cached in the instruction cache 252 and dispatched for execution by the instruction unit 254. The instruction unit 254 may dispatch instructions as thread groups (e.g., warps), with each thread of a thread group being assigned to various execution units in the GPGPU cores 262. An instruction may access either the local address space, the shared address space, or the global address space by specifying an address in the unified address space. The address mapping unit 256 may be used to translate addresses in the unified address space into different memory addresses that the load / store unit 266 can access.

[0072] Register file 258 provides a set of registers to the functional units of graphics multiprocessor 234. Register file 258 provides temporary storage for operands connected to the data paths of the functional units (e.g., GPGPU cores 262, load / store unit 266) of graphics multiprocessor 234. Register file 258 may be partitioned among the functional units such that each functional unit is assigned a dedicated portion of register file 258. For example, register file 258 may be partitioned among the various warps that graphics multiprocessor 234 executes.

[0073] Each of these GPGPU cores 262 may include functional units (e.g., a floating-point unit 264 (FPU) and / or an integer arithmetic logic unit (ALU)) used to execute instructions of the graphics multiprocessor 234. In some implementations, the GPGPU cores 262 may include hardware logic that may otherwise be present in the tensor and / or ray tracing cores 263. These GPGPU cores 262 may be similar in architecture or may be different in architecture. For example, in one embodiment, a first portion of the GPGPU cores 262 includes a single-precision FPU and an integer ALU, and a second portion of the GPGPU cores includes a double-precision FPU. If desired, the FPU may implement the IEEE 754-2008 standard for floating-point arithmetic or enable variable-precision floating-point arithmetic. The graphics multiprocessor 234 may further include one or more fixed-function or special-function units that perform specific functions, such as rectangular copy operations or pixel blending operations. One or more of the multiple GPGPU cores may also include fixed function logic or special function logic.

[0074] GPGPU core 262 may include SIMD logic capable of executing a single instruction on multiple sets of data. Optionally, GPGPU core 262 may physically execute SIMD4, SIMD8, and SIMD16 instructions, and logically execute SIMD1, SIMD2, and SIMD32 instructions. SIMD instructions for the GPGPU core may be generated by a shader compiler at compile time, or may be generated automatically when executing a program written and compiled for a single program, multiple data (SPMD) or SIMT architecture. Multiple threads of a program configured for a SIMT execution model may be executed by one SIMD instruction. For example, in one embodiment, eight SIMT threads performing the same or similar operations may be executed in parallel by one SIMD8 logical unit.

[0075] Memory and cache interconnect 268 is an interconnect network that connects each functional unit of graphics multiprocessor 234 to register file 258 and shared memory 270. For example, memory and cache interconnect 268 is a crossbar interconnect that allows load / store unit 266 to perform load and store operations between shared memory 270 and register file 258. Because register file 258 can operate at the same frequency as GPGPU cores 262, data transfer latency between GPGPU cores 262 and register file 258 is very low. Shared memory 270 may be used to enable communication between threads executing on functional units within graphics multiprocessor 234. Cache memory 272 may be used as a data cache, for example, to cache texture data communicated between the functional units and texture unit 236. Shared memory 270 may also be used as a program-managed cache. Threads executing on GPGPU cores 262 can programmatically store data in the shared memory in addition to the automatically cached data stored in cache memory 272.

[0076] 3A-3C illustrate another graphics multiprocessor according to embodiments. FIGS. 3A-3B illustrate graphics multiprocessors 325, 350, which are related to and may be used in place of one of the graphics multiprocessors 234 of FIG. 2C. Thus, any disclosure herein of any functionality associated with the graphics multiprocessor 234 also discloses a corresponding combination with the graphics multiprocessor 325, 350, but is not limited to such. FIG. 3C illustrates a graphics processing unit (GPU) 380 that includes a set of dedicated graphics processing resources arranged in multicore groups 365A-365N corresponding to the graphics multiprocessors 325, 350. The illustrated graphics multiprocessors 325, 350 and multicore groups 365A-365N may be streaming multiprocessors (SMs) capable of simultaneous execution of multiple execution threads.

[0077] The graphics multiprocessor 325 of FIG. 3A includes multiple separate instances of execution resource units compared to the graphics multiprocessor 234 of FIG. 2D. For example, the graphics multiprocessor 325 may include multiple instances of instruction units 332A-332B, register files 334A-334B, and texture units 344A-344B. The graphics multiprocessor 325 also includes multiple sets of graphics or compute execution units (e.g., GPGPU cores 336A-336B, tensor cores 337A-337B, ray tracing cores 338A-338B), and multiple sets of load / store units 340A-340B. The execution resource units have a common instruction cache 330, texture and / or data cache memory 342, and shared memory 346.

[0078] The various components can communicate via interconnect fabric 327. Interconnect fabric 327 may include one or more crossbar switches that enable communication between the various components of graphics multiprocessor 325. Interconnect fabric 327 may be a separate high-speed network fabric layer on which each component of graphics multiprocessor 325 is stacked. Each component of graphics multiprocessor 325 communicates with remote components via interconnect fabric 327. For example, GPGPU cores 336A-336B, 337A-337B, and 338A-338B can each communicate with shared memory 346 via interconnect fabric 327. Interconnect fabric 327 can arbitrate communications within graphics multiprocessor 325 to ensure fair bandwidth allocation between components.

[0079] Graphics multiprocessor 350 of FIG. 3B includes multiple sets of execution resources 356A-356D, each set including multiple instruction units, register files, GPGPU cores, and load store units, as shown in FIG. 2D and FIG. 3A. Execution resources 356A-356D can work in conjunction with texture units 360A-360D for texture manipulation, while sharing instruction cache 354 and shared memory 353. For example, execution resources 356A-356D can share instruction cache 354 and shared memory 353, as well as multiple instances of texture and / or data cache memories 358A-358B. The various components can communicate via interconnect fabric 352, similar to interconnect fabric 327 of FIG. 3A.

[0080] Those skilled in the art will appreciate that the architectures described in Figures 1, 2A-2D, and 3A-3B are descriptive and not limiting with respect to the scope of the present embodiments. Thus, the techniques described herein may be implemented in any suitably configured processing unit. Such processing units include, but are not limited to, one or more mobile application processors, one or more desktop or server CPUs, including multi-core central processing units (CPUs), one or more parallel processing units, such as parallel processing unit 202 of Figure 2A, and one or more graphics processors or special purpose processing units, without departing from the scope of the embodiments described herein.

[0081] A parallel processor or GPGPU as described herein may be communicatively coupled to a host / processor core that accelerates graphics operations, machine learning operations, pattern analysis operations, and various functions of a general-purpose GPU (GPGPU). The GPU may be communicatively coupled to the host processor / core by a bus or other interconnect (e.g., a high-speed interconnect such as PCIe or NVLink). In other embodiments, the GPU may be integrated into the same package or chip as the core and communicatively coupled to the core by an internal processor bus / interconnect (i.e., internal to the package or chip). Regardless of the manner in which the GPU is connected, the processor core may assign work to the GPU in the form of a sequence of commands / instructions contained in a work descriptor. The GPU then uses dedicated circuitry / logic to efficiently process these commands / instructions.

[0082] 3C illustrates a graphics processing unit (GPU) 380 that includes a set of dedicated graphics processing resources arranged into multi-core groups 365A-365N. Although details of only one multi-core group 365A are provided, it will be understood that the other multi-core groups 365B-365N may include the same or similar sets of graphics processing resources. Details described with respect to multi-core groups 365A-365N may also apply to any of the graphics multiprocessors 234, 325, 350 described herein.

[0083] As shown, multi-core group 365A may include a set of graphics cores 370, a set of tensor cores 371, and a set of ray tracing cores 372. A scheduler / dispatcher 368 schedules and dispatches the graphics threads to execute on the various cores 370, 371, 372. A set of register files 369 store operand values ​​for use by the cores 370, 371, 372 when executing the graphics threads. These register files may include, for example, integer registers for storing integer values, floating point registers for storing floating point values, vector registers for storing packed data elements (integer data elements and / or floating point data elements), and tile registers for storing tensor / matrix values. The tile registers may be implemented as a set of combined vector registers.

[0084] One or more combined level cache 1 (L1) and shared memory units 373 store graphics data such as texture data, vertex data, pixel data, ray data, bounding volume data, etc., local to each multicore group 365A. One or more texture units 374 may also be used to perform texture operations such as texture mapping and sampling. A level 2 (L2) cache 375, shared by all or a subset of the multicore groups 365A-365N, stores graphics data and / or instructions for multiple simultaneous graphics threads. As shown, the L2 cache 375 may be shared across multiple multicore groups 365A-365N. One or more memory controllers 367 couple the GPU 380 to memory 366, which may be system memory (e.g., DRAM) and / or dedicated graphics memory (e.g., GDDR6 memory).

[0085] Input / output (I / O) circuitry 363 couples GPU 380 to one or more I / O devices 362, such as a digital signal processor (DSP), a network controller, or a user input device. An on-chip interconnect may be used to couple I / O device 362 to GPU 380 and memory 366. One or more I / O memory management units (IOMMUs) 364 of I / O circuitry 363 directly couple I / O device 362 to system memory 366. Optionally, IOMMU 364 manages multiple sets of page tables that map virtual addresses to physical addresses in system memory 366. I / O device 362, CPU 361, and GPU 380 may then share the same virtual address space.

[0086] In one implementation of IOMMU 364, IOMMU 364 supports virtualization. In this case, IOMMU 364 may manage a first set of page tables that map guest / graphics virtual addresses to guest / graphics physical addresses and a second set of page tables that map guest / graphics physical addresses to system / host physical addresses (e.g., in system memory 366). The base addresses of each of the first and second sets of page tables may be stored in control registers and swapped out on a context switch (e.g., so that a new context is provided access to the associated set of page tables). Although not shown in FIG. 3C, each of cores 370, 371, 372, and / or multicore groups 365A-365N may include a translation lookaside buffer (TLB) for caching guest virtual to guest physical translations, guest physical to host physical translations, and guest virtual to host physical translations.

[0087] The CPU 361, GPU 380, and I / O devices 362 may be integrated into a single semiconductor chip and / or chip package. The illustrated memory 366 may be integrated into the same chip or may be coupled to a memory controller 367 via an off-chip interface. In one implementation, the memory 366 includes GDDR6 memory that shares the same virtual address space as other physical system-level memory, although the underlying principles described herein are not limited to this particular implementation.

[0088] Tensor cores 371 may include multiple execution units specifically designed to perform matrix operations, which are fundamental compute operations used to perform deep learning operations. For example, simultaneous matrix multiplication operations may be used in neural network training and inference. Tensor cores 371 may perform matrix processing using a variety of operand precisions, including single-precision floating point (e.g., 32 bits), half-precision floating point (e.g., 16 bits), integer word (16 bits), byte (8 bits), and half-byte (4 bits). For example, a neural network implementation may extract feature points from each rendered scene and potentially combine detailed information from multiple frames to construct a high-quality final image.

[0089] In a deep learning implementation, parallel matrix multiplication operations may be scheduled for execution on tensor cores 371. Training a neural network, in particular, requires many matrix dot product operations. To process the inner product formula of an N×N×N matrix multiplication, tensor cores 371 may include at least N dot product processing elements. Before starting the matrix multiplication, an entire matrix is ​​loaded into a tile register, and at least one column of a second matrix is ​​loaded every N cycles. Every cycle, N dot products are processed.

[0090] Matrix elements may be stored in various precisions, including 16-bit words, 8-bit bytes (e.g., INT8), and 4-bit half-bytes (e.g., INT4), depending on the particular implementation. Different precision modes may be specified for tensor core 371 to ensure that the most efficient precision is used for different workloads (e.g., inference workloads that can tolerate quantization into bytes and half-bytes).

[0091] Ray tracing core 372 may accelerate ray tracing operations for both real-time and non-real-time ray tracing implementations. In particular, ray tracing core 372 may include ray traversal / intersection circuitry for performing ray traversal using a bounding volume hierarchy (BVH) and identifying intersections of rays and primitives enclosed in the BVH volume. Ray tracing core 372 may also include circuitry for performing depth testing and culling (e.g., using a Z-buffer or similar configuration). In one implementation, ray tracing core 372 performs traversal and intersection operations in conjunction with image denoising techniques described herein, at least some of which may be executed by tensor core 371. For example, tensor core 371 may implement a deep learning neural network to denoise frames generated by ray tracing core 372. However, the CPU 361, the graphics core 370, and / or the ray tracing core 372 may also implement all or part of the denoising algorithm and / or the deep learning algorithm.

[0092] Additionally, as mentioned above, a distributed approach to denoising may be used where the GPU 380 is in a computing device that is coupled to other computing devices by a network or high speed interconnect. In this distributed approach, the interconnected computing devices may share neural network learning / training data to improve the speed at which the entire system learns to denoise different types of image frames and / or different graphics applications.

[0093] The ray tracing core 372 may handle all BVH traversal and / or ray-primitive intersection to prevent the graphics core 370 from being overloaded with thousands of instructions per ray. For example, each ray tracing core 372 includes a first set of dedicated circuitry to perform bounding box tests (e.g., for traversal operations) and / or a second set of dedicated circuitry to perform ray-triangle intersection tests (e.g., intersecting traversed rays). Thus, for example, the multi-core group 365A may simply launch a ray probe, and the ray tracing core 372 may independently perform ray traversal and intersection, returning hit data (e.g., hit, no hit, multiple hits, etc.) to the thread context. The other cores 370, 371 are freed to perform other graphics or compute work while the ray tracing core 372 performs traversal and intersection operations.

[0094] Optionally, each ray tracing core 372 may include a traversal unit to perform BVH test operations and / or an intersection unit to perform ray-primitive intersection tests. The intersection unit generates and provides a "hit" response, a "no hit" response, or a "multiple hits" response to the appropriate thread. During traversal and intersection operations, the execution resources of other cores (e.g., graphics core 370 and tensor core 371) are freed to perform other forms of graphics work.

[0095] In one optional embodiment described below, a hybrid rasterization / ray tracing approach is used, in which work is distributed between a graphics core 370 and a ray tracing core 372.

[0096] The ray tracing core 372 (and / or other cores 370, 371) may include hardware support for a ray tracing instruction set such as Microsoft's DirectX Ray Tracing (DXR), including the DispatchRays command, as well as ray generation shaders, nearest hit shaders, any hit shaders, and miss shaders, which allow for a unique set of shaders and textures to be assigned per object. Another ray tracing platform that may be supported by the ray tracing core 372, graphics core 370, and tensor core 371 is Vulkan 1.1.85. However, it should be noted that the underlying principles described herein are not limited to any particular ray tracing ISA.

[0097] In general, the various cores 372, 371, 370 may support a ray tracing instruction set that includes one or more of the following instructions / functions: ray generation, nearest hit, any hit, ray-primitive intersection, per-primitive and hierarchical bounding box construction, miss, visit, and exception. More specifically, preferred embodiments include ray tracing instructions that perform one or more of the following functions:

[0098] Ray Generation: Ray generation instructions may be executed per pixel, per sample, or other user-defined work allocation.

[0099] Nearest Hit: A nearest hit instruction may be executed to identify the nearest intersection point between a ray and a primitive in a scene.

[0100] Any Hit: The any hit command identifies multiple intersection points between rays and primitives in the scene, and in some cases identifies a new closest intersection point.

[0101] Intersection: The intersection instruction performs a ray-primitive intersection test and outputs the result.

[0102] Per-primitive bounding box construction: This instruction creates a bounding box around a given primitive or group of primitives (e.g., when creating a new BVH or other acceleration data structure).

[0103] Miss: The ray did not hit any geometry in the Scene or in the specified area of ​​the Scene.

[0104] Visit: Indicates the child volume that the ray traverses.

[0105] Exceptions: Contains different types of exception handlers (e.g., invoked for various error conditions). [GPU-host processor interconnection techniques]

[0106] 4A illustrates an exemplary architecture in which multiple GPUs 410-413, such as the parallel processor 200 illustrated in FIG. 2A, are communicatively coupled to multiple multi-core processors 405-406 by high-speed links 440A-440D (e.g., buses, point-to-point interconnects, etc.). Depending on the implementation, the high-speed links 440A-440D may support communication throughputs of 4 GB / s, 30 GB / s, 80 GB / s, or higher. A variety of interconnect protocols may be used, including, but not limited to, PCIe 4.0 or 5.0 and NVLink 2.0. However, the underlying principles described herein are not limited to any particular communication protocol or throughput.

[0107] Two or more of the GPUs 410-413 may be interconnected by high-speed links 442A-442B, which may be implemented using the same or different protocol / link as used for the high-speed links 440A-440D. Similarly, two or more of the multi-core processors 405-406 may be connected by high-speed link 443, which may be a symmetric multi-processor (SMP) bus operating at 20 GB / s, 30 GB / s, 120 GB / s, or faster. Alternatively, communication between the various system components shown in FIG. 4A may all be achieved using the same protocol / link (e.g., by a common interconnect fabric). However, as previously mentioned, the underlying principles described herein are not limited to any particular type of interconnect technology.

[0108] Each multi-core processor 405-406 may be communicatively coupled to a processor memory 401-402, respectively, via a memory interconnect 430A-430B, and each GPU 410-413 may be communicatively coupled to a GPU memory 420-423, respectively, by a GPU memory interconnect 450A-450D. The memory interconnects 430A-430B and 450A-450D may utilize the same or different memory access technologies. By way of example and not limitation, the processor memory 401-402 and the GPU memory 420-423 may be volatile memory, such as dynamic random access memory (DRAM) (including stacked DRAM), graphics DDR SDRAM (GDDR) (e.g., GDDR5, GDDR6), or high bandwidth memory (HBM), and / or may be non-volatile memory, such as 3D XPoint or Nano-Ram. For example, some of these memories may be volatile memory and other portions may be non-volatile memory (eg, using a two level memory (2LM) hierarchy).

[0109] As described below, the various processors 405-406 and GPUs 410-413 may each be physically coupled to a particular memory 401-402, 420-423, but a unified memory architecture may be implemented in which the same virtual system address space (also called the "effective address" space) is allocated to all of the various physical memories. For example, the processor memories 401-402 may each include 64 GB of system memory address space, and the GPU memories 420-423 may each include 32 GB of system memory address space (for a total of 256 GB of addressable memory in this example).

[0110] 4B shows further optional details regarding the interconnection between multi-core processor 407 and graphics acceleration module 446. Graphics acceleration module 446 may include one or more GPU chips integrated into a line card that is coupled to processor 407 via high-speed link 440. Alternatively, graphics acceleration module 446 may be integrated into the same package or chip as processor 407.

[0111] The illustrated processor 407 includes multiple cores 460A-460D, each having a translation lookaside buffer 461A-461D and one or more caches 462A-462D. These cores may include various other components that execute instructions and process data. These other components (e.g., instruction fetch units, branch prediction units, decoders, execution units, reorder buffers, etc.) are not shown to avoid obscuring the underlying principles of the components described herein. The caches 462A-462D may include a level 1 (L1) cache and a level 2 (L2) cache. Additionally, one or more shared caches 456 may be included in the cache hierarchy and shared by the set of cores 460A-460D. For example, one embodiment of the processor 407 includes 24 cores, each with its own L1 cache, 12 shared L2 caches, and 12 shared L3 caches. In this embodiment, one of the L2 cache and the L3 cache is shared by two adjacent cores. The processor 407 and the graphics accelerator integrated module 446 are connected to a system memory 441. This system memory may include the processor memories 401-402.

[0112] Coherency of data and instructions stored in the various caches 462A-462D, 456 and system memory 441 is maintained by inter-core communication over coherence bus 464. For example, each cache may have associated cache coherency logic / circuitry that communicates over coherence bus 464 in response to detecting a read or write to a particular cache line. In one implementation, a cache snooping protocol is implemented on coherence bus 464 to snoop cache accesses. Cache snooping / coherency techniques are well understood by those skilled in the art and will not be described in detail here to avoid obscuring the underlying principles described herein.

[0113] A proxy circuit 425 may be provided that communicatively couples the graphics acceleration module 446 to the coherence bus 464, allowing the graphics acceleration module 446 to participate in cache coherence protocols as a peer of these cores. In particular, an interface 435 provides a connection to the proxy circuit 425 by a high-speed link 440 (e.g., a PCIe bus, an NVLink, etc.), and an interface 437 connects the graphics acceleration module 446 to the high-speed link 440.

[0114] In one implementation, the accelerator integrated circuit 436 provides cache management services, memory access services, context management services, and interrupt management services on behalf of the multiple graphics processing engines 431, 432, N of the graphics acceleration module 446. Each of the graphics processing engines 431, 432, N may have a separate graphics processing unit (GPU). Alternatively, the graphics processing engines 431, 432, N may include different types of graphics processing engines in a GPU, such as a graphics execution unit, a media processing engine (e.g., a video encoder / decoder), a sampler, and a blit engine. In other words, the graphics acceleration module may be a GPU having multiple graphics processing engines 431-432, N, and the graphics processing engines 431-432, N may be individual GPUs integrated in a common package, line card, or chip.

[0115] The accelerator integrated circuitry 436 may include a memory management unit (MMU) 439 for performing various memory management functions such as virtual-to-physical memory translation (also referred to as effective-to-real memory translation), and a memory access protocol for accessing the system memory 441. The MMU 439 may also include a translation lookaside buffer (TLB) (not shown) for caching translations between virtual / effective addresses and physical / real addresses. In one implementation, a cache 438 stores commands and data for efficient access by the graphics processing engines 431-432, N. Data stored in the cache 438 and the graphics memories 433-434, M may be kept coherent with the core caches 462A-462D, 456 and the system memory 441. As already mentioned, this may be achieved by a proxy circuit 425 participating in a cache coherency mechanism on behalf of cache 438 and memories 433-434, M (e.g., sending updates to cache 438 related to modifications / accesses of cache lines in processor caches 462A-462D, 456, and receiving updates from cache 438).

[0116] A set of registers 445 stores context data for threads executed by the graphics processing engines 431-432,N, and a context management circuit 448 manages the thread contexts. For example, the context management circuit 448 may perform save and restore operations to save and restore the context of various threads during a context switch (e.g., a first thread is saved and a second thread is stored so that the graphics processing engine can execute the second thread). For example, in a context switch, the context management circuit 448 may store current register values ​​in a designated area of ​​memory (e.g., identified by a context pointer). The context management circuit may then restore the register values ​​when returning to the context. An interrupt management circuit 447 may, for example, receive interrupts from system devices and handle the received interrupts.

[0117] In one implementation, virtual / effective addresses from the graphics processing engine 431 are translated by the MMU 439 to real / physical addresses in the system memory 441. Optionally, the accelerator integration circuitry 436 supports multiple (e.g., 4, 8, 16) graphics accelerator modules 446 and / or other accelerator devices. The graphics accelerator modules 446 may be dedicated to a single application executing on the processor 407 or may be shared among multiple applications. Optionally, a virtualized graphics execution environment is provided in which the resources of the graphics processing engines 431-432, N are shared with multiple applications or virtual machines (VMs). These resources may be subdivided into "slices" that are assigned to different VMs and / or applications based on the processing requirements and priorities associated with the VMs and / or applications.

[0118] Thus, the accelerator integration circuitry 436 acts as a bridge to the system for the graphics acceleration module 446, providing address translation and system memory caching services. In one embodiment, to facilitate the bridging function, the accelerator integration circuitry 436 may also include shared I / O 497 (e.g., PCIe, USB), as well as hardware that allows system control of voltage, clocking, performance, thermal, and security. The shared I / O 497 may utilize a separate physical connection or may go through high speed link 440. Additionally, the accelerator integration circuitry 436 may provide virtualization capabilities to the host processor that manages the virtualization, interrupts, and memory management of the graphics processing engine.

[0119] The hardware resources of the graphics processing engines 431-432, N are explicitly mapped into the real address space seen by the host processor 407, so that any host processor can directly address these resources using effective address values. One optional feature of the accelerator integrated circuitry 436 is the physical separation of the graphics processing engines 431-432, N, so that the graphics processing engines appear to the system as independent units.

[0120] One or more graphics memories 433-434, M may be coupled to each of the graphics processing engines 431-432, N, respectively. The graphics memories 433-434, M store instructions and data to be processed by each of the graphics processing engines 431-432, N, respectively. The graphics memories 433-434, M may be volatile memory, such as DRAM (including stacked DRAM), GDDR memory (e.g., GDDR5, GDDR6), or HBM, and / or may be non-volatile memory, such as 3D XPoint or Nano-Ram.

[0121] To reduce data traffic over high-speed link 440, biasing techniques may be used to ensure that the data stored in graphics memories 433-434,M is the data most frequently used by graphics processing engines 431-432,N, and preferably is not used (or at least is not frequently used) by cores 460A-460D. Similarly, the biasing mechanism attempts to keep data needed by a core (preferably not graphics processing engines 431-432,N) in the core's caches 462A-462D,456 and system memory 441.

[0122] According to the variant shown in Fig. 4C, the accelerator integration circuit 436 is integrated into the processor 407. The graphics processing engines 431-432, N communicate directly over high-speed link 440 to the accelerator integration circuit 436 via interface 437 and interface 435 (which again may utilize any type of bus or interface protocol). The accelerator integration circuit 436 may perform the same operations as described with respect to Fig. 4B, but with potentially higher throughput given its proximity to the coherency bus 464 and caches 462A-462D, 456.

[0123] The described embodiments may support different programming models, including a dedicated process programming model (without graphics acceleration module virtualization) and a shared programming model (with virtualization), which may include a programming model controlled by the accelerator integration circuitry 436 and a programming model controlled by the graphics acceleration module 446.

[0124] In a dedicated process model embodiment, the graphics processing engines 431-432,N may be dedicated to a single application or process under a single operating system. The single application can send other application requests to the graphics engines 431-432,N to provide virtualization within a VM / partition.

[0125] In a dedicated process programming model, the graphics processing engines 431-432, N may be shared by multiple VM / application partitions. The shared model requires a system hypervisor to virtualize the graphics processing engines 431-432, N and make them accessible to each operating system. In a single partition system without a hypervisor, the graphics processing engines 431-432, N are owned by the operating system. In either case, the operating system can virtualize the graphics processing engines 431-432, N and provide access to each process or application.

[0126] In the shared programming model, the graphics acceleration module 446 or the individual graphics processing engines 431-432,N select a process element using a process handle. The process element may be stored in the system memory 441 and be addressable using the effective-to-real address translation techniques described herein. The process handle may be an implementation-specific value provided to the host process when it registers its context with the graphics processing engines 431-432,N (i.e., calls the system software that adds the process element to the process element linked list). The low-order 16 bits of the process handle may be the offset of the process element in the process element linked list.

[0127] 4D illustrates an exemplary accelerator integration slice 490. As used herein, a "slice" includes a designated portion of the processing resources of the accelerator integration circuitry 436. An application effective address space 482 in the system memory 441 stores a process element 483. The process element 483 may be stored in response to a GPU call 481 from an application 480 executing on the processor 407. The process element 483 includes a process state of the corresponding application 480. A work descriptor (WD) 484 included in the process element 483 may be a job requested by the application and may include a pointer to a queue of jobs. In the latter case, the WD 484 is a pointer to a job request queue in the application's address space 482.

[0128] The graphics acceleration module 446 and / or the individual graphics processing engines 431-432, N may be shared by all or a subset of the processes in the system. For example, the techniques described herein may include an infrastructure for setting process state and sending a WD484 to the graphics acceleration module 446 to start a job in a virtualized environment.

[0129] In one implementation, the dedicated process programming model is implementation specific, in which a single process owns the graphics acceleration module 446 or an individual graphics processing engine 431. Because the graphics acceleration module 446 is owned by a single process, the hypervisor initializes the accelerator integrated circuitry 436 of the owning partition, and the operating system initializes the accelerator integrated circuitry 436 of the owning process when the graphics acceleration module 446 is allocated.

[0130] In operation, a WD fetch unit 491 in the accelerator integrated slice 490 fetches the next WD 484, which includes an indication of work to be performed by one of the multiple graphics processing engines of the graphics acceleration module 446. As shown, data from the WD 484 may be stored in a register 445 and used by the MMU 439, the interrupt management circuitry 447, and / or the context management circuitry 448. For example, the MMU 439 may include a segment / page walk circuit to access a segment / page table 486 in the OS virtual address space 485. The interrupt management circuitry 447 may process an interrupt event 492 received from the graphics acceleration module 446. When performing a graphics operation, an effective address 493 generated by the graphics processing engines 431-432,N is translated to a real address by the MMU 439.

[0131] The same set of registers 445 may be replicated for each graphics processing engine 431-432, N and / or graphics acceleration module 446 and may be initialized by the hypervisor or operating system. Each of these replicated registers may be included in the accelerator integration slice 490. Exemplary registers that may be initialized by the hypervisor are shown in Table 1. [Table 1]

[0132] Exemplary registers that may be initialized by the operating system are shown in Table 2. [Table 2]

[0133] Each WD 484 may be specific to a particular graphics acceleration module 446 and / or graphics processing engine 431-432, N. WD 484 may contain all the information needed for the graphics processing engines 431-432, N to do its work, or may be a pointer to a memory location where an application has set up a command queue for work to be completed.

[0134] 4E shows further optional details of the sharing model, including hypervisor real address space 498, in which process element list 499 is stored, accessible via hypervisor 496, which virtualizes the graphics acceleration module engine for operating system 495.

[0135] A shared programming model allows all or a subset of processes from all or a subset of the partitions in a system to use the graphics acceleration module 446. There are two programming models in which the graphics acceleration module 446 is shared by multiple processes and partitions: time sliced ​​shared and graphics oriented shared.

[0136] In this model, the system hypervisor 496 owns the graphics acceleration module 446 and makes its functionality available to all operating systems 495. For the graphics acceleration module 446 to support virtualization by the system hypervisor 496, the graphics acceleration module 446 may adhere to the following requirements: 1) the application's job requests must be autonomous (i.e., the state does not need to be maintained between jobs) or the graphics acceleration module 446 must provide a context save and restore mechanism; 2) the application's job requests are guaranteed by the graphics acceleration module 446 to complete in a constant amount of time including any transformation faults or the graphics acceleration module 446 provides the ability to preempt the processing of jobs; and 3) the graphics acceleration module 446 must ensure process fairness when operating in a oriented sharing programming model.

[0137] In the case of the shared model, the application 480 may be required to make a system call to the operating system 495 with the type of the graphics acceleration module 446, a working descriptor (WD), an authority mask register (AMR) value, and a context save / restore area pointer (CSRP). The type of the graphics acceleration module 446 indicates the target acceleration function of the system call. The type of the graphics acceleration module 446 may be a system specific value. The WD is formatted specifically for the graphics acceleration module 446 and may be in the form of a graphics acceleration module 446 command, an effective address pointer to a user defined type structure, an effective address pointer to a queue of commands, or any other data structure indicating the work to be done by the graphics acceleration module 446. In one embodiment, the AMR value is the AMR state to use for the current process. The value sent to the operating system is similar to the application setting the AMR. If the implementation of the accelerator integration circuitry 436 and the graphics acceleration module 446 does not support a user authority mask override register (UAMOR), the operating system may apply the current UAMOR value to the AMR value before sending the AMR in the hypervisor call. The hypervisor 496 may optionally apply the current authority mask override register (AMOR) value before placing the AMR in the process element 483. The CSRP may be one of several registers 445 that contain the effective address of a region in the application's address space 482 where the graphics acceleration module 446 saves and restores the context state. This pointer is optional if state does not need to be saved between jobs or when a job is being preempted. The context save / restore region may be pinned system memory.

[0138] When operating system 495 receives the system call, it may verify that application 480 is registered and authorized to use graphics acceleration module 446. Operating system 495 then calls hypervisor 496 with the information shown in Table 3. [Table 3]

[0139] When the hypervisor 496 receives the hypervisor call, it may verify that the operating system 495 is registered and authorized to use the graphics acceleration module 446. The hypervisor 496 then places the process element 483 in a process element linked list for the type of the corresponding graphics acceleration module 446. The process element may include the information shown in Table 4. [Table 4]

[0140] The hypervisor may initialize multiple registers 445 of the accelerator integrated slice 490.

[0141] As shown in FIG. 4F, one optional implementation uses a unified memory addressable via a common virtual memory address space used to access the physical processor memories 401-402 and the GPU memories 420-423. In this implementation, operations performed on the GPUs 410-413 use the same virtual / effective memory address space to access the processor memories 401-402 and vice versa. This simplifies programmability. A first portion of the virtual / effective address space may be assigned to the processor memory 401, a second portion to the second processor memory 402, a third portion to the GPU memory 420, and so on. This allows any processor or GPU to access any physical memory using virtual addresses mapped to the memory, with the entire virtual / effective memory space (sometimes referred to as the effective address space) being allocated to each of the processor memories 401-402 and the GPU memories 420-423.

[0142] Bias / coherence management circuits 494A-494E may be provided within one or more of the MMUs 439A-439E to ensure cache coherence between the host processor (e.g., 405) and the caches of the GPUs 410-413 and to implement biasing techniques to indicate the physical memory in which particular types of data should be stored. Although multiple instances of bias / coherence management circuits 494A-494E are shown in FIG. 4F, the bias / coherence circuits may be implemented within the MMUs of one or more of the host processors 405 and / or within the accelerator integration circuitry 436.

[0143] The GPU-attached memory 420-423 may be mapped as part of the system memory and accessed using shared virtual memory (SVM) techniques, but without suffering the typical performance drawbacks associated with full-system cache coherence. The ability of the GPU-attached memory 420-423 to be accessed as system memory without the burdensome cache coherence overhead provides a beneficial operating environment for GPU offload. This configuration allows host processor 405 software to set up operands and access computation results without the overhead of traditional I / O DMA data copies. Such typical copies require driver calls, interrupts, and memory-mapped I / O (MMIO) accesses that are quite inefficient compared to simple memory accesses. At the same time, the ability to access the GPU-attached memory 420-423 without cache coherence overhead can be critical to the execution time of the offloaded computation. In the presence of a large amount of streaming write memory traffic, for example, cache coherence overhead can significantly reduce the effective write bandwidth seen by the GPUs 410-413. The efficiency of operand setup, the efficiency of result access, and the efficiency of GPU computation all play a role in determining the effectiveness of GPU offload.

[0144] The selection between the GPU bias and the host processor bias may be determined by a bias tracker data structure. A bias table may be used, which may be, for example, a page granular structure that includes one or two bits per GPU-attached memory page (i.e., controlled at the granularity of a memory page). The bias table may be implemented in stolen memory regions of one or more GPU-attached memories 420-423, with or without a bias cache (e.g., caching frequently / recently used entries of the bias table) in the GPU 410-413. Alternatively, the entire bias table may be maintained within the GPU.

[0145] In one implementation, the bias table entry associated with each access to the GPU attached memory 420-423 is accessed before the actual access to the GPU memory, resulting in the following operations: First, local requests from the GPUs 410-413 to find a page for each GPU in the GPU bias are forwarded directly to the corresponding GPU memory 420-423. Local requests from the GPUs to find a page in the host bias are forwarded to the processor 405 (e.g., by the high-speed link described above). If necessary, the processor 405 request to find the requested page in the host processor bias completes the request as if it were a normal memory read. Alternatively, requests targeting the GPU bias page may be forwarded to the GPUs 410-413. The GPU may then migrate the page to the host processor bias if the host processor bias is not currently using the page.

[0146] The bias state of the page may be changed either by software-based mechanisms, hardware-assisted software-based mechanisms, or in limited combinations, entirely hardware-based mechanisms.

[0147] One mechanism for changing the bias state is to use an API call (e.g., OpenCL) that then calls the GPU's device driver, which then sends a message (or queues a command descriptor) to the GPU instructing it to change the bias state and, depending on the transition, to perform a cache flush operation on the host. A cache flush operation is required to transition from host processor 405 bias to GPU bias, but is not required for the reverse transition.

[0148] Cache coherency may be maintained by temporarily rendering GPU bias pages that are not cacheable by the host processor 405. To access these pages, the processor 405 may request access from the GPU 410, which may or may not immediately grant access, depending on the implementation. Thus, to reduce communication between the host processor 405 and the GPU 410, it is beneficial to ensure that the GPU bias pages are pages that are needed by the GPU and not the host processor 405, and vice versa. [Graphics Processing Pipeline]

[0149] FIG. 5 illustrates a graphics processing pipeline 500. A graphics multiprocessor, such as the graphics multiprocessor 234 in FIG. 2D, the graphics multiprocessor 325 in FIG. 3A, or the graphics multiprocessor 350 in FIG. 3B, may implement the illustrated graphics processing pipeline 500. The graphics multiprocessor may be included in a parallel processing subsystem described herein, such as the parallel processor 200 in FIG. 2A. The parallel processor 200 may be related to or may be used in place of one of the parallel processors 112 in FIG. 1. Various parallel processing systems may implement the graphics processing pipeline 500 with one or more instances of the parallel processing units described herein (e.g., the parallel processing unit 202 in FIG. 2A). For example, a shader unit (e.g., the graphics multiprocessor 234 in FIG. 2C) may be configured to perform one or more functions of the vertex processing unit 504, the tessellation control processing unit 508, the tessellation evaluation processing unit 512, the geometry processing unit 516, and the fragment / pixel processing unit 524. The functions of data assembler 502, primitive assemblers 506, 514, 518, tessellation unit 510, rasterizer 522, and raster operation unit 526 may also be performed by other processing engines in a processing cluster (e.g., processing cluster 214 of FIG. 2A) and corresponding partition unit (e.g., partition units 220A-220N of FIG. 2A). Graphics processing pipeline 500 may be implemented using dedicated processing units for one or more functions. It is also possible that one or more portions of graphics processing pipeline 500 are performed by parallel processing logic within a general-purpose processor (e.g., CPU). If necessary, one or more portions of graphics processing pipeline 500 can access on-chip memory (e.g., parallel processor memory 222 as in FIG. 2A) through memory interface 528. Memory interface 528 may be an instance of memory interface 218 of FIG. 2A.Graphics processor pipeline 500 may be implemented by multi-core group 365A as in FIG. 3C.

[0150] Data assembler 502 is a processing unit that may collect vertex data for surfaces and primitives. Data assembler 502 then outputs vertex data, including vertex attributes, to vertex processing unit 504. Vertex processing unit 504 is a programmable execution unit that runs vertex shader programs and performs lighting and transformation of the vertex data as specified by the vertex shader programs. Vertex processing unit 504 may be programmed to read data stored in a cache, local memory, or system memory for use in processing the vertex data, and to transform the vertex data from an object-based coordinate representation to a world space coordinate space or a normalized device coordinate space.

[0151] A first instance of a primitive assembler 506 receives vertex attributes from the vertex processing unit 504. The primitive assembler 506 reads the stored vertex attributes as needed and constructs graphics primitives for processing by the tessellation control processing unit 508. Graphics primitives include triangles, lines, points, patches, etc. that are supported by various graphics processing application programming interfaces (APIs).

[0152] The tessellation control processing unit 508 treats the input vertices as control points of a geometric patch. These control points are converted from the input representation of the patch (e.g., the base of the patch) to a representation suitable for surface evaluation by the tessellation evaluation processing unit 512. The tessellation control processing unit 508 can also calculate tessellation coefficients for the edges of the geometric patch. The tessellation coefficients are applied to a single edge to quantify a viewpoint-dependent level of detail associated with that edge. A tessellation unit 510 is configured to receive the tessellation coefficients for the edges of the patch and tessellate the patch into a number of geometric primitives, such as line primitives, triangle primitives, or quadrilateral primitives, which are sent to the tessellation evaluation processing unit 512. The tessellation evaluation processing unit 512 processes the parameterized coordinates of the subdivided patch and generates a surface representation and vertex attributes for each vertex associated with the geometric primitive.

[0153] A second instance of primitive assembler 514 receives vertex attributes from tessellation evaluation processing unit 512, reads stored vertex attributes as necessary, and constructs graphics primitives for processing by geometry processing unit 516. Geometry processing unit 516 is a programmable execution unit that executes a geometry shader program to transform the graphics primitives received from primitive assembler 514 as specified by the geometry shader program. Geometry processing unit 516 may be programmed to subdivide graphics primitives into one or more new graphics primitives and calculate parameters used to rasterize the new graphics primitives.

[0154] The geometry processing unit 516 may be capable of adding elements to or removing elements from the geometry stream. The geometry processing unit 516 outputs parameters and vertices that specify new graphics primitives to a primitive assembler 518. The primitive assembler 518 receives the parameters and vertices from the geometry processing unit 516 and builds graphics primitives for processing by the viewport scaling, culling, and clipping unit 520. The geometry processing unit 516 reads data stored in a parallel processor memory or in system memory for use in processing the geometry data. The viewport scaling, culling, and clipping unit 520 performs clipping, culling, and viewport scaling, and outputs the processed graphics primitives to a rasterizer 522.

[0155] The rasterizer 522 may perform depth sculling and other depth-based optimizations. The rasterizer 522 also performs scan conversion on new graphics primitives, generating fragments and outputting these fragments and associated coverage data to the fragment / pixel processing unit 524. The fragment / pixel processing unit 524 is a programmable execution unit configured to execute a fragment shader program or a pixel shader program. The fragment / pixel processing unit 524 transforms fragments or pixels received from the rasterizer 522 as specified by the fragment shader program or the pixel shader program. For example, the fragment / pixel processing unit 524 may be programmed to perform operations including, but not limited to, texture mapping, shading, blending, texture correction, and perspective correction to generate shaded fragments or pixels that are output to the raster operation unit 526. The fragment / pixel processing unit 524 may read data stored in either the parallel processor memory or the system memory for use in processing the fragment data. The fragment shader program or pixel shader program may be configured to shade at a sample, pixel, tile, or other granularity, depending on the sampling rate set for the processing unit.

[0156] Raster operations unit 526 is a processing unit that performs raster operations including, but not limited to, stencil, z-test, and blending, and outputs pixel data as processed graphics data to be stored in a graphics memory (e.g., parallel processor memory 222 as in FIG. 2A and / or system memory 104 as in FIG. 1), to be displayed on one or more display devices 110, or for further processing by one or more of processors 102 or parallel processors 112. Raster operations unit 526 may be configured to compress z data or color data written to memory and to decompress z data or color data read from memory. [Machine Learning Overview]

[0157] The above architectures may be applied to perform training and inference operations with machine learning models. Machine learning has been successful in solving many types of tasks. The computations that occur when training and using machine learning algorithms (e.g., neural networks) are inherently suitable for efficient parallel implementation. Thus, parallel processors such as general-purpose graphics processing units (GPGPUs) have played a key role in practical implementations of deep neural networks. Parallel graphics processors with single instruction multiple threads (SIMT) architectures are designed to maximize the amount of parallelism in the graphics pipeline. In SIMT architectures, a group of parallel threads synchronously executes program instructions together as often as possible in an attempt to increase processing efficiency. The efficiency provided by parallel machine learning algorithm implementations allows the use of high-capacity networks and allows these networks to be trained on large datasets.

[0158] Machine learning algorithms are algorithms that can learn based on a set of data. For example, machine learning algorithms may be designed to model higher-level abstractions within a dataset. For example, image recognition algorithms can be used to determine which of several categories a given input belongs to, regression algorithms can output a numerical value given an input, and pattern recognition algorithms can be used to generate converted text or perform text-to-speech and / or speech recognition.

[0159] An exemplary type of machine learning algorithm is a neural network. Among the many types of neural networks, a simple type of neural network is a feedforward network. A feedforward network may be implemented as an acyclic graph with nodes arranged in layers. Typically, a feedforward network topology includes an input layer and an output layer separated by at least one hidden layer. The hidden layer converts the inputs received by the input layer into a representation that helps generate the outputs at the output layer. The network nodes are fully connected via edges to the nodes of adjacent layers, but there are no edges between the nodes within each layer. Data received at the nodes of the input layer of a feedforward network is propagated (i.e., "feeded forward") to the nodes of the output layer through an activation function that calculates the state of the nodes of each subsequent layer in the network based on the coefficients ("weights") associated with each edge connecting each layer. Depending on the particular model represented by the algorithm being executed, the output from a neural network algorithm can take a variety of forms.

[0160] Before a machine learning algorithm can be used to model a particular problem, the algorithm is trained using a training dataset. Training a neural network involves selecting a network topology, using a set of training data representative of the problem to be modeled by the network, and adjusting the weights until the network model performs with minimal error for all instances in the training dataset. For example, in a supervised learning training process for a neural network, outputs generated by the network in response to inputs representing instances in the training dataset are compared with "ground-truth" labeled outputs for those instances, an error signal representing the difference between the outputs and the labeled outputs is calculated, and the weights associated with the connections are adjusted to minimize the error as the error signal is back-propagated through each layer of the network. The network is considered "trained" when the error for each of the outputs generated from each instance in the training dataset is minimized.

[0161] The accuracy of machine learning algorithms can be greatly affected by the quality of the data set used to train the algorithm. The training process can be computationally intensive and can require a lot of time on a conventional general-purpose processor. Therefore, parallel processing hardware is used to train many types of machine learning algorithms. This is particularly useful for optimizing the training of neural networks, since the calculations performed in adjusting coefficients in neural networks are naturally suited to parallel implementation. In particular, many machine learning algorithms and software applications are adapted to take advantage of parallel processing hardware in general-purpose graphics processing devices.

[0162] 6 is a generalized diagram of a machine learning software stack 600. A machine learning application 602 may be configured to train a neural network with a training dataset or to implement machine intelligence using a trained deep neural network. The machine learning application 602 may include training and inference functions for the neural network and / or specialized software that may be used to train the neural network before it is deployed. The machine learning application 602 may implement any type of machine intelligence, including, but not limited to, image recognition, mapping and localization, autonomous navigation, speech synthesis, medical imaging, or language conversion.

[0163] Hardware acceleration of the machine learning application 602 may be enabled by the machine learning framework 604. The machine learning framework 604 may provide a library of machine learning primitives. Machine learning primitives are basic operations commonly performed by machine learning algorithms. Without the machine learning framework 604, developers of machine learning algorithms would need to create and optimize the core computational logic associated with the machine learning algorithm and reoptimize this computational logic as new parallel processors are developed. Instead, machine learning applications may be configured to perform the necessary computations using primitives provided by the machine learning framework 604. Exemplary primitives include tensor convolutions, activation functions, and pooling, which are computational operations performed when training a convolutional neural network (CNN). The machine learning framework 604 may also provide primitives for implementing basic linear algebra subprograms, such as matrix and vector operations, performed by many machine learning algorithms.

[0164] The machine learning framework 604 can process input data received from the machine learning application 602 and generate appropriate inputs to the compute framework 606. The compute framework 606 can abstract the underlying instructions provided to the GPGPU driver 608 to enable the machine learning framework 604 to leverage hardware acceleration via the GPGPU hardware 610 without the machine learning framework 604 needing to have detailed knowledge of the architecture of the GPGPU hardware 610. Furthermore, the compute framework 606 can enable hardware acceleration of the machine learning framework 604 across different types and generations of GPGPU hardware 610. [GPGPU Machine Learning Acceleration]

[0165] Figure 7 illustrates a general purpose graphics processing unit 700, which may be parallel processor 200 of Figure 2A or parallel processor 112 of Figure 1. General purpose processing unit (GPGPU) 700 may be configured to be particularly efficient at handling the types of computational workloads associated with training deep neural networks. Furthermore, GPGPU 700 may be directly connected to other instances of GPGPU to create a multi-GPU cluster to specifically improve the training speed of deep neural networks.

[0166] The GPGPU 700 includes a host interface 702 that allows for connection to a host processor. The host interface 702 may be a PCI Express interface. However, the host interface may also be a vendor-specific communication interface or fabric. The GPGPU 700 receives commands from the host processor and uses a global scheduler 704 to distribute execution threads associated with those commands to a set of processing clusters 706A-706H. The processing clusters 706A-706H share a cache memory 708. The cache memory 708 may act as a higher-level cache for the cache memories in the processing clusters 706A-706H. The illustrated processing clusters 706A-706H may correspond to the processing clusters 214A-214N as in FIG. 2A.

[0167] The GPGPU 700 includes memories 714A-B coupled to the processing clusters 706A-H via a set of memory controllers 712A-B. The memories 714A-B may include various types of memory devices including dynamic random access memory (DRAM) or graphics random access memory, such as synchronous graphics random access memory (SGRAM), including graphics double data rate (GDDR) memory. The memories 714A-B may also include 3D stacked memory, including but not limited to high bandwidth memory (HBM).

[0168] Each of the processing clusters 706A-H may include a set of graphics multiprocessors, such as the graphics multiprocessor 234 of FIG. 2D, the graphics multiprocessor 325 of FIG. 3A, the graphics multiprocessor 350 of FIG. 3B, or may include multicore groups 365A-N as in FIG. 3C. The graphics multiprocessors of the compute cluster include multiple types of integer and floating point logic units capable of performing computational operations with a range of precision suitable for machine learning computations. For example, at least a subset of the floating point units 264 in each of the processing clusters 706A-H may be configured to perform 16-bit or 32-bit floating point operations, and another subset of the floating point units 264 may be configured to perform 64-bit floating point operations.

[0169] Multiple instances of GPGPU 700 may be configured to operate as a compute cluster. The communication mechanisms used by the compute cluster for synchronization and data exchange vary depending on the embodiment. For example, multiple instances of GPGPU 700 communicate over host interface 702. In one embodiment, GPGPU 700 includes an I / O hub 709 that couples GPGPU 700 to GPU link 710 that allows direct connection to other instances of GPGPU. GPU link 710 may be coupled to a dedicated GPU-to-GPU bridge that allows communication and synchronization between multiple instances of GPGPU 700. If necessary, GPU link 710 couples to other GPGPUs or parallel processors using a high-speed interconnect to send and receive data. Multiple instances of GPGPU 700 may be located in separate data processing systems and communicate over a network device accessible via host interface 702. GPU link 710 may be configured to allow connection to a host processor in addition to or as an alternative to host interface 702.

[0170] While the illustrated configuration of GPGPU 700 may be configured to train a neural network, alternative configurations of GPGPU 700 may be configured for deployment on high performance or low power inference platforms. In an inference configuration, GPGPU 700 includes a smaller number of processing clusters 706A-706H as compared to the training configuration. Additionally, the memory technology associated with memories 714A-714B may differ between the inference and training configurations. In one embodiment, the inference configuration of GPGPU 700 may support inference of certain instructions. For example, an inference configuration may provide support for one or more 8-bit integer dot product instructions, which are commonly used during inference operations of deployed neural networks.

[0171] FIG. 8 illustrates a multi-GPU computing system 800. The multi-GPU computing system 800 may include a processor 802 coupled to a plurality of GPGPUs 806A-806D via a host interface switch 804. The host interface switch 804 may be a PCI Express switch device that couples the processor 802 to a PCI Express bus by which the processor 802 can communicate with the set of GPGPUs 806A-806D. Each of the plurality of GPGPUs 806A-806D may be an instance of the GPGPU 700 of FIG. 7. The GPGPUs 806A-806D may be interconnected via a set of high-speed point-to-point inter-GPU links 816. The high-speed inter-GPU links may be connected to each of the GPGPUs 806A-806D via a dedicated GPU link, such as the GPU link 710 in FIG. 7. The P2P GPU link 816 allows direct communication between each of the GPGPUs 806A-806D without requiring communication over the host interface bus to which the processor 802 is connected. With inter-GPU traffic directed over the P2P GPU link, the host interface bus is still available for system memory access or for communication with other instances of the multi-GPU computing system 800, e.g., via one or more network devices. While FIG. 8 shows GPGPUs 806A-D connecting to the processor 802 through the host interface switch 804, the processor 802 may alternatively include direct support for the P2P GPU link 816 and connect directly to the GPGPUs 806A-806D. [Implementation of machine learning neural network]

[0172] The computer processing architecture described herein may be configured to perform a type of parallel processing that is particularly suitable for training and deploying neural networks for machine learning. A neural network may be generalized as a network of functions with graph relationships. As is well known in the art, there are various types of implementations of neural networks used for machine learning. One exemplary type of neural network is the feedforward network, as described above.

[0173] Another exemplary type of neural network is the Convolutional Neural Network (CNN). CNNs are feed-forward neural networks specialized for processing data, such as image data, that have a known grid-like topology. Thus, CNNs are often used in computer vision and image recognition applications, but may also be used for other types of pattern recognition, such as speech and language processing. Each node in the CNN input layer is organized into a set of "filters" (feature detectors inspired by the receptive fields found in the retina), and the output of each set of filters is propagated to the nodes of each subsequent layer of the network. CNN computation involves applying a convolution mathematical operation to each filter to generate the output of that filter. A convolution is a special type of mathematical operation performed by two functions to generate a third function, where the third function is a modified version of one of the two original functions. In convolutional network terminology, the first function for a convolution may be called the input, and the second function may be called the convolution kernel. The output may be called a feature map. For example, the input to a convolutional layer may be a multidimensional array of data defining the various color components of the input image, and the convolution kernel may be a multidimensional array of parameters, which are adapted by a neural network training process.

[0174] Recurrent neural networks (RNNs) are a family of feedforward neural networks that contain feedback connections between layers. RNNs allow for the modeling of sequential data by sharing parameter data across different parts of the neural network. The architecture of an RNN contains cycles. At least a portion of the output data from the RNN is used as feedback to process the next input in the sequence, so the cycle represents the effect that the current value of a variable has on its own value at a future time. This feature makes RNNs particularly useful for language processing due to the variability in which language data can be constructed.

[0175] The figures described below show exemplary feedforward, CNN, and RNN networks, along with general processes for training and deploying each type of network, respectively. It will be understood that these descriptions are exemplary and non-limiting with respect to any particular embodiment described herein, and that the concepts shown may be generally applied to deep neural networks and machine learning techniques in general.

[0176] The above-mentioned exemplary neural network may be used to perform deep learning. Deep learning is machine learning using deep neural networks. The deep neural networks used in deep learning are artificial neural networks that are composed of multiple hidden layers, and are different from shallow neural networks that only contain one hidden layer. The deeper the neural network, the more computationally intensive the training generally becomes. However, the additional hidden layers of the network enable multi-stage pattern recognition, which results in a reduction in output error compared to shallow machine learning techniques.

[0177] A deep neural network used in deep learning typically includes a front-end network for feature recognition coupled to a back-end network representing a mathematical model that can perform multiple operations (e.g., object classification, speech recognition, etc.) based on feature representations provided to the model. Deep learning allows machine learning to be performed without requiring hand-crafted feature processing on the model. Instead, deep neural networks can learn features based on statistical structures or correlations in the input data. The learned features may be provided to a mathematical model that can map the detected features to an output. The mathematical model used by the network is generally specialized for the particular task being performed, and different models will be used to perform different tasks.

[0178] Once a neural network is constructed, a learning model may be applied to the network to train the network to perform a particular task. The learning model indicates how to adjust the weights in the model to reduce the network's output error. Backpropagation of error is a common method used to train neural networks. An input vector is provided to the network for processing. The output of the network is compared to the desired output using a loss function, and an error value is calculated for each of multiple neurons in the output layer. The error values ​​are then propagated backwards until each neuron has a correlated error value that approximately represents its contribution to the original output. The network may then learn from these errors using an algorithm such as a stochastic gradient descent algorithm to update the neural network weights.

[0179] 9A-9B show an exemplary convolutional neural network. FIG. 9A shows various layers in a CNN. As shown in FIG. 9A, an exemplary CNN used to model image processing may receive an input 902 representing red, green, and blue (RGB) components of an input image. The input 902 may be processed by multiple convolutional layers (e.g., convolutional layer 904, convolutional layer 906). The output from the multiple convolutional layers may be processed by a set of fully connected layers 908, if necessary. As described above for feedforward networks, neurons in a fully connected layer are fully connected to all activations of the previous layer. The output from the fully connected layer 908 may be used to generate an output result from the network. The activations in the fully connected layer 908 may be calculated using matrix multiplication instead of convolution. Not all CNN implementations utilize a fully connected layer 908. For example, in some implementations, the convolutional layer 906 may generate the output of the CNN.

[0180] Convolutional layers are sparsely connected, unlike typical neural network configurations found in the fully connected layer 908. Typical neural network layers are fully connected, such that all output units communicate with all input units. However, as shown, convolutional layers are sparsely connected because the outputs of the convolutions in the field are input to the nodes in the next layer (instead of the respective state values ​​of each node in the field). The kernels associated with the convolutional layers perform the convolution operations, and the outputs are sent to the next layer. The dimensionality reduction that occurs in the convolutional layers is one aspect that allows CNNs to scale to processing large images.

[0181] 9B illustrates exemplary computational stages within a convolutional layer of a CNN. An input 912 to a convolutional layer of a CNN may be processed in three stages in a convolutional layer 914. The three stages may include a convolution stage 916, a detector stage 918, and a pooling stage 920. The convolutional layer 914 may then output data to a subsequent convolutional layer. The final convolutional layer of the network may generate output feature map data or provide an input to a fully connected layer to, for example, generate a classification value for the input to the CNN.

[0182] The convolution stage 916 performs several convolutions in parallel to generate a set of linear activations. The convolution stage 916 may include an affine transformation, which is any transformation that can be specified as a linear transformation plus a translation. Affine transformations include rotation, translation, scaling, and combinations of these transformations. The convolution stage calculates the output of a function (e.g., a neuron) that is connected to a particular region of the input. This region may be determined as the local region associated with the neuron. The neuron calculates the dot product of the neuron's weight and the region in the local input to which the neuron is connected. The output from the convolution stage 916 defines a set of linear activations that are processed by each subsequent stage of the convolution layer 914.

[0183] The linear activations may be processed by a detector stage 918, where each linear activation is processed with a nonlinear activation function. The nonlinear activation function improves the nonlinear properties of the entire network without affecting the receptive fields of the convolutional layers. Several types of nonlinear activation functions may be used. One particular type is the rectified linear unit (ReLU), which uses an activation function defined as f(x)=max(0,x), resulting in dividing the activations by zero.

[0184] The pooling stage 920 uses a pooling function that replaces the output of the convolutional layer 906 with a summary statistic of nearby outputs. The pooling function may be used to introduce translation invariance into the neural network so that small translations to the input do not change the pooled output. Invariance to local translations may be useful in scenarios where the presence of a feature in the input data is more important than the exact location of the feature. Various types of pooling functions may be used in the pooling stage 920, including max pooling, average pooling, and L2 norm pooling. Furthermore, some CNN implementations do not include a pooling stage. Instead, such implementations substitute another convolution stage with an increased stride compared to previous convolution stages.

[0185] The output from the convolutional layer 914 may then be processed by the next layer 922, which may be another convolutional layer or one of the fully connected layers 908. For example, the first convolutional layer 904 in FIG. 9A may output to the second convolutional layer 906, which may output to the first of the fully connected layers 908.

[0186] FIG. 10 illustrates an exemplary recurrent neural network 1000. In a recurrent neural network (RNN), the previous state of the network influences the output of the current state of the network. RNNs may be constructed in a variety of ways, using a variety of functions. The use of RNNs generally revolves around using mathematical models to predict the future based on a previous sequence of inputs. For example, an RNN may be used to implement a statistical language model that predicts the next word given a previous sequence of words. The illustrated RNN 1000 may be described as having an input layer 1002 that receives input vectors, a hidden layer 1004 that implements a recurrent function, a feedback mechanism 1005 that allows for "memory" of the previous state, and an output layer 1006 that outputs the result. The RNN 1000 operates based on time steps. The state of the RNN at a given time step is influenced via the feedback mechanism 1005 based on the previous time step. At a given time step, the state of the hidden layer 1004 is determined by the previous state and the input at the current time step. An initial input (X1) at a first time step may be processed by the hidden layer 1004. A second input (X2) may be processed by the hidden layer 1004 using state information determined when processing the initial input (X1). A given state is s t =f(Ux t +Ws t-1 ), where U and W are parameter matrices. The function f is typically a nonlinearity, such as a variant of the hyperbolic tangent function (Tanh) or the normalization function f(x)=max(0,x). However, the particular mathematical functions used in the hidden layer 1004 may vary depending on the details of the particular implementation of the RNN 1000.

[0187] In addition to the basic CNN- and RNN-type networks described, variations on these networks may be possible. One exemplary RNN variation is the long short-term memory (LSTM) RNN. LSTM RNNs can learn long-term dependencies that may be necessary to process long sequences of language. A variation of CNN is the convolutional deep belief network, which has a similar structure to CNNs and is trained in a similar manner to deep belief networks. A deep belief network (DBN) is a generative adversarial neural network composed of multiple layers of random (random) variables. DBNs may be trained layer-by-layer using greedy unsupervised learning. The learned weights of the DBN may then be used to provide a pre-trained neural network by determining an optimal initial set of weights for the neural network.

[0188] FIG. 11 illustrates the training and deployment of deep neural networks. Once a given network is constructed for a task, the neural network is trained using a training dataset 1102. Various training frameworks 1104 have been developed to enable hardware acceleration of the training process. For example, the machine learning framework 604 of FIG. 6 may be configured as the training framework 604. The training framework 604 may connect to a pre-trained neural network 1106 and enable training of the pre-trained neural net using parallel processing resources as described herein to generate a trained neural net 1108.

[0189] To begin the training process, initial weights may be selected randomly or by pre-training with a deep belief network. Training cycles may then be performed in either a supervised or unsupervised manner.

[0190] Supervised learning is a learning method in which training is performed as an intermediate operation, such as when a training data set 1102 includes inputs combined with desired outputs for the inputs, or when a training data set includes inputs with known outputs and the neural network's output is manually graded. The network processes the inputs and compares the resulting output to a set of expected or desired outputs. Errors are then propagated backwards through the system. The training framework 1104 can be adjusted to adjust the weights that control the pre-trained neural network 1106. The training framework 1104 can provide tools to monitor how well the pre-trained neural network 1106 is converging toward a model suitable for generating correct answers based on known input data. Once the network weights have been adjusted to improve the output generated by the neural network, the training process is repeated. The training process may continue until the neural network reaches a desired statistical accuracy associated with the trained neural net 1108. The trained neural network 1108 may then be deployed to implement any number of machine learning operations that generate inference results 1114 based on the input of new data 1112.

[0191] Unsupervised learning is a learning method in which a network attempts to train itself using unlabeled data. Thus, in the case of unsupervised learning, the training data set 1102 will contain input data without any associated output data. The pre-trained neural network 1106 can learn groupings of unlabeled inputs and determine how each input is related to the entire data set. Unsupervised training can be used to generate self-organizing maps, which are a type of trained neural network 1108 that can perform operations that are useful in reducing the dimensionality of the data. Unsupervised training can also be used to perform anomaly detection, which allows for the identification of data points in an input data set that fall outside of the normal pattern of the data.

[0192] Variations of supervised and unsupervised training may also be used. Semi-supervised learning is a technique in which the training dataset 1102 contains a combination of the same distribution of labeled and unlabeled data. Incremental learning is a variation of supervised learning in which input data is continually used to further train the model. Incremental learning allows the trained neural network 1108 to adapt to new data 1112 without forgetting the knowledge imparted to the network during initial training.

[0193] The training process, especially of deep neural networks, whether supervised or unsupervised, may be too computationally intensive for a single compute node. Instead of using a single compute node, a distributed network of compute nodes can be used to accelerate the training process.

[0194] 12 is a block diagram illustrating distributed learning, a training model that uses multiple distributed computer processing nodes for supervised or unsupervised training of neural networks. Each distributed computing node may include one or more host processors and one or more of multiple general purpose processing nodes, such as the highly parallel general purpose graphics processing unit 700 in FIG. 7. As shown, distributed learning may be performed with model parallelism 1202, data parallelism 1204, or a combination of model and data parallelism 1206.

[0195] In model parallelism 1202, different computational nodes in a distributed system can perform training calculations for different parts of a single network. For example, each layer of a neural network may be trained by a different processing node of the distributed system. Advantages of model parallelism include the ability to scale, especially to large models. Splitting up the multiple computations associated with different layers of a neural network enables the training of very large neural networks where not all layer weights fit into the memory of a single computational node. In some examples, model parallelism can be particularly useful when performing unsupervised training of large neural networks.

[0196] In data parallelism 1204, each node of the distributed network has a complete instance of the model, with each node receiving a different portion of the data. The results from the individual nodes are then combined. Data parallelism can be done in a variety of ways, but all data parallel training approaches require a technique to combine the results and synchronize the model parameters between each node. Exemplary approaches for combining data include parameter averaging and update-based data parallelism. Parameter averaging trains each node on a subset of the training data and sets the global parameters (e.g., weights, biases) to the average of the parameters from each node. Parameter averaging uses a central parameter server that maintains the parameter data. Update-based data parallelism is similar to parameter averaging, except that instead of transferring parameters from each node to the parameter server, modeling updates are transferred. Additionally, update-based data parallelism may be done in a distributed manner, where updates are compressed and transferred between nodes.

[0197] Concurrent parallel processing of models and data 1206 may be performed, for example, in a distributed system where each computing node contains multiple GPUs. Each node may have a complete instance of the model, and independent GPUs within each node are used to train different parts of the model.

[0198] Distributed training incurs increased overhead compared to training on a single machine. However, the parallel processors and GPGPUs described herein can each implement various techniques to reduce the overhead of distributed training, including techniques that enable high-bandwidth inter-GPU data transfers and accelerated remote data synchronization. [Example machine learning application]

[0199] Machine learning can be applied to solve a variety of technical problems, including, but not limited to, computer vision, autonomous driving, navigation, speech recognition, and language processing. Computer vision has been one of the most active research areas of machine learning applications to date. Computer vision applications range from replicating human visual capabilities, such as face recognition, to creating new categories of visual capabilities. For example, a computer vision application may be configured to recognize sound waves from vibrations induced in a visible object in a video. Accelerated machine learning with parallel processors allows computer vision applications to be trained with significantly larger training datasets than previously feasible, and allows inference systems to be deployed using low-power parallel processors.

[0200] Accelerated machine learning with parallel processors has automated driving applications including lane and road sign recognition, obstacle avoidance, navigation, and driving control. Accelerated machine learning techniques may be used to train driving models based on data sets that define appropriate responses to specific training inputs. The parallel processors described herein may enable rapid training of increasingly complex neural networks used in automated driving solutions and may enable deployment of low-power inference processors on mobile platforms suitable for incorporation into autonomous vehicles.

[0201] Parallel processor accelerated deep neural networks have enabled machine learning approaches to automatic speech recognition (ASR). ASR involves the creation of a function that computes the most likely language sequence given an input acoustic sequence. Accelerated machine learning using deep neural networks has enabled the replacement of hidden Markov models (HMMs) and Gaussian mixture models (GMMs) previously used for ASR.

[0202] Parallel processor-accelerated machine learning may also be used to accelerate natural language processing. Automated learning procedures can utilize statistical inference algorithms to generate models that are robust to erroneous or unknown inputs. Exemplary natural language processor applications include automatic machine translation between human languages.

[0203] Parallel processing platforms used for machine learning may be divided into training platforms and deployment platforms. Training platforms are generally highly parallelized and include optimizations to accelerate multi-GPU single-node training and multi-node multi-GPU training. Exemplary parallel processors suitable for training include the general-purpose graphics processing unit 700 of FIG. 7 and the multi-GPU computer processing system 800 of FIG. 8. On the other hand, deployment machine learning platforms generally include low-power parallel processors suitable for use in products such as cameras, autonomous robots, and autonomous vehicles.

[0204] FIG. 13 illustrates an exemplary inference system-on-chip (SoC) 1300 suitable for performing inference using trained models. The SoC 1300 may integrate processing components including a media processor 1302, a vision processor 1304, a GPGPU 1306, and a multi-core processor 1308. The GPGPU 1306 may be a GPGPU such as the GPGPU 700 described herein, and the multi-core processor 1308 may be a multi-core processor described herein such as the multi-core processors 405-406. The SoC 1300 may further include an on-chip memory 1305 that may enable a shared on-chip data pool accessible by each of these processing components. Each processing component may be optimized for low-power operation to enable deployment in a variety of machine learning platforms, including autonomous vehicles and autonomous robots. For example, one implementation of the SoC 1300 may be used as part of the main control system of an autonomous vehicle. When the SoC1300 is configured for use in an autonomous vehicle, the SoC is designed and constructed to comply with the relevant functional safety standards of the deployment jurisdiction.

[0205] In operation, the media processor 1302 and the vision processor 1304 can work in conjunction to accelerate computer vision operations. The media processor 1302 can enable low latency decoding of multiple high resolution (e.g., 4K, 8K) video streams. The decoded video streams can be written to a buffer in the on-chip memory 1305. The vision processor 1304 can then analyze the decoded video and perform pre-processing operations on frames of the decoded video in preparation for processing the frames using a trained image recognition model. For example, the vision processor 1304 can accelerate convolution operations for CNNs used to perform image recognition on high resolution video data, with back-end model computations being performed by the GPGPU 1306.

[0206] The multicore processor 1308 may include control logic that assists in sequencing and synchronization of data transfer and shared memory operations performed by the media processor 1302 and the vision processor 1304. The multicore processor 1308 may also function as an application processor that executes software applications that can utilize the inference computational power of the GPGPU 1306. For example, at least a portion of the navigation and driving logic may be implemented in software that executes on the multicore processor 1308. Such software may issue computational workloads directly to the GPGPU 1306, or the computational workloads may be issued to the multicore processor 1308. The multicore processor may offload at least a portion of these operations to the GPGPU 1306.

[0207] GPGPU 1306 may include a compute cluster, such as processing clusters 706A-706H in a low-power configuration in general-purpose graphics processing unit 700. The compute clusters in GPGPU 1306 may support instructions specifically optimized to perform inference calculations on trained neural networks. For example, GPGPU 1306 may support instructions to perform low-precision calculations, such as 8-bit and 4-bit integer vector operations. [Another system overview]

[0208] FIG. 14 is a block diagram of a processing system 1400. Elements of FIG. 14 having the same or similar names as elements of any other figures herein refer to the same elements as in the other figures, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but are not limited to such. System 1400 may be used in a single processor desktop system, a multiprocessor workstation system, or a server system having multiple processors 1402 or processor cores 1407. System 1400 may be a processing platform incorporated in a system-on-chip (SoC) integrated circuit used in a mobile device, a handheld device, or an embedded device, such as an "Internet of Things" (IoT) device with wired or wireless connectivity to a local or wide area network.

[0209] System 1400 may be a processing system having components corresponding to those of Figure 1. For example, in a different configuration, processor 1402 or processor core 1407 may correspond to processor 102 of Figure 1. Graphics processor 1408 may correspond to parallel processor 112 of Figure 1. External graphics processor 1418 may be one of add-in devices 120 of Figure 1.

[0210] System 1400 may include, be coupled to, or be integrated with a server-based gaming platform, a gaming console, including a game and media console, a portable gaming console, a handheld gaming console, or an online gaming console. System 1400 may be part of a portable Internet-connected device, such as a mobile phone, a smart phone, a tablet computing device, or a laptop with low internal storage capacity. Processing system 1400 may also include, be coupled to, or be integrated with a wearable device, such as a smart watch wearable device, smart eyewear or clothing that provides visual, audio, or haptic output to supplement a real-world visual, audio, or haptic experience, or that is otherwise augmented with augmented reality (AR) or virtual reality (VR) capabilities that provide text, audio, graphics, video, holographic or holographic images, or haptic feedback, other augmented reality (AR) devices, or other virtual reality (VR) devices. Processing system 1400 may include, be part of, or be a part of a television or set-top box device. System 1400 may be included in, coupled with, or integrated into an autonomous vehicle, such as a bus, a tractor trailer, an automobile, a motorized or electric motorcycle, an airplane or glider (or any combination thereof), etc. The autonomous vehicle may use system 1400 to process the sensed environment around the vehicle.

[0211] The one or more processors 1402 may include one or more processor cores 1407 that process instructions that, when executed, perform operations of the system or user software. The smallest of the one or more processor cores 1407 may be configured to process a particular instruction set 1409. The instruction set 1409 may facilitate complex instruction set computing (CISC), reduced instruction set computing (RISC), or very long instruction word (VLIW) computing. The one or more processor cores 1407 may process another instruction set 1409, which may include instructions that facilitate emulation of other instruction sets. The processor cores 1407 may also include other processing devices, such as digital signal processors (DSPs).

[0212] The processor 1402 may include cache memory 1404. Depending on the architecture, the processor 1402 may have a single internal cache or multiple levels of internal cache. In some embodiments, the cache memory is shared among various components of the processor 1402. In some embodiments, the processor 1402 also uses an external cache (e.g., a level 3 (L3) cache or a last level cache (LLC)) (not shown). This external cache may be shared among multiple processor cores 1407 using known cache coherency techniques. A register file 1406 may also be included in the processor 1402, which may include different types of registers (e.g., integer registers, floating point registers, status registers, and instruction pointer registers) for storing different types of data. Some registers may be general purpose registers and other registers may be specific to the processor 1402 architecture.

[0213] The one or more processors 1402 may be coupled to one or more interface buses 1410 that carry communication signals, such as address, data, or control signals, between the processors 1402 and other components of the system 1400. In one of these embodiments, the interface bus 1410 may be a processor bus, such as a version of a Direct Media Interface (DMI) bus. However, the processor bus is not limited to a DMI bus and may include one or more peripheral component interconnect buses (e.g., PCI, PCI Express), memory buses, or other types of interface buses. For example, the processor 1402 may include an integrated memory controller 1416 and a platform controller hub 1430. The memory controller 1416 facilitates communication between memory devices and other components of the system 1400, and the platform controller hub (PCH) 1430 provides connectivity to I / O devices via a local I / O bus.

[0214] The memory device 1420 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, a flash memory device, a phase change memory device, or any other memory device with suitable performance to act as a process memory. The memory device 1420 may, for example, act as a system memory for the system 1400 and store data 1422 and instructions 1421 for use by the one or more processors 1402 in executing applications or processes. The memory controller 1416 may also couple to an optional external graphics processor 1418, which may communicate with the one or more graphics processors 1408 of the processor 1402 to perform graphics and media operations. In some embodiments, graphics, media, and / or compute operations may be assisted by an accelerator 1412, which is a co-processor that may be configured to perform a specialized set of graphics, media, or compute operations. For example, accelerator 1412 may be a matrix multiplication accelerator used to optimize machine learning or compute operations. Accelerator 1412 may be a ray tracing accelerator that may be used to perform ray tracing operations in conjunction with graphics processor 1408. In one embodiment, an external accelerator 1419 may be used in place of or in conjunction with accelerator 1412.

[0215] A display device 1411 may be provided that may be connected to the processor 1402. The display device 1411 may be one or both of an internal display device, such as in a portable electronic device or laptop device, or an external display device attached via a display interface (e.g., DisplayPort, etc.). The display device 1411 may be a head mounted display (HMD), such as a stereoscopic display device for use in virtual reality (VR) or augmented reality (AR) applications.

[0216] The platform controller hub 1430 may enable peripherals to connect to the memory device 1420 and the processor 1402 via a high-speed I / O bus. The I / O peripherals include, but are not limited to, an audio controller 1446, a network controller 1434, a firmware interface 1428, a wireless transceiver 1426, a touch sensor 1425, and a data storage device 1424 (e.g., non-volatile memory, volatile memory, hard disk drive, flash memory, NAND, 3D NAND, 3D XPoint, etc.). The data storage device 1424 can be connected via a storage interface (e.g., SATA) or via a peripheral bus such as a peripheral component interconnect bus (e.g., PCI, PCI Express). The touch sensor 1425 may include a touch screen sensor, a pressure sensor, or a fingerprint sensor. The wireless transceiver 1426 may be a Wi-Fi transceiver, a Bluetooth transceiver, or a mobile network transceiver such as a 3G, 4G, 5G, or Long Term Evolution (LTE) transceiver. Firmware interface 1428 allows for communication with system firmware and may be, for example, a Unified Extensible Firmware Interface (UEFI). Network controller 1434 may allow for network connectivity to a wired network. In some embodiments, a high performance network controller (not shown) couples to interface bus 1410. Audio controller 1446 may be a multi-channel high definition audio controller. In some of these embodiments, system 1400 includes an optional legacy I / O controller 1440 for coupling legacy (e.g., Personal System / 2 (PS / 2)) devices to the system. Platform controller hub 1430 may also connect to one or more Universal Serial Bus (USB) controllers 1442 for connecting input devices, such as a keyboard and mouse 1443 combination, a camera 1444, or other USB input devices.

[0217] It will be understood that the system 1400 shown is illustrative and not limiting, as other types of data processing systems configured differently may also be used. For example, instances of the memory controller 1416 and platform controller hub 1430 may be integrated into a separate external graphics processor, such as the external graphics processor 1418. The platform controller hub 1430 and / or memory controller 1416 may be external to one or more processors 1402. For example, the system 1400 may include an external memory controller 1416 and platform controller hub 1430, which may be configured as a memory controller hub and a peripheral controller hub in a system chipset that communicates with the processor 1402.

[0218] For example, a circuit board ("sled") on which components such as CPUs, memory, and other components are located may be used and designed to enhance thermal performance. Processing components such as processors may be located on the front side of the sled, and near memory such as DIMMs are located on the back side of the sled. As a result of the improved airflow provided in this manner, the components can operate at higher frequencies and power levels than those of a typical system, thereby improving performance. Furthermore, the sleds are configured to easily couple with power and data communication cables in the rack, improving the ability to quickly remove, upgrade, reinstall, and / or replace the sleds. Similarly, the individual components located on the sleds, such as processors, accelerators, memory, and data storage drives, are configured to be easily upgraded by increasing the spacing between each other. In an exemplary embodiment, these components further include hardware authentication features to attest to their authenticity.

[0219] Data centers may utilize a single network architecture ("fabric") that supports multiple other network architectures, including Ethernet and Omnipath. The threads may be coupled to switches via optical fiber, which provides higher bandwidth and lower latency than typical twisted pair cables (e.g., Cat 5, Cat 5e, Cat 6, etc.). The high-bandwidth, low-latency interconnect and network architecture may enable data centers to pool resources, such as memory, accelerators (e.g., GPUs, graphics accelerators, FPGAs, ASICs, neural networks, and / or artificial intelligence accelerators, etc.), and data storage drives that are physically disaggregated, when in use, and provide these resources to compute resources (e.g., processors) as needed, allowing the compute resources to access these resources as if the pooled resources were local.

[0220] A power supply or source may provide voltage and / or current to system 1400 or any component or system described herein. In one example, the power supply includes an AC / DC (alternating current / direct current) adapter that plugs into a wall outlet. Such an AC source may be a renewable energy (e.g., solar power) source. In one example, the source includes a DC source, such as an external AC / DC converter. The source or power supply may also include wireless charging hardware that charges by proximity to a charging field. The source may include an internal battery, an AC source, a motion-based source, a solar source, or a fuel cell source.

[0221] Figures 15A-15C show a computer processing system and a graphics processor. Elements in Figures 15A-15C that have the same or similar names as elements in any other figure herein refer to the same element as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but are not limited to such.

[0222] FIG. 15A is a block diagram of a processor 1500, which may be a variation of one of the processors 1402 and may be used in place of one of the processors 1402. Thus, disclosure of any feature in combination with the processor 1500 herein also discloses the corresponding combination with the processor 1402, but is not limited to such. The processor 1500 may have one or more processor cores 1502A-1502N, an integrated memory controller 1514, and an integrated graphics processor 1508. With the exception of the integrated graphics processor 1508, a system including the processor would include a graphics processor device in a system chipset or coupled via a system bus. The processor 1500 may include additional cores up to and including the additional core 1502N, which is represented by a dashed box. Each of the processor cores 1502A-1502N includes one or more built-in cache units 1504A-1504N. In some embodiments, each processor core 1502A-1502N may also have access to one or more shared cache units 1506. The internal cache units 1504A-1504N and the shared cache unit 1506 represent a cache memory hierarchy within the processor 1500. The cache memory hierarchy may include at least one level of instruction and data cache within each processor core, and one or more levels of shared mid-level cache, such as a level 2 (L2), level 3 (L3), level 4 (L4), or other level cache, with the highest level cache before an external memory classified as an LLC. In some embodiments, cache coherency logic maintains consistency between the various cache units 1506 and 1504A-1504N.

[0223] Processor 1500 may also include a set of one or more bus controller units 1516, which manage a set of peripheral buses, such as one or more PCI or PCI Express buses. System agent core 1510 provides management functions for various processor components. System agent core 1510 may include one or more integrated memory controllers 1514 that manage access to various external memory devices (not shown).

[0224] For example, one or more of the processor cores 1502A-1502N may include support for simultaneous multithreading. The system agent core 1510 includes components for coordinating and operating the cores 1502A-1502N during multithreaded processing. The system agent core 1510 may further include a power control unit (PCU), which includes logic and components for regulating the power state of the processor cores 1502A-1502N and the graphics processor 1508.

[0225] The processor 1500 may further include a graphics processor 1508 that performs graphics processing operations. In some of these embodiments, the graphics processor 1508 couples to a system agent core 1510 that includes a set of shared cache units 1506 and one or more integrated memory controllers 1514. The system agent core 1510 may also include a display controller 1511 that drives output of the graphics processor to one or more coupled displays. The display controller 1511 may be a separate module coupled to the graphics processor via at least one interconnect or may be integrated into the graphics processor 1508.

[0226] A ring-based interconnect unit 1512 may be used to couple the internal components of the processor 1500. However, alternative interconnect units may be used, such as a point-to-point interconnect, a switched interconnect, or other techniques, including techniques well known in the art. In some of these embodiments using a ring-based interconnect 1512, the graphics processor 1508 couples to the ring-based interconnect 1512 via an I / O link 1513.

[0227] The exemplary I / O link 1513 represents at least one of a number of different I / O interconnects, including on-package I / O interconnects that facilitate communication between various processor components and a high-performance embedded memory module 1518, such as an eDRAM module. If desired, each of the processor cores 1502A-1502N and the graphics processor 1508 may use the embedded memory module 1518 as a shared last-level cache.

[0228] The processor cores 1502A-1502N may be homogenous cores, for example, executing the same instruction set architecture. Alternatively, the processor cores 1502A-1502N may be heterogeneous in terms of instruction set architecture (ISA), where one or more of the processor cores 1502A-1502N execute a first instruction set and at least one of the other cores executes a subset of the first instruction set or another instruction set. The processor cores 1502A-1502N may be heterogeneous in terms of microarchitecture, where one or more cores having relatively higher power consumption are combined with one or more cores having lower power consumption. As another example, the processor cores 1502A-1502N are heterogeneous in terms of computational capabilities. Additionally, the processor 1500 may be implemented on one or more chips or as an SoC integrated circuit having the illustrated components among other components.

[0229] 15B is a block diagram of hardware logic of a graphics processor core 1519 according to some embodiments described herein. The graphics processor core 1519 may be referred to as a core slice and may be one or more graphics cores in a modular graphics processor. The graphics processor core 1519 is a typical graphics core slice, and the graphics processor described herein may include multiple graphics core slices based on a target power and performance envelope. Each graphics processor core 1519 may include a fixed function block 1530 coupled with multiple sub-cores 1521A-1521F, also referred to as sub-slices, which include general purpose logic and modular blocks of fixed function logic.

[0230] Fixed function block 1530 may include a geometry / fixed function pipeline 1531 that may be shared by all sub-cores in graphics processor core 1519, for example, in a low-performance and / or low-power graphics processor implementation. Geometry / fixed function pipeline 1531 may include a 3D fixed function pipeline (e.g., 3D pipeline 1612 as in FIG. 16A described below), a video front end unit, a thread spawner and thread dispatcher, and a unified return buffer manager that manages a unified return buffer (e.g., unified return buffer 1718 of FIG. 17 described below).

[0231] The fixed function block 1530 may also include a graphics SoC interface 1532, a graphics microcontroller 1533, and a media pipeline 1534. The graphics SoC interface 1532 provides an interface between the graphics processor core 1519 and other processor cores in the system-on-chip integrated circuit. The graphics microcontroller 1533 is a programmable sub-processor that can be configured to manage various functions of the graphics processor core 1519, including thread dispatch, scheduling, and preemption. The media pipeline 1534 (e.g., the media pipeline 1616 in Figures 16A and 17) includes logic that facilitates decoding, encoding, pre-processing, and / or post-processing of multimedia data, including image data and video data. The media pipeline 1534 performs media operations upon requests to the compute logic or sampling logic in the sub-cores 1521-1521F.

[0232] The SoC interface 1532 may enable the graphics processor core 1519 to communicate with a general-purpose application processor core (e.g., a CPU) and / or other components in the SoC (including memory hierarchy elements such as shared last-level cache memory, system RAM, and / or embedded on-chip or on-package DRAM). The SoC interface 1532 may also enable communication with fixed function devices in the SoC, such as a camera imaging pipeline, and enables use of and / or implements global memory atomics that may be shared between the graphics processor core 1519 and CPUs in the SoC. The SoC interface 1532 may also provide power management control for the graphics processor core 1519 and enable an interface between the clock domain of the graphics core 1519 and other clock domains in the SoC. If necessary, the SoC interface 1532 enables receipt of command buffers from a command streamer and a global thread dispatcher configured to provide commands and instructions to each of one or more graphics cores in the graphics processor. Commands and instructions may be dispatched to the media pipeline 1534 when media operations are to be performed, or to the geometry and fixed function pipelines (e.g., geometry and fixed function pipelines 1531, geometry and fixed function pipelines 1537) when graphics processing operations are to be performed.

[0233] The graphics microcontroller 1533 may be configured to perform various scheduling and management tasks for the graphics processor core 1519. In one configuration, the graphics microcontroller 1533 may perform graphics and / or compute workload scheduling on various graphics parallel engines, for example, in the execution unit (EU) arrays 1522A-1522F, 1524A-1524F in the sub-cores 1521A-1521F. In this workload scheduling, host software executing on a CPU core of the SoC including the graphics processor core 1519 may submit a workload to one of multiple graphics processor doorbells. The doorbell invokes a scheduling operation on the appropriate graphics engine. The scheduling operation includes determining which workload to run next, submitting the workload to the command streamer, preempting existing workloads running on an engine, monitoring the progress of the workload, and notifying the host software when the workload is completed. If desired, the graphics microcontroller 1533 may also facilitate low power or hibernation states for the graphics processor core 1519 and provide the graphics processor core 1519 with the ability to save and restore registers within the graphics processor core 1519 during low power state transitions independent of the operating system and / or graphics driver software on the system.

[0234] Graphics processor core 1519 may have more or less than the illustrated sub-cores 1521A-1521F, up to N modular sub-cores. For each set of N sub-cores, graphics processor core 1519 may also include shared function logic 1535, shared and / or cache memory 1536, geometry / fixed function pipeline 1537, and other fixed function logic 1538 to accelerate various graphics and compute processing operations. Shared function logic 1535 may include logic units associated with shared function logic 1720 of FIG. 17 (e.g., sampler logic, math logic, and / or inter-thread communication logic), which may be shared by each of the N sub-cores in graphics processor core 1519. Shared and / or cache memory 1536 may be a last level cache for the set of N sub-cores 1521A-1521F in graphics processor core 1519, and may also act as a shared memory accessible by multiple sub-cores. Geometry / fixed function pipeline 1537 may be included in place of geometry / fixed function pipeline 1531 in fixed function block 1530 and may include the same or similar logical units.

[0235] Graphics processor core 1519 may include separate fixed function logic 1538, which may include various fixed function acceleration logic for use by graphics processor core 1519. Optionally, separate fixed function logic 1538 includes a separate geometry pipeline for use with position-only shading. There are two geometry pipelines for position-only shading: a full geometry pipeline in geometry / fixed function pipeline 1538, 1531, and a culling pipeline, which is a separate geometry pipeline that may be included in separate fixed function logic 1538. For example, the culling pipeline may be a scaled-down version of the full geometry pipeline. The full pipeline and the culling pipeline may run separate instances of the same application, each instance having a separate context. Position-only shading may hide lengthy culling steps of discarded triangles, allowing shading to complete sooner in some instances. For example, the culling pipeline logic in separate fixed function logic 1538 may run multiple position shaders in parallel with the main application, typically producing significant results faster than the full pipeline. This is because the culling pipeline only fetches and shades vertex position attributes without rasterizing and rendering pixels to the framebuffer. The culling pipeline can use the generated nontrivial results to calculate visibility information for all triangles, regardless of whether these triangles have been culled or not. The full pipeline (this instance is sometimes called the replay pipeline) can use the visibility information to skip culled triangles and shade only visible triangles that are eventually sent to the rasterization stage.

[0236] Optionally, the separate fixed function logic 1538 may also include machine learning acceleration logic, such as fixed function matrix multiplication logic, for implementations involving machine learning training or inference optimization.

[0237] Each of the graphics sub-cores 1521A-1521F includes a set of execution resources that can be used to perform graphics, media, and compute operations in response to requests from a graphics pipeline, a media pipeline, or a shader program. The graphics sub-cores 1521A-1521F include a plurality of EU arrays 1522A-1522F, 1524A-1524F, thread dispatch and inter-thread communication (TD / IC) logic 1523A-1523F, 3D (e.g., texture) samplers 1525A-1525F, media samplers 1506A-1506F, shader processors 1527A-1527F, and shared local memories (SLMs) 1528A-1528F. Each of the EU arrays 1522A-1522F, 1524A-1524F includes multiple execution units, which are general-purpose graphics processing units capable of performing floating-point and integer / fixed-point logical operations in service of graphics, media, or compute operations, and which include graphics shader programs, media shader programs, or compute shader programs. The TD / IC logic 1523A-1523F performs local thread dispatch and thread control operations for the execution units in the sub-cores, and facilitates communication between threads executing on the execution units of the sub-cores. The 3D samplers 1525A-1525F can read textures or other 3D graphics related data into memory. The 3D samplers can separately read texture data based on the configured sampling state and the texture format associated with a given texture. The media samplers 1506A-1506F can perform similar read operations based on the type and format associated with the media data. For example, each graphics sub-core 1521A-1521F may alternately include an integrated 3D and media sampler.Threads executing on execution units included in each of the sub-cores 1521A-1521F may utilize shared local memory 1528A-1528F within each sub-core, allowing threads executing within a thread group to execute using a common pool of on-chip memory.

[0238] 15C is a block diagram of a general purpose graphics processing unit (GPGPU) 1570 that may be configured as a graphics processor, e.g., graphics processor 1508 and / or compute accelerator according to embodiments described herein. GPGPU 1570 may be interconnected with a host processor (e.g., one or more CPUs 1546) and memories 1571, 1572 via one or more system buses and / or memory buses. Memory 1571 may be system memory that may be shared with one or more CPUs 1546, and memory 1572 is device memory dedicated to GPGPU 1570. For example, each component in GPGPU 1570 and device memory 1572 may be mapped to memory addresses accessible by one or more CPUs 1546. Access to memories 1571 and 1572 may be facilitated via memory controller 1568. Memory controller 1568 may include an integrated direct memory access (DMA) controller 1569 or may contain logic to perform operations otherwise performed by a DMA controller.

[0239] The GPGPU 1570 includes multiple cache memories, including an L2 cache 1553, an L1 cache 1554, an instruction cache 1555, and a shared memory 1556, at least some of which may also be partitioned as cache memories. The GPGPU 1570 also includes multiple compute units 1560A-1560N. Each compute unit 1560A-1560N includes a set of vector registers 1561, scalar registers 1562, a vector logic unit 1563, and a scalar logic unit 1564. The compute units 1560A-1560N may also include a local shared memory 1565 and a program counter 1566. The compute units 1560A-1560N may be coupled to a constant cache 1567 that may be used to store constant data, which is data that does not change during execution of a kernel or shader program executed by the GPGPU 1570. The constant cache 1567 may be a scalar data cache, and cached data may be fetched directly into the scalar registers 1562.

[0240] In operation, one or more CPUs 1546 may write commands to registers or memory in the GPGPU 1570 that are mapped to an accessible address space. The command processor 1557 may read commands from the registers or memory and determine how to process these commands in the GPGPU 1570. The thread dispatcher 1558 may then be used to dispatch threads to the compute units 1560A-1560N that execute these commands. Each compute unit 1560A-1560N may execute threads independently of the other compute units. Additionally, each compute unit 1560A-1560N may be independently configured for conditional computations and may conditionally output computation results to memory. The command processor 1557 may interrupt one or more CPUs 1546 upon completion of submitted commands.

[0241] Figures 16A-16C show block diagrams of another graphics processor and compute accelerator architecture provided by an embodiment described herein, for example, according to Figures 15A-15C. Elements in Figures 16A-16C having the same or similar names as elements in any other figure herein refer to the same element as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, may be connected to other entities, but are not limited to such.

[0242] FIG. 16A is a block diagram of a graphics processor 1600, which may be a separate graphics processing unit or may be a graphics processor integrated with multiple processing cores or other semiconductor devices (such as, but not limited to, a memory device or a network interface). The graphics processor 1600 may be a variation of the graphics processor 1508 or may be used in place of the graphics processor 1508. Thus, the disclosure herein of any feature combined with the graphics processor 1508 also discloses the corresponding combination with the graphics processor 1600, but is not limited to such. The graphics processor may communicate via a memory-mapped I / O interface to registers on the graphics processor and with commands located in the processor memory. The graphics processor 1600 may include a memory interface 1614 to access memory. The memory interface 1614 may be an interface to local memory, one or more internal caches, one or more shared external caches, and / or a system memory.

[0243] Optionally, graphics processor 1600 also includes a display controller 1602 that drives the display of output data to a display device 1618. Display controller 1602 includes hardware for one or more overlay planes used to display and composite multiple layers of video or user interface elements. Display device 1618 may be an internal or external display device. In one embodiment, display device 1618 is a head-mounted display device, such as a virtual reality (VR) display device or an augmented reality (AR) display device. Graphics processor 1600 may include a video codec engine 1606 that encodes, decodes, or transcodes media to or from one or more media encoding formats. These formats include, but are not limited to, Moving Picture Experts Group (MPEG) formats (such as MPEG-2), Advanced Video Coding (AVC) formats (such as H.264 / MPEG-4 AVC, H.265 / HEVC), Alliance for Open Media's (AOMedia) VP8, VP9, ​​and Society of Motion Picture and Television Engineers' (SMPTE) 421M / VC-1, and Joint Photographic Experts Group (JPEG) formats (such as the JPEG format, Motion JPEG (MJPEG) format, etc.).

[0244] Graphics processor 1600 may include a block image transfer (BLIT) engine 1604 for performing two-dimensional (2D) rasterizer operations, including, for example, bit-boundary block transfers. Alternatively, however, 2D graphics operations may be performed using one or more components of a graphics processing engine (GPE) 1610. In some embodiments, GPE 1610 is a compute engine for performing graphics operations, including three-dimensional (3D) graphics operations and media operations.

[0245] The GPE 1610 may include a 3D pipeline 1612 for performing 3D operations such as rendering three-dimensional images and scenes using processing functions that operate on 3D primitive shapes (e.g., rectangles, triangles, etc.). The 3D pipeline 1612 includes programmable and fixed function elements that perform various tasks within the element and / or generate execution threads to the 3D / media subsystem 1615. While the 3D pipeline 1612 may be used to perform media operations, one embodiment of the GPE 1610 also includes a media pipeline 1616 that is used specifically to perform media operations such as video post-processing and image enhancement.

[0246] The media pipeline 1616 may include fixed function or programmable logic units that perform one or more specialized media operations, such as video decode acceleration, video deinterlacing, and video encode acceleration, in place of or on behalf of the video codec engine 1606. The media pipeline 1616 may also include a thread generation unit that generates threads to execute in the 3D / media subsystem 1615. The generated threads perform calculations for the media operations in one or more graphics execution units included in the 3D / media subsystem 1615.

[0247] The 3D / Media subsystem 1615 may include logic for executing threads created by the 3D pipeline 1612 and the media pipeline 1616. These pipelines may send thread execution requests to the 3D / Media subsystem 1615, which includes thread dispatch logic for arbitrating the various requests and dispatching them to available thread execution resources. The execution resources include an array of graphics execution units for processing the 3D and media threads. The 3D / Media subsystem 1615 may include one or more built-in caches for thread instructions and data. Additionally, the 3D / Media subsystem 1615 may also include shared memory, including registers and addressable memory, for sharing data between threads and storing output data.

[0248] FIG. 16B illustrates a graphics processor 1620 that is a variation of graphics processor 1600, where graphics processor 1620 may be substituted for graphics processor 1600, or vice versa. Thus, any disclosure herein of any feature in combination with graphics processor 1600 also discloses the corresponding combination with graphics processor 1620, but is not limited to such. Graphics processor 1620 has a tiled architecture according to embodiments described herein. Graphics processor 1620 may include a graphics processing engine cluster 1622 having multiple instances of graphics processing engine 1610 of FIG. 16A in graphics engine tiles 1610A-1610D. Each graphics engine tile 1610A-1610D may be interconnected via a set of tile interconnects 1623A-1623F. Each graphics engine tile 1610A-1610D may also be coupled to a memory module or memory device 1626A-1626D via a memory interconnect 1625A-1625D. The memory devices 1626A-1626D may use any graphics memory technology. For example, the memory devices 1626A-1626D may be graphics double data rate (GDDR) memory. The memory devices 1626A-1626D may be high bandwidth memory (HBM) modules that may be on-die with the respective graphics engine tile 1610A-1610D. The memory devices 1626A-1626D may be stacked memory devices that may be stacked on top of the respective graphics engine tile 1610A-1610D. Each graphics engine tile 1610A-1610D and associated memory 1626A-1626D may reside on a separate chiplet, which is bonded to a base die or base substrate, as described in further detail in Figures 24B-24D.

[0249] The graphics processor 1620 may be configured with a non-uniform memory access (NUMA) system in which memory devices 1626A-1626D are coupled to associated graphics engine tiles 1610A-1610D. A given memory device may be accessed by graphics engine tiles other than the tile to which it is directly coupled. However, access latency to memory devices 1626A-1626D may be lowest when accessing a local tile. In one embodiment, a cache coherent NUMA (ccNUMA) system is enabled, using tile interconnects 1623A-1623F to enable communication between cache controllers in the graphics engine tiles 1610A-1610D to maintain a consistent memory image when more than one cache stores the same memory location.

[0250] The graphics processing engine cluster 1622 may connect with an on-chip or on-package fabric interconnect 1624. The fabric interconnect 1624 may enable communication between the graphics engine tiles 1610A-1610D and components such as the video codec 1606 and one or more copy engines 1604. The copy engines 1604 may be used to move data between the memory devices 1626A-1626D and between the memory devices 1626A-1626D and memory external to the graphics processor 1620 (e.g., system memory). The fabric interconnect 1624 may also be used to interconnect the graphics engine tiles 1610A-1610D. The graphics processor 1620 may include a display controller 1602 that enables connection to an external display device 1618, if desired. The graphics processor may also be configured as a graphics accelerator or a compute accelerator. In an accelerator configuration, the display controller 1602 and the display device 1618 may be omitted.

[0251] Graphics processor 1620 may be connected to a host system via host interface 1628. Host interface 1628 may allow communication between graphics processor 1620, system memory, and / or other system components. Host interface 1628 may be, for example, a PCI Express bus or another type of host system interface.

[0252] FIG. 16C illustrates a compute accelerator 1630 according to an embodiment described herein. The compute accelerator 1630 may include architectural similarities to the graphics processor 1620 of FIG. 16B and is optimized for compute acceleration. A compute engine cluster 1632 may include a set of compute engine tiles 1640A-1640D that include execution logic optimized for general-purpose parallel or vector-based compute operations. The compute engine tiles 1640A-1640D may not include fixed-function graphics processing logic, although in some embodiments, one or more of the compute engine tiles 1640A-1640D may include logic for media acceleration. The compute engine tiles 1640A-1640D may be connected to memory 1626A-1626D via memory interconnects 1625A-1625D. The memories 1626A-1626D and memory interconnects 1625A-1625D may be of similar technology as found in the graphics processor 1620 or may be different. The graphics compute engine tiles 1640A-1640D may be interconnected via a set of tile interconnects 1623A-1623F and may be connected to and / or by the fabric interconnect 1624. In one embodiment, the compute accelerator 1630 includes a large L3 cache 1636 that may be configured as a device-wide cache. The compute accelerator 1630 may also be connected to a host processor and memory via a host interface 1628 in a manner similar to the graphics processor 1620 of FIG. 16B. [Graphics processing engine]

[0253] FIG. 17 is a block diagram of a graphics processing engine 1710 of a graphics processor according to some embodiments. The graphics processing engine (GPE) 1710 may be a version of the GPE 1610 shown in FIG. 16A and may also represent the graphics engine tiles 1610A-1610D of FIG. 16B. Elements of FIG. 17 having the same or similar names as elements of any other figure herein refer to the same elements as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but are not limited to such. For example, the 3D pipeline 1612 and media pipeline 1616 of FIG. 16A are also shown in FIG. 17. The media pipeline 1616 is optional in some embodiments of the GPE 1710 and may not be explicitly included in the GPE 1710. For example, in at least one embodiment, a separate media and / or image processor is coupled to the GPE 1710.

[0254] The GPE 1710 may be coupled to or include a command streamer 1703, which provides a command stream to the 3D pipeline 1612 and / or the media pipeline 1616. Alternatively or additionally, the command streamer 1703 may be directly coupled to a unified return buffer 1718, which may be communicatively coupled to the graphics core array 1714. Optionally, the command streamer 1703 may be coupled to a memory, which may be a system memory or one or both of an internal cache memory and a shared cache memory. The command streamer 1703 may receive commands from the memory and send the commands to the 3D pipeline 1612 and / or the media pipeline 1616. The commands are instructions fetched from a ring buffer, which stores commands for the 3D pipeline 1612 and the media pipeline 1616. The ring buffer may further include a batch command buffer, which stores a batch of multiple commands. Commands for the 3D pipeline 1612 may also include references to data stored in memory, such as, but not limited to, vertex and geometry data for the 3D pipeline 1612 and / or image data and memory objects for the media pipeline 316. The 3D pipeline 1612 and the media pipeline 1616 process the commands and data by operating through logic within the respective pipelines or by dispatching one or more execution threads to the graphics core array 1714. The graphics core array 1714 may include one or more blocks of graphics cores (e.g., graphics core 1715A, graphics core 1715B), each block including one or more graphics cores. Each graphics core includes a set of graphics execution resources including general-purpose and graphics-specific execution logic for performing graphics and compute operations, as well as fixed-function texture processing and / or machine learning and artificial intelligence acceleration logic.

[0255] In various embodiments, the 3D pipeline 1612 may include fixed function logic and programmable logic to process one or more shader programs, such as vertex shaders, geometry shaders, pixel shaders, fragment shaders, compute shaders, or other shader programs, by processing instructions and dispatching execution threads to the graphics core array 1714. The graphics core array 1714 provides a unified block of execution resources for use in processing these shader programs. General purpose execution logic (e.g., execution units) within the graphics cores 1715A-1715B of the graphics core array 1714 includes support for a variety of 3D API shader languages ​​and can execute multiple concurrent threads associated with multiple shaders.

[0256] Graphics core array 1714 may include execution logic to perform media functions such as video processing and / or image processing. In addition to graphics processing operations, the execution units may include general purpose logic that is programmable to perform parallel general purpose computing operations. The general purpose logic may perform processing operations in parallel or in conjunction with general purpose logic in processor core 1407 of FIG. 14 or cores 1502A-1502N as in FIG. 15A.

[0257] Output data generated by threads executing on the graphics core array 1714 can output data to memory in a unified return buffer (URB) 1718. The URB 1718 can store data for multiple threads. The URB 1718 can be used to transmit data between individual threads executing on the graphics core array 1714. The URB 1718 can also be used for synchronization between threads on the graphics core array 1714 and fixed function logic in the shared function logic 1720.

[0258] Optionally, the graphics core array 1714 may be scalable, including a variable number of graphics cores, each having a variable number of execution units based on the target power and performance levels of the GPE 1710. The execution resources may be dynamically scalable, such that execution resources may be enabled or disabled as needed.

[0259] The graphics core array 1714 couples to shared function logic 1720 that includes multiple resources shared among the graphics cores in the graphics core array. The shared functions in the shared function logic 1720 are hardware logic units that provide specialized auxiliary functions to the graphics core array 1714. In various embodiments, the shared function logic 1720 includes, but is not limited to, sampler logic 1721, math logic 1722, and inter-thread communication (ITC) logic 1723. Additionally, one or more caches 1725 in the shared function logic 1720 may be implemented.

[0260] Shared functionality is implemented when the requirements of at least a given specialized functionality are insufficient for inclusion within the graphics core array 1714. Instead, a single instantiation of this specialized functionality is implemented in the shared functionality logic 1720 as a standalone entity and shared among the execution resources within the graphics core array 1714. The exact set of functionality shared among and included within the graphics core array 1714 varies from embodiment to embodiment. Certain shared functionality within the shared functionality logic 1720 that is widely used by the graphics core array 1714 may be included in the shared functionality logic 1716 within the graphics core array 1714. If desired, the shared functionality logic 1716 within the graphics core array 1714 may include some or all of the logic within the shared functionality logic 1720. All logic elements within the shared functionality logic 1720 may be duplicated in the shared functionality logic 1716 of the graphics core array 1714. Alternatively, the shared functionality logic 1720 is omitted in favor of the shared functionality logic 1716 within the graphics core array 1714. [Execution unit]

[0261] 18A-18B illustrate thread execution logic 1800 including an array of processing elements for use in a graphics processor core according to embodiments described herein. Elements in FIG. 18A-18B having the same or similar names as elements in any other figure herein refer to the same element as in the other figure and may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but are not limited to such. FIG. 18A-18B illustrates an overview of thread execution logic 1800, which may represent the hardware logic illustrated with each of the sub-cores 1521A-1521F in FIG. 15B. FIG. 18A illustrates an execution unit in a general-purpose graphics processor, and FIG. 18B illustrates an execution unit that may be used in a compute accelerator.

[0262] 18A, thread execution logic 1800 may include a shader processor 1802, a thread dispatcher 1804, an instruction cache 1806, a scalable execution unit array including multiple execution units 1808A-1808N, a sampler 1810, a shared local memory 1811, a data cache 1812, and a data port 1814. If desired, the scalable execution unit array may be dynamically scaled by enabling or disabling one or more execution units (e.g., any of execution units 1808A, 1808B, 1808C, 1808D, ..., 1808N-1, and 1808N) based on the computational requirements of the workload. The included components may be interconnected via an interconnect fabric that connects to each of these components. The thread execution logic 1800 may include one or more connections to memory, such as system memory or cache memory, through an instruction cache 1806, a data port 1814, a sampler 1810, and one or more of the execution units 1808A-1808N. Each execution unit (e.g., 1808A) may be a standalone programmable general-purpose computational unit capable of executing multiple simultaneous hardware threads, processing multiple data elements in parallel per thread. In various embodiments, the array of execution units 1808A-1808N is scalable to include any number of individual execution units.

[0263] The execution units 1808A-1808N may be primarily used to execute shader programs. The shader processor 1802 may process various shader programs and dispatch execution threads associated with the shader programs via a thread dispatcher 1804. The thread dispatcher may include logic to arbitrate thread initiation requests from the graphics and media pipelines and instantiate the requested threads on one or more execution units 1808A-1808N. For example, the geometry pipeline may dispatch vertex shaders, tessellation shaders, or geometry shaders to thread execution logic for processing. If necessary, the thread dispatcher 1804 may also handle run-time thread creation requests from executing shader programs.

[0264] The execution units 1808A-N may support an instruction set that includes native support for many standard 3D graphics shader instructions, resulting in shader programs from graphics libraries (e.g., Direct 3D and OpenGL) being executed with minimal translation. The execution units support vertex and geometry processing (e.g., vertex programs, geometry programs, vertex shaders), pixel processing (e.g., pixel shaders, fragment shaders), and general-purpose processing (e.g., compute shaders and media shaders). Each of the execution units 1808A-N is capable of multiple issue of single instruction multiple data (SIMD) execution, allowing multi-threaded operation to enable an efficient execution environment despite high-latency memory accesses. Each hardware thread within each execution unit has a dedicated high-bandwidth register file and associated independent thread state. Execution is multiple issue per clock to a pipeline capable of integer arithmetic, single and double precision floating point arithmetic, SIMD branch functions, logical operations, transcendental operations, and other miscellaneous operations. While waiting for data from memory or one of the shared functions, dependency logic within the execution units 1808A-1808N causes the waiting thread to sleep until the requested data is returned. While the waiting thread is asleep, hardware resources may be dedicated to processing other threads. For example, during delays associated with a vertex shader operation, the execution units may operate a pixel shader, fragment shader, or another type of shader program (including another vertex shader, such as the vertex shader 2107 shown in FIG. 21). Various embodiments may be applied to employ execution using single instruction multiple threads (SIMT) as an alternative to or in addition to using SIMD. Reference to a SIMD core or operation may also apply to SIMT or to SIMD in combination with SIMT.

[0265] Each execution unit in execution units 1808A-1808N processes an array of data elements. The number of data elements is the "execution size" or number of channels for an instruction. An execution channel is a logical unit of execution for data element access, masking, and flow control within an instruction. The number of channels may be independent of the number of physical arithmetic logic units (ALUs), floating point units (FPUs), or other logical units (e.g., tensor cores, ray tracing cores, etc.) for a particular graphics processor. Additionally, execution units 1808A-1808N may support integer and floating point data types.

[0266] The execution unit instruction set includes SIMD instructions. Various data elements may be stored in registers as packed data types, and the execution unit will process the various elements based on the data size of these elements. For example, when processing a 256-bit wide vector, the 256 bits of the vector are stored in registers, and the execution unit processes the vector as four separate 64-bit packed data elements (quadword (QW) sized data elements), eight separate 32-bit packed data elements (doubleword (DW) sized data elements), sixteen separate 16-bit packed data elements (word (W) sized data elements), or thirty-two separate 8-bit data elements (byte (B) sized data elements). However, other vector widths and register sizes are possible.

[0267] If necessary, one or more execution units may be combined into a fused execution unit 1809A-1809N, which has thread control logic (1807A-1807N) common to each EU. Multiple EUs may also be fused into an EU group. Each EU in a fused EU group may be configured to execute a separate SIMD hardware thread. The number of EUs in a fused EU group may vary according to the embodiment. Furthermore, various SIMD widths may be executed per EU, including but not limited to SIMD8, SIMD16, and SIMD32. Each fused graphics execution unit 1809A-1809N includes at least two execution units. For example, the fused execution unit 1809A includes a first EU (1808A), a second EU (1808B), and thread control logic 1807A common to the first EU (1808A) and the second EU (1808B). Thread control logic 1807A controls the threads executing in fused graphics execution unit 1809A and enables each EU in fused execution units 1809A-1809N to execute using a common instruction pointer register.

[0268] One or more built-in instruction caches (e.g., 1806) are included in the thread execution logic 1800 for caching thread instructions for the execution units. One or more data caches (e.g., 1812) may be included in the thread execution logic 1800 for caching thread data during thread execution. Threads executing in the execution logic 1800 may also store explicitly managed data in the shared local memory 1811. A sampler 1810 may be included to provide texture sampling for 3D operations and media sampling for media operations. The sampler 1810 may include specialized texture or media sampling functions to process the texture or media data in a sampling step before providing the sampled data to the execution units.

[0269] At run time, the graphics and media pipelines send thread start requests to the thread execution logic 1800 via thread creation and dispatch logic. Once a group of geometric objects has been processed and rasterized into pixel data, pixel processor logic (e.g., pixel shader logic, fragment shader logic, etc.) in the shader processor 1802 is invoked to further compute output information and write the results to an output surface (e.g., color buffer, depth buffer, stencil buffer, etc.). The pixel shader or fragment shader may compute values ​​for various vertex attributes that will be interpolated across the rasterized object. The pixel processor logic in the shader processor 1802 may then execute pixel shader or fragment shader programs provided by an application programming interface (API). To execute the shader programs, the shader processor 1802 dispatches threads to the execution units (e.g., 1808A) via the thread dispatcher 1804. Shader processor 1802 may use texture sampling logic in sampler 1810 to access texture data in texture maps stored in memory. Arithmetic operations on the texture data and the input geometry data compute pixel color data for each geometric fragment, or discard one or more pixels from further processing.

[0270] Additionally, data port 1814 may provide a memory access mechanism to thread execution logic 1800 to output processed data to memory for further processing in the output pipeline of the graphics processor. Data port 1814 may include or be coupled to one or more cache memories (e.g., data cache 1812) that cache data for memory access via data port 1814.

[0271] Optionally, the execution logic 1800 may also include a ray tracer 1805 that can provide ray tracing acceleration functionality. The ray tracer 1805 can support a ray tracing instruction set that includes instructions / functions for ray generation. This ray tracing instruction set may be similar to or different from the ray tracing instruction set supported by the ray tracing core 372 of FIG. 3C.

[0272] FIG. 18B illustrates example internal details of the execution unit 1808. The GSU 1808 may include an instruction fetch unit 1837, a general purpose register file array (GRF) 1824, an architectural register file array (ARF) 1826, a thread arbiter 1822, a send unit 1830, a branch unit 1832, a set of SIMD floating point units (FPUs) 1834, and an optional set of dedicated integer SIMD ALUs 1835. The GRF 1824 and ARF 1826 include a set of general purpose register files and architectural register files associated with each concurrent hardware thread that may be active in the GSU 1808. For each thread, architectural state may be maintained in the ARF 1826, and data used during thread execution is stored in the GRF 1824. The execution state of each thread may be held in thread-specific registers in the ARF 1826, including an instruction pointer for each thread.

[0273] The GSUs 1808 may have an architecture that combines simultaneous multithreading (SMT) and fine-grained interleaved multithreading (IMT). This architecture may have a modular configuration that can be tuned at design time based on the number of targeted simultaneous threads and the number of registers per execution unit, and the resources of the execution unit are divided across the logic used to execute multiple simultaneous threads. The number of logical threads that may be executed by the GSUs 1808 is not limited to the number of hardware threads, and multiple logical threads may be assigned to each hardware thread.

[0274] If desired, the graphics execution unit 1808 can issue multiple instructions simultaneously, each of which may be a different instruction. The thread arbiter 1822 of the graphics execution unit thread 1808 can dispatch instructions for execution to one of the send unit 1830, the branch unit 1832, or the SIMD FPU 1834. Each execution thread can access 128 general purpose registers in the GRF 1824. Each register can store 32 bytes accessible as a SIMD 8-element vector of 32-bit data elements. Each execution unit thread can access 4K bytes of the GRF 1824, although embodiments are not so limited and more or less register resources may be provided in other embodiments. The graphics execution unit 1808 can be divided into seven hardware threads that can independently perform computational operations, although the number of threads per execution unit may also vary according to the embodiment, for example, up to 16 hardware threads may be supported. In an exemplary embodiment where 7 threads may access 4K bytes, GRF 1824 can store a total of 28K bytes. In another exemplary embodiment where 16 threads may access 4K bytes, GRF 1824 can store a total of 64K bytes. However, the number of threads per execution unit is not limited to these examples and may be more or less than the given number. A flexible addressing scheme may allow multiple registers to be addressed together to effectively build wide range registers or represent strided rectangular block data structures.

[0275] Additionally or alternatively, memory operations, sampler operations, and other long latency system communications may be dispatched via "send" instructions executed by a messaging send unit 1830. Branch instructions may be dispatched to a dedicated branch unit 1832 to facilitate SIMD divergence and eventual convergence.

[0276] The graphics execution units 1808 may include one or more SIMD floating point units (FPUs) 1834 for performing floating point operations. The FPUs 1834 may also support integer calculations. In some examples, the FPUs 1834 may perform up to M 32-bit floating point (or integer) operations in SIMD, or up to 2M 16-bit integer or 16-bit floating point operations in SIMD. Optionally, at least one of the FPUs provides extended math capabilities to support high-throughput transcendental math functions and double-precision 184-bit floating point. There may also be a set of 8-bit integer SIMD ALUs 1835, which may be specifically optimized for performing operations associated with machine learning calculations.

[0277] If desired, an array of multiple instances of the graphics execution unit 1808 may be instantiated into a graphics sub-core group (e.g., a sub-slice). For scalability, the exact number of execution units per sub-core group can be selected by the product designer. The execution units 1808 may execute instructions across multiple execution channels. Furthermore, each thread executing on the graphics execution unit 1808 may be executed on a different channel.

[0278] FIG. 19 illustrates a further exemplary execution unit 1900. Elements of FIG. 19 having the same or similar names as elements of any other figure herein refer to the same element as in the other figure and may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but is not limited to such. The execution unit 1900 may be, for example, a compute-optimized execution unit for use in the compute engine tiles 1640A-1640D as in FIG. 16C, but is not limited to such. The execution unit 1900 may also be used in the graphics engine tiles 1610A-1610D as in FIG. 16B. The execution unit 1900 may include a thread control unit 1901, a thread state unit 1902, an instruction fetch / prefetch unit 1903, and an instruction decode unit 1904. The execution unit 1900 may further include a register file 1906 that stores registers that may be assigned to hardware threads within the execution unit. The execution units 1900 may further include a sending unit 1907 and a branching unit 1908. The sending unit 1907 and branching unit 1908 may operate similarly to the sending unit 1830 and branching unit 1832 of the graphics execution unit 1808 in FIG. 18B.

[0279] The execution unit 1900 may also include a compute unit 1910 that includes multiple different types of functional units. The compute unit 1910 may also include an ALU unit 1911 that includes an array of arithmetic logic units. The ALU unit 1911 may be configured to perform 64-bit, 32-bit, and 16-bit integer and floating point operations. The integer and floating point operations may be performed simultaneously. The compute unit 1910 may also include a systolic array 1912 and a math unit 1913. The systolic array 1912 includes a W by D network of data processing units that may be used to perform vector operations or other data parallel operations in a systolic manner. The systolic array 1912 may be configured to perform matrix operations, such as matrix dot product operations. The systolic array 1912 may support 16-bit floating point operations, as well as 8-bit and 4-bit integer operations. The systolic array 1912 may be configured to accelerate machine learning operations. The systolic array 1912 may be configured to support Bfloat16, a 16-bit floating point format. A math unit 1913 may be included, which performs a specific subset of math operations in a more efficient low-power manner than the ALU unit 1911 at the time. The math unit 1913 may include math logic found in the shared function logic of the graphics processing engine provided by other described embodiments, such as math logic 1722 of the shared function logic 1720 of FIG. 17. The math unit 1913 may be configured to perform 32-bit and 64-bit floating point operations.

[0280] The thread control unit 1901 includes logic to control the execution of threads in the execution units. The thread control unit 1901 may include thread arbitration logic to start, suspend, and preempt the execution of threads in the execution units 1900. The thread state unit 1902 may be used to store thread states for threads assigned to execute in the execution units 1900. Storing the thread states in the execution units 1900 allows for rapid preemption of threads when they are blocked or paused. The instruction fetch / prefetch unit 1903 may fetch instructions from an instruction cache of higher level execution logic (e.g., instruction cache 1806 as in FIG. 18A). The instruction fetch / prefetch unit 1903 may also issue prefetch requests for instructions to be loaded into the instruction cache based on an analysis of currently executing threads. The instruction decode unit 1904 may be used to decode instructions to be executed by the compute units. The instruction decode unit 1904 may be used as a secondary decoder to decode compound instructions into their constituent micro-operations.

[0281] The execution unit 1900 further includes a register file 1906 that may be used by hardware threads executing in the execution unit 1900. Registers in the register file 1906 may be divided throughout logic used to execute multiple simultaneous threads within the compute units 1910 of the execution unit 1900. The number of logical threads that may be executed by the graphics execution unit 1900 is not limited to the number of hardware threads, and multiple logical threads may be assigned to each hardware thread. The size of the register file 1906 may vary between embodiments based on the number of hardware threads supported. Register renaming may be used to dynamically assign registers to hardware threads.

[0282] 20 is a block diagram illustrating a graphics processor instruction format 2000. The graphics processor execution units support an instruction set having instructions in multiple formats. The solid outlines indicate components that are typically included in execution unit instructions, while the dashed outlines include optional components or components that are only included in a subset of the instructions. The instruction format 2000 described and illustrated is a macro-instruction, which differs from the micro-operations that result from instruction decode as the instruction is processed, in that it is an instruction that is provided to the execution units.

[0283] The graphics processor execution units described herein may natively support instructions in a 128-bit instruction format 2010. A 64-bit compressed instruction format 2030 is available for some instructions based on the selected instruction, instruction options, and number of operands. The native 128-bit instruction format 2010 provides access to all instruction options, with some options and operations restricted to the 64-bit format 2030. The native instructions available in the 64-bit format 2030 vary from embodiment to embodiment. The instruction is partially compressed using a set of index values ​​in index field 2013. The execution unit hardware looks up a set of compression tables based on the index value and uses the output of the compression tables to reconstruct the native instruction into the 128-bit instruction format 2010. Instructions of other sizes and formats may be used.

[0284] For each format, the instruction opcode 2012 defines the operation that the execution unit performs. The execution unit executes each instruction in parallel across multiple data elements of each operand. For example, in response to an add instruction, the execution unit performs a simultaneous add operation on each color channel representing a texture or picture element. By default, the execution unit performs each instruction on all data channels of an operand. The instruction control field 2014 may allow control of certain execution options such as channel selection (e.g., prediction) and data channel order (e.g., swizzle). For instructions in the 128-bit instruction format 2010, the execution size field 2016 limits the number of data channels that are executed in parallel. The execution size field 2016 may not be available for use with the 64-bit compressed instruction format 2030.

[0285] Some execution unit instructions have up to three operands, including two source operands, source 0 (2020) and source 1 (2022), and one destination 2018. An execution unit may support dual destination instructions, where one of the destinations is implied. Data manipulation instructions may have a third source operand (e.g., source 2 (2024)), with the instruction opcode 2012 determining the number of source operands. The final source operand of an instruction may be an immediate (e.g., hard-coded) value sent with the instruction.

[0286] The 128-bit instruction format 2010 may include, for example, an access / address mode field 2026 that specifies whether a direct or indirect register addressing mode is used. When a direct register addressing mode is used, the register addresses of one or more operands are provided directly by bits in the instruction.

[0287] The 128-bit instruction format 2010 may also include an access / address mode field 2026 that specifies the address mode and / or access mode of the instruction. The access mode may be used to define the data access alignment of the instruction. Access modes including 16-byte aligned and 1-byte aligned access modes may be supported, and the byte alignment of the access mode determines the access alignment of the instruction operands. For example, in a first mode, the instruction may use byte-aligned addressing for source and destination operands, and in a second mode, the instruction may use 16-byte aligned addressing for all source and destination operands.

[0288] The address mode portion of the access / address mode field 2026 may determine whether the instruction uses direct or indirect addressing. If the direct register addressing mode is used, bits in the instruction directly provide the register address of one or more operands. If the indirect register addressing mode is used, the register address of one or more operands may be calculated based on the value of an address register and an address immediate field in the instruction.

[0289] Instructions may be grouped based on opcode 2012 bit fields to simplify opcode decode 2040. For 8-bit opcodes, bits 4, 5, and 6 allow the execution unit to determine the type of opcode. The exact opcode group shown is merely an example. The move / logical opcode group 2042 may include data movement instructions and logical instructions (e.g., move (mov), compare (cmp)). The move / logical group 2042 may share the five most significant bits (MSBs), with move (mov) instructions being of the form 0000xxxxb and logical instructions being of the form 0001xxxxb. The flow control instruction group 2044 (e.g., call, jump (jmp)) includes instructions of the form 0010xxxxb (e.g., 0x20). A miscellaneous instruction group 2046 includes a combination of instructions including synchronization instructions (e.g., wait, send) in the form of 0011xxxxb (e.g., 0x30). A parallel math instruction group 2048 includes component-wise arithmetic instructions (e.g., add, multiply (mul)) in the form of 0100xxxxb (e.g., 0x40). The parallel math group 2048 performs arithmetic operations in parallel across data channels. A vector math group 2050 includes arithmetic instructions (e.g., dp4) in the form of 0101xxxxb (e.g., 0x50). The vector math group performs arithmetic such as a dot product calculation on vector operands. The illustrated opcode decode 2040 may, in one embodiment, be used to determine which portion of the execution units to use to execute the decoded instruction. For example, some instructions may be designated as systolic instructions that execute in a systolic array. Other instructions, such as ray tracing instructions (not shown), may be sent to a ray tracing core or ray tracing logic within a slice or partition of the execution logic. [Graphics Pipeline]

[0290] Figure 21 is a block diagram of a graphics processor 2100 according to another embodiment. Each element in Figure 21 having the same or similar name as an element in any other figure herein denotes the same element as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but is not limited to such.

[0291] The graphics processor 2100 may include different types of graphics processing pipelines, such as a geometry pipeline 2120, a media pipeline 2130, a display engine 2140, thread execution logic 2150, and a render output pipeline 2170. The graphics processor 2100 may be a graphics processor in a multi-core processing system that includes one or more general-purpose processing cores. The graphics processor may be controlled by register writes to one or more control registers (not shown) or through commands issued to the graphics processor 2100 via a ring interconnect 2102. The ring interconnect 2102 may couple the graphics processor 2100 to other processing components, such as other graphics processors or general-purpose processors. Commands from the ring interconnect 2102 are translated in a command streamer 2103, which provides instructions to individual components of the geometry pipeline 2120 or the media pipeline 2130.

[0292] The command streamer 2103 may direct the operation of the vertex fetcher 2105, which reads vertex data from memory and executes vertex processing commands provided by the command streamer 2103. The vertex fetcher 2105 may provide the vertex data to a vertex shader 2107, which performs coordinate space transformation and lighting computations for each vertex. The vertex fetcher 2105 and vertex shader 2107 may execute the vertex processing instructions by dispatching execution threads to the execution units 2152A-2152B via the thread dispatcher 2131.

[0293] The execution units 2152A-B may be an array of vector processors with an instruction set for performing graphics and media operations. The execution units 2152A-B may have an associated L1 cache 2151 that is specific to each array or shared between the arrays. This cache may be configured as a data cache, an instruction cache, or one cache that is partitioned to contain data and instructions in separate partitions.

[0294] The geometry pipeline 2120 may include a tessellation component that performs hardware accelerated tessellation of 3D objects. A programmable hull shader 2111 may configure the tessellation operations. A programmable domain shader 2117 may provide back-end evaluation of the tessellation output. A tessellator 2113 operates at the direction of the hull shader 2111 and may include dedicated logic to generate a set of detailed geometric objects based on a coarse geometric model provided as input to the geometry pipeline 2120. Additionally, if tessellation is not used, the tessellation components (e.g., hull shader 2111, tessellator 2113, domain shader 2117) may be skipped.

[0295] The entire geometric object may be processed by the geometry shader 2119 via one or more threads dispatched to the execution units 2152A-2152B, or may proceed directly to the clipper 2129. The geometry shader may process the entire geometric object, rather than vertices or patches of vertices as in earlier stages of the graphics pipeline. If tessellation is disabled, the geometry shader 2119 receives input from the vertex shader 2107. The geometry shader 2119 may be programmable such that if the tessellation unit is disabled, the geometry shader program performs geometry tessellation.

[0296] A clipper 2129 processes the vertex data before rasterization. The clipper 2129 may be a fixed-function clipper or a programmable clipper with clipping and geometry shader functions. A rasterizer and depth test component 2173 in the render output pipeline 2170 may dispatch pixel shaders that convert geometric objects into per-pixel representations. The pixel shader logic may be included in the thread execution logic 2150. If necessary, an application can bypass the rasterizer and depth test component 2173 and access pre-rasterized vertex data via the stream-out unit 2123.

[0297] The graphics processor 2100 has an interconnect bus, interconnect fabric, or some other interconnect mechanism that can pass data and messages between the major components of the processor. In some embodiments, the execution units 2152A-2152B and associated logic units (e.g., L1 cache 2151, sampler 2154, texture cache 2158, etc.) interconnect via data port 2156 for memory access and communication with the processor's render output pipeline components. The sampler 2154, caches 2151, 2158, and execution units 2152A-2152B may each have a separate memory access path. If desired, the texture cache 2158 may also be configured as a sampler cache.

[0298] The render output pipeline 2170 may include a rasterizer and depth test component 2173 that converts vertex-based objects into an associated pixel-based representation. The rasterizer logic may include a windower / masker unit that performs fixed-function triangle and line rasterization. An associated render cache 2178 and depth cache 2179 are also available in some embodiments. A pixel manipulation component 2177 performs pixel-based operations on the data, although in some examples pixel operations associated with 2D operations (e.g. bit-block image transfer with blending) are performed by the 2D engine 2141 or substituted at display time by the display controller 2143 using overlay display planes. A shared L3 cache 2175 may be available to all graphics components, allowing data sharing without using main system memory.

[0299] The graphics processor's media pipeline 2130 may include a media engine 2137 and a video front end 2134. The video front end 2134 may receive pipeline commands from the command streamer 2103. The media pipeline 2130 may include a separate command streamer. The video front end 2134 may process the media commands before sending the commands to the media engine 2137. The media engine 2137 may include a thread generation function that generates threads for dispatching to the thread execution logic 2150 via the thread dispatcher 2131.

[0300] The graphics processor 2100 may include a display engine 2140, which may be external to the processor 2100 and may be coupled to the graphics processor via a ring interconnect 2102 or some other interconnect bus or fabric. The display engine 2140 may include a 2D engine 2141 and a display controller 2143. The display engine 2140 may include dedicated logic capable of operating independently of the 3D pipeline. The display controller 2143 may be coupled to a display device (not shown), which may be a system-integrated display device such as found in a laptop computer or an external display device attached via a display device connector.

[0301] The geometry pipeline 2120 and the media pipeline 2130 may be configurable to operate according to multiple graphics and media programming interfaces and are not specific to any one application programming interface (API). The graphics processor's driver software may translate API calls specific to a particular graphics or media library into commands that the graphics processor can process. The Open Graphics Library (OpenGL), Open Computing Language (OpenCL), and / or Vulkan graphics and compute APIs may all be supported by the Khronos Group. The Direct3D library may also be supported by Microsoft Corporation. Combinations of these libraries may also be supported. Support may also be provided for the open source computer vision library (OpenCV). Future APIs, including compatible 3D pipelines, may also be supported if a mapping from the future API's pipeline to the graphics processor's pipeline can be made. [Graphics Pipeline Programming]

[0302] FIG. 22A is a block diagram illustrating a graphics processor command format 2200 that may be used to program a graphics processing pipeline, such as those described herein with respect to FIG. 16A, FIG. 17, and FIG. 21. FIG. 22B is a block diagram illustrating a graphics processor command sequence 2210 according to one embodiment. The solid outline in FIG. 22A illustrates components that are typically included in a graphics command, while the dashed outline includes optional components or components that are included in only a subset of the graphics commands. The example graphics processor command format 2200 of FIG. 22A includes data fields that identify the client 2202 of the command, the command operation code (opcode) 2204, and data 2206. Sub-opcodes 2205 and command size 2208 are also included in some commands.

[0303] The client 2202 may specify a client unit of the graphics device that will process the command data. A graphics processor command parser may inspect the client field of each command, condition further processing of the command, and send the command data to the appropriate client unit. The graphics processor client units may include a memory interface unit, a render unit, a 2D unit, a 3D unit, and a media unit. Each client unit may have a corresponding processing pipeline that processes the command. When a command is received by a client unit, the client unit reads the opcode 2204 and, if present, the sub-opcode 2205 to determine the operation to perform. The client unit executes the command using the information in the data field 2206. For some commands, an explicit command size 2208 is expected to specify the size of the command. The command parser may automatically determine the size of at least a portion of the command based on the command opcode. The command may be aligned on a multiple of a double word. Other command formats may also be used.

[0304] The flow diagram of Figure 22B illustrates an exemplary graphics processor command sequence 2210. Software or firmware of a data processing system featuring an exemplary graphics processor may use some version of the command sequence shown to set up, execute, and terminate a set of graphics operations. The sample command sequence is shown and described for illustrative purposes only, and is not limited to these particular commands or this command sequence. Furthermore, commands may be issued as a batch of commands in a command sequence such that the graphics processor will process the sequence of commands at least partially concurrently.

[0305] The graphics processor command sequence 2210 may begin with a pipeline flush command 2212 to cause any active graphics pipelines to complete commands currently pending for that pipeline. If desired, the 3D pipeline 2222 and the media pipeline 2224 may not operate simultaneously. The pipeline flush is performed to cause any pending commands to be completed for the active graphics pipelines. In response to the pipeline flush, the graphics processor's command parser will suspend command processing until the active drawing engines have completed pending operations and the associated read caches are invalidated. If desired, any data in the render cache that is marked as "dirty" may be flushed to memory. The pipeline flush command 2212 may be used for pipeline synchronization or prior to placing the graphics processor in a low power state.

[0306] The pipeline select command 2213 may be used when a command sequence requires the graphics processor to explicitly switch pipelines. The pipeline select command 2213 may only be required once within an execution context before issuing a pipeline command, unless the context issues commands to both pipelines. A pipeline flush command 2212 may be required immediately before the pipeline switches via the pipeline select command 2213.

[0307] Pipeline control commands 2214 may configure the graphics pipeline for operation and may be used to program the 3D pipeline 2222 and the media pipeline 2224. Pipeline control commands 2214 may configure the pipeline state for an active pipeline. Pipeline control commands 2214 may be used for pipeline synchronization and to remove data from one or more cache memories in an active pipeline before processing a batch of commands.

[0308] Return buffer state commands 2216 may be used to configure a set of return buffers for each pipeline to write data to. Some pipeline operations require allocation, selection, or configuration of one or more return buffers to which each operation writes intermediate data during processing. Graphics processors may also use one or more return buffers to store output data and for inter-thread communication. Return buffer state commands 2216 may include selecting the size and number of return buffers to use for a set of pipeline operations.

[0309] The other commands in the command sequence differ based on the active pipeline for the operation. The command sequence is tailored to either the 3D pipeline 2222 starting at the 3D pipeline state 2230 or the media pipeline 2224 starting at the media pipeline state 2240 based on a pipeline decision 2220.

[0310] The commands that configure the 3D pipeline state 2230 include 3D state setting commands for vertex buffer state, vertex element state, constant color state, depth buffer state, and other state variables that are configured before 3D primitive commands are processed. The values ​​of these commands are determined at least in part based on the particular 3D API being used. The 3D pipeline state 2230 commands may also selectively disable or skip certain pipeline elements if those elements are not used.

[0311] The 3D primitive 2232 commands may be used to populate 3D primitives to be processed in the 3D pipeline. These commands and associated parameters sent to the graphics processor via the 3D primitive 2232 commands are forwarded to a vertex fetch function in the graphics pipeline. The vertex fetch function uses the 3D primitive 2232 command data to generate a vertex data structure. The vertex data structure is stored in one or more return buffers. The 3D primitive 2232 commands may be used to perform vertex operations on the 3D primitives via vertex shaders. To process the vertex shaders, the 3D pipeline 2222 dispatches shader execution threads to the graphics processor execution units.

[0312] The 3D pipeline 2222 may be triggered with an execute 2234 command or event. A register may be written to trigger command execution. Execution may be triggered with a "go" command or a "kick" command in the command sequence. Command execution may be triggered with a pipeline synchronization command to flush the command sequence through the graphics pipeline. The 3D pipeline will perform geometry processing of the 3D primitives. Once the operations are complete, the resulting geometric objects are rasterized and the pixel engine colors the resulting pixels. Separate commands to control pixel shading and pixel backend operations may also be included for these operations.

[0313] The graphics processor command sequence 2210 may follow the path of the media pipeline 2224 when performing media operations. In general, the particular use or manner of programming for the media pipeline 2224 depends on the media or compute operations being performed. Certain media decode operations may be offloaded to the media pipeline during media decoding. The media pipeline may be ignored and media decoding may be performed in whole or in part using resources provided by one or more general purpose processing cores. The media pipeline may also include elements for general purpose graphics processor unit (GPGPU) operations, where the graphics processor is used to perform SIMD vector operations using compute shader programs that are not explicitly related to rendering graphics primitives.

[0314] The media pipeline 2224 may be configured in a similar manner to the 3D pipeline 2222. A set of commands that configure the media pipeline state 2240 are dispatched or placed in a command queue before the media object commands 2242. Commands for the media pipeline state 2240 may contain data that configures the media pipeline elements used to process the media object. This includes data such as the encoding format or decoding format that configures the video decode and video encode logic in the media pipeline. Commands for the media pipeline state 2240 may also support the use of one or more pointers to "indirect" state elements that contain a batch of state settings.

[0315] A media object command 2242 may provide a pointer to a media object that the media pipeline processes. A media object includes a memory buffer containing the video data to be processed. If necessary, all media pipeline state must be valid before issuing a media object command 2242. Once the pipeline state is configured and the media object command 2242 is queued, the media pipeline 2224 is triggered with an execute command 2244 or an equivalent execute event (e.g., a register write). Output from the media pipeline 2224 may then be post-processed with operations provided by the 3D pipeline 2222 or the media pipeline 2224. GPGPU operations may be configured and executed in a similar manner to the media operations. [Graphics Software Architecture]

[0316] FIG. 23 illustrates an exemplary graphics software architecture for a data processing system 2300. Such software architecture may include a 3D graphics application 2310, an operating system 2320, and at least one processor 2330. The processor 2330 may include a graphics processor 2332 and one or more general-purpose processor cores 2334. The processor 2330 may be a variation of the processor 1402 or any of the processors described herein. The processor 2330 may be substituted for the processor 1402 or any of the processors described herein. Thus, the disclosure of any feature in combination with the processor 1402 or any of the processors described herein also discloses the corresponding combination with the graphics processor 2330, but is not limited to such. Additionally, each element in FIG. 23 having the same or similar name as an element in any other figure herein represents the same element as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but is not limited to such. The graphics application 2310 and the operating system 2320 each execute in the system memory 2350 of the data processing system.

[0317] The 3D graphics application 2310 may include one or more shader programs that include shader instructions 2312. The shader language instructions may be in a high-level shader language, such as Direct3D High Level Shader Language (HLSL) and OpenGL Shader Language (GLSL). The application may also include executable instructions 2314 in a machine language suitable for execution by a general-purpose processor core 2334. The application may also include graphics objects 2316 defined by vertex data.

[0318] The operating system 2320 may be Microsoft Corporation's Microsoft® Windows® operating system, a proprietary Unix®-like operating system, or an open source Unix-like operating system that uses a variation of the Linux® kernel. The operating system 2320 may support a graphics API 2322, such as the Direct3D API, the OpenGL API, or the Vulkan API. If the Direct3D API is used, the operating system 2320 compiles any shader instructions 2312 in HLSL into a low-level shader language using a front-end shader compiler 2324. The compilation may be just-in-time (JIT) compilation, or the application may perform shader pre-compilation. High-level shaders may be compiled into low-level shaders at compile time of the 3D graphics application 2310. The shader instructions 2312 may be provided in an intermediate form, such as some version of the standard portable intermediate representation (SPIR) used by the Vulkan API.

[0319] A user mode graphics driver 2326 may include a back-end shader compiler 2327 that converts the shader instructions 2312 into a hardware-specific representation. If the OpenGL API is used, the shader instructions 2312 in the GLSL high-level language are sent to the user mode graphics driver 2326 for compilation. The user mode graphics driver 2326 may communicate with a kernel mode graphics driver 2329 using kernel mode functions 2328 of the operating system. The kernel mode graphics driver 2329 may communicate with the graphics processor 2332 to dispatch commands and instructions. [IP core implementation form]

[0320] One or more aspects may be implemented in representative code stored on a machine-readable medium that represents and / or defines logic in an integrated circuit, such as a processor. For example, the machine-readable medium may include instructions that represent various logic in a processor. The instructions, when read by a machine, may cause the machine to manufacture logic that performs the techniques described herein. Such representations, known as "IP cores," are reusable units of logic for an integrated circuit that may be stored on a tangible machine-readable medium as a hardware model that represents the structure of the integrated circuit. The hardware model may be provided to various customers or manufacturing facilities that load the hardware model into a manufacturing machine that produces the integrated circuit. The integrated circuit may be manufactured such that the circuit performs the operations described in connection with any of the embodiments described herein.

[0321] FIG. 24A is a block diagram illustrating an IP core development system 2400 that may be used to manufacture an integrated circuit that performs operations according to an embodiment. The IP core development system 2400 may be used to generate modular, reusable designs that can be incorporated into a larger design or used to build an entire integrated circuit (e.g., an SoC integrated circuit). A design facility 2430 may generate a software simulation 2410 of the IP core design in a high-level programming language (e.g., C / C++). The software simulation 2410 may be used to design, test, and verify the operation of the IP core using a simulation model 2412. The simulation model 2412 may include functional, behavioral, and / or timing simulation. A register transfer level (RTL) design 2415 may then be created or synthesized from the simulation model 2412. The RTL design 2415 is an abstraction of the behavior of the integrated circuit that models the flow of digital signals between hardware registers, including associated logic that is performed using the modeled digital signals. In addition to the RTL design 2415, a lower level design at the logic or transistor level may also be created, designed, or synthesized. Thus, the specific details of the initial design and simulation may vary.

[0322] The RTL design 2415 or equivalent may be further synthesized at the design facility into a hardware model 2420, which may be in a hardware description language (HDL) or some other representation of the physical design data. The HDL may be further simulated or tested to verify the IP core design. The IP core design may be stored using a non-volatile memory 2440 (e.g., a hard disk, a flash memory, or any non-volatile storage medium) for delivery to a third-party manufacturing facility 2465. Alternatively, the IP core design may be sent over a wired connection 2450 or a wireless connection 2460 (e.g., via the Internet). The manufacturing facility 2465 may then manufacture an integrated circuit that is based at least in part on the IP core design. The manufactured integrated circuit may be configured to perform operations according to at least one embodiment described herein.

[0323] FIG. 24B illustrates a vertical cross-sectional view of an integrated circuit package assembly 2470. The integrated circuit package assembly 2470 illustrates one implementation of one or more processors or accelerator devices described herein. The package assembly 2470 includes multiple units of hardware logic 2472, 2474 coupled to a substrate 2480. The logic 2472, 2474 may be at least partially implemented in configurable logic hardware or fixed function logic hardware and may include one or more portions of any of the processor cores, graphics processors, or other accelerator devices described herein. Each unit of logic 2472, 2474 may be implemented on a semiconductor die and coupled to the substrate 2480 via an interconnect structure 2473. The interconnect structure 2473 may be configured to pass electrical signals between the logic 2472, 2474 and the substrate 2480 and may include interconnects such as, but not limited to, bumps or pillars. The interconnect structure 2473 may be configured to pass electrical signals, such as, for example, input / output (I / O) signals and / or power or ground signals associated with the operation of the logic 2472, 2474. If desired, the substrate 2480 may be an epoxy-based laminate substrate. The substrate 2480 may also include other suitable types of substrates. The package assembly 2470 may be connected to other electrical devices via the package interconnect 2483. The package interconnect 2483 may be coupled to a surface of the substrate 2480 to route electrical signals to other electrical devices, such as a motherboard, other chipset, or multi-chip module.

[0324] The units of logic 2472, 2474 may be electrically coupled to a bridge 2482 configured to pass electrical signals between logic 2472 and 2474. Bridge 2482 may be a high density interconnect structure that provides a route for the electrical signals. Bridge 2482 may include a bridge substrate constructed of glass or a suitable semiconductor material. Electrical routing structures may be formed in the bridge substrate to provide chip-to-chip connections between logic 2472 and 2474.

[0325] Although two units of logic 2472, 2474 and one bridge 2482 are shown, embodiments described herein may include more or fewer logic units on one or more dies. One or more dies may be connected without a bridge or with one or more bridges (as bridge 2482 may be omitted if the logic is included on a single die). Alternatively, multiple dies or units of logic may be connected with one or more bridges. Additionally, multiple logic units, dies, and bridges may be connected together in other possible configurations, including three-dimensional configurations.

[0326] FIG. 24C illustrates a package assembly 2490 including multiple hardware logic chiplet units connected to a substrate 2480 (e.g., a base die). The graphics processing units, parallel processors, and / or compute accelerators described herein may be composed of different silicon chiplets that are manufactured separately. In this context, a chiplet is an at least partially packaged integrated circuit that includes multiple separate logic units and may be assembled with other chiplets into a larger package. A diverse set of chiplets with different IP core logic may be assembled into one device. Furthermore, these chiplets may be integrated into a base die or base chiplet using active interposer technology. The concepts described herein allow interconnection and communication between different types of IP in a GPU. These IP cores can be manufactured using different process technologies and can be configured during manufacturing, thus avoiding the complexities of integrating multiple IPs in the same manufacturing process, especially in a large SoC with several types of IPs. Enabling the use of multiple process technologies improves time to market and provides a cost-effective way to create multiple product SKUs. Additionally, decomposable IP is better suited to individual power gating, allowing components not used for a given workload to be powered down, reducing overall power consumption.

[0327] The hardware logic chiplets may include dedicated hardware logic chiplets 2472, logic or I / O chiplets 2474, and / or memory chiplets 2475. The hardware logic chiplets 2472 and logic or I / O chiplets 2474 may be at least partially implemented in configurable logic or fixed function logic hardware and may include one or more portions of any of the processor cores, graphics processors, parallel processors, or other accelerator devices described herein. The memory chiplets 2475 may be DRAM (e.g., GDDR, HBM) memory or cache (SRAM) memory.

[0328] Each chiplet may be fabricated as a separate semiconductor die and may be coupled to substrate 2480 via interconnect structures 2473. Interconnect structures 2473 may be configured to pass electrical signals between the various chiplets and logic within substrate 2480. Interconnect structures 2473 may include interconnects such as, but not limited to, bumps or pillars. In some embodiments, interconnect structures 2473 may be configured to pass electrical signals, such as, for example, input / output (I / O) signals and / or power or ground signals associated with operation of the logic, I / O, and memory chiplets.

[0329] Substrate 2480 may be an epoxy-based laminate substrate. However, substrate 2480 is not limited thereto and may include other suitable types of substrates. Package assembly 2490 may be connected to other electrical devices via package interconnect 2483. Package interconnect 2483 may be coupled to a surface of substrate 2480 to route electrical signals to other electrical devices such as a motherboard, other chipset, or multi-chip module.

[0330] The logic or I / O chiplets 2474 and the memory chiplets 2475 may be electrically coupled via a bridge 2487 configured to pass electrical signals between the logic or I / O chiplets 2474 and the memory chiplets 2475. The bridge 2487 may be a high density interconnect structure that provides a route for the electrical signals. The bridge 2487 may include a bridge substrate constructed of glass or a suitable semiconductor material. Electrical routing structures may be formed in the bridge substrate to provide inter-chip connections between the logic or I / O chiplets 2474 and the memory chiplets 2475. The bridge 2487 may also be referred to as a silicon bridge or an interconnect bridge. For example, the bridge 2487 is an embedded multi-die interconnect bridge (EMIB). Alternatively, the bridge 2487 may simply be a direct connection from one chiplet to another.

[0331] Substrate 2480 may include hardware components for I / O 2491, cache memory 2492, and other hardware logic 2493. Fabric 2485 may be embedded in substrate 2480 to enable communication between the various logic chiplets and logic 2491, 2493 in substrate 2480. If desired, I / O 2491, fabric 2485, caches, bridges, and other hardware logic 2493 may be integrated into a base die layered over substrate 2480. Fabric 2485 may be a network-on-chip type interconnect or another form of packet-switched fabric that switches data packets between components of a package assembly.

[0332] Additionally, package assembly 2490 may also include a smaller or larger number of components and chiplets interconnected with fabric 2485 or one or more bridges 2487. Chiplets in package assembly 2490 may be arranged in a 3D or 2.5D configuration. In general, bridge structures 2487 may be used to facilitate point-to-point interconnections between, for example, logic or I / O chiplets and memory chiplets. Fabric 2485 may be used to interconnect various logic and / or I / O chiplets (e.g., chiplets 2472, 2474, 2491, 2493) with other logic and / or I / O chiplets. Cache memory 2492 within the substrate may act as a global cache for package assembly 2490, a part of a distributed global cache, or a dedicated cache for fabric 2485.

[0333] FIG. 24D illustrates a package assembly 2494 including a replaceable chiplet 2495, according to one embodiment. The replaceable chiplet 2495 may be assembled into a standard slot of one or more base chiplets 2496, 2498. The base chiplets 2496, 2498 may be coupled via a bridge interconnect 2497, which may be similar to other bridge interconnects described herein, for example, an EMIB. Memory chiplets may also be connected to logic or I / O chiplets via a bridge interconnect. I / O and logic chiplets may communicate via an interconnect fabric. Each base chiplet may support one or more slots in one of the standard formats for logic or I / O, or memory / cache.

[0334] SRAM and power supply circuits may be fabricated in one or both of the base chiplets 2496, 2498, which may be fabricated using a different process technology than the replaceable chiplet 2495 stacked on top of the base chiplet. For example, the base chiplets 2496, 2498 may be fabricated using a large scale process technology and the replaceable chiplets may be fabricated using a small scale process technology. One or more of the replaceable chiplets 2495 may be memory (e.g., DRAM) chiplets. Different memory densities may be selected for the package assembly 2494 based on the power and / or performance targeted for the product using the package assembly 2494. Furthermore, logic chiplets with different numbers and types of functional units may be selected at assembly based on the power and / or performance targeted for the product. Furthermore, chiplets containing different types of IP logic cores may be inserted into the slots of the replaceable chiplets, allowing for a hybrid processor approach where IP blocks from different technologies can be mixed and matched. [Exemplary System-on-Chip Integrated Circuit]

[0335] 25-26B show an example integrated circuit and associated graphics processor that may be made using one or more IP cores. In addition to what is shown, other logic and circuitry may be included, including another graphics processor / core, a peripheral interface controller, or a general-purpose processor core. Each element in FIG. 25-26B that has the same or similar name as an element in any other figure herein represents the same element as in the other figure, may operate or function in a similar manner as described elsewhere herein, may have the same components, and may be connected to other entities, but is not limited to such.

[0336] FIG. 25 is a block diagram illustrating an exemplary system-on-chip integrated circuit 2500 that may be made using one or more IP cores. The exemplary integrated circuit 2500 includes one or more application processors 2505 (e.g., CPUs) and at least one graphics processor 2510, which may be a variation of the graphics processors 1408, 1508, 2510 or any of the graphics processors described herein, or may be used in place of any of the graphics processors described. Thus, any disclosure herein of any feature in combination with a graphics processor also discloses the corresponding combination with the graphics processor 2510, but is not limited to such. The integrated circuit 2500 may further include an image processor 2515 and / or a video processor 2520, any of which may be modular IP cores from the same or multiple different design facilities. The integrated circuit 2500 includes a USB controller 2525, a UART controller 2530, an SPI / SDIO controller 2535, and an I / O controller 2540. 2 S / I 2 The integrated circuits may include peripheral or bus logic including a C controller 2540. Additionally, the integrated circuits may include a display device 2545 coupled to one or both of a High-Definition Multimedia Interface (HDMI®) controller 2550 and a Mobile Industry Processor Interface (MIPI) display interface 2555. Storage may be provided by a flash memory subsystem 2560 including a flash memory and a flash memory controller. A memory interface may be provided via a memory controller 2565 for access to an SDRAM or SRAM memory device. Some integrated circuits further include an embedded security engine 2570.

[0337] 26A-26B are block diagrams illustrating an exemplary graphics processor for use in an SoC, according to an embodiment described herein. The graphics processor may be a variation of graphics processor 1408, 1508, 2510, or any other graphics processor described herein. The graphics processor may be used in place of graphics processor 1408, 1508, 2510, or any of the graphics processors described herein. Thus, disclosure of any feature in combination with graphics processor 1408, 1508, 2510, or any of the graphics processors described herein also discloses the corresponding combination with the graphics processor of FIGS. 26A-26B, but is not so limited. FIG. 26A illustrates an exemplary graphics processor 2610 of a system-on-chip integrated circuit that may be made using one or more IP cores, according to one embodiment. FIG. 26B illustrates another exemplary graphics processor 2640 of a system-on-chip integrated circuit that may be made using one or more IP cores, according to one embodiment. The graphics processor 2610 of FIG. 26A is an example of a low power graphics processor core. The graphics processor 2640 of Figure 26B is an example of a more powerful graphics processor core. For example, each of the graphics processors 2610, 2640 may be a variation of the graphics processor 2510 of Figure 25, as previously described at the beginning of this paragraph.

[0338] As shown in FIG. 26A, the graphics processor 2610 includes a vertex processor 2605 and one or more fragment processors 2615A-2615N (e.g., 2615A, 2615B, 2615C, 2615D, ..., 2615N-1, and 2615N). Because the graphics processor 2610 can execute different shader programs via separate logic, the vertex processor 2605 is optimized to execute the operations of a vertex shader program, and the one or more fragment processors 2615A-2615N execute fragment (e.g., pixel) shading operations for a fragment shader program or a pixel shader program. The vertex processor 2605 performs the vertex processing stage of the 3D graphics pipeline and generates primitive and vertex data. The fragment processors 2615A-2615N use the primitive and vertex data generated by the vertex processor 2605 to generate a frame buffer that is displayed on a display device. The fragment processors 2615A-2615N may be optimized to execute fragment shader programs provided in an OpenGL API, which may be used to perform operations similar to pixel shader programs provided in a Direct 3D API.

[0339] The graphics processor 2610 further includes one or more memory management units (MMUs) 2620A-2620B, caches 2625A-2625B, and circuit interconnects 2630A-2630B. The one or more MMUs (2620A-2620B) provide virtual-to-physical address mapping for the graphics processor 2610, including the vertex processor 2605 and / or fragment processors 2615A-2615N. The graphics processor 2610 may reference vertex data or image / texture data stored in memory in addition to vertex data or image / texture data stored in one or more caches 2625A-2625B. The one or more MMUs (2620A-2620B) may be synchronized with other MMUs in the system, including one or more MMUs associated with one or more application processors 2505, image processor 2515, and / or video processor 2520 of FIG. 25, such that each processor 2505-2520 may participate in a shared or unified virtual memory system. Components of the graphics processor 2610 may correspond to components of other graphics processors described herein. The one or more MMUs (2620A-2620B) may correspond to MMU 245 of FIG. 2C. The vertex processor 2605 and fragment processors 2615A-2615N may correspond to the graphics multiprocessor 234. One or more circuit interconnects 2630A-B enable the graphics processor 2610 to interface with other IP cores in an SoC according to an embodiment, either via the SoC's internal bus or via a direct connection. One or more circuit interconnects 2630A-B may correspond to data crossbar 240 of FIG. 2C. Furthermore, correspondences may be found between similar components of the graphics processor 2610 and the various graphics processor architectures described herein.

[0340] As shown in FIG. 26B, the graphics processor 2640 includes one or more MMUs (2620A-2620B), caches 2625A-2625B, and circuit interconnects 2630A-2630B of the graphics processor 2610 shown in FIG. 26A. The graphics processor 2640 includes one or more shader cores 2655A-2655N (e.g., 2655A, 2655B, 2655C, 2655D, 2655E, 2655F, ..., 2655N-1, and 2655N) that provide a unified shader core architecture in which a single core or type of core can execute all types of programmable shader code, including shader program code for implementing vertex shaders, fragment shaders, and / or compute shaders. The exact number of shader cores present may vary depending on the embodiment and implementation. Furthermore, the graphics processor 2640 includes an inter-core task manager 2645, which acts as a thread dispatcher that dispatches execution threads to one or more shader cores 2655A-2655N and tiling units 2658 to accelerate tiling operations for tile-based rendering, where rendering operations for a scene are subdivided in image space, for example, to exploit local spatial coherence within the scene or to optimize the use of built-in caches. The shader cores 2655A-2655N may correspond, for example, to the graphics multiprocessor 234 as in FIG. 2D, or the graphics multiprocessors 325, 350 of FIGS. 3A and 3B, respectively, or the multi-core group 365A of FIG. 3C.

[0341] [Graphics processor with hybrid floating-point data type dot-product-accumulate instruction]

[0342] Computers use a scientific (binary) representation of numbers, known as floating point in computer science. In scientific notation, a number is represented by a rational number called the mantissa, followed by a multiplication of this mantissa with a base number raised to some power. Floating-point number formats for computers commonly include binary32 (or FP32) and binary64 (or FP64). The number here refers to the total number of digits used to represent the number, with most digits assigned to the mantissa, since it provides the highest precision.

[0343] This variant includes a generic DP2A instruction that performs dot products using the bfloat (BF16) format, which is useful for products that do not support constructs such as matrix multiplication.

[0344] The Bfloat16 format is a shortened 16-bit version of the 32-bit IEEE754 single-precision floating-point format that reserves 8 exponent bits, but reduces the mantissa precision from 24 bits to 8 bits to save memory, bandwidth, and processing resources while still maintaining the same precision. The Bfloat16 format was designed primarily for machine learning applications and proximity sensor computing applications, where precision is needed close to 0, but not so much at the maximum range.

[0345] [DP2A instruction for BF16 data type]

[0346] An embodiment of the present invention includes a generic dot-product-accumulate (DP2A) instruction provided to perform dot products using the Bfloat16 (BF16) format. BF16 is an intermediate format between half-precision floating-point numeric format (FP16) and single-precision floating-point numeric format (FP32). BF16 is 16-bit like FP16, but has the same number of exponent bits as FP32. This simplifies conversion between BF16 and FP32, except for some edge cases involving denormalized numbers. Each number has one sign bit.

[0347] | format | bits | exponent | fraction |

[0348] |--------+------+----------+----------|

[0349] |FP32 | 32 | 8 | 23 |

[0350] | FP16 | 16 | 5 | 10 |

[0351] | BF16 | 16 | 8 | 7 |

[0352] The precision of BF16 is much lower than other formats, closer to 1. The exponent in BF16 has a larger range than the same 16-bit FP16.

[0353] In one embodiment, a generic DP2A instruction is provided to perform a dot product using the BF16 format. The instructions listed below are part of the operations of a processing resource (e.g., a normal part of FPU operations).

[0354] Destination (output 2722) = source 0 operand FP32 + (source 1 operand BF16) x (source 2 operand BF16) + (source 1 operand BF16) x (source 2 operand BF16)

[0355] 27 shows a block diagram of a circuit 2700 for executing this DP2A instruction in BF16 format on a processing resource (e.g., functional unit, FPU) according to one embodiment. This instruction becomes part of the normal FPU operation (not in the systolic pipeline).

[0356] Each of the processing clusters 214 of the parallel processor 200 of Figure 2A may include a floating point unit 264 (FPU). Each of the GPGPU cores 262 of Figure 2D may include a floating point unit 264 (FPU) and / or an integer arithmetic logic unit (ALU) used to execute instructions for the graphics multiprocessor 234. Each of the processing clusters 706A-706H of Figure 7 may include a floating point unit 264 (FPU). A set of SIMD floating point units (FPUs) 1834 is shown in Figure 18B.

[0357] GPGPU core 262 may include SIMD logic capable of executing a single instruction (e.g., a DP2A instruction in BF16 format) on multiple sets of data. If desired, GPGPU core 262 may physically execute SIMD4, SIMD8, and SIMD16 instructions, and logically execute SIMD1, SIMD2, and SIMD32 instructions.

[0358] Circuit 2700 includes multipliers 2710 and 2711 (e.g., BF16 multipliers) that each multiply an input operand (e.g., source 1, source 2). An accumulator or adder 2720 adds a first operand (source 0) and the output from each of multipliers 2710 and 2711. Accumulator or adder 2720 generates an output 2722 for the destination. Circuit 2700 may be repeated multiple times for each lane of the SIMD.

[0359] When the mode feature is set, denormalized operations are supported by BF16: denormalized numbers may be flushed to zero.

[0360] In another embodiment, the destination (output 2722) and source 0 operands may be in FP16, FP32, or BF16 floating-point numeric format.

[0361] A processing resource may represent a processing element associated with a graphics processor (e.g., a GPGPU core, a ray tracing core, a tensor core, an execution resource, an execution unit (EU), a stream processor, a streaming multiprocessor (SM), a graphics multiprocessor), or a graphics processor structure within a GPU as described herein (e.g., a parallel processing unit, a graphics processing engine, a multicore group, a compute unit, a compute unit of a graphics core next). For example, the processing resource may be one of GPGPU cores 262 with FPU, or tensor / ray tracing core 263 of graphics multiprocessor 234, ray tracing cores 338A-338B, tensor cores 337A-337B, or GPGPU cores 336A-336B of graphics multiprocessor 325, execution resources 356A-356D of graphics multiprocessor 350, one of GFX core 370, tensor core 371, or ray tracing core 372 of multi-core groups 365A-365N, one of vector logic units 1563 or scalar logic units 1564 of compute units 1506A-1506N, an execution unit with EU arrays 1522A-1522F or EU arrays 1524A-1524F, execution units 1808A-1808N of execution logic 1800, and / or execution unit 1900. The processing resources may be, for example, execution resources within graphics processing engines 431-432, processing clusters 706A-706H, GPGPUs 806A-806D, GPGPU 1306, graphics processing engine 1610, graphics processing engine cluster 1622, and / or graphics processing engine 1710. The processing resources may be processing resources within graphics processor 2510m, graphics processor 2610, and / or graphics processor 2640.

[0362] In another embodiment, a generic DP2A instruction is provided to perform dot products using the BF16 format. The instructions listed below are part of the normal FPU operations.

[0363] Destination (output 2822) = source 0 operand FP32 + (source 1 operand BF16) x (source 2 operand BF16) + (source 1 operand BF16) x (source 2 operand BF16) + (source 1 operand BF16) x (source 2 operand BF16)

[0364] Figure 28 shows a block diagram of a circuit 2800 for executing this DP2A instruction in BF16 format in an FPU according to another embodiment. This instruction becomes part of the normal FPU operation (not the systolic pipeline). This instruction is useful when the product does not support structures such as large matrices.

[0365] Circuit 2800 includes multipliers 2810, 2811, 2812 (e.g., BF16 multipliers), each of which multiplies multiple input operands (e.g., source 1, source 2, or other input operands). An accumulator or adder 2820 adds a first operand (source 0) and the output from each of multipliers 2810, 2811, and 2812. The accumulator or adder 2820 generates an output 2822 for the destination. Circuit 2800 may be repeated multiple times for each lane of the SIMD.

[0366] Circuit 2800 may be modified to have additional multipliers or fewer multipliers.

[0367] [Cascaded dot product accumulation]

[0368] Circuit 2700 or circuit 2800 may be repeated, with the output of the first stage provided as the input of the second stage. The output of the second stage may be provided as the input of the third stage, and so on.

[0369] In one embodiment, a generic DP2A instruction is provided to perform dot products using the BF16 format. The instructions listed below are part of the normal FPU operations.

[0370] Destination (output 2922) = source 0 operand FP32 + (source 1 operand BF16) x source 2 operand BF16) + (source 1 operand BF16) x (source 2 operand BF16)

[0371] Destination (output 2962) = output 2922 operand FP32 + (source 1 operand BF16) x source 2 operand BF16) + (source 1 operand BF16) x (source 2 operand BF16)

[0372] Figure 29 shows a block diagram of a circuit 2900 for executing this DP2A instruction in BF16 format in an FPU according to one embodiment. This instruction becomes part of the normal FPU operation (not in the systolic pipeline).

[0373] Circuit 2900 includes a first stage and a second stage, although other stages may also be included. First stage 2901 includes multipliers 2910 and 2911 (e.g., BF16 multipliers) that multiply input operands (e.g., source 1, source 2), respectively. An accumulator or adder 2920 in first stage 2901 adds a first operand (source 0) and the outputs from each of multipliers 2910 and 2911. Accumulator or adder 2920 generates a first output 2922.

[0374] The second stage 2902 includes multipliers 2960 and 2961 (e.g., BF16 multipliers) that each multiply an input operand (e.g., source 1, source 2). An accumulator or adder 2960 in the second stage 2902 adds a first operand (output 2922) and the output from each of the multipliers 2960 and 2961. The accumulator or adder 2960 generates a second output 2962 for the destination, or another stage N may follow the second stage.

[0375] While many of the methods have been described in their most basic form, steps could be added to or deleted from any of these methods, and information could be added to or subtracted from any of the messages described, without departing from the basic scope of the present embodiments. It will be apparent to those skilled in the art that many further modifications and adaptations may be made. The detailed embodiments are provided to illustrate concepts, not to limit the concepts. The scope of these embodiments is to be determined only by the claims that follow, rather than by the specific examples provided above.

[0376] When an element "A" is said to be coupled to or with an element "B," the element "A" may be directly coupled to element "B" or indirectly coupled, for example, through element "C." When the specification or claims state that a component, feature, structure, process, or characteristic A results in a component, feature, structure, process, or characteristic B, what is meant is that "A" is at least partially responsible for "B," but there may also be at least one other component, feature, structure, process, or characteristic that helps to result in "B." When the specification or claims indicate that a component, feature, structure, process, or characteristic "may," "might," or "may" be included, it is not necessary that this particular component, feature, structure, process, or characteristic be included. When the specification or claims refer to "one" element, this does not mean that there is only one of the described elements.

[0377] An embodiment is one implementation or example. References in this specification to "one embodiment," "one embodiment," "some embodiments," or "other embodiments" mean that certain features, structures, or characteristics described in connection with these embodiments are included in at least some of the embodiments, but not necessarily all of the embodiments. Although "one embodiment," "one embodiment," or "some embodiments" appear in various places, they do not necessarily all refer to the same embodiment. In the foregoing description of exemplary embodiments, it should be understood that various features may be grouped together in one embodiment, figure, or description thereof for the purpose of streamlining the disclosure and facilitating understanding of one or more of the various novel aspects. However, this method of disclosure should not be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as the following claims reflect, not all features of a single disclosed embodiment described above have novel aspects. Thus, the claims are hereby expressly incorporated into this description, with each claim standing on its own as a separate embodiment.

[0378] Some embodiments relate to Example 1, which includes a graphics multiprocessor having an instruction unit that dispatches instructions and a processing resource coupled to the instruction unit. The processing resource is configured to receive a dot product accumulate instruction from the instruction unit and process the dot product accumulate instruction using a Bfloat16 numeric (BF16) format.

[0379] Example 2 includes the subject matter of example 1, wherein the dot product accumulate instruction causes the first BF16 multiplier and the second BF16 multiplier to multiply the second source operand and the third source operand, respectively, and an accumulator adds the first source operand and outputs from each of the first BF16 multiplier and the second BF16 multiplier.

[0380] Example 3 includes the subject matter of any of Examples 1-2, wherein the accumulator generates an output for a destination.

[0381] Example 4 includes the subject matter of any of Examples 1 to 3, wherein the first source operand has a single-precision floating-point format, and the second source operand and the third source operand have a BF16 format.

[0382] Example 5 includes the subject matter of any of Examples 1 to 4, wherein the first source operand and the destination are in half-precision floating-point format, single-precision floating-point format, or BF16 format.

[0383] Example 6 includes the subject matter of any of examples 1-5, wherein the processing resource includes a floating point unit (FPU) that executes a dot product accumulate instruction using a BF16 format.

[0384] Example 7 includes the subject matter of any of examples 1-6, wherein the instruction unit dispatches instructions including single instruction multiple data (SIMD) instructions.

[0385] Some embodiments relate to example 8, which includes a general purpose graphics processing unit core (GPGPU core) having a single precision floating point unit for single precision floating point operations and a half precision floating point unit for half precision floating point operations. The half precision floating point unit is configured to execute a dot product accumulate instruction using a Bfloat16 (BF16) format.

[0386] Example 9 includes the subject matter of example 8, in which the dot product accumulate instruction causes the first BF16 multiplier and the second BF16 multiplier to multiply the second source operand and the third source operand, respectively, and an accumulator adds the first source operand and outputs from each of the first BF16 multiplier and the second BF16 multiplier.

[0387] Example 10 includes the subject matter of any of Examples 8 to 9, wherein the accumulator generates an output for a destination.

[0388] Example 11 includes the subject matter of any of Examples 8 to 10, wherein the first source operand has a single-precision floating-point format, and the second source operand and the third source operand have a BF16 format.

[0389] Example 12 includes the subject matter of any of Examples 8 to 11, wherein the first source operand and the destination are in half-precision floating-point format, single-precision floating-point format, or BF16 format.

[0390] Example 13 includes the subject matter of any of Examples 8-12, wherein the dot product accumulate instruction causes the first BF16 multiplier and the second BF16 multiplier of the first stage to multiply the second source operand and the third source operand, respectively, and the accumulator adds the first source operand and outputs from the first BF16 multiplier and the second BF16 multiplier, respectively, to generate an output of the first stage.

[0391] Example 14 includes the subject matter of any of Examples 8-12, wherein a dot product accumulate instruction for a cascade configuration having N stages of multipliers and accumulators causes a first BF16 multiplier and a second BF16 multiplier in the second stage to multiply a second source operand and a third source operand, respectively, and an accumulator adds an output from the first stage and an output from each of the first BF16 multiplier and the second BF16 multiplier in the second stage.

[0392] Example 15 relates to some embodiments including a parallel processing unit having a first processing cluster for performing parallel processing operations and a second processing cluster coupled to the first processing cluster. The first processing cluster includes a floating point unit for performing floating point operations. The floating point unit is configured to process a dot product accumulate instruction using a Bfloat16 (BF16) format.

[0393] Example 16 includes the subject matter of example 15, in which the dot product accumulate instruction causes the first BF16 multiplier and the second BF16 multiplier to multiply the second source operand and the third source operand, respectively, and the accumulator adds the first source operand and outputs from each of the first BF16 multiplier and the second BF16 multiplier.

[0394] Example 17 includes the subject matter of any of examples 15-16, wherein the accumulator generates an output for a destination.

[0395] Example 18 includes the subject matter of any of Examples 15 to 17, wherein the first source operand has a single-precision floating-point format, and the second source operand and the third source operand have a BF16 format.

[0396] Example 19 includes the subject matter of any of Examples 15 to 18, wherein the first source operand and the destination are in half-precision floating-point format, single-precision floating-point format, or BF16 format.

[0397] Example 20 includes the subject matter of any of Examples 15-19, wherein the dot product accumulate instruction causes the first BF16 multiplier and the second BF16 multiplier of the first stage to multiply the second source operand and the third source operand, respectively, and the accumulator adds the first source operand and the output from each of the first BF16 multiplier and the second BF16 multiplier; and the dot product accumulate instruction for a cascade configuration having N stages of multipliers and accumulators causes the first BF16 multiplier and the second BF16 multiplier of the second stage to multiply the second source operand and the third source operand, respectively, and the accumulator adds the output from the first stage and the output from each of the first BF16 multiplier and the second BF16 multiplier of the second stage.

[0398] Example 21 includes the subject matter of any of examples 15-20, wherein the floating point unit has a cascade configuration with N stages of BF16 multipliers and accumulators.

[0399] Example 22 includes the subject matter of any of Examples 15 to 21, and the N stages include: a first stage having a first BF16 multiplier and a second BF16 multiplier that respectively multiply a second source operand and a third source operand, and an accumulator that adds the first source operand and outputs from the first BF16 multiplier and the second BF16 multiplier to generate an output of the first stage; and a second stage having a first BF16 multiplier and a second BF16 multiplier that respectively multiply the second source operand and a third source operand, and an accumulator that adds the output from the first stage and outputs from the first BF16 multiplier and the second BF16 multiplier of the second stage.

[0400] The foregoing description and drawings should be treated in an illustrative, and not a restrictive, sense. Those skilled in the art will understand that various modifications and changes may be made to the embodiments described herein without departing from the broader spirit and scope of the invention as set forth in the appended claims. (Item 1) 1. A graphics processing unit comprising a single instruction multiple thread (SIMT) multiprocessor, comprising: The SIMT multiprocessor includes: An instruction cache; a shared memory coupled to the instruction cache; a circuit coupled to the shared memory and the instruction cache, the circuit comprising: A plurality of texture units; a first core including hardware for accelerating matrix operations; A second core, receiving an instruction having a plurality of operands in a Bfloat16 (BF16) numeric format, the plurality of operands including a first source operand, a second source operand, and a third source operand, the BF16 numeric format being a 16-bit floating-point format with an 8-bit exponent; processing the instruction, the processing comprising multiplying the second source operand and the third source operand and adding the first source operand to a result of the multiplication; a second core configured to A circuit including: A graphics processing unit having (Item 2) 2. The graphics processing unit of claim 1, wherein the instruction is a single instruction, multiple data (SIMD) instruction. (Item 3) 3. The graphics processing unit of claim 2, wherein the instruction causes the second core to apply a normalized linear unit function to a result of the addition. (Item 4) 4. A graphics processing unit according to any one of items 1 to 3, wherein the SIMT multiprocessor executes threads of one warp in response to the instruction. (Item 5) 5. A graphics processing unit according to any one of items 1 to 4, wherein the SIMT multiprocessor performs a parallel matrix multiplication operation via the first core, the parallel matrix multiplication operation being performed on an input having the BF16 numeric format. (Item 6) 6. The graphics processing unit of claim 1, further comprising a third core including hardware for accelerating ray tracing operations. (Item 7) 7. A graphics processing unit according to any one of claims 1 to 6, wherein the graphics processing unit further comprises a texture processing circuit external to the SIMT multiprocessor and coupled to the SIMT multiprocessor. (Item 8) 8. The graphics processing unit of claim 1, wherein the instruction causes the second core to perform a dot product operation. (Item 9) fetching an instruction from an instruction cache of a graphics processing unit (GPU), the instruction having a plurality of operands in a Bfloat16 (BF16) numeric format, the plurality of operands including a first source operand, a second source operand, and a third source operand, the BF16 numeric format being a 16-bit floating-point format with an 8-bit exponent, the GPU including: a shared memory coupled to the instruction cache; and circuitry coupled to the shared memory and the instruction cache; In response to the instruction, dispatching threads of a warp to a single instruction multiple thread (SIMT) multiprocessor of the GPU, the SIMT multiprocessor including a first core including multiple texture units, hardware for accelerating matrix operations, and a second core configured to execute the threads of the instruction; processing the instruction using the second core, the processing including multiplying the second source operand by the third source operand and adding a first source operand to a result of the multiplication; A method for providing the above. (Item 10) 10. The method of claim 9, wherein the instruction is a single instruction, multiple data (SIMD) instruction. (Item 11) Item 11. The method of item 10, further comprising applying, via the second core, a normalized linear unit function to a result of the addition. (Item 12) 12. The method of any one of claims 9 to 11, further comprising performing a parallel matrix multiplication operation on an input having the BF16 numeric format via the first core. (Item 13) 13. The method according to any one of claims 9 to 12, wherein the SIMT multiprocessor includes a third core that accelerates ray tracing operations, and the method further comprises accelerating ray tracing operations in parallel with processing the instructions via the third core. (Item 14) 14. The method according to any one of claims 9 to 13, further comprising performing texture processing operations via a texture processing circuit external to and coupled to the SIMT multiprocessor. (Item 15) 15. The method of any one of claims 9 to 14, further comprising performing a dot product operation via the second core in response to the instruction. (Item 16) A memory device; a graphics processor coupled to the memory device, the graphics processor having a single instruction multiple threads (SIMT) multiprocessor including an instruction cache, a shared memory coupled to the instruction cache, and circuitry coupled to the shared memory and the instruction cache; 1. A graphics processing system comprising: The circuit comprises: A plurality of texture units; a first core including hardware for accelerating matrix operations; a second core configured to receive an instruction having a plurality of operands A, B, C, and D in a Bfloat16 (BF16) numeric format, the BF16 numeric format being a 16-bit floating-point format with an 8-bit exponent, and to process the instruction, the processing of the instruction including performing an operation D=A×B+C; a graphics processing system including (Item 17) Item 17. The graphics processing system of item 16, wherein the instruction is a single instruction multiple data (SIMD) instruction. (Item 18) Item 18. The graphics processing system of item 17, wherein the second core is further configured to apply a normalized linear unit function to a result of the operation D=A×B+C. (Item 19) 19. The graphics processing system of any one of items 16 to 18, wherein the SIMT multiprocessor executes threads of one warp in response to the instruction. (Item 20) 20. The graphics processing system of any one of items 16 to 19, wherein the SIMT multiprocessor performs a parallel matrix multiplication operation via the first core, the parallel matrix multiplication operation being performed on an input having the BF16 numeric format. (Item 21) 21. The graphics processing system of any one of items 16 to 20, further comprising a third core including hardware for accelerating ray tracing operations. (Item 22) 22. The graphics processing system of any one of items 16 to 21, further comprising a texture processing circuit external to and coupled to the SIMT multiprocessor. (Item 23) 23. The graphics processing system of any one of items 16 to 22, wherein the instruction causes the second core to perform a dot product operation. (Item 24) a first processing cluster for performing parallel processing operations, the parallel processing operations including ray tracing operations and matrix multiplication operations; a second processing cluster coupled to the first processing cluster, the first processing cluster including a floating point unit configured to perform floating point operations, the floating point unit configured to process instructions using a Bfloat16 (BF16) format, a BF16 multiplier multiplying a second source operand and a third source operand, and an accumulator adding an output from the BF16 multiplier and the first source operand; A parallel processing unit comprising: (Item 25) 25. The parallel processing unit of claim 24, wherein the instruction causes the BF16 multiplier to multiply a second source operand and a third source operand having the BF16 format, and the accumulator adds an output from the BF16 multiplier and a first source operand. (Item 26) 26. The parallel processing unit of claim 24 or 25, wherein the first source operand has a single-precision floating-point format, and the second source operand and the third source operand have a BF16 format. (Item 27) A program for causing a machine to execute the method according to any one of items 9 to 15. (Item 28) 28. A machine-readable storage medium storing the program according to item 27. (Item 29) 16. An apparatus comprising means for carrying out the method according to any one of items 9 to 15.

Claims

1. a first processing cluster including a plurality of processing resources for performing parallel processing operations, the parallel processing operations including ray tracing operations and matrix multiplication operations, the plurality of processing resources being interconnected via a data interconnect, the first processing cluster including a ray tracing circuit for performing the ray tracing operations and a first matrix processing circuit for performing the matrix multiplication operations; a second processing cluster coupled to the first processing cluster, the second processing cluster including a second matrix processing circuit including a floating point unit for performing floating point operations associated with the matrix multiplication operation, the floating point unit configured to process instructions using a Bfloat16 (BF16) format, a multiplier multiplying a second source operand and a third source operand, and an accumulator adding an output from the multiplier and the first source operand; A graphics processor comprising:

2. 2. The graphics processor of claim 1, wherein the instruction causes the multiplier to multiply a second source operand and a third source operand having the BF16 format, and the accumulator adds an output from the multiplier and a first source operand.

3. 3. The graphics processor of claim 2, wherein the first source operand has a single precision floating point format, and the second source operand and the third source operand have the BF16 format.

4. 2. The graphics processor of claim 1, wherein the ray tracing circuitry is configured to perform the ray tracing operations in response to a request received from a processing resource of the first processing cluster.

5. The graphics processor of claim 1 , wherein each of the first matrix processing circuit and the second matrix processing circuit includes a systolic array.

6. fetching a first instruction from an instruction cache of a graphics processing unit (GPU), the first instruction being a single instruction, multiple data (SIMD) instruction having multiple operands, at least one of the multiple operands being in a Bfloat16 (BF16) numeric format, the multiple operands including a first source operand, a second source operand, and a third source operand, the first source operand having a single precision floating-point format, the second source operand and the third source operand having the BF16 numeric format, the GPU including: a shared memory coupled to the instruction cache; and circuitry coupled to the shared memory and the instruction cache; In response to the first instruction, dispatching a thread of one warp to processing resources of the GPU, the processing resources including a first core including hardware for accelerating matrix operations and a second core configured to execute the thread of the instruction; executing the thread of the first instruction using the second core, the executing the thread of the first instruction including multiplying elements of the second source operand by elements of the third source operand and adding elements of the first source operand to a result of the multiplication; A method for providing the above.

7. 7. The method of claim 6, further comprising performing, via the first core, a parallel matrix multiplication operation on an input having the BF16 numeric format, the parallel matrix multiplication operation being performed in response to a second instruction, the second instruction being executed contemporaneously with the first instruction.

8. 8. The method of claim 7, wherein the processing resource is a single instruction multiple thread (SIMT) multiprocessor, the method comprising dispatching threads of the one warp to the SIMT multiprocessor.

9. 9. The method of claim 8, wherein the SIMT multiprocessor includes a third core that accelerates ray tracing operations, the method further comprising accelerating ray tracing operations in parallel with execution of the first instruction and the second instruction via the third core.

10. 9. The method of claim 8, further comprising performing texture processing operations via texture processing circuitry external to and coupled with the SIMT multiprocessor.

11. The method of claim 8 , further comprising: performing a dot product operation via the second core based on the first instruction.

12. 10. The method of claim 8, further comprising: performing a dot product operation via the first core based on a third instruction.

13. a memory device configured to store instructions; a graphics processor coupled to the memory device, a first processing cluster including a plurality of processing resources for performing parallel processing operations, the parallel processing operations including ray tracing operations and matrix multiplication operations, the plurality of processing resources being interconnected via a data interconnect, the first processing cluster including a ray tracing circuit for performing the ray tracing operations and a first matrix processing circuit for performing the matrix multiplication operations; a second processing cluster coupled to the first processing cluster, the second processing cluster including a second matrix processing circuit including a floating point unit for performing floating point operations associated with the matrix multiplication operation, the floating point unit configured to process instructions using a Bfloat16 (BF16) format, a multiplier multiplying a second source operand and a third source operand, and an accumulator adding an output from the multiplier and the first source operand; the graphics processor having A data processing system comprising:

14. 14. The data processing system of claim 13, wherein said instruction causes said multiplier to multiply a second source operand and a third source operand having said BF16 format, and said accumulator adds an output from said multiplier and a first source operand.

15. 15. The data processing system of claim 14, wherein the first source operand has a single precision floating point format, and the second source operand and the third source operand have the BF16 format.

16. 14. The data processing system of claim 13, wherein the ray tracing circuitry is configured to perform the ray tracing operation in response to a request received from a processing resource of the first processing cluster.

Images