COMPUTE OPTIMIZATIONS FOR LOW PRECISION MACHINE LEARNING OPERATIONS

Abstract

One embodiment provides a multi-chip module accelerator usable to execute tensor data processing operations a multi-chip module. The multi-chip module may include a memory stack including multiple memory dies and parallel processor circuitry communicatively coupled to the memory stack. The parallel processor circuitry may include multiprocessor cores to execute matrix multiplication and accumulate operations. The matrix multiplication and accumulate operations may include floating-point operations that are configurable to include two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions. The floating-point operations may include a first operation at a first precision and a second operation at a second precision. The first operation may include a multiply having at least one 16-bit floating-point input and the second operation may include an accumulate having a 32-bit floating-point input.

Description

FIELD

Embodiments relate generally to data processing and more particularly to data processing via a general-purpose graphics processing unit.

BACKGROUND OF THE DESCRIPTION

Current parallel graphics data processing includes systems and methods developed to perform specific operations on graphics data such as, for example, linear interpolation, tessellation, rasterization, texture mapping, depth testing, etc. Traditionally, graphics processors used fixed function computational units to process graphics data; however, more recently, portions of graphics processors have been made programmable, enabling such processors to support a wider variety of operations for processing vertex and fragment data.

To further increase performance, graphics processors typically implement processing techniques such as pipelining that attempt to process, in parallel, as much graphics data as possible throughout the different parts of the graphics pipeline. Parallel graphics processors with single instruction, multiple thread (SIMT) architectures are designed to maximize the amount of parallel processing in the graphics pipeline. In an SIMT architecture, groups of parallel threads attempt to execute program instructions synchronously together as often as possible to increase processing efficiency. A general overview of software and hardware for SIMT architectures can be found in Shane Cook, CUDA Programming, Chapter 3, pages 37-51 (2013) and/or Nicholas Wilt, CUDA Handbook, A Comprehensive Guide to GPU Programming, Sections 2.6.2 to 3.1.2 (June 2013).

BRIEF DESCRIPTION OF THE DRAWINGS

So that the features of the present invention can be understood in detail, a more particular description of the invention may be had by reference to embodiments, some of which are illustrated in the appended drawings. It is to be noted, however, that the appended drawings illustrate only typical embodiments and arm therefore not to be considered limiting of the scope of all embodiments.

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

FIG. 2A-2D illustrate parallel processor components, according to an embodiment;

FIG. 3A-3B are block diagrams of graphics multiprocessors, according to embodiments;

FIG. 4A-4F illustrate an exemplary architecture in which a plurality of GPUs are communicatively coupled to a plurality of multi-core processors;

FIG. 5 illustrates a graphics processing pipeline, according to an embodiment;

FIG. 6 illustrates a machine learning software stack, according to an embodiment;

FIG. 7 illustrates a highly-parallel general-purpose graphics processing unit, according to an embodiment;

FIG. 8 illustrates a multi-GPU computing system, according to an embodiment;

FIG. 9A-9B illustrate layers of exemplary deep neural networks;

FIG. 10 illustrates an exemplary recurrent neural network;

FIG. 11 illustrates training and deployment of a deep neural network;

FIG. 12 is a block diagram illustrating distributed learning;

FIG. 13 illustrates an exemplary inferencing system on a chip (SOC) suitable for performing inferencing using a trained model;

FIG. 14 illustrates components of a dynamic precision floating point unit, according to an embodiment;

FIG. 15 provides additional details with respect to a dynamic precision floating point unit, according to an embodiment;

FIG. 16 illustrates thread assignments for a dynamic precision processing system, according to an embodiment;

FIG. 17 illustrates logic to perform a numerical operation at less than a requested precision, according to an embodiment;

FIG. 18 illustrates loop vectorization for SIMD units, according to an embodiment;

FIG. 19 illustrates a thread processing system, according to an embodiment;

FIG. 20 illustrates logic to assign threads for computation, according to an embodiment;

FIG. 21 illustrates a deep neural network that may be processed using compute logic provided by embodiments described herein;

FIG. 22 is a flow diagram of logic to prevent error or significant precision loss when performing low precision operations for machine learning, according to an embodiment;

FIG. 23 is a block diagram of a processing system, according to an embodiment;

FIG. 24 is a block diagram of an embodiment of a processor having one or more processor cores, an integrated memory controller, and an integrated graphics processor;

FIG. 25 is a block diagram of a graphics processor, which may be a discrete graphics processing unit, or may be a graphics processor integrated with a plurality of processing cores;

FIG. 26 is a block diagram of a graphics processing engine of a graphics processor in accordance with some embodiments;

FIG. 27 is a block diagram of a graphics processor provided by an additional embodiment;

FIG. 28 illustrates thread execution logic including an array of processing elements employed in some embodiments;

FIG. 29 is a block diagram illustrating graphics processor instruction formats according to some embodiments;

FIG. 30 is a block diagram of a graphics processor according to another embodiment.

FIG. 31A-31B illustrate a graphics processor command format and command sequence, according to some embodiments;

FIG. 32 illustrates exemplary graphics software architecture for a data processing system according to some embodiments;

FIG. 33 is a block diagram illustrating an IP core development system, according to an embodiment;

FIG. 34 is a block diagram illustrating an exemplary system on a chip integrated circuit, according to an embodiment;

FIG. 35 is a block diagram illustrating an additional graphics processor, according to an embodiment; and

FIG. 36 is a block diagram illustrating an additional exemplary graphics processor of a system on a chip integrated circuit, according to an embodiment.

DETAILED DESCRIPTION

In some embodiments, a graphics processing unit (GPU) is communicatively coupled to host/processor cores to accelerate graphics operations, machine-learning operations, pattern analysis operations, and various general purpose GPU (GPGPU) functions. The GPU may be communicatively coupled to the host processor/cores over a bus or another interconnect (e.g., a high-speed interconnect such as PCIe or NVLink). In other embodiments, the GPU may be integrated on the same package or chip as the cores and communicatively coupled to the cores over 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 cores may allocate work to the GPU in the form of sequences of commands/instructions contained in a work descriptor. The GPU then uses dedicated circuitry/logic for efficiently processing these commands/instructions.

In the following description, numerous specific details are set forth to provide a more thorough understanding. However, it will be apparent to one of 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 to avoid obscuring the details of the present embodiments.

System Overview

FIG. 1 is a block diagram illustrating a computing system 100 configured to implement one or more aspects of the embodiments described herein. The computing system 100 includes a processing subsystem 101 having one or more processor(s) 102 and a system memory 104 communicating via an interconnection path that may include a memory hub 105. The memory hub 105 may be a separate component within a chipset component or may be integrated within the one or more processor(s) 102. The memory hub 105 couples with an I/O subsystem 111 via a communication link 106. The I/O subsystem 111 includes an I/O hub 107 that can enable the computing system 100 to receive input from one or more input device(s) 108. Additionally, the I/O hub 107 can enable a display controller, which may be included in the one or more processor(s) 102, to provide outputs to one or more display device(s) 110A. In one embodiment the one or more display device(s) 110A coupled with the I/O hub 107 can include a local, internal, or embedded display device.

In one embodiment the processing subsystem 101 includes one or more parallel processor(s) 112 coupled to 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 communications interface or communications fabric. In one embodiment the one or more parallel processor(s) 112 form a computationally focused parallel or vector processing system that can include a large number of processing cores and/or processing clusters, such as a many integrated core (MIC) processor. In one embodiment the one or more parallel processor(s) 112 form a graphics processing subsystem that can output pixels to one of the one or more display device(s) 110A coupled via the I/O hub 107. The one or more parallel processor(s) 112 can also include a display controller and display interface (not shown) to enable a direct connection to one or more display device(s) 110B.

Within the I/O subsystem 111, a system storage unit 114 can connect to the I/O hub 107 to provide a storage mechanism for the computing system 100. An I/O switch 116 can be used to provide an interface mechanism to enable connections between the I/O hub 107 and other components, such as a network adapter 118 and/or wireless network adapter 119 that may be integrated into the platform, and various other devices that can be added via one or more add-in device(s) 120. The network adapter 118 can be an Ethernet adapter or another wired network adapter. The wireless network adapter 119 can include one or more of a Wi-Fi, Bluetooth, near field communication (NFC), or other network device that includes one or more wireless radios.

The computing system 100 can include other components not explicitly shown, including USB or other port connections, optical storage drives, video capture devices, and the like, may also be connected to the I/O hub 107. Communication paths interconnecting the various components in FIG. 1 may be implemented using any suitable protocols, such as PCI (Peripheral Component Interconnect) based protocols (e.g., PCI-Express), or any other bus or point-to-point communication interfaces and/or protocol(s), such as the NV-Link high-speed interconnect, or interconnect protocols known in the art.

In one embodiment, the one or more parallel processor(s) 112 incorporate circuitry optimized for graphics and video processing, including, for example, video output circuitry, and constitutes a graphics processing unit (GPU). In another embodiment, the one or more parallel processor(s) 112 incorporate circuitry optimized for general purpose processing, while preserving the underlying computational architecture, described in greater detail herein. In yet another embodiment, components of the computing system 100 may be integrated with one or more other system elements on a single integrated circuit. For example, the one or more parallel processor(s) 112, memory hub 105, processor(s) 102, and I/O hub 107 can be integrated into a system on chip (SoC) integrated circuit. Alternatively, the components of the computing system 100 can be integrated into a single package to form a system in package (SIP) configuration. In one embodiment at least a portion of the components of the computing system 100 can be integrated into a multi-chip module (MCM), which can be interconnected with other multi-chip modules into a modular computing system.

It will be appreciated that the computing system 100 shown herein is illustrative and that variations and modifications are possible. The connection topology, including the number and arrangement of bridges, the number of processor(s) 102, and the number of parallel processor(s) 112, may be modified as desired. For instance, in some embodiments, system memory 104 is connected to the processor(s) 102 directly rather than through a bridge, while other devices communicate with system memory 104 via the memory hub 105 and the processor(s) 102. In other alternative topologies, the parallel processor(s) 112 are connected to the I/O hub 107 or directly to one of the one or more processor(s) 102, rather than to the memory hub 105. In other embodiments, the I/O hub 107 and memory hub 105 may be integrated into a single chip. Some embodiments may include two or more sets of processor(s) 102 attached via multiple sockets, which can couple with two or more instances of the parallel processor(s) 112.

Some of the particular components shown herein are optional and may not be included in all implementations of the computing system 100. For example, any number of add-in cards or peripherals may be supported, or some components may be eliminated. Furthermore, some architectures may use different terminology for components similar to those illustrated in FIG. 1. For example, the memory hub 105 may be referred to as a Northbridge in some architectures, while the I/O hub 107 may be referred to as a Southbridge.

FIG. 2A illustrates a parallel processor 200, according to an embodiment. The various components of the parallel processor 200 may be implemented using one or more integrated circuit devices, such as programmable processors, application specific integrated circuits (ASICs), or field programmable gate arrays (FPGA). The illustrated parallel processor 200 is a variant of the one or more parallel processor(s) 112 shown in FIG. 1, according to an embodiment.

In one embodiment the parallel processor 200 includes a parallel processing unit 202. The parallel processing unit includes an I/O unit 204 that enables communication with other devices, including other instances of the parallel processing unit 202. The I/O unit 204 may be directly connected to other devices. In one embodiment the I/O unit 204 connects with other devices via the use of a hub or switch interface, such as memory hub 105. The connections between the memory hub 105 and the I/O unit 204 form a communication link 113. Within the parallel processing unit 202, the I/O unit 204 connects with a host interface 206 and a memory crossbar 216, where the host interface 206 receives commands directed to performing processing operations and the memory crossbar 216 receives commands directed to performing memory operations.

When the host interface 206 receives a command buffer via the I/O unit 204, the host interface 206 can direct work operations to perform those commands to a front end 208. In one embodiment the front end 208 couples with a scheduler 210, which is configured to distribute commands or other work items to a processing cluster array 212. In one embodiment the scheduler 210 ensures that the processing cluster array 212 is properly configured and in a valid state before tasks are distributed to the processing clusters of the processing cluster array 212. In one embodiment the scheduler 210 is implemented via firmware logic executing on a microcontroller. The microcontroller implemented scheduler 210 is configurable to perform complex scheduling and work distribution operations at coarse and fine granularity, enabling rapid preemption and context switching of threads executing on the processing cluster array 212. In one embodiment, the host software can prove workloads for scheduling on the processing cluster array 212 via one of multiple graphics processing doorbells. The workloads can then be automatically distributed across the processing cluster array 212 by the scheduler 210 logic within the scheduler microcontroller.

The processing cluster array 212 can include up to “N” processing clusters (e.g., cluster 214A, cluster 214B, through cluster 214N). Each cluster 214A-214N of the processing cluster array 212 can execute a large number of concurrent threads. The scheduler 210 can allocate work to the clusters 214A-214N of the processing cluster array 212 using various scheduling and/or work distribution algorithms, which may vary depending on the workload arising for each type of program or computation. The scheduling can be handled dynamically by the scheduler 210, or can be assisted in part by compiler logic during compilation of program logic configured for execution by the processing cluster array 212. In one embodiment, different clusters 214A-214N of the processing cluster array 212 can be allocated for processing different types of programs or for performing different types of computations.

The processing cluster array 212 can be configured to perform various types of parallel processing operations. In one embodiment the processing cluster array 212 is configured to perform general-purpose parallel compute operations. For example, the processing cluster array 212 can include logic to execute processing tasks including filtering of video and/or audio data, performing modeling operations, including physics operations, and performing data transformations.

In one embodiment the processing cluster array 212 is configured to perform parallel graphics processing operations. In embodiments in which the parallel processor 200 is configured to perform graphics processing operations, the processing cluster array 212 can include additional 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 can be configured to execute graphics processing related shader programs such as, but not limited to vertex shaders, tessellation shaders, geometry shaders, and pixel shaders. The parallel processing unit 202 can transfer data from system memory via the I/O unit 204 for processing. During processing the transferred data can be stored to on-chip memory (e.g., parallel processor memory 222) during processing, then written back to system memory.

In one embodiment, when the parallel processing unit 202 is used to perform graphics processing, the scheduler 210 can be configured to divide the processing workload into approximately equal sized tasks, to better enable distribution of the graphics processing operations to multiple clusters 214A-214N of the processing cluster array 212. In some embodiments, portions of the processing cluster array 212 can be configured to perform different types of processing. For example 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, to produce a rendered image for display. Intermediate data produced by one or more of the clusters 214A-214N may be stored in buffers to allow the intermediate data to be transmitted between clusters 214A-214N for further processing.

During operation, the processing cluster array 212 can receive processing tasks to be executed via the scheduler 210, which receives commands defining processing tasks from front end 208. For graphics processing operations, processing tasks can include indices of data to be processed, e.g., surface (patch) data, primitive data, vertex data, and/or pixel data, as well as state parameters and commands defining how the data is to be processed (e.g., what program is to be executed). The scheduler 210 may be configured to fetch the indices corresponding to the tasks or may receive the indices from the front end 208. The front end 208 can be configured to ensure the processing cluster array 212 is configured to a valid state before the workload specified by incoming command buffers (e.g., batch-buffers, push buffers, etc.) is initiated.

Each of the one or more instances of the parallel processing unit 202 can couple with parallel processor memory 222. The parallel processor memory 222 can be accessed via the memory crossbar 216, which can receive memory requests from the processing cluster array 212 as well as 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 multiple partition units (e.g., partition unit 220A, partition unit 220B, through partition unit 220N) that can each couple to a portion (e.g., memory unit) of parallel processor memory 222. In one implementation the number of partition units 220A-220N is configured to be equal to the number of memory units, such that a first partition unit 220A has a corresponding first memory unit 224A, a second partition unit 220B has a corresponding memory unit 224B, and an 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.

In various embodiments, the memory units 224A-224N can 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. In one embodiment, the memory units 224A-224N may also include 3D stacked memory, including but not limited to high bandwidth memory (HBM). Persons skilled in the art will appreciate that the specific implementation of the memory units 224A-224N can vary and can be selected from one of various conventional designs. Render targets, such as frame buffers or texture maps may be stored across the memory units 224A-224N, allowing partition units 220A-220N to write portions of each render target in parallel to efficiently use the available bandwidth of parallel processor memory 222. In some embodiments, a local instance of the parallel processor memory 222 may be excluded in favor of a unified memory design that utilizes system memory in conjunction with local cache memory.

In one embodiment, any one of the clusters 214A-214N of the processing cluster array 212 can process data that will be written to any of the memory units 224A-224N within parallel processor memory 222. The memory crossbar 216 can be configured to transfer the output of each cluster 214A-214N to any partition unit 220A-220N or to another cluster 214A-214N, which can perform additional processing operations on the output. Each cluster 214A-214N can communicate with the memory interface 218 through the memory crossbar 216 to read from or write to various external memory devices. In one embodiment the memory crossbar 216 has a connection to the memory interface 218 to communicate with the I/O unit 204, as well as a connection to a local instance of the parallel processor memory 222, enabling the processing units within the different processing clusters 214A-214N to communicate with system memory or other memory that is not local to the parallel processing unit 202. In one embodiment the memory crossbar 216 can use virtual channels to separate traffic streams between the clusters 214A-214N and the partition units 220A-220N.

While a single instance of the parallel processing unit 202 is illustrated within the parallel processor 200, any number of instances of the parallel processing unit 202 can be included. For example, multiple instances of the parallel processing unit 202 can be provided on a single add-in card, or multiple add-in cards can be interconnected. The different instances of the parallel processing unit 202 can be configured to inter-operate even if the different instances have different numbers of processing cores, different amounts of local parallel processor memory, and/or other configuration differences. For example, in one embodiment some instances of the parallel processing unit 202 can include higher precision floating point units relative to other instances. Systems incorporating one or more instances of the parallel processing unit 202 or the parallel processor 200 can be implemented in a variety of configurations and form factors, including but not limited to desktop, laptop, or handheld personal computers, servers, workstations, game consoles, and/or embedded systems.

FIG. 2B is a block diagram of a partition unit 220, according to an embodiment. In one embodiment the partition unit 220 is an instance of one of the partition units 220A-220N of FIG. 2A. As illustrated, the partition unit 220 includes an L2 cache 221, a frame buffer interface 225, and a ROP 226 (raster operations unit). The L2 cache 221 is a read/write cache that is configured to perform load and store operations received from the memory crossbar 216 and ROP 226. Read misses and urgent write-back requests are output by L2 cache 221 to frame buffer interface 225 for processing. Updates can also be sent to the frame buffer via the frame buffer interface 225 for processing. In one embodiment the frame buffer interface 225 interfaces with one of the memory units in parallel processor memory, such as the memory units 224A-224N of FIG. 2A (e.g., within parallel processor memory 222).

In graphics applications, the ROP 226 is a processing unit that performs raster operations such as stencil, z test, blending, and the like. The ROP 226 then outputs processed graphics data that is stored in graphics memory. In some embodiments the ROP 226 includes compression logic to compress data (e.g., depth, color, etc.) that is written to memory and decompress depth or color data that is read from memory. The compression logic can be lossless compression logic that makes use of one or more of multiple compression algorithms. The type of compression that is performed by the ROP 226 can vary based on the statistical characteristics of the data to be compressed. For example, in one embodiment, delta color compression is performed on depth and color data on a per-tile basis.

In some embodiments, the ROP 226 is included within each processing cluster (e.g., cluster 214A-214N of FIG. 2A) instead of within the partition unit 220. In such embodiment, read and write requests for pixel data are transmitted over the memory crossbar 216 instead of pixel fragment data. The processed graphics data may be displayed on a display device, such as one of the one or more display device(s) 110A-110B of FIG. 1, routed for further processing by the processor(s) 102, or routed for further processing by one of the processing entities within the parallel processor 200 of FIG. 2A.

FIG. 2C is a block diagram of a processing cluster 214 within a parallel processing unit, according to an embodiment. In one embodiment the processing cluster is an instance of one of the processing clusters 214A-214N of FIG. 2A. The processing cluster 214 can be configured to execute many threads in parallel, where the term “thread” refers to an instance of a particular program executing on a particular set of input data. In some embodiments, single-instruction, multiple-data (SIMD) instruction issue techniques are used to support parallel execution of a large number of threads without providing multiple independent instruction units. In other embodiments, single-instruction, multiple-thread (SIMT) techniques are used to support parallel execution of a large number of generally synchronized threads, using a common instruction unit configured to issue instructions to a set of processing engines within each one of the processing clusters. Unlike a SIMD execution regime, where all processing engines typically execute identical instructions, SIMT execution allows different threads to more readily follow divergent execution paths through a given thread program. Persons skilled in the art will understand that a SIMD processing regime represents a functional subset of a SIMT processing regime.

Operation of the processing cluster 214 can be controlled via a pipeline manager 232 that distributes processing tasks to SIMT parallel processors. The pipeline manager 232 receives instructions from the scheduler 210 of FIG. 2A and manages execution of those instructions via a graphics multiprocessor 234 and/or a 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 differing architectures may be included within the processing cluster 214. One or more instances of the graphics multiprocessor 234 can be included within a processing cluster 214. The graphics multiprocessor 234 can process data and a data crossbar 240 can be used to distribute the processed data to one of multiple possible destinations, including other shader units. The pipeline manager 232 can facilitate the distribution of processed data by specifying destinations for processed data to be distributed via the data crossbar 240.

Each graphics multiprocessor 234 within the processing cluster 214 can include an identical set of functional execution logic (e.g., arithmetic logic units, load-store units, etc.). The functional execution logic can be configured in a pipelined manner in which new instructions can be issued before previous instructions are complete. The functional execution logic supports a variety of operations including integer and floating point arithmetic, comparison operations, Boolean operations, bit-shifting, and computation of various algebraic functions. In one embodiment the same functional-unit hardware can be leveraged to perform different operations and any combination of functional units may be present.

The instructions transmitted to the processing cluster 214 constitutes a thread. A set of threads executing across the set of parallel processing engines is a thread group. A thread group executes the same program on different input data. Each thread within a thread group can be assigned to a different processing engine within a graphics multiprocessor 234. A thread group may include fewer threads than the number of processing engines within the graphics multiprocessor 234. When a thread group includes fewer threads than the number of processing engines, one or more of the processing engines may be idle during cycles in which that thread group is being processed. A thread group may also include more threads than the number of processing engines within the graphics multiprocessor 234. When the thread group includes more threads than the number of processing engines within the graphics multiprocessor 234, processing can be performed over consecutive clock cycles. In one embodiment multiple thread groups can be executed concurrently on a graphics multiprocessor 234.

In one embodiment the graphics multiprocessor 234 includes an internal cache memory to perform load and store operations. In one embodiment, the graphics multiprocessor 234 can forego an internal cache and use a cache memory (e.g., L1 cache 248) within the processing cluster 214. Each graphics multiprocessor 234 also has access to L2 caches within the partition units (e.g., partition units 220A-220N of FIG. 2A) that are shared among 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 can include one or more of local parallel processor memory and/or system memory. Any memory external to the parallel processing unit 202 may be used as global memory. Embodiments in which the processing cluster 214 includes multiple instances of the graphics multiprocessor 234 can share common instructions and data, which may be stored in the L1 cache 248.

Each processing cluster 214 may include an MMU 245 (memory management unit) that is configured to map virtual addresses into physical addresses. In other embodiments, one or more instances of the MMU 245 may reside within the memory interface 218 of FIG. 2A. The MMU 245 includes a set of page table entries (PTEs) used to map a virtual address to a physical address of a tile and optionally a cache line index. The MMU 245 may include address translation lookaside buffers (TLB) or caches that may reside within the graphics multiprocessor 234 or the L1 cache or processing cluster 214. The physical address is processed to distribute surface data access locality to allow 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 miss.

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

It will be appreciated that the core architecture described herein is illustrative and that variations and modifications are possible. Any number of processing units, e.g., graphics multiprocessor 234, texture units 236, preROPs 242, etc., may be included within a processing cluster 214. Further, while only one processing cluster 214 is shown, a parallel processing unit as described herein may include any number of instances of the processing cluster 214. In one embodiment, each processing cluster 214 can be configured to operate independently of other processing clusters 214 using separate and distinct processing units, L1 caches, etc.

FIG. 2D shows a graphics multiprocessor 234, according to one embodiment. In such embodiment the graphics multiprocessor 234 couples with the pipeline manager 232 of the 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 load/store units 266 are coupled with cache memory 272 and shared memory 270 via a memory and cache interconnect 268.

In one embodiment, the instruction cache 252 receives a stream of instructions to execute from the pipeline manager 232. The instructions are cached in the instruction cache 252 and dispatched for execution by the instruction unit 254. The instruction unit 254 can dispatch instructions as thread groups (e.g., warps), with each thread of the thread group assigned to a different execution unit within GPGPU core 262. An instruction can access any of a local, shared, or global address space by specifying an address within a unified address space. The address mapping unit 256 can be used to translate addresses in the unified address space into a distinct memory address that can be accessed by the load/store units 266.

The register file 258 provides a set of registers for the functional units of the graphics multiprocessor 234. The register file 258 provides temporary storage for operands connected to the data paths of the functional units (e.g., GPGPU cores 262, load/store units 266) of the graphics multiprocessor 234. In one embodiment, the register file 258 is divided between each of the functional units such that each functional unit is allocated a dedicated portion of the register file 258. In one embodiment, the register file 258 is divided between the different warps being executed by the graphics multiprocessor 234.

The GPGPU cores 262 can each include floating point units (FPUs) and/or integer arithmetic logic units (ALUs) that are used to execute instructions of the graphics multiprocessor 234. The GPGPU cores 262 can be similar in architecture or can differ in architecture, according to embodiments. For example and in one embodiment, a first portion of the GPGPU cores 262 include a single precision FPU and an integer ALU while a second portion of the GPGPU cores include a double precision FPU. In one embodiment the FPUs can implement the IEEE 754-2008 standard for floating point arithmetic or enable variable precision floating point arithmetic. The graphics multiprocessor 234 can additionally include one or more fixed function or special function units to perform specific functions such as copy rectangle or pixel blending operations. In one embodiment one or more of the GPGPU cores can also include fixed or special function logic.

In one embodiment the GPGPU cores 262 include SIMD logic capable of performing a single instruction on multiple sets of data. In one embodiment GPGPU cores 262 can physically execute SIMD4, SIMD8, and SIMD16 instructions and logically execute SIMD1, SIMD2, and SIMD32 instructions. The SIMD instructions for the GPGPU cores can be generated at compile time by a shader compiler or automatically generated when executing programs written and compiled for single program multiple data (SPMD) or SIMT architectures. Multiple threads of a program configured for the SIMT execution model can be executed via a single SIMD instruction. For example and in one embodiment, eight SIMT threads that perform the same or similar operations can be executed in parallel via a single SIMD8 logic unit.

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

FIG. 3A-3B illustrate additional graphics multiprocessors, according to embodiments. The illustrated graphics multiprocessors 325, 350 are variants of the graphics multiprocessor 234 of FIG. 2C. The illustrated graphics multiprocessors 325, 350 can be configured as a streaming multiprocessor (SM) capable of simultaneous execution of a large number of execution threads.

FIG. 3A shows a graphics multiprocessor 325 according to an additional embodiment. The graphics multiprocessor 325 includes multiple additional instances of execution resource units relative to the graphics multiprocessor 234 of FIG. 2D. For example, the graphics multiprocessor 325 can include multiple instances of the instruction unit 332A-332B, register file 334A-334B, and texture unit(s) 344A-344B. The graphics multiprocessor 325 also includes multiple sets of graphics or compute execution units (e.g., GPGPU core 336A-336B, GPGPU core 337A-337B, GPGPU core 338A-338B) and multiple sets of load/store units 340A-340B. In one embodiment the execution resource units have a common instruction cache 330, texture and/or data cache memory 342, and shared memory 346.

The various components can communicate via an interconnect fabric 327. In one embodiment the interconnect fabric 327 includes one or more crossbar switches to enable communication between the various components of the graphics multiprocessor 325. In one embodiment the interconnect fabric 327 is a separate, high-speed network fabric layer upon which each component of the graphics multiprocessor 325 is stacked. The components of the graphics multiprocessor 325 communicate with remote components via the interconnect fabric 327. For example, the GPGPU cores 336A-336B, 337A-337B, and 3378A-338B can each communicate with shared memory 346 via the interconnect fabric 327. The interconnect fabric 327 can arbitrate communication within the graphics multiprocessor 325 to ensure a fair bandwidth allocation between components.

FIG. 3B shows a graphics multiprocessor 350 according to an additional embodiment. The graphics multiprocessor 350 includes multiple sets of execution resources 356A-356D, where each set of execution resource includes multiple instruction units, register files, GPGPU cores, and load store units, as illustrated in FIG. 2D and FIG. 3A. The execution resources 356A-356D can work in concert with texture unit(s) 360A-360D for texture operations, while sharing an instruction cache 354, and shared memory 362. In one embodiment the execution resources 356A-356D can share an instruction cache 354 and shared memory 362, as well as multiple instances of a texture and/or data cache memory 358A-358B. The various components can communicate via an interconnect fabric 352 similar to the interconnect fabric 327 of FIG. 3A.

Persons skilled in the art will understand that the architecture described in FIGS. 1, 2A-2D, and 3A-3B are descriptive and not limiting as to the scope of the present embodiments. Thus, the techniques described herein may be implemented on any properly configured processing unit, including, without limitation, one or more mobile application processors, one or more desktop or server central processing units (CPUs) including multi-core CPUs, one or more parallel processing units, such as the parallel processing unit 202 of FIG. 2A, as well as one or more graphics processors or special purpose processing units, without departure from the scope of the embodiments described herein.

In some embodiments a parallel processor or GPGPU as described herein is communicatively coupled to host/processor cores to accelerate graphics operations, machine-learning operations, pattern analysis operations, and various general purpose GPU (GPGPU) functions. The GPU may be communicatively coupled to the host processor/cores over a bus or other interconnect (e.g., a high speed interconnect such as PCIe or NVLink). In other embodiments, the GPU may be integrated on the same package or chip as the cores and communicatively coupled to the cores over 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 cores may allocate work to the GPU in the form of sequences of commands/instructions contained in a work descriptor. The GPU then uses dedicated circuitry/logic for efficiently processing these commands/instructions.

Techniques for GPU to Host Processor Interconnection

FIG. 4A illustrates an exemplary architecture in which a plurality of GPUs 410-413 are communicatively coupled to a plurality of multi-core processors 405-406 over high-speed links 440A-440D (e.g., buses, point-to-point interconnects, etc.). In one embodiment, the high-speed links 440A-440D support a communication throughput of 4 GB/s, 30 GB/s, 80 GB/s or higher, depending on the implementation. Various interconnect protocols may be used including, but not limited to, PCIe 4.0 or 5.0 and NVLink 2.0. However, the underlying principles of the invention are not limited to any particular communication protocol or throughput.

In addition, in one embodiment, two or more of the GPUs 410-413 are interconnected over high-speed links 442A-442B, which may be implemented using the same or different protocols/links than those used for high-speed links 440A-440D. Similarly, two or more of the multi-core processors 405-406 may be connected over high speed link 443 which may be symmetric multi-processor (SMP) buses operating at 20 GB/s, 30 GB/s, 120 GB/s or higher. Alternatively, all communication between the various system components shown in FIG. 4A may be accomplished using the same protocols/links (e.g., over a common interconnection fabric). As mentioned, however, the underlying principles of the invention are not limited to any particular type of interconnect technology.

In one embodiment, each multi-core processor 405-406 is communicatively coupled to a processor memory 401-402, via memory interconnects 430A-430B, respectively, and each GPU 410-413 is communicatively coupled to GPU memory 420-423 over GPU memory interconnects 450A-450D, respectively. 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 memories 401-402 and GPU memories 420-423 may be volatile memories such as dynamic random access memories (DRAMs) (including stacked DRAMs), Graphics DDR SDRAM (GDDR) (e.g., GDDR5, GDDR6), or High Bandwidth Memory (HBM) and/or may be non-volatile memories such as 3D XPoint or Nano-Ram. In one embodiment, some portion of the memories may be volatile memory and another portion may be non-volatile memory (e.g., using a two-level memory (2LM) hierarchy).

As described below, although the various processors 405-406 and GPUs 410-413 may be physically coupled to a particular memory 401-402, 420-423, respectively, a unified memory architecture may be implemented in which the same virtual system address space (also referred to as the “effective address” space) is distributed among all of the various physical memories. For example, processor memories 401-402 may each comprise 64 GB of the system memory address space and GPU memories 420-423 may each comprise 32 GB of the system memory address space (resulting in a total of 256 GB addressable memory in this example).

FIG. 4B illustrates additional details for an interconnection between a multi-core processor 407 and a graphics acceleration module 446 in accordance with one embodiment. The graphics acceleration module 446 may include one or more GPU chips integrated on a line card which is coupled to the processor 407 via the high-speed link 440. Alternatively, the graphics acceleration module 446 may be integrated on the same package or chip as the processor 407.

The illustrated processor 407 includes a plurality of cores 460A-460D, each with a translation lookaside buffer 461A-461D and one or more caches 462A-462D. The cores may include various other components for executing instructions and processing data which are not illustrated to avoid obscuring the underlying principles of the invention (e.g., instruction fetch units, branch prediction units, decoders, execution units, reorder buffers, etc.). The caches 462A-462D may comprise level 1 (L1) and level 2 (L2) caches. In addition, one or more shared caches 456 may be included in the caching hierarchy and shared by sets of the cores 460A-460D. For example, one embodiment of the processor 407 includes 24 cores, each with its own L1 cache, twelve shared L2 caches, and twelve shared L3 caches. In this embodiment, one of the L2 and L3 caches are shared by two adjacent cores. The processor 407 and the graphics accelerator integration module 446 connect with system memory 441, which may include processor memories 401-402.

Coherency is maintained for data and instructions stored in the various caches 462A-462D, 456 and system memory 441 via inter-core communication over a coherence bus 464. For example, each cache may have cache coherency logic/circuitry associated therewith to communicate to over the coherence bus 464 in response to detected reads or writes to particular cache lines. In one implementation, a cache snooping protocol is implemented over the coherence bus 464 to snoop cache accesses. Cache snooping/coherency techniques are well understood by those of skill in the art and will not be described in detail here to avoid obscuring the underlying principles of the invention.

In one embodiment, a proxy circuit 425 communicatively couples the graphics acceleration module 446 to the coherence bus 464, allowing the graphics acceleration module 446 to participate in the cache coherence protocol as a peer of the cores. In particular, an interface 435 provides connectivity to the proxy circuit 425 over high-speed link 440 (e.g., a PCIe bus, NVLink, etc.) and an interface 437 connects the graphics acceleration module 446 to the high-speed link 440.

In one implementation, an accelerator integration circuit 436 provides cache management, memory access, context management, and interrupt management services on behalf of a plurality of graphics processing engines 431, 432, N of the graphics acceleration module 446. The graphics processing engines 431, 432, N may each comprise a separate graphics processing unit (GPU). Alternatively, the graphics processing engines 431, 432, N may comprise different types of graphics processing engines within a GPU such as graphics execution units, media processing engines (e.g., video encoders/decoders), samplers, and blit engines. In other words, the graphics acceleration module may be a GPU with a plurality of graphics processing engines 431-432, N or the graphics processing engines 431-432, N may be individual GPUs integrated on a common package, line card, or chip.

In one embodiment, the accelerator integration circuit 436 includes a memory management unit (MMU) 439 for performing various memory management functions such as virtual-to-physical memory translations (also referred to as effective-to-real memory translations) and memory access protocols for accessing system memory 441. The MMU 439 may also include a translation lookaside buffer (TLB) (not shown) for caching the virtual/effective to physical/real address translations. In one implementation, a cache 438 stores commands and data for efficient access by the graphics processing engines 431-432, N. In one embodiment, the data stored in cache 438 and graphics memories 433-434, M is kept coherent with the core caches 462A-462D, 456 and system memory 411. As mentioned, this may be accomplished via proxy circuit 425 which takes part in the cache coherency mechanism on behalf of cache 438 and memories 433-434, M (e.g., sending updates to the cache 438 related to modifications/accesses of cache lines on processor caches 462A-462D, 456 and receiving updates from the cache 438).

A set of registers 445 store 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 contexts of the various threads during contexts switches (e.g., where a first thread is saved and a second thread is stored so that the second thread can be execute by a graphics processing engine). For example, on a context switch, the context management circuit 448 may store current register values to a designated region in memory (e.g., identified by a context pointer). It may then restore the register values when returning to the context. In one embodiment, an interrupt management circuit 447 receives and processes interrupts received from system devices.

In one implementation, virtual/effective addresses from a graphics processing engine 431 are translated to real/physical addresses in system memory 411 by the MMU 439. One embodiment of the accelerator integration circuit 436 supports multiple (e.g., 4, 8, 16) graphics accelerator modules 446 and/or other accelerator devices. The graphics accelerator module 446 may be dedicated to a single application executed on the processor 407 or may be shared between multiple applications. In one embodiment, a virtualized graphics execution environment is presented in which the resources of the graphics processing engines 431-432, N are shared with multiple applications or virtual machines (VMs). The resources may be subdivided into “slices” which are allocated to different VMs and/or applications based on the processing requirements and priorities associated with the VMs and/or applications.

Thus, the accelerator integration circuit acts as a bridge to the system for the graphics acceleration module 446 and provides address translation and system memory cache services. In addition, the accelerator integration circuit 436 may provide virtualization facilities for the host processor to manage virtualization of the graphics processing engines, interrupts, and memory management.

Because hardware resources of the graphics processing engines 431-432, N are mapped explicitly to the real address space seen by the host processor 407, any host processor can address these resources directly using an effective address value. One function of the accelerator integration circuit 436, in one embodiment, is the physical separation of the graphics processing engines 431-432, N so that they appear to the system as independent units.

As mentioned, in the illustrated embodiment, one or more graphics memories 433-434, M are coupled to each of the graphics processing engines 431-432, N, respectively. The graphics memories 433-434, M store instructions and data being processed by each of the graphics processing engines 431-432, N. The graphics memories 433-434, M may be volatile memories such as DRAMs (including stacked DRAMs), GDDR memory (e.g., GDDR5, GDDR6), or HBM, and/or may be non-volatile memories such as 3D XPoint or Nano-Ram.

In one embodiment, to reduce data traffic over the high-speed link 440, biasing techniques are used to ensure that the data stored in graphics memories 433-434, M is data which will be used most frequently by the graphics processing engines 431-432, N and preferably not used by the cores 460A-460D (at least not frequently). Similarly, the biasing mechanism attempts to keep data needed by the cores (and preferably not the graphics processing engines 431-432, N) within the caches 462A-462D, 456 of the cores and system memory 411.

FIG. 4C illustrates another embodiment in which the accelerator integration circuit 436 is integrated within the processor 407. In this embodiment, the graphics processing engines 431-432, N communicate directly over the high-speed link 440 to the accelerator integration circuit 436 via interface 437 and interface 435 (which, again, may be utilize any form of bus or interface protocol). The accelerator integration circuit 436 may perform the same operations as those described with respect to FIG. 4B, but potentially at a higher throughput given its close proximity to the coherence bus 464 and caches 462A-462D, 456.

One embodiment supports different programming models including a dedicated-process programming model (no graphics acceleration module virtualization) and shared programming models (with virtualization). The latter may include programming models which are controlled by the accelerator integration circuit 436 and programming models which are controlled by the graphics acceleration module 446.

In one embodiment of the dedicated process model, graphics processing engines 431-432, N are dedicated to a single application or process under a single operating system. The single application can funnel other application requests to the graphics engines 431-432, N, providing virtualization within a VM/partition.

In the dedicated-process programming models, the graphics processing engines 431-432, N, may be shared by multiple VM/application partitions. The shared models require a system hypervisor to virtualize the graphics processing engines 431-432, N to allow access by each operating system. For single-partition systems without a hypervisor, the graphics processing engines 431-432, N are owned by the operating system. In both cases, the operating system can virtualize the graphics processing engines 431-432, N to provide access to each process or application.

For the shared programming model, the graphics acceleration module 446 or an individual graphics processing engine 431-432, N selects a process element using a process handle. In one embodiment, process elements are stored in system memory 411 and are addressable using the effective address to real address translation techniques described herein. The process handle may be an implementation-specific value provided to the host process when registering its context with the graphics processing engine 431-432, N (that is, calling system software to add the process element to the process element linked list). The lower 16-bits of the process handle may be the offset of the process element within the process element linked list.

FIG. 4D illustrates an exemplary accelerator integration slice 490. As used herein, a “slice” comprises a specified portion of the processing resources of the accelerator integration circuit 436. Application effective address space 482 within system memory 411 stores process elements 483. In one embodiment, the process elements 483 are stored in response to GPU invocations 481 from applications 480 executed on the processor 407. A process element 483 contains the process state for the corresponding application 480. A work descriptor (WD) 484 contained in the process element 483 can be a single job requested by an application or may contain a pointer to a queue of jobs. In the latter case, the WD 484 is a pointer to the job request queue in the application's address space 482.

The graphics acceleration module 446 and/or the individual graphics processing engines 431-432, N can be shared by all or a subset of the processes in the system. Embodiments of the invention include an infrastructure for setting up the process state and sending a WD 484 to a graphics acceleration module 446 to start a job in a virtualized environment.

In one implementation, the dedicated-process programming model is implementation-specific. In this model, 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 integration circuit 436 for the owning partition and the operating system initializes the accelerator integration circuit 436 for the owning process at the time when the graphics acceleration module 446 is assigned.

In operation, a WD fetch unit 491 in the accelerator integration slice 490 fetches the next WD 484 which includes an indication of the work to be done by one of the graphics processing engines of the graphics acceleration module 446. Data from the WD 484 may be stored in registers 445 and used by the MMU 439, interrupt management circuit 447 and/or context management circuit 448 as illustrated. For example, one embodiment of the MMU 439 includes segment/page walk circuitry for accessing segment/page tables 486 within the OS virtual address space 485. The interrupt management circuit 447 may process interrupt events 492 received from the graphics acceleration module 446. When performing graphics operations, an effective address 493 generated by a graphics processing engine 431-432, N is translated to a real address by the MMU 439.

In one embodiment, the same set of registers 445 are duplicated 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 duplicated registers may be included in an accelerator integration slice 490. Exemplary registers that may be initialized by the hypervisor are shown in Table 1.

TABLE 1
Hypervisor Initialized Registers
1 Slice Control Register
2 Real Address (RA) Scheduled Processes Area Pointer
3 Authority Mask Override Register
4 Interrupt Vector Table Entry Offset
5 Interrupt Vector Table Entry Limit
6 State Register
7 Logical Partition ID
8 Real address (RA) Hypervisor Accelerator Utilization Record Pointer
9 Storage Description Register

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

TABLE 2
Operating System Initialized Registers
1 Process and Thread Identification
2 Effective Address (EA) Context Save/Restore Pointer
3 Virtual Address (VA) Accelerator Utilization Record Pointer
4 Virtual Address (VA) Storage Segment Table Pointer
5 Authority Mask
6 Work descriptor

In one embodiment, each WD 484 is specific to a particular graphics acceleration module 446 and/or graphics processing engine 431-432, N. It contains all the information a graphics processing engine 431-432, N requires to do its work or it can be a pointer to a memory location where the application has set up a command queue of work to be completed.

FIG. 4E illustrates additional details for one embodiment of a shared model. This embodiment includes a hypervisor real address space 498 in which a process element list 499 is stored. The hypervisor real address space 498 is accessible via a hypervisor 496 which virtualizes the graphics acceleration module engines for the operating system 495.

The shared programming models allow for all or a subset of processes from all or a subset of partitions in the system to use a graphics acceleration module 446. There are two programming models where the graphics acceleration module 446 is shared by multiple processes and partitions: time-sliced shared and graphics directed shared.

In this model, the system hypervisor 496 owns the graphics acceleration module 446 and makes its function available to all operating systems 495. For a graphics acceleration module 446 to support virtualization by the system hypervisor 496, the graphics acceleration module 446 may adhere to the following requirements: 1) An application's job request must be autonomous (that is, 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) An application's job request is guaranteed by the graphics acceleration module 446 to complete in a specified amount of time, including any translation faults, or the graphics acceleration module 446 provides the ability to preempt the processing of the job. 3) The graphics acceleration module 446 must be guaranteed fairness between processes when operating in the directed shared programming model.

In one embodiment, for the shared model, the application 480 is required to make an operating system 495 system call with a graphics acceleration module 446 type, a work descriptor (WD), an authority mask register (AMR) value, and a context save/restore area pointer (CSRP). The graphics acceleration module 446 type describes the targeted acceleration function for the system call. The graphics acceleration module 446 type may be a system-specific value. The WD is formatted specifically for the graphics acceleration module 446 and can be in the form of a graphics acceleration module 446 command, an effective address pointer to a user-defined structure, an effective address pointer to a queue of commands, or any other data structure to describe 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 passed to the operating system is similar to an application setting the AMR. If the accelerator integration circuit 436 and graphics acceleration module 446 implementations do not support a User Authority Mask Override Register (UAMOR), the operating system may apply the current UAMOR value to the AMR value before passing the AMR in the hypervisor call. The hypervisor 496 may optionally apply the current Authority Mask Override Register (AMOR) value before placing the AMR into the process element 483. In one embodiment, the CSRP is one of the registers 445 containing the effective address of an area in the application's address space 482 for the graphics acceleration module 446 to save and restore the context state. This pointer is optional if no state is required to be saved between jobs or when a job is preempted. The context save/restore area may be pinned system memory.

Upon receiving the system call, the operating system 495 may verify that the application 480 has registered and been given the authority to use the graphics acceleration module 446. The operating system 495 then calls the hypervisor 496 with the information shown in Table 3.

TABLE 3
OS to Hypervisor Call Parameters
1 A work descriptor (WD)
2 An Authority Mask Register (AMR) value (potentially masked)
3 An effective address (EA) Context Save/Restore Area Pointer (CSRP)
4 A process ID (PID) and optional thread ID (TID)
5 A virtual address (VA) accelerator utilization record pointer (AURP)
6 Virtual address of storage segment table pointer (SSTP)
7 A logical interrupt service number (LISN)

Upon receiving the hypervisor call, the hypervisor 496 verifies that the operating system 495 has registered and been given the authority to use the graphics acceleration module 446. The hypervisor 496 then puts the process element 483 into the process element linked list for the corresponding graphics acceleration module 446 type. The process element may include the information shown in Table 4.

TABLE 4
Process Element Information
1  A work descriptor (WD)
2  An Authority Mask Register (AMR) value (potentially masked).
3  An effective address (EA) Context Save/Restore Area Pointer (CSRP)
4  A process ID (PID) and optional thread ID (TID)
5  A virtual address (VA) accelerator utilization record pointer (AURP)
6  Virtual address of storage segment table pointer (SSTP)
7  A logical interrupt service number (LISN)
8  Interrupt vector table, derived from hypervisor call parameters
9  A state register (SR) value
10 A logical partition ID (LPID)
11 A real address (RA) hypervisor accelerator utilization record pointer
12 Storage Descriptor Register (SDR)

In one embodiment, the hypervisor initializes a plurality of accelerator integration slice 490 registers 445.

As illustrated in FIG. 4F, one embodiment of the invention employs a unified memory addressable via a common virtual memory address space used to access the physical processor memories 401-402 and GPU memories 420-423. In this implementation, operations executed on the GPUs 410-413 utilize the same virtual/effective memory address space to access the processors memories 401-402 and vice versa, thereby simplifying programmability. In one embodiment, a first portion of the virtual/effective address space is allocated 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. The entire virtual/effective memory space (sometimes referred to as the effective address space) is thereby distributed across each of the processor memories 401-402 and GPU memories 420-423, allowing any processor or GPU to access any physical memory with a virtual address mapped to that memory.

In one embodiment, bias/coherence management circuitry 494A-494E within one or more of the MMUs 439A-439E ensures cache coherence between the caches of the host processors (e.g., 405) and the GPUs 410-413 and implements biasing techniques indicating the physical memories in which certain types of data should be stored. While multiple instances of bias/coherence management circuitry 494A-494E are illustrated in FIG. 4F, the bias/coherence circuitry may be implemented within the MMU of one or more host processors 405 and/or within the accelerator integration circuit 436.

One embodiment allows GPU-attached memory 420-423 to be mapped as part of system memory, and accessed using shared virtual memory (SVM) technology, but without suffering the typical performance drawbacks associated with full system cache coherence. The ability to GPU-attached memory 420-423 to be accessed as system memory without onerous cache coherence overhead provides a beneficial operating environment for GPU offload. This arrangement allows the host processor 405 software to setup operands and access computation results, without the overhead of tradition I/O DMA data copies. Such traditional copies involve driver calls, interrupts and memory mapped I/O (MMIO) accesses that are all inefficient relative to simple memory accesses. At the same time, the ability to access GPU attached memory 420-423 without cache coherence overheads can be critical to the execution time of an offloaded computation. In cases with substantial streaming write memory traffic, for example, cache coherence overhead can significantly reduce the effective write bandwidth seen by a GPU 410-413. The efficiency of operand setup, the efficiency of results access, and the efficiency of GPU computation all play a role in determining the effectiveness of GPU offload.

In one implementation, the selection of between GPU bias and host processor bias is driven by a bias tracker data structure. A bias table may be used, for example, which may be a page-granular structure (i.e., controlled at the granularity of a memory page) that includes 1 or 2 bits per GPU-attached memory page. The bias table may be implemented in a stolen memory range of one or more GPU-attached memories 420-423, with or without a bias cache in the GPU 410-413 (e.g., to cache frequently/recently used entries of the bias table). Alternatively, the entire bias table may be maintained within the GPU.

In one implementation, the bias table entry associated with each access to the GPU-attached memory 420-423 is accessed prior the actual access to the GPU memory, causing the following operations. First, local requests from the GPU 410-413 that find their page in GPU bias are forwarded directly to a corresponding GPU memory 420-423. Local requests from the GPU that find their page in host bias are forwarded to the processor 405 (e.g., over a high-speed link as discussed above). In one embodiment, requests from the processor 405 that find the requested page in host processor bias complete the request like a normal memory read. Alternatively, requests directed to a GPU-biased page may be forwarded to the GPU 410-413. The GPU may then transition the page to a host processor bias if it is not currently using the page.

The bias state of a page can be changed either by a software-based mechanism, a hardware-assisted software-based mechanism, or, for a limited set of cases, a purely hardware-based mechanism.

One mechanism for changing the bias state employs an API call (e.g. OpenCL), which, in turn, calls the GPU's device driver which, in turn, sends a message (or enqueues a command descriptor) to the GPU directing it to change the bias state and, for some transitions, perform a cache flushing operation in the host. The cache flushing operation is required for a transition from host processor 405 bias to GPU bias, but is not required for the opposite transition.

In one embodiment, cache coherency is maintained by temporarily rendering GPU-biased pages uncacheable by the host processor 405. To access these pages, the processor 405 may request access from the GPU 410 which may or may not grant access right away, depending on the implementation. Thus, to reduce communication between the processor 405 and GPU 410 it is beneficial to ensure that GPU-biased pages are those which are required by the GPU but not the host processor 405 and vice versa.

Graphics Processing Pipeline

FIG. 5 illustrates a graphics processing pipeline 500, according to an embodiment. In one embodiment a graphics processor can implement the illustrated graphics processing pipeline 500. The graphics processor can be included within the parallel processing subsystems as described herein, such as the parallel processor 200 of FIG. 2A, which, in one embodiment, is a variant of the parallel processor(s) 112 of FIG. 1. The various parallel processing systems can implement the graphics processing pipeline 500 via one or more instances of the parallel processing unit (e.g., parallel processing unit 202 of FIG. 2A) as described herein. For example, a shader unit (e.g., graphics multiprocessor 234 of FIG. 2C) may be configured to perform the functions of one or more of a vertex processing unit 504, a tessellation control processing unit 508, a tessellation evaluation processing unit 512, a geometry processing unit 516, and a fragment/pixel processing unit 524. The functions of data assembler 502, primitive assemblers 506, 514, 518, tessellation unit 510, rasterizer 522, and raster operations unit 526 may also be performed by other processing engines within a processing cluster (e.g., processing cluster 214 of FIG. 2A) and a corresponding partition unit (e.g., partition unit 220A-220N of FIG. 2A). The graphics processing pipeline 500 may also be implemented using dedicated processing units for one or more functions. In one embodiment, one or more portions of the graphics processing pipeline 500 can be performed by parallel processing logic within a general purpose processor (e.g., CPU). In one embodiment, one or more portions of the graphics processing pipeline 500 can access on-chip memory (e.g., parallel processor memory 222 as in FIG. 2A) via a memory interface 528, which may be an instance of the memory interface 218 of FIG. 2A.

In one embodiment the data assembler 502 is a processing unit that collects vertex data for surfaces and primitives. The data assembler 502 then outputs the vertex data, including the vertex attributes, to the vertex processing unit 504. The vertex processing unit 504 is a programmable execution unit that executes vertex shader programs, lighting and transforming vertex data as specified by the vertex shader programs. The vertex processing unit 504 reads data that is stored in cache, local or system memory for use in processing the vertex data and may be programmed to transform the vertex data from an object-based coordinate representation to a world space coordinate space or a normalized device coordinate space.

A first instance of a primitive assembler 506 receives vertex attributes from the vertex processing unit 504. The primitive assembler 506 readings stored vertex attributes as needed and constructs graphics primitives for processing by tessellation control processing unit 508. The graphics primitives include triangles, line segments, points, patches, and so forth, as supported by various graphics processing application programming interfaces (APIs).

The tessellation control processing unit 508 treats the input vertices as control points for a geometric patch. The control points are transformed from an input representation from the patch (e.g., the patch's bases) to a representation that is suitable for use in surface evaluation by the tessellation evaluation processing unit 512. The tessellation control processing unit 508 can also compute tessellation factors for edges of geometric patches. A tessellation factor applies to a single edge and quantifies a view-dependent level of detail associated with the edge. A tessellation unit 510 is configured to receive the tessellation factors for edges of a patch and to tessellate the patch into multiple geometric primitives such as line, triangle, or quadrilateral primitives, which are transmitted to a tessellation evaluation processing unit 512. The tessellation evaluation processing unit 512 operates on parameterized coordinates of the subdivided patch to generate a surface representation and vertex attributes for each vertex associated with the geometric primitives.

A second instance of a primitive assembler 514 receives vertex attributes from the tessellation evaluation processing unit 512, reading stored vertex attributes as needed, and constructs graphics primitives for processing by the geometry processing unit 516. The geometry processing unit 516 is a programmable execution unit that executes geometry shader programs to transform graphics primitives received from primitive assembler 514 as specified by the geometry shader programs. In one embodiment the geometry processing unit 516 is programmed to subdivide the graphics primitives into one or more new graphics primitives and calculate parameters used to rasterize the new graphics primitives.

In some embodiments the geometry processing unit 516 can add or delete elements in the geometry stream. The geometry processing unit 516 outputs the parameters and vertices specifying new graphics primitives to primitive assembler 518. The primitive assembler 518 receives the parameters and vertices from the geometry processing unit 516 and constructs graphics primitives for processing by a viewport scale, cull, and clip unit 520. The geometry processing unit 516 reads data that is stored in parallel processor memory or system memory for use in processing the geometry data. The viewport scale, cull, and clip unit 520 performs clipping, culling, and viewport scaling and outputs processed graphics primitives to a rasterizer 522.

The rasterizer 522 can perform depth culling and other depth-based optimizations. The rasterizer 522 also performs scan conversion on the new graphics primitives to generate fragments and output those fragments and associated coverage data to the fragment/pixel processing unit 524. The fragment/pixel processing unit 524 is a programmable execution unit that is configured to execute fragment shader programs or pixel shader programs. The fragment/pixel processing unit 524 transforming fragments or pixels received from rasterizer 522, as specified by the fragment or pixel shader programs. For example, the fragment/pixel processing unit 524 may be programmed to perform operations included but not limited to texture mapping, shading, blending, texture correction and perspective correction to produce shaded fragments or pixels that are output to a raster operations unit 526. The fragment/pixel processing unit 524 can read data that is stored in either the parallel processor memory or the system memory for use when processing the fragment data. Fragment or pixel shader programs may be configured to shade at sample, pixel, tile, or other granularities depending on the sampling rate configured for the processing units.

The raster operations unit 526 is a processing unit that performs raster operations including, but not limited to stencil, z test, blending, and the like, and outputs pixel data as processed graphics data to be stored in 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 the one or more display device(s) 110A-110B or for further processing by one of the one or more processor(s) 102 or parallel processor(s) 112. In some embodiments the raster operations unit 526 is configured to compress z or color data that is written to memory and decompress z or color data that is read from memory.

Machine Learning Overview

A machine learning algorithm is an algorithm that can learn based on a set of data. Embodiments of machine learning algorithms can be designed to model high-level abstractions within a data set. For example, image recognition algorithms can be used to determine which of several categories to which a given input belong; regression algorithms can output a numerical value given an input; and pattern recognition algorithms can be used to generate translated text or perform text to speech and/or speech recognition.

An exemplary type of machine learning algorithm is a neural network. There are 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 in which the nodes are arranged in layers. Typically, a feedforward network topology includes an input layer and an output layer that are separated by at least one hidden layer. The hidden layer transforms input received by the input layer into a representation that is useful for generating output in the output layer. The network nodes are fully connected via edges to the nodes in adjacent layers, but there are no edges between nodes within each layer. Data received at the nodes of an input layer of a feedforward network are propagated (i.e., “fed forward”) to the nodes of the output layer via an activation function that calculates the states of the nodes of each successive layer in the network based on coefficients (“weights”) respectively associated with each of the edges connecting the layers. Depending on the specific model being represented by the algorithm being executed, the output from the neural network algorithm can take various forms.

Before a machine learning algorithm can be used to model a particular problem, the algorithm is trained using a training data set. Training a neural network involves selecting a network topology, using a set of training data representing a problem being modeled by the network, and adjusting the weights until the network model performs with a minimal error for all instances of the training data set. For example, during a supervised learning training process for a neural network, the output produced by the network in response to the input representing an instance in a training data set is compared to the “correct” labeled output for that instance, an error signal representing the difference between the output and the labeled output is calculated, and the weights associated with the connections are adjusted to minimize that error as the error signal is backward propagated through the layers of the network. The network is considered “trained” when the errors for each of the outputs generated from the instances of the training data set are minimized.

The accuracy of a machine learning algorithm can be affected significantly by the quality of the data set used to train the algorithm. The training process can be computationally intensive and may require a significant amount of time on a conventional general-purpose processor. Accordingly, parallel processing hardware is used to train many types of machine learning algorithms. This is particularly useful for optimizing the training of neural networks, as the computations performed in adjusting the coefficients in neural networks lend themselves naturally to parallel implementations. Specifically, many machine learning algorithms and software applications have been adapted to make use of the parallel processing hardware within general-purpose graphics processing devices.

FIG. 6 is a generalized diagram of a machine learning software stack 600. A machine learning application 602 can be configured to train a neural network using a training dataset or to use a trained deep neural network to implement machine intelligence. The machine learning application 602 can include training and inference functionality for a neural network and/or specialized software that can be used to train a neural network before deployment. The machine learning application 602 can implement any type of machine intelligence including but not limited to image recognition, mapping and localization, autonomous navigation, speech synthesis, medical imaging, or language translation.

Hardware acceleration for the machine learning application 602 can be enabled via a machine learning framework 604. The machine learning framework 604 can provide a library of machine learning primitives. Machine learning primitives are basic operations that are commonly performed by machine learning algorithms. Without the machine learning framework 604, developers of machine learning algorithms would be required to create and optimize the main computational logic associated with the machine learning algorithm, then re-optimize the computational logic as new parallel processors are developed. Instead, the machine learning application can be configured to perform the necessary computations using the primitives provided by the machine learning framework 604. Exemplary primitives include tensor convolutions, activation functions, and pooling, which are computational operations that are performed while training a convolutional neural network (CNN). The machine learning framework 604 can also provide primitives to implement basic linear algebra subprograms performed by many machine-learning algorithms, such as matrix and vector operations.

The machine learning framework 604 can process input data received from the machine learning application 602 and generate the appropriate input to a 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 take advantage of hardware acceleration via the GPGPU hardware 610 without requiring the machine learning framework 604 to have intimate knowledge of the architecture of the GPGPU hardware 610. Additionally, the compute framework 606 can enable hardware acceleration for the machine learning framework 604 across a variety of types and generations of the GPGPU hardware 610.

GPGPU Machine Learning Acceleration

FIG. 7 illustrates a highly-parallel general-purpose graphics processing unit 700, according to an embodiment. In one embodiment the general-purpose processing unit (GPGPU) 700 can be configured to be particularly efficient in processing the type of computational workloads associated with training deep neural networks. Additionally, the GPGPU 700 can be linked directly to other instances of the GPGPU to create a multi-GPU cluster to improve training speed for particularly deep neural networks.

The GPGPU 700 includes a host interface 702 to enable a connection with a host processor. In one embodiment the host interface 702 is a PCI Express interface. However, the host interface can also be a vendor specific communications interface or communications 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 compute clusters 706A-706H. The compute clusters 706A-706H share a cache memory 708. The cache memory 708 can serve as a higher-level cache for cache memories within the compute clusters 706A-706H.

The GPGPU 700 includes memory 714A-714B coupled with the compute clusters 706A-H via a set of memory controllers 712A-712B. In various embodiments, the memory 714A-714B can 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, or 3D stacked memory, including but not limited to high bandwidth memory (HBM).

In one embodiment each compute cluster 706A-706H includes a set of graphics multiprocessors, such as the graphics multiprocessor 234 of FIG. 2C. The graphics multiprocessors of the compute cluster include multiple types of integer and floating point logic units that can perform computational operations at a range of precisions including suited for machine learning computations. For example and in one embodiment at least a subset of the floating point units in each of the compute clusters 706A-706H can be configured to perform 16-bit or 32-bit floating point operations, while a different subset of the floating point units can be configured to perform 64-bit floating point operations.

Multiple instances of the GPGPU 700 can be configured to operate as a compute cluster. The communication mechanism used by the compute cluster for synchronization and data exchange varies across embodiments. In one embodiment the multiple instances of the GPGPU 700 communicate over the host interface 702. In one embodiment the GPGPU 700 includes an I/O hub 709 that couples the GPGPU 700 with a GPU link 710 that enables a direct connection to other instances of the GPGPU. In one embodiment the GPU link 710 is coupled to a dedicated GPU-to-GPU bridge that enables communication and synchronization between multiple instances of the GPGPU 700. In one embodiment the GPU link 710 couples with a high speed interconnect to transmit and receive data to other GPGPUs or parallel processors. In one embodiment the multiple instances of the GPGPU 700 are located in separate data processing systems and communicate via a network device that is accessible via the host interface 702. In one embodiment the GPU link 710 can be configured to enable a connection to a host processor in addition to or as an alternative to the host interface 702.

While the illustrated configuration of the GPGPU 700 can be configured to train neural networks, one embodiment provides alternate configuration of the GPGPU 700 that can be configured for deployment within a high performance or low power inferencing platform. In an inferencing configuration the GPGPU 700 includes fewer of the compute clusters 706A-H relative to the training configuration. Additionally memory technology associated with the memory 714A-714B may differ between inferencing and training configurations. In one embodiment the inferencing configuration of the GPGPU 700 can support inferencing specific instructions. For example, an inferencing configuration can provide support for one or more 8-bit integer dot product instructions, which are commonly used during inferencing operations for deployed neural networks.

FIG. 8 illustrates a multi-GPU computing system 800, according to an embodiment. The multi-GPU computing system 800 can include a processor 802 coupled to multiple GPGPUs 806A-806D via a host interface switch 804. The host interface switch 804, in one embodiment, is a PCI express switch device that couples the processor 802 to a PCI express bus over which the processor 802 can communicate with the set of GPGPUs 806A-806D. Each of the multiple GPGPUs 806A-806D can be an instance of the GPGPU 700 of FIG. 7. The GPGPUs 806A-806D can interconnect via a set of high-speed point-to-point GPU to GPU links 816. The high-speed GPU to GPU links can connect to each of the GPGPUs 806A-806D via a dedicated GPU link, such as the GPU link 710 as in FIG. 7. The P2P GPU links 816 enable 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 GPU-to-GPU traffic directed to the P2P GPU links, the host interface bus remains available for system memory access or to communicate with other instances of the multi-GPU computing system 800, for example, via one or more network devices. While in the illustrated embodiment the GPGPUs 806A-806D connect to the processor 802 via the host interface switch 804, in one embodiment the processor 802 includes direct support for the P2P GPU links 816 and can connect directly to the GPGPUs 806A-806D.

Machine Learning Neural Network Implementations

The computing architecture provided by embodiments described herein can be configured to perform the types of parallel processing that is particularly suited for training and deploying neural networks for machine learning. A neural network can be generalized as a network of functions having a graph relationship. As is well-known in the art, there are a variety of types of neural network implementations used in machine learning. One exemplary type of neural network is the feedforward network, as previously described.

A second exemplary type of neural network is the Convolutional Neural Network (CNN). A CNN is a specialized feedforward neural network for processing data having a known, grid-like topology, such as image data. Accordingly, CNNs are commonly used for compute vision and image recognition applications, but they also may be used for other types of pattern recognition such as speech and language processing. The nodes in the CNN input layer are 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 nodes in successive layers of the network. The computations for a CNN include applying the convolution mathematical operation to each filter to produce the output of that filter. Convolution is a specialized kind of mathematical operation performed by two functions to produce a third function that is a modified version of one of the two original functions. In convolutional network terminology, the first function to the convolution can be referred to as the input, while the second function can be referred to as the convolution kernel. The output may be referred to as the feature map. For example, the input to a convolution layer can be a multidimensional array of data that defines the various color components of an input image. The convolution kernel can be a multidimensional array of parameters, where the parameters are adapted by the training process for the neural network.

Recurrent neural networks (RNNs) are a family of feedforward neural networks that include feedback connections between layers. RNNs enable modeling of sequential data by sharing parameter data across different parts of the neural network. The architecture for a RNN includes cycles. The cycles represent the influence of a present value of a variable on its own value at a future time, as at least a portion of the output data from the RNN is used as feedback for processing subsequent input in a sequence. This feature makes RNNs particularly useful for language processing due to the variable nature in which language data can be composed.

The figures described below present exemplary feedforward, CNN, and RNN networks, as well as describe a general process for respectively training and deploying each of those types of networks. It will be understood that these descriptions are exemplary and non-limiting as to any specific embodiment described herein and the concepts illustrated can be applied generally to deep neural networks and machine learning techniques in general.

The exemplary neural networks described above can 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 composed of multiple hidden layers, as opposed to shallow neural networks that include only a single hidden layer. Deeper neural networks are generally more computationally intensive to train. However, the additional hidden layers of the network enable multistep pattern recognition that results in reduced output error relative to shallow machine learning techniques.

Deep neural networks used in deep learning typically include a front-end network to perform feature recognition coupled to a back-end network which represents a mathematical model that can perform operations (e.g., object classification, speech recognition, etc.) based on the feature representation provided to the model. Deep learning enables machine learning to be performed without requiring hand crafted feature engineering to be performed for the model. Instead, deep neural networks can learn features based on statistical structure or correlation within the input data. The learned features can be provided to a mathematical model that can map detected features to an output. The mathematical model used by the network is generally specialized for the specific task to be performed, and different models will be used to perform different task.

Once the neural network is structured, a learning model can be applied to the network to train the network to perform specific tasks. The learning model describes how to adjust the weights within the model to reduce the output error of the network. Backpropagation of errors is a common method used to train neural networks. An input vector is presented 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 the neurons in the output layer. The error values are then propagated backwards until each neuron has an associated error value which roughly represents its contribution to the original output. The network can then learn from those errors using an algorithm, such as the stochastic gradient descent algorithm, to update the weights of the of the neural network.

FIG. 9A-9B illustrate an exemplary convolutional neural network. FIG. 9A illustrates various layers within a CNN. As shown in FIG. 9A, an exemplary CNN used to model image processing can receive input 902 describing the red, green, and blue (RGB) components of an input image. The input 902 can be processed by multiple convolutional layers (e.g., convolutional layer 904, convolutional layer 906). The output from the multiple convolutional layers may optionally be processed by a set of fully connected layers 908. Neurons in a fully connected layer have full connections to all activations in the previous layer, as previously described for a feedforward network. The output from the fully connected layers 908 can be used to generate an output result from the network. The activations within the fully connected layers 908 can be computed using matrix multiplication instead of convolution. Not all CNN implementations are configured to make use of fully connected layers 908. For example, in some implementations the convolutional layer 906 can generate output for the CNN.

The convolutional layers are sparsely connected, which differs from traditional neural network configuration found in the fully connected layers 908. Traditional neural network layers are fully connected, such that every output unit interacts with every input unit. However, the convolutional layers are sparsely connected because the output of the convolution of a field is input (instead of the respective state value of each of the nodes in the field) to the nodes of the subsequent layer, as illustrated. The kernels associated with the convolutional layers perform convolution operations, the output of which is sent to the next layer. The dimensionality reduction performed within the convolutional layers is one aspect that enables the CNN to scale to process large images.

FIG. 9B illustrates exemplary computation stages within a convolutional layer of a CNN. Input to a convolutional layer 912 of a CNN can be processed in three stages of a convolutional layer 914. The three stages can include a convolution stage 916, a detector stage 918, and a pooling stage 920. The convolutional layer 914 can then output data to a successive convolutional layer. The final convolutional layer of the network can generate output feature map data or provide input to a fully connected layer, for example, to generate a classification value for the input to the CNN.

In the convolution stage 916 performs several convolutions in parallel to produce a set of linear activations. The convolution stage 916 can include an affine transformation, which is any transformation that can be specified as a linear transformation plus a translation. Affine transformations include rotations, translations, scaling, and combinations of these transformations. The convolution stage computes the output of functions (e.g., neurons) that are connected to specific regions in the input, which can be determined as the local region associated with the neuron. The neurons compute a dot product between the weights of the neurons and the region in the local input to which the neurons are connected. The output from the convolution stage 916 defines a set of linear activations that are processed by successive stages of the convolutional layer 914.

The linear activations can be processed by a detector stage 918. In the detector stage 918, each linear activation is processed by a non-linear activation function. The non-linear activation function increases the nonlinear properties of the overall network without affecting the receptive fields of the convolution layer. Several types of non-linear activation functions may be used. One particular type is the rectified linear unit (ReLU), which uses an activation function defined as ƒ(x)=max (0,x), such that the activation is thresholded at zero.

The pooling stage 920 uses a pooling function that replaces the output of the convolutional layer 906 with a summary statistic of the nearby outputs. The pooling function can be used to introduce translation invariance into the neural network, such that small translations to the input do not change the pooled outputs. Invariance to local translation can be useful in scenarios where the presence of a feature in the input data is more important than the precise location of the feature. Various types of pooling functions can be used during the pooling stage 920, including max pooling, average pooling, and l2-norm pooling. Additionally, some CNN implementations do not include a pooling stage. Instead, such implementations substitute and additional convolution stage having an increased stride relative to previous convolution stages.

The output from the convolutional layer 914 can then be processed by the next layer 922. The next layer 922 can be an additional convolutional layer or one of the fully connected layers 908. For example, the first convolutional layer 904 of FIG. 9A can output to the second convolutional layer 906, while the second convolutional layer can output to a first layer of the fully connected layers 908.

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 can be built 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 prior sequence of inputs. For example, an RNN may be used to perform statistical language modeling to predict an upcoming word given a previous sequence of words. The illustrated RNN 1000 can be described has having an input layer 1002 that receives an input vector, hidden layers 1004 to implement a recurrent function, a feedback mechanism 1005 to enable a ‘memory’ of previous states, and an output layer 1006 to output a result. The RNN 1000 operates based on time-steps. The state of the RNN at a given time step is influenced based on the previous time step via the feedback mechanism 1005. For a given time step, the state of the hidden layers 1004 is defined by the previous state and the input at the current time step. An initial input (x₁) at a first time step can be processed by the hidden layer 1004. A second input (x₂) can be processed by the hidden layer 1004 using state information that is determined during the processing of the initial input (x₁). A given state can be computed as s_t=ƒ(Ux_t+Ws_t-1), where U and W are parameter matrices. The function ƒ is generally a nonlinearity, such as the hyperbolic tangent function (Tanh) or a variant of the rectifier function ƒ(x)=max(0,x). However, the specific mathematical function used in the hidden layers 1004 can vary depending on the specific implementation details of the RNN 1000.

In addition to the basic CNN and RNN networks described, variations on those networks may be enabled. One example RNN variant is the long short term memory (LSTM) RNN. LSTM RNNs are capable of learning long-term dependencies that may be necessary for processing longer sequences of language. A variant on the CNN is a convolutional deep belief network, which has a structure similar to a CNN and is trained in a manner similar to a deep belief network. A deep belief network (DBN) is a generative neural network that is composed of multiple layers of stochastic (random) variables. DBNs can be trained layer-by-layer using greedy unsupervised learning. The learned weights of the DBN can then be used to provide pre-train neural networks by determining an optimal initial set of weights for the neural network.

FIG. 11 illustrates training and deployment of a deep neural network. Once a given network has been structured 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 a training framework 1104. The training framework 1104 can hook into an untrained neural network 1106 and enable the untrained neural net to be trained using the parallel processing resources described herein to generate a trained neural network 1108.

To start the training process the initial weights may be chosen randomly or by pre-training using a deep belief network. The training cycle then be performed in either a supervised or unsupervised manner.

Supervised learning is a learning method in which training is performed as a mediated operation, such as when the training dataset 1102 includes input paired with the desired output for the input, or where the training dataset includes input having known output and the output of the neural network is manually graded. The network processes the inputs and compares the resulting outputs against a set of expected or desired outputs. Errors are then propagated back through the system. The training framework 1104 can adjust to adjust the weights that control the untrained neural network 1106. The training framework 1104 can provide tools to monitor how well the untrained neural network 1106 is converging towards a model suitable to generating correct answers based on known input data. The training process occurs repeatedly as the weights of the network are adjusted to refine the output generated by the neural network. The training process can continue until the neural network reaches a statistically desired accuracy associated with a trained neural network 1108. The trained neural network 1108 can then be deployed to implement any number of machine learning operations to generate an inference result 1114 based on input of new data 1112.

Unsupervised learning is a learning method in which the network attempts to train itself using unlabeled data. Thus, for unsupervised learning the training dataset 1102 will include input data without any associated output data. The untrained neural network 1106 can learn groupings within the unlabeled input and can determine how individual inputs are related to the overall dataset. Unsupervised training can be used to generate a self-organizing map, which is a type of trained neural network 1108 capable of performing operations useful in reducing the dimensionality of data. Unsupervised training can also be used to perform anomaly detection, which allows the identification of data points in an input dataset that deviate from the normal patterns of the data.

Variations on supervised and unsupervised training may also be employed. Semi-supervised learning is a technique in which in the training dataset 1102 includes a mix of labeled and unlabeled data of the same distribution. Incremental learning is a variant of supervised learning in which input data is continuously used to further train the model. Incremental learning enables the trained neural network 1108 to adapt to the new data 1112 without forgetting the knowledge instilled within the network during initial training.

Whether supervised or unsupervised, the training process for particularly deep neural networks may be too computationally intensive for a single compute node. Instead of using a single compute node, a distributed network of computational nodes can be used to accelerate the training process.

FIG. 12 is a block diagram illustrating distributed learning. Distributed learning is a training model that uses multiple distributed computing nodes to perform supervised or unsupervised training of a neural network. The distributed computational nodes can each include one or more host processors and one or more of the general-purpose processing nodes, such as the highly-parallel general-purpose graphics processing unit 700 as in FIG. 7. As illustrated, distributed learning can be performed model parallelism 1202, data parallelism 1204, or a combination of model and data parallelism 1204.

In model parallelism 1202, different computational nodes in a distributed system can perform training computations for different parts of a single network. For example, each layer of a neural network can be trained by a different processing node of the distributed system. The benefits of model parallelism include the ability to scale to particularly large models. Splitting the computations associated with different layers of the neural network enables the training of very large neural networks in which the weights of all layers would not fit into the memory of a single computational node. In some instances, model parallelism can be particularly useful in performing unsupervised training of large neural networks.

In data parallelism 1204, the different nodes of the distributed network have a complete instance of the model and each node receives a different portion of the data. The results from the different nodes are then combined. While different approaches to data parallelism are possible, data parallel training approaches all require a technique of combining results and synchronizing the model parameters between each node. Exemplary approaches to 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 the nodes to the parameter server, the updates to the model are transferred. Additionally, update-based data parallelism can be performed in a decentralized manner, where the updates are compressed and transferred between nodes.

Combined model and data parallelism 1206 can be implemented, for example, in a distributed system in which each computational node includes multiple GPUs. Each node can have a complete instance of the model with separate GPUs within each node are used to train different portions of the model.

Distributed training has increased overhead relative 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 to enable high bandwidth GPU-to-GPU data transfer and accelerated remote data synchronization.

Exemplary Machine Learning Applications

Machine learning can be applied to solve a variety of technological problems, including but not limited to computer vision, autonomous driving and navigation, speech recognition, and language processing. Computer vision has traditionally been one of the most active research areas for machine learning applications. Applications of computer vision range from reproducing human visual abilities, such as recognizing faces, to creating new categories of visual abilities. For example, computer vision applications can be configured to recognize sound waves from the vibrations induced in objects visible in a video. Parallel processor accelerated machine learning enables computer vision applications to be trained using significantly larger training dataset than previously feasible and enables inferencing systems to be deployed using low power parallel processors.

Parallel processor accelerated machine learning has autonomous driving applications including lane and road sign recognition, obstacle avoidance, navigation, and driving control. Accelerated machine learning techniques can be used to train driving models based on datasets that define the appropriate responses to specific training input. The parallel processors described herein can enable rapid training of the increasingly complex neural networks used for autonomous driving solutions and enables the deployment of low power inferencing processors in a mobile platform suitable for integration into autonomous vehicles.

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

Parallel processor accelerated machine learning can also be used to accelerate natural language processing. Automatic learning procedures can make use of statistical inference algorithms to produce models that are robust to erroneous or unfamiliar input. Exemplary natural language processor applications include automatic machine translation between human languages.

The parallel processing platforms used for machine learning can be divided into training platforms and deployment platforms. Training platforms are generally highly parallel and include optimizations to accelerate multi-GPU single node training and multi-node, multi-GPU training. Exemplary parallel processors suited for training include the highly-parallel general-purpose graphics processing unit 700 of FIG. 7 and the multi-GPU computing system 800 of FIG. 8. On the contrary, deployed machine learning platforms generally include lower power parallel processors suitable for use in products such as cameras, autonomous robots, and autonomous vehicles.

FIG. 13 illustrates an exemplary inferencing system on a chip (SOC) 1300 suitable for performing inferencing using a trained model. The SOC 1300 can integrate processing components including a media processor 1302, a vision processor 1304, a GPGPU 1306 and a multi-core processor 1308. The SOC 1300 can additionally include on-chip memory 1305 that can enable a shared on-chip data pool that is accessible by each of the processing components. The processing components can be optimized for low power operation to enable deployment to a variety of machine learning platforms, including autonomous vehicles and autonomous robots. For example, one implementation of the SOC 1300 can be used as a portion of the main control system for an autonomous vehicle. Where the SOC 1300 is configured for use in autonomous vehicles the SOC is designed and configured for compliance with the relevant functional safety standards of the deployment jurisdiction.

During operation, the media processor 1302 and vision processor 1304 can work in concert to accelerate computer vision operations. The media processor 1302 can enable low latency decode 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 parse the decoded video and perform preliminary processing operations on the frames of the decoded video in preparation of processing the frames using a trained image recognition model. For example, the vision processor 1304 can accelerate convolution operations for a CNN that is used to perform image recognition on the high-resolution video data, while back end model computations are performed by the GPGPU 1306.

The multi-core processor 1308 can include control logic to assist with sequencing and synchronization of data transfers and shared memory operations performed by the media processor 1302 and the vision processor 1304. The multi-core processor 1308 can also function as an application processor to execute software applications that can make use of the inferencing compute capability of the GPGPU 1306. For example, at least a portion of the navigation and driving logic can be implemented in software executing on the multi-core processor 1308. Such software can directly issue computational workloads to the GPGPU 1306 or the computational workloads can be issued to the multi-core processor 1308, which can offload at least a portion of those operations to the GPGPU 1306.

The GPGPU 1306 can include compute clusters such as a low power configuration of the compute clusters 706A-706H within the highly-parallel general-purpose graphics processing unit 700. The compute clusters within the GPGPU 1306 can support instruction that are specifically optimized to perform inferencing computations on a trained neural network. For example, the GPGPU 1306 can support instructions to perform low precision computations such as 8-bit and 4-bit integer vector operations.

Dynamic Floating Point Unit Accuracy Reduction for Machine Learning Operations

The IEEE 754 single-precision binary floating point format specifies a 32-bit binary representation having a 1-bit sign, an 8-bit exponent, and a 24-bit significand, of which 23 bits are explicitly stored. The IEEE 754 half-precision binary floating point format specifies a 16-bit binary representation having a 1-bit sign, 5-bit exponent, and 11-bit significand, of which 10-bits are explicitly stored. The implicit significand bits are defined to be one for non-zero exponent values and zero when all exponent bits are zero. Floating point units capable of performing arithmetic operations at single and half precision are known in the art. For example, existing floating point units can perform 32-bit single precision floating point operations (FP32) or dual 16-bit half precision floating point operations (FP16).

Embodiments described herein extend this capability by providing support for instruction and associated logic to enable variable precision operations. Floating point instructions that allow variable precision operations can dynamically increase throughput by performing operations at lower precision when possible. In one embodiment a set of instructions and associated logic is provided in which throughput is increased by performing floating point operations at the lowest precision possible without significant loss of data. In one embodiment a set of instructions and associated logic is provided in which the floating point logic will verify lower precision results against results performed at a higher precision to determine if any significant loss of data has occurred.

FIG. 14 illustrates components of a dynamic precision floating point unit 1400, according to an embodiment. In one embodiment the dynamic precision floating point unit 1400 includes a control unit 1402, a set of internal registers 1404, an exponent block 1406, and a significand block 1408. In addition to floating point control logic known in the art, in one embodiment the control unit 1402 additionally includes precision tracking logic 1412 and a numerical transform unit 1422.

In one embodiment the precision tracking logic 1412 is hardware logic configured track an available number of bits of precision for computed data relative to a target precision. The precision tracking logic 1412 can track precision registers within the exponent block 1406 and the significand block 1408 to track precision metrics such as the minimum number of bits of precision required to store computed values that are generated by the exponent block 1406 and the significand block 1408. In one embodiment the precision metrics include a running average of numerical precision required to represent data over a set of computations. In one embodiment the precision metrics include a maximum required precision within a given set of data. In one embodiment the dynamic precision floating point unit 1400 supports instructions to read or reset the register data used by the precision tracking logic 1412 to generate the precision metrics described herein. In one embodiment the compute unit housing the dynamic precision floating point unit supports instructions to set or reset the register data used by the precision tracking logic 1412. In one embodiment the precision tracking logic 1412 monitors an error accumulator 1434 in the set of internal registers 1404. The error accumulator can be used to track an accumulated error (e.g., rounding error) over a set of floating point operations. In one embodiment the dynamic precision floating point unit 1400 supports a set of instructions including an instruction to reset the error accumulator 1434 and an instruction to read the error accumulator 1434. In one embodiment the error accumulator can be reset in response to a bit or flag that is supplied as an operand to an instruction.

In one embodiment the numerical transform unit 1422 can be used to perform intermediate numerical transforms on data when performing lower precision operations to prevent or mitigate the possibility of overflow or underflow while performing operations. For example, when approaching the precision limits of a given datatype, the numerical transform unit 1422 can perform multiplication or division operations using logarithms and transform the resulting value via exponentiation. Further details with respect to the precision tracking logic 1412 and the numerical transform unit 1422 are provided in FIG. 22.

The internal registers 1404 includes a set of operand registers 1414 that store input values for the dynamic precision floating point unit 1400. In one embodiment the operand registers 1414 includes two operands (A, B). For floating point input data, the input data values can be divided into exponent portions (EXA, EXB) and significand portions (SIGA, SIGB). In various embodiments the operand registers 1414 are not limited to supporting two floating point inputs. In one embodiment the operand registers 1414 include three input operands, for example, to support fused multiply-add, multiply-subtract, multiply-accumulate, or related operations. In one embodiment the operand registers 1414 can also store integer values, as in one embodiment the dynamic precision floating point unit supports 32-bit, 16-bit, and 8-bit integer operations. The specific data-type and baseline precision is configurable, in one embodiment, via an input to the control unit 1402.

In one embodiment, floating point operations are performed at dynamic precision using the exponent block 1406 and the significand block 1408. In one embodiment, integer operations can be performed via the significand block 1408. In one embodiment, dual 8-bit integer operations can be performed using the exponent block 1406 and significand block 1408.

In one embodiment the exponent block 1406 includes a comparator 1416 and a dynamic precision exponent adder 1426. The comparator determines the difference between exponents and determines the smaller of the two exponents. During floating point addition, the exponent of the smaller number is adjusted to match the exponent of the larger number. The dynamic precision exponent adder 1426 can be used to add exponent values for FP16 or FP32 values. The significand block 1408 includes a dynamic precision multiplier 1418, a shift unit 1428, a dynamic precision significand adder 1438, and an accumulator register 1448.

In one embodiment an FP16 or FP32 data type can be specified for an operation. Where FP16 is specified, the dynamic precision floating point unit 1400 can power gate elements that are unnecessary for performing FP32 operations while maintaining logic to track precision loss or error (e.g., via error accumulator 1434). For example and in one embodiment, the error accumulator 1434 can be used to track a number of rounding operations within a period of instructions. In one embodiment the error accumulator maintains a value of the total accumulated rounding error over a set of instructions. The dynamic precision floating point unit 1400 can enable support for an instruction to clear or read the error accumulator 1434 from software. Where FP32 is specified, the dynamic precision floating point unit 1400 can attempt to perform FP32 operations at FP16 precision, while power gating elements and components beyond those required to perform operations at FP16 precision. Based on the input or intermediate values, where the dynamic precision floating point unit 1400 is requested to performed operations at FP32, the dynamic precision floating point unit 1400 can initially attempt to perform operations at FP16 and expand precision as needed up to FP32. Where FP32 operations can be performed at FP16 precision, the power demand per operation is reduced, allowing a larger number of compute elements to be enabled simultaneously. For example, dynamic capacitance and/or power budget limitations for a given configuration, such as a battery powered configuration or a passive-only cooling configuration, may not allow all floating point units or other compute elements within a GPGPU to be enabled simultaneously. Reducing the dynamic power of a set of floating point units by enabling dynamic lower precision compute can increase the overall throughput of the compute units of a GPGPU within a given power envelope, as a greater number of threads can be processed on a per-cycle basis without exceeding dynamic power limitations.

FIG. 15 provides additional details with respect to the dynamic precision floating point unit 1400 of FIG. 14, according to an embodiment. In one embodiment the dynamic precision multiplier 1418 includes a set of input buffers 1502 to store significand data. In one embodiment the set of input buffers include two buffers to store two input values for a multiply or divide operation. For a fused operation (e.g., multiply-add, multiply-subtract) the product of the operation can be added to a third input via an adder and/or stored in an accumulator register.

In one embodiment some configurations of the dynamic precision multiplier 1418 include input buffers that are 24-bit inputs that can explicitly store 24-bits of significand data for single precision floating point inputs or 11-bits of significand data for half precision floating point values. In some configurations the input buffers 1502 may also be 32-bit buffers to enable multiplication of 32-bit integer values. In one embodiment a single configuration of the input buffers 1502 are present that is selectable or configurable between 32-bits and 24-bits. In one embodiment the output buffer 1510 is similarly configurable or selectable between 24-bits and 32-bits to selectively enable storage of a full precision 32-bit integer or a 24-bit and/or 11-bit significand value for a 32-bit or 16-bit floating point number.

In one embodiment the dynamic precision multiplier 1418 includes a multiplier 1506 and an overflow multiplier 1504. The multiplier 1506 is configurable to perform a multiply or divide operation at half precision for a data type. For example, the multiplier 1506 can perform an 11-bit multiply operation for the significand of an FP16 floating point value and/or a P multiply operation for a 16-bit integer operation. The multiplier 1506 can also perform an 8-bit multiply operation for an INT8 integer value. For a 32-bit floating point value or a 32-bit integer value, the multiplier 1506 can perform a multiplication operation for a 24-bit significand at 11-bits (e.g., FP16 precision). The multiplier 1506 can, if necessary, perform a multiplication value a 16-bits of significand precision for a 24-bit FP32 significand. In one embodiment the required and resulting precision for an operation on a given set of inputs can be tracked via the precision register 1508. In one embodiment the required and resulting precision can be represented within the precision register 1508 via a loss of precision that would result should the output of the multiplier 1506 be output via the output buffer 1510. In such embodiment the precision register 1508 can track the precision loss associated with the use of lower-precision data types as well as the precision loss associated with performing operations at a lower than requested precision.

In one embodiment control logic associated with the dynamic precision multiplier 1418 (e.g., within control unit 1402 of FIG. 14), can monitor precision loss associated with performing operations for higher precision (e.g., FP32, INT32) operations at lower precision (e.g., FP16, INT16, INT8). If precision loss will be significant the control logic can enable the overflow multiplier 1504 to perform operations for the additional bits of precision. Additionally, if the control logic determines that an overflow or underflow will occur based on the current inputs, the overflow multiplier 1504 is enabled and the multiplication operation is performed using the overflow multiplier 1504 and the multiplier 1506.

Similar control operations are performed for the dynamic precision exponent adder 1426 and the dynamic precision significand adder 1438. The dynamic precision exponent adder 1426 includes a set of 8-bit input buffers that can store exponent data for FP32 (8-bit) and FP16 (5-bit). The 8-bit input buffers 1512 can also store a set of INT-8 inputs. The output buffer 1520 for the dynamic precision exponent adder 1426 can be similarly configured. The dynamic precision significand adder 1438 includes a set of input buffers 1522 that can be selected from one of a set of 24-bit and 32-bit buffers or can be dynamically configurable to store input data of either 24-bits or 32-bits. In one embodiment the input buffers 1522 are simply 32-bit buffers that can also store 24-bit input data. The output buffer 1530 for the dynamic precision significand adder 1438 can be similarly configured. Precision register 1518 within the dynamic precision exponent adder 1426 and precision register 1528 within the dynamic precision significand adder 1438 can be configured to track precision loss for performed operations. Control logic can enable overflow adder 1514 and/or overflow adder 1524 as needed to prevent overflow or underflow conditions or to prevent precision loss exceeding a threshold.

Returning to FIG. 14, in one embodiment, dual INT8 operations can be performed by the dynamic precision floating point unit 1400 using the dynamic precision exponent adder 1426 and the dynamic precision significand adder 1438. For example, instead of disabling the exponent block 1406 during integer operations, the exponent block 1406 can be configured to perform an operation on a first set of 8-bit integer operands while the significand block 1408 can be configured to perform an operation on a second set of 8-bit operands. To enable support for dual 8-bit multiply, dual fused multiply-add, dual fused multiply-subtract, and/or other multiplication based operations, in one embodiment the exponent block 1406 can include an additional multiplier 1436. The multiplier can be a fixed 8-bit multiplier to enable simultaneous dual 8-bit multiply operations using the exponent block 1406 and the significand block 1408.

FIG. 16 illustrates thread assignments for a dynamic precision processing system 1600, according to an embodiment. In one embodiment the dynamic precision processing system 1600 includes a set of dynamic floating point units 1608A-1608D. The dynamic floating point units 1608A-1608D can execute a set of operation threads 1606A-1606D that can perform mixed precision operations and generate output data at variable precisions. In one embodiment a first operation (e.g., add, subtract, multiply, divide, etc.) can be performed by a first operational thread 1606A on a first dynamic floating point unit 1608A, where the first operational thread 1606A accepts as input two 16-bit floating point values 1602A-1602B and outputs a 16-bit floating point value (FP16). The first operation can be performed as a dual operation in which a single instruction executed by a GPGPU allows a mixed precision FP16/FP32 dual operation. The second operation of the dual operation can be performed by a second operation thread 1606B that is performed by a second dynamic floating point unit 1608B, which can generate a second output 1612 that is a 32-bit floating point output. The second operation thread 1606B configures the second dynamic floating point unit 1608B to receive two 32-bit floating point input values 1603A-1603B. In one embodiment the operation on the two 32-bit floating point operation can be performed at 16-bits of precision if the operation can be performed without an underflow, overflow, or excessive precision loss would not occur by performing the operation at a lower precision.

In one embodiment the dynamic precision processing system 1600 can execute a single instruction having a 16-bit operand 1604A and a 32-bit operand 1604B. Operation thread 1606C can be executed on a dynamic floating point unit 1608C. The dynamic floating point unit 1608C will attempt to perform a mixed precision 16-bit/32-bit operation at 16-bits of precision unless significant precision loss or error will occur. In one embodiment the dynamic precision processing system 1600 can also be configured to perform integer operations. For example, an operation on a pair of 8-bit integer inputs 1605A-1605B can be performed via an operation thread 1606D via a dynamic floating point unit 1608D to generate an 8-bit integer output 1616. In one embodiment the dynamic floating point unit 1608D is configurable to perform dual 8-bit integer operations in which two 8-bit integer operations can be performed in a single cycle.

FIG. 17 illustrates logic 1700 to perform a numerical operation at less than a requested precision, according to an embodiment. In one embodiment the logic 1700 is implemented via hardware integrated within the dynamic precision floating point unit 1400 of FIG. 14. In one embodiment the logic 1700 is performed in part via a control unit 1402 within the dynamic precision floating point unit 1400 of FIG. 14.

In one embodiment the logic 1700 can receive a request to perform a numerical operation at a first precision, as shown at block 1702. The numerical operation can be a floating point operation or an integer operation. The first precision can be, for example, a 32-bit precision. In one embodiment the numerical operation may be an operation at the first precision that is performed upon operations having mixed precision. The logic 1700 can then perform the numerical operation using a number of bits associated with a second precision that is lower than the first precision, as shown at block 1704. For example and in one embodiment the number of bits used to perform the operation can be a number of bits associated with a 16-bit operation, while the first precision is a 32-bit precision. The logic 1700 can generate an intermediate result at the second precision at block 1706. The logic 1700 can then determine a precision loss of the intermediate result relative to the first precision. The precision loss can be read from a register that stores an indicator of precision loss that is stored during the operation.

The logic 1700 can determine whether the precision loss is less than a threshold at block 1709. In one embodiment the threshold associated with the precision loss can be software configurable, although a hardware default threshold is used in some embodiments. In one embodiment the degree of precision loss can also be determined via execution of full precision operations in parallel on unused compute units. The reduced precision results can then be compared to the full precision results. If the precision loss is less than the threshold, the logic 1700 can output the result at the second precision, as shown at block 1712. If the precision loss is not less than the threshold at block 1709, the logic 1700 can compute the remaining bits of the result at block 1710 and output the result at the first precision, as shown at block 1714. Computing the remaining bits of the result at block 1710 can be performed, in one embodiment, via overflow logic units, such as the overflow multiplier 1504, overflow adder 1514, and/or overflow adder 1524, as in FIG. 15.

Vertical Stacking of Operations for 16-bit Floating Point Operations

When performing single instruction multiple thread (SIMT) operations at lower precision, in some circumstances it may be difficult to maintain full utilization of underlying single instruction multiple data (SIMD) logic due to the larger number of elements required to fill all SIMD lanes. For example, a SIMD logic unit configured for FP32 operation on 128-bit input registers can perform a single operation on four sets of input data. If that logic unit is configured to perform FP16 operations on the same four sets of input data, the underlying throughput for the operations may be increased due to the lower precision of the operation, but SIMD utilization is reduced by half. One solution to SIMD underutilization is to perform the operation on eight sets of input data. However, the software executing on the logic units may not require as much parallelism as the underlying hardware can provide.

For example, a loop that performs iterative operations on input arrays can be vectorized such that each iteration of the array is performed in parallel as a separate SIMT thread. The separate SIMT threads can be performed in a single operation on underlying SIMD/vector logic within a compute unit. When performing parallel instructions derived via compiler loop vectorization logic, a loop shorter than eight iterations will not fill all eight SIMD lanes available to execute the threads spawned for those operations, reducing overall utilization of the compute units. Additionally, where the underlying hardware has N SIMD lanes, any number of vectorized iterations that are not multiple of N will require the execution of remainder iterations on a less than full SIMD unit. Furthermore, vectorization may require the separate execution of peel loops before executing the main body of the vectorized operations.

Some embodiments described herein can increase SIMD utilization by stacking multiple unrelated FP16 operations into a single SIMD unit for execution. Where a SIMD unit has 8 lanes available for execution, thread scheduling logic can dispatch threads in units of N/2 or N/4, and allow unrelated sets threads that are to perform the same or compatible operations to share a single SIMD unit. Additionally, one embodiment enables SIMD lane scheduling that allows the intermingling of dynamically assembled SIMT thread groups with vector SIMD threads.

FIG. 18 illustrates loop vectorization for SIMD units, according to an embodiment. In one embodiment, software logic can include a loop that is automatically vectorized by compiler software executing on a data processing system. The loop can include a peel loop 1802, a vectorized main loop 1804, and a remainder loop 1806. In some configurations loop vectorization is most efficient when performed on data that accesses aligned memory. For example, a GPGPU can be configured such that vector memory accesses may be most efficiently performed in 64-byte chunks 1801A-1801F. In such configuration, a peel loop 1802 includes a subset of loop iterations that are peeled from the main loop to enable unaligned memory accesses to be sectioned off from the main loop. The vectorized main loop 1804 includes the majority of the iterations of the loop. Each iteration of the vectorized main loop can be performed in parallel and the memory accesses for each element are aligned on a specific memory boundary. The remainder loop 1806 includes the set of iterations that follow the vectorized main loop 1804. The iterations in the remainder loop 1806 generally may not be performed in parallel as efficiently as the main loop.

In one embodiment the peel loop 1802 and the remainder loop 1806 can also be vectorized. In one embodiment each of the peel loop 1802, main loop 1804, and remainder loop 1806 can be executed on FP16 SIMD8 units, where eight instances of the same operation can be performed in parallel. Loop iterations can be executed in parallel on SIMD hardware (e.g., FP16 SIMD8 units 1808A-1808C) using execution mask 1812, execution mask 1814, and execution mask 1816 that each enable and disable SIMD lanes for an operation cycle. For the illustrated peel loop 1802 and remainder loop 1806, a subset of elements is selected in execution mask 1812 and execution mask 1816. All lanes are selected in the execution mask 1814 of the vectorized main loop 1804.

In one embodiment, for a SIMD unit with inactive lanes can be configured to perform other operations on those inactive lanes. For a given cycle, where scheduling logic configure a set of inactive lanes for a SIMD unit (e.g., FP16 SIMD8 1808A, FP16 SIMD8 1808C), instead of idling those lanes during a cycle, the scheduler can stack other multi-element SIMD threads or assign SIMT threads to the otherwise idle SIMD lanes.

FIG. 19 illustrates a thread processing system 1900, according to an embodiment. In one embodiment the thread processing system 1900 includes a SIMD compute unit, such as a SIMD8 floating point unit 1920 that includes multiple dynamic floating point units 1922A-1922H. Depending on the operation, the SIMD8 floating point unit 1920 can execute eight or more of the same or similar operations in a single cycle. For example and in one embodiment, each of the eight dynamic floating point units 1922A-1922H can execute a single operation at FP16 precision. In one embodiment, each of the eight dynamic floating point units 1922A-1922H can perform two paired INT8 operations in a single cycle.

Under some circumstances, such as with peel or remainder loops as illustrated in FIG. 18, not all lanes of a SIMD floating point unit will be active during a cycle. To increase utilization, SIMD slots can be assigned at smaller granularities, to enable otherwise unused SIMD lanes to be utilized. For example, the SIMD8 floating point unit 1920 would generally be assigned threads or operations at an eight-operational granularity, where less than eight operations present a potential loss of computational efficiency. In one embodiment, SIMD lanes can be occupied a single vector SIMD thread that includes an execution mask that selects at least eight elements or a SIMT thread group having at least eight elements.

To increase SIMD utilization, one embodiment divides eight SIMD lanes into two SIMD4 slots (e.g., SIMD4 slot 1910, SIMD4 slot 1912). The SIMD4 slots can be filled in a variety of ways. In one embodiment, two separate SIMD threads (SIMD thread 1902, SIMD thread 1904) that combine to cover a total of four SIMD lanes are assigned to a SIMD4 slot (e.g., SIMD4 slot 1910). In one embodiment, the SIMT thread group 1906 can be assigned to a SIMD4 slot 1912. The SIMT thread group 1906 can includes any number of threads that is a multiple of four threads (e.g., 4, 8, 12, 16, etc.). The threads within the SIMT thread group 1906 can be processed four threads at a time, with the number of cycles required to process all threads within the SIMT thread group 1906 dependent upon the number of threads in the group.

FIG. 20 illustrates logic 2000 to assign threads for computation, according to an embodiment. In one embodiment the logic 2000 is performed via a thread processing system 1900 as in FIG. 19. In one embodiment the logic 2000 can receive a first set of threads at a SIMD unit having a first number of lanes, as shown at block 2002. The logic 2000 can then determine if the first set of threads fill all SIMD lanes of the SIMD unit, as shown at block 2003. If the first set of threads include enough SIMT threads or the threads of the first set of threads include enough SIMD vector elements to fill all SIMD lanes, the logic 2000 can assign the first set of threads to the SIMD unit, as shown at block 2004.

If the first set of threads do not fill all SIMD lanes, as determined at block 2003, the logic 2000 can assign a first set of threads to a second number of lanes that is less than the first number of lanes at block 2006. The assignment can be performed by assigning a SIMD thread to the SIMD unit and masking out the inactive lanes. The assignment can also be performed by assigning a set of SIMT threads to the SIMD unit. The logic can then stack one or more additional sets of threads to fill all SIMD lanes, as shown at bock 2008. The additional set of threads can specify active SIMD lanes that occupy lanes that are not occupied by the initial threads.

System to Enable Normalization and Transformations for Low Precision Data

When performing operations with low precision data types, care must be taken to avoid overflow or underflow of data during numerical operations. This responsibility typically falls upon the data scientist that is developing the low precision algorithm. Due to the limitations of low precision arithmetic, many neural networks have been adapted to use binary and/or ternary values that only occupy one or two bits per element. However, there is a need for integer and floating point arithmetic logic units that can enable N-bit low precision arithmetic with guard logic to warn against or attempt to prevent significant loss of precision during arithmetic operations. N-bit low precision arithmetic can be performed at arbitrary precisions depending on the dynamic range or precision of the dataset being processed. In one embodiment the dynamic precision floating point units described herein include logic to warn when numerical calculations are approaching the limits of low precision calculations.

As shown in FIG. 14, a dynamic precision floating point unit 1400 can include precision tracking logic 1412 and a numerical transform unit 1422. In one embodiment the precision tracking logic 1412 tracks the available bits of precision remaining for computed data relative to a target precision. The available bits if precision can be tracked for intermediate data to determine whether an intermediate data value, which, in one embodiment is computed at a higher precision relative to input data or output data, can be stored at an output precision without significant precision loss or rounding error. For example and in one embodiment, specific low precision operations can be efficiently performed at a higher precision and the precision tracking logic 1412 can determine whether a result of a calculation would overflow a given output precision. In one embodiment the logic units described herein can output status information that indicates the degree of lost precision due to rounding error. In one embodiment the logic units can perform intermediate numeric transformations on data to prevent significant data loss. The logic units can then output the transformed value. In one embodiment a full precision or near-full precision output value can be programmatically derived based on the output and status information provided with the output.

FIG. 21 illustrates a deep neural network 2100 that may be processed using compute logic provided by embodiments described herein. The deep neural network (DNN) is an artificial neural network including multiple neural network layers 2102A-2102N. Each layer represents a set of non-linear compute operations to perform feature extraction and transformation in a manner consistent with the machine learning neural networks described herein. Each successive layer uses the output from the previous layer as input. In the case of a convolutional neural network, fused multiply add logic (e.g., FMA logic 2104A, 2104B) can be used to compute dot products between feature map and filter data to generate activation map data that is provided as input for a successive layer.

Low precision neural networks may be implemented using binary or ternary weights in combination with binary, ternary, or N-bit feature maps. Some neural networks may still benefit from the added precision of calculations using N-bit feature maps and N-bit filters. In some implementations, N-bit features and weights for a neural network can be processed at low precision without significant reduction in output error. However, a data scientist implementing a low precision N-bit neural network (e.g., FP16, FP8) generally should be aware of rounding errors or out of bounds data that may arise due to successive calculations at low precision. Should precision tracking logic (e.g., precision tracking logic 1412 of FIG. 14) in the FMA logic 2104A-2104B determine that weight or feature map data is approaching the limits of the available precision of the data type, a status bit can be set by the FMA logic 2104A-2104B. The status bit can serve as an indicator to a data scientist that is developing the neural network model that is present within the neural network layers 2102A-2102N that the model may require optimization or greater numerical precision may be warranted.

In one embodiment normalize and transform logic 2106A-2106B can be enabled to perform weight normalization or numerical transforms on feature map data before providing the feature map data to the next neural network layer for input. The application of the normalize and transform logic 2106A-2106B is optional at each stage and may be performed only if significant precision loss, overflow, or underflow conditions are likely during processing of an upcoming layer. In one embodiment weights or feature maps output from a layer of a neural network can be automatically normalized via an instance of the normalize and transform logic 2106A-2106B.

In one embodiment the normalize and transform logic 2106A-2106B can use the numerical transform unit 1422 of FIG. 14 to transform feature map data or weight data. Feature map data output from a neural layer can be based on a set of data output from a set of functions. In such embodiment a specific set of low precision instructions are provided that enable automatic adjustment of N-bit neural network data to prevent catastrophic loss of precision. In some circumstances, tradeoffs can be made dynamically between range and precision based on the statistical characteristics of the data. Exemplary transformations or normalizations that may be performed by the normalize and transform logic 2106A-2106B include weight normalization to a range of values or a set of persistent and reversible feature data transformation. In one embodiment, weight normalization can be performed to compress the dynamic range of a set of filter weights to within a pre-determined range. Weight data can be normalized, for example, within a range of [−1,1], which can preserve the relative differences between weight values while reducing the overall magnitude of the weight values. In one embodiment, neural network weight or feature map data can be normalized via the mean value of the dataset.

In one embodiment, neural network calculations using data that approaches the range limits of the data type can be transformed prior to the data is used in calculations. For example, a multiply operation using large values that may result in an overflow can be performed as an addition of logarithms instead of a multiply operation. While such transformations may result in some degree of precision loss, the calculations will be able to be performed without overflowing the number of bits allocated to perform the operation. For example, a series of operations may be presented as in equation (1).

$\begin{array}{cc}f=\frac{A\times B\times C}{D\times E}& \left(1\right)\end{array}$

Should precision tracking logic within a compute unit determine that such operation may overflow or underflow, the operation can be transformed to equation (2).

ƒ=10^{(log(A)+log(B)+log(C)−log(D)−log(E))}  (2)

Equation (2) can be performed to produce a result without triggering an overflow of the datatype. In one embodiment, the normalize and transform logic 2106A-2016B can transform the output values to logarithmic values for storage and transform the values via exponentiation before using the values for machine learning calculations described herein.

FIG. 22 is a flow diagram of logic 2200 to prevent error or significant precision loss when performing low precision operations for machine learning, according to an embodiment. In one embodiment the logic 2200 can be implemented via precision tracking logic 1412 and a numerical transform unit 1422 within a dynamic precision floating point unit 1400, as in FIG. 14.

In one embodiment the logic 2200 can compute an activation map based on filter and feature map data associated with a layer of a neural network, as shown at block 2202. The logic 2200 can then track loss of precision that occurs during computation of an activation map for the layer of the neural network. The logic 2200 can then determine if the loss of precision is approaching a threshold at block 2205. If the loss of precision does not approach the default or configured threshold at block 2205, the logic 2200 can continue to compute activation maps (and apply activation functions) for successive layers until and unless a loss of precision approaching a threshold occurs at block 2205. When the loss of precision approaches the threshold, the logic 2200 can determine if automatic numerical transformations are enabled at block 2207. If automatic transforms are enabled at block 2207, for example, via the instructions used to perform the set of numerical operations, the logic 2200 can transform the neural network data to reduce error due to precision loss at block 2208. The logic 2200 can perform any of the numerical transformations described herein, including data normalization within a range or via a mean value. Without regard to whether automatic transformations are enabled at block 2207, the logic 2200 can output status indicating that precision loss is approaching the threshold at block 2210. The status can be output as a status flag that is output from a compute unit as a result of a performed operation. A programmer can configure software logic to response to such status by performing algorithmic adjustments to an executing program or adjusting the neural network model used to perform machine learning.

Additional Exemplary Graphics Processing System

Details of the embodiments described above can be incorporated within graphics processing systems and devices described below. The graphics processing system and devices of FIG. 23 through FIG. 36 illustrate alternative systems and graphics processing hardware that can implement any and all of the techniques described above.

Additional Exemplary Graphics Processing System Overview

FIG. 23 is a block diagram of a processing system 2300, according to an embodiment. In various embodiments the system 2300 includes one or more processors 2302 and one or more graphics processors 2308, and may be a single processor desktop system, a multiprocessor workstation system, or a server system having a large number of processors 2302 or processor cores 2307. In one embodiment, the system 2300 is a processing platform incorporated within a system-on-a-chip (SoC) integrated circuit for use in mobile, handheld, or embedded devices.

An embodiment of system 2300 can include, or be incorporated within a server-based gaming platform, a game console, including a game and media console, a mobile gaming console, a handheld game console, or an online game console. In some embodiments system 2300 is a mobile phone, smart phone, tablet computing device or mobile Internet device. Data processing system 2300 can also include, couple with, or be integrated within a wearable device, such as a smart watch wearable device, smart eyewear device, augmented reality device, or virtual reality device. In some embodiments, data processing system 2300 is a television or set top box device having one or more processors 2302 and a graphical interface generated by one or more graphics processors 2308.

In some embodiments, the one or more processors 2302 each include one or more processor cores 2307 to process instructions which, when executed, perform operations for system and user software. In some embodiments, each of the one or more processor cores 2307 is configured to process a specific instruction set 2309. In some embodiments, instruction set 2309 may facilitate Complex Instruction Set Computing (CISC), Reduced Instruction Set Computing (RISC), or computing via a Very Long Instruction Word (VLIW). Multiple processor cores 2307 may each process a different instruction set 2309, which may include instructions to facilitate the emulation of other instruction sets. Processor core 2307 may also include other processing devices, such a Digital Signal Processor (DSP).

In some embodiments, the processor 2302 includes cache memory 2304. Depending on the architecture, the processor 2302 can 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 2302. In some embodiments, the processor 2302 also uses an external cache (e.g., a Level-3 (L3) cache or Last Level Cache (LLC)) (not shown), which may be shared among processor cores 2307 using known cache coherency techniques. A register file 2306 is additionally included in processor 2302 which may include different types of registers for storing different types of data (e.g., integer registers, floating point registers, status registers, and an instruction pointer register). Some registers may be general-purpose registers, while other registers may be specific to the design of the processor 2302.

In some embodiments, processor 2302 is coupled with a processor bus 2310 to transmit communication signals such as address, data, or control signals between processor 2302 and other components in system 2300. In one embodiment the system 2300 uses an exemplary ‘hub’ system architecture, including a memory controller hub 2316 and an Input Output (10) controller hub 2330. A memory controller hub 2316 facilitates communication between a memory device and other components of system 2300, while an I/O Controller Hub (ICH) 2330 provides connections to I/O devices via a local I/O bus. In one embodiment, the logic of the memory controller hub 2316 is integrated within the processor.

Memory device 2320 can be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, flash memory device, phase-change memory device, or some other memory device having suitable performance to serve as process memory. In one embodiment the memory device 2320 can operate as system memory for the system 2300, to store data 2322 and instructions 2321 for use when the one or more processors 2302 executes an application or process. Memory controller hub 2316 also couples with an optional external graphics processor 2312, which may communicate with the one or more graphics processors 2308 in processors 2302 to perform graphics and media operations.

In some embodiments, ICH 2330 enables peripherals to connect to memory device 2320 and processor 2302 via a high-speed I/O bus. The I/O peripherals include, but are not limited to, an audio controller 2346, a firmware interface 2328, a wireless transceiver 2326 (e.g., Wi-Fi, Bluetooth), a data storage device 2324 (e.g., hard disk drive, flash memory, etc.), and a legacy I/O controller 2340 for coupling legacy (e.g., Personal System 2 (PS/2)) devices to the system. One or more Universal Serial Bus (USB) controllers 2342 connect input devices, such as keyboard and mouse 2344 combinations. A network controller 2334 may also couple with ICH 2330. In some embodiments, a high-performance network controller (not shown) couples with processor bus 2310. It will be appreciated that the system 2300 shown is exemplary and not limiting, as other types of data processing systems that are differently configured may also be used. For example, the I/O controller hub 2330 may be integrated within the one or more processor 2302, or the memory controller hub 2316 and I/O controller hub 2330 may be integrated into a discreet external graphics processor, such as the external graphics processor 2312.

FIG. 24 is a block diagram of an embodiment of a processor 2400 having one or more processor cores 2402A-2402N, an integrated memory controller 2414, and an integrated graphics processor 2408. Those elements of FIG. 24 having the same reference numbers (or names) as the elements of any other figure herein can operate or function in any manner similar to that described elsewhere herein, but are not limited to such. Processor 2400 can include additional cores up to and including additional core 2402N represented by the dashed lined boxes. Each of processor cores 2402A-2402N includes one or more internal cache units 2404A-2404N. In some embodiments each processor core also has access to one or more shared cached units 2406.

The internal cache units 2404A-2404N and shared cache units 2406 represent a cache memory hierarchy within the processor 2400. 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 levels of cache, where the highest level of cache before external memory is classified as the LLC. In some embodiments, cache coherency logic maintains coherency between the various cache units 2406 and 2404A-2404N.

In some embodiments, processor 2400 may also include a set of one or more bus controller units 2416 and a system agent core 2410. The one or more bus controller units 2416 manage a set of peripheral buses, such as one or more Peripheral Component Interconnect buses (e.g., PCI, PCI Express). System agent core 2410 provides management functionality for the various processor components. In some embodiments, system agent core 2410 includes one or more integrated memory controllers 2414 to manage access to various external memory devices (not shown).

In some embodiments, one or more of the processor cores 2402A-2402N include support for simultaneous multi-threading. In such embodiment, the system agent core 2410 includes components for coordinating and operating cores 2402A-2402N during multi-threaded processing. System agent core 2410 may additionally include a power control unit (PCU), which includes logic and components to regulate the power state of processor cores 2402A-2402N and graphics processor 2408.

In some embodiments, processor 2400 additionally includes graphics processor 2408 to execute graphics processing operations. In some embodiments, the graphics processor 2408 couples with the set of shared cache units 2406, and the system agent core 2410, including the one or more integrated memory controllers 2414. In some embodiments, a display controller 2411 is coupled with the graphics processor 2408 to drive graphics processor output to one or more coupled displays. In some embodiments, display controller 2411 may be a separate module coupled with the graphics processor via at least one interconnect, or may be integrated within the graphics processor 2408 or system agent core 2410.

In some embodiments, a ring-based interconnect 2412 is used to couple the internal components of the processor 2400. However, an alternative interconnect unit 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 embodiments, graphics processor 2408 couples with the ring-based interconnect 2412 via an I/O link 2413.

The exemplary I/O link 2413 represents at least one of multiple varieties of I/O interconnects, including an on package I/O interconnect which facilitates communication between various processor components and a high-performance embedded memory module 2418, such as an eDRAM module. In some embodiments, each of the processor cores 2402A-2402N and graphics processor 2408 use embedded memory modules 2418 as a shared Last Level Cache.

In some embodiments, processor cores 2402A-2402N are homogenous cores executing the same instruction set architecture. In another embodiment, processor cores 2402A-2402N are heterogeneous in terms of instruction set architecture (ISA), where one or more of processor cores 2402A-2402N execute a first instruction set, while at least one of the other cores executes a subset of the first instruction set or a different instruction set. In one embodiment processor cores 2402A-2402N are heterogeneous in terms of microarchitecture, where one or more cores having a relatively higher power consumption couple with one or more power cores having a lower power consumption. Additionally, processor 2400 can be implemented on one or more chips or as an SoC integrated circuit having the illustrated components, in addition to other components.

FIG. 25 is a block diagram of a graphics processor 2500, which may be a discrete graphics processing unit, or may be a graphics processor integrated with a plurality of processing cores. In some embodiments, the graphics processor communicates via a memory mapped I/O interface to registers on the graphics processor and with commands placed into the processor memory. In some embodiments, graphics processor 2500 includes a memory interface 2514 to access memory. Memory interface 2514 can be an interface to local memory, one or more internal caches, one or more shared external caches, and/or to system memory.

In some embodiments, graphics processor 2500 also includes a display controller 2502 to drive display output data to a display device 2520. Display controller 2502 includes hardware for one or more overlay planes for the display and composition of multiple layers of video or user interface elements. In some embodiments, graphics processor 2500 includes a video codec engine 2506 to encode, decode, or transcode media to, from, or between one or more media encoding formats, including, but 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, as well as the Society of Motion Picture & Television Engineers (SMPTE) 421M/VC-1, and Joint Photographic Experts Group (JPEG) formats such as JPEG, and Motion JPEG (MJPEG) formats.

In some embodiments, graphics processor 2500 includes a block image transfer (BLIT) engine 2504 to perform two-dimensional (2D) rasterizer operations including, for example, bit-boundary block transfers. However, in one embodiment, 2D graphics operations are performed using one or more components of graphics processing engine (GPE) 2510. In some embodiments, GPE 2510 is a compute engine for performing graphics operations, including three-dimensional (3D) graphics operations and media operations.

In some embodiments, GPE 310 includes a 3D pipeline 2512 for performing 3D operations, such as rendering three-dimensional images and scenes using processing functions that act upon 3D primitive shapes (e.g., rectangle, triangle, etc.). The 3D pipeline 2512 includes programmable and fixed function elements that perform various tasks within the element and/or spawn execution threads to a 3D/Media sub-system 2515. While 3D pipeline 2512 can be used to perform media operations, an embodiment of GPE 2510 also includes a media pipeline 2516 that is specifically used to perform media operations, such as video post-processing and image enhancement.

In some embodiments, media pipeline 2516 includes fixed function or programmable logic units to perform one or more specialized media operations, such as video decode acceleration, video de-interlacing, and video encode acceleration in place of, or on behalf of video codec engine 2506. In some embodiments, media pipeline 2516 additionally includes a thread spawning unit to spawn threads for execution on 3D/Media sub-system 2515. The spawned threads perform computations for the media operations on one or more graphics execution units included in 3D/Media sub-system 2515.

In some embodiments, 3D/Media sub-system 2515 includes logic for executing threads spawned by 3D pipeline 2512 and media pipeline 2516. In one embodiment, the pipelines send thread execution requests to 3D/Media sub-system 2515, which includes thread dispatch logic for arbitrating and dispatching the various requests to available thread execution resources. The execution resources include an array of graphics execution units to process the 3D and media threads. In some embodiments, 3D/Media sub-system 2515 includes one or more internal caches for thread instructions and data. In some embodiments, the subsystem also includes shared memory, including registers and addressable memory, to share data between threads and to store output data.

Exemplary Additional Graphics Processing Engine

FIG. 26 is a block diagram of a graphics processing engine 2610 of a graphics processor in accordance with some embodiments. In one embodiment, the graphics processing engine (GPE) 2610 is a version of the GPE 2510 shown in FIG. 25. Elements of FIG. 26 having the same reference numbers (or names) as the elements of any other figure herein can operate or function in any manner similar to that described elsewhere herein, but are not limited to such. For example, the 3D pipeline 2512 and media pipeline 2516 of FIG. 25 are illustrated. The media pipeline 2516 is optional in some embodiments of the GPE 2610 and may not be explicitly included within the GPE 2610. For example and in at least one embodiment, a separate media and/or image processor is coupled to the GPE 2610.

In some embodiments, GPE 2610 couples with or includes a command streamer 2603, which provides a command stream to the 3D pipeline 2512 and/or media pipelines 2516. In some embodiments, command streamer 2603 is coupled with memory, which can be system memory, or one or more of internal cache memory and shared cache memory. In some embodiments, command streamer 2603 receives commands from the memory and sends the commands to 3D pipeline 2512 and/or media pipeline 2516. The commands are directives fetched from a ring buffer, which stores commands for the 3D pipeline 2512 and media pipeline 2516. In one embodiment, the ring buffer can additionally include batch command buffers storing batches of multiple commands. The commands for the 3D pipeline 2512 can also include references to data stored in memory, such as but not limited to vertex and geometry data for the 3D pipeline 2512 and/or image data and memory objects for the media pipeline 2516. The 3D pipeline 2512 and media pipeline 2516 process the commands and data by performing operations via logic within the respective pipelines or by dispatching one or more execution threads to a graphics core array 2614.

In various embodiments the 3D pipeline 2512 can execute one or more shader programs, such as vertex shaders, geometry shaders, pixel shaders, fragment shaders, compute shaders, or other shader programs, by processing the instructions and dispatching execution threads to the graphics core array 2614. The graphics core array 2614 provides a unified block of execution resources. Multi-purpose execution logic (e.g., execution units) within the graphics core array 2614 includes support for various 3D API shader languages and can execute multiple simultaneous execution threads associated with multiple shaders.

In some embodiments the graphics core array 2614 also includes execution logic to perform media functions, such as video and/or image processing. In one embodiment, the execution units additionally include general-purpose logic that is programmable to perform parallel general-purpose computational operations, in addition to graphics processing operations. The general-purpose logic can perform processing operations in parallel or in conjunction with general purpose logic within the processor core(s) 2307 of FIG. 23, core 2402A-2402N as in FIG. 24, or any other processor described herein.

Output data generated by threads executing on the graphics core array 2614 can output data to memory in a unified return buffer (URB) 2618. The URB 2618 can store data for multiple threads. In some embodiments the URB 2618 may be used to send data between different threads executing on the graphics core array 2614. In some embodiments the URB 2618 may additionally be used for synchronization between threads on the graphics core array and fixed function logic within the shared function logic 2620.

In some embodiments, graphics core array 2614 is scalable, such that the array includes a variable number of graphics cores, each having a variable number of execution units based on the target power and performance level of GPE 2610. In one embodiment the execution resources are dynamically scalable, such that execution resources may be enabled or disabled as needed.

The graphics core array 2614 couples with shared function logic 2620 that includes multiple resources that are shared between the graphics cores in the graphics core array. The shared functions within the shared function logic 2620 are hardware logic units that provide specialized supplemental functionality to the graphics core array 2614. In various embodiments, shared function logic 2620 includes but is not limited to sampler 2621, math 2622, and inter-thread communication (ITC) 2623 logic. Additionally, some embodiments implement one or more cache(s) 2625 within the shared function logic 2620. A shared function is implemented where the demand for a given specialized function is insufficient for inclusion within the graphics core array 2614. Instead a single instantiation of that specialized function is implemented as a stand-alone entity in the shared function logic 2620 and shared among the execution resources within the graphics core array 2614. The precise set of functions that are shared between the graphics core array 2614 and included within the graphics core array 2614 varies between embodiments.

FIG. 27 is a block diagram of another embodiment of a graphics processor 2700. Elements of FIG. 27 having the same reference numbers (or names) as the elements of any other figure herein can operate or function in any manner similar to that described elsewhere herein but are not limited to such.

In some embodiments, graphics processor 2700 includes a ring interconnect 2702, a pipeline front-end 2704, a media engine 2737, and graphics cores 2780A-2780N. In some embodiments, ring interconnect 2702 couples the graphics processor to other processing units, including other graphics processors or one or more general-purpose processor cores. In some embodiments, the graphics processor is one of many processors integrated within a multi-core processing system.

In some embodiments, graphics processor 2700 receives batches of commands via ring interconnect 2702. The incoming commands are interpreted by a command streamer 2703 in the pipeline front-end 2704. In some embodiments, graphics processor 2700 includes scalable execution logic to perform 3D geometry processing and media processing via the graphics core(s) 2780A-2780N. For 3D geometry processing commands, command streamer 2703 supplies commands to geometry pipeline 2736. For at least some media processing commands, command streamer 2703 supplies the commands to a video front end 2734, which couples with a media engine 2737. In some embodiments, media engine 2737 includes a Video Quality Engine (VQE) 2730 for video and image post-processing and a multi-format encode/decode (MFX) 2733 engine to provide hardware-accelerated media data encode and decode. In some embodiments, geometry pipeline 2736 and media engine 2737 each generate execution threads for the thread execution resources provided by at least one graphics core 2780A.

In some embodiments, graphics processor 2700 includes scalable thread execution resources featuring modular cores 2780A-2780N (sometimes referred to as core slices), each having multiple sub-cores 2750A-2750N, 2760A-2760N (sometimes referred to as core sub-slices). In some embodiments, graphics processor 2700 can have any number of graphics cores 2780A through 2780N. In some embodiments, graphics processor 2700 includes a graphics core 2780A having at least a first sub-core 2750A and a second sub-core 2760A. In other embodiments, the graphics processor is a low power processor with a single sub-core (e.g., 2750A). In some embodiments, graphics processor 2700 includes multiple graphics cores 2780A-2780N, each including a set of first sub-cores 2750A-2750N and a set of second sub-cores 2760A-2760N. Each sub-core in the set of first sub-cores 2750A-2750N includes at least a first set of execution units 2752A-2752N and media/texture samplers 2754A-2754N. Each sub-core in the set of second sub-cores 2760A-2760N includes at least a second set of execution units 2762A-2762N and samplers 2764A-2764N. In some embodiments, each sub-core 2750A-2750N, 2760A-2760N shares a set of shared resources 2770A-2770N. In some embodiments, the shared resources include shared cache memory and pixel operation logic. Other shared resources may also be included in the various embodiments of the graphics processor.

Exemplary Additional Execution Units

FIG. 28 illustrates thread execution logic 2800 including an array of processing elements employed in some embodiments of a GPE. Elements of FIG. 28 having the same reference numbers (or names) as the elements of any other figure herein can operate or function in any manner similar to that described elsewhere herein but are not limited to such.

In some embodiments, thread execution logic 2800 includes a shader processor 2802, a thread dispatcher 2804, instruction cache 2806, a scalable execution unit array including a plurality of execution units 2808A-2808N, a sampler 2810, a data cache 2812, and a data port 2814. In one embodiment the scalable execution unit array can dynamically scale by enabling or disabling one or more execution units (e.g., any of execution unit 2808A, 2808B, 2808C, 2808D, through 2808N-1 and 2808N) based on the computational requirements of a workload. In one embodiment the included components are interconnected via an interconnect fabric that links to each of the components. In some embodiments, thread execution logic 2800 includes one or more connections to memory, such as system memory or cache memory, through one or more of instruction cache 2806, data port 2814, sampler 2810, and execution units 2808A-2808N. In some embodiments, each execution unit (e.g. 2808A) is a stand-alone programmable general-purpose computational unit that is capable of executing multiple simultaneous hardware threads while processing multiple data elements in parallel for each thread. In various embodiments, the array of execution units 2808A-2808N is scalable to include any number individual execution units.

In some embodiments, the execution units 2808A-2808N are primarily used to execute shader programs. A shader processor 2802 can process the various shader programs and dispatch execution threads associated with the shader programs via a thread dispatcher 2804. In one embodiment the thread dispatcher includes logic to arbitrate thread initiation requests from the graphics and media pipelines and instantiate the requested threads on one or more execution unit in the execution units 2808A-2808N. For example, the geometry pipeline (e.g., 2736 of FIG. 27) can dispatch vertex, tessellation, or geometry shaders to the thread execution logic 2800 (FIG. 28) for processing. In some embodiments, thread dispatcher 2804 can also process runtime thread spawning requests from the executing shader programs.

In some embodiments, the execution units 2808A-2808N support an instruction set that includes native support for many standard 3D graphics shader instructions, such that shader programs from graphics libraries (e.g., Direct 3D and OpenGL) are executed with a 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 and media shaders). Each of the execution units 2808A-2808N is capable of multi-issue single instruction multiple data (SIMD) execution and multi-threaded operation enables an efficient execution environment in the face of higher latency memory accesses. Each hardware thread within each execution unit has a dedicated high-bandwidth register file and associated independent thread-state. Execution is multi-issue per clock to pipelines capable of integer, single and double precision floating point operations, SIMD branch capability, 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 2808A-2808N causes a waiting thread to sleep until the requested data has been returned. While the waiting thread is sleeping, hardware resources may be devoted to processing other threads. For example, during a delay associated with a vertex shader operation, an execution unit can perform operations for a pixel shader, fragment shader, or another type of shader program, including a different vertex shader.

Each execution unit in execution units 2808A-2808N operates on arrays of data elements. The number of data elements is the “execution size,” or the number of channels for the instruction. An execution channel is a logical unit of execution for data element access, masking, and flow control within instructions. The number of channels may be independent of the number of physical Arithmetic Logic Units (ALUs) or Floating Point Units (FPUs) for a particular graphics processor. In some embodiments, execution units 2808A-2808N support integer and floating point data types.

The execution unit instruction set includes SIMD instructions. The various data elements can be stored as a packed data type in a register and the execution unit will process the various elements based on the data size of the elements. For example, when operating on a 256-bit wide vector, the 256 bits of the vector are stored in a register and the execution unit operates on the vector as four separate 64-bit packed data elements (Quad-Word (QW) size data elements), eight separate 32-bit packed data elements (Double Word (DW) size data elements), sixteen separate 16-bit packed data elements (Word (W) size data elements), or thirty-two separate 8-bit data elements (byte (B) size data elements). However, different vector widths and register sizes are possible.

One or more internal instruction caches (e.g., 2806) are included in the thread execution logic 2800 to cache thread instructions for the execution units. In some embodiments, one or more data caches (e.g., 2812) are included to cache thread data during thread execution. In some embodiments, a sampler 2810 is included to provide texture sampling for 3D operations and media sampling for media operations. In some embodiments, sampler 2810 includes specialized texture or media sampling functionality to process texture or media data during the sampling process before providing the sampled data to an execution unit.

During execution, the graphics and media pipelines send thread initiation requests to thread execution logic 2800 via thread spawning 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.) within the shader processor 2802 is invoked to further compute output information and cause results to be written to output surfaces (e.g., color buffers, depth buffers, stencil buffers, etc.). In some embodiments, a pixel shader or fragment shader calculates the values of the various vertex attributes that are to be interpolated across the rasterized object. In some embodiments, pixel processor logic within the shader processor 2802 then executes an application programming interface (API)-supplied pixel or fragment shader program. To execute the shader program, the shader processor 2802 dispatches threads to an execution unit (e.g., 2808A) via thread dispatcher 2804. In some embodiments, shader processor 2802 uses texture sampling logic in the sampler 2810 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 discards one or more pixels from further processing.

In some embodiments, the data port 2814 provides a memory access mechanism for the thread execution logic 2800 output processed data to memory for processing on a graphics processor output pipeline. In some embodiments, the data port 2814 includes or couples to one or more cache memories (e.g., data cache 2812) to cache data for memory access via the data port.

FIG. 29 is a block diagram illustrating graphics processor instruction formats 2900 according to some embodiments. In one or more embodiment, the graphics processor execution units support an instruction set having instructions in multiple formats. The solid lined boxes illustrate the components that are generally included in an execution unit instruction, while the dashed lines include components that are optional or that are only included in a sub-set of the instructions. In some embodiments, the instruction formats 2900 described and illustrated are macro-instructions, in that they are instructions supplied to the execution unit, as opposed to micro-operations resulting from instruction decode once the instruction is processed.

In some embodiments, the graphics processor execution units natively support instructions in a 128-bit instruction format 2910. A 64-bit compacted instruction format 2930 is available for some instructions based on the selected instruction, instruction options, and number of operands. The native 128-bit instruction format 2910 provides access to all instruction options, while some options and operations are restricted in the 64-bit format 2930. The native instructions available in the 64-bit format 2930 vary by embodiment. In some embodiments, the instruction is compacted in part using a set of index values in an index field 2913. The execution unit hardware references a set of compaction tables based on the index values and uses the compaction table outputs to reconstruct a native instruction in the 128-bit instruction format 2910.

For each format, instruction opcode 2912 defines the operation that the execution unit is to perform. The execution units execute each instruction in parallel across the multiple data elements of each operand. For example, in response to an add instruction the execution unit performs a simultaneous add operation across each color channel representing a texture element or picture element. By default, the execution unit performs each instruction across all data channels of the operands. In some embodiments, instruction control field 2914 enables control over certain execution options, such as channels selection (e.g., predication) and data channel order (e.g., swizzle). For instructions in the 128-bit instruction format 2910 an exec-size field 2916 limits the number of data channels that will be executed in parallel. In some embodiments, exec-size field 2916 is not available for use in the 64-bit compact instruction format 2930.

Some execution unit instructions have up to three operands including two source operands, src02920, src12922, and one destination 2918. In some embodiments, the execution units support dual destination instructions, where one of the destinations is implied. Data manipulation instructions can have a third source operand (e.g., SRC22924), where the instruction opcode 2912 determines the number of source operands. An instruction's last source operand can be an immediate (e.g., hard-coded) value passed with the instruction.

In some embodiments, the 128-bit instruction format 2910 includes an access/address mode field 2926 specifying, for example, whether direct register addressing mode or indirect register addressing mode is used. When direct register addressing mode is used, the register address of one or more operands is directly provided by bits in the instruction.

In some embodiments, the 128-bit instruction format 2910 includes an access/address mode field 2926, which specifies an address mode and/or an access mode for the instruction. In one embodiment the access mode is used to define a data access alignment for the instruction. Some embodiments support access modes including a 16-byte aligned access mode and a 1-byte aligned access mode, where the byte alignment of the access mode determines the access alignment of the instruction operands. For example, when in a first mode, the instruction may use byte-aligned addressing for source and destination operands and when in a second mode, the instruction may use 16-byte-aligned addressing for all source and destination operands.

In one embodiment, the address mode portion of the access/address mode field 2926 determines whether the instruction is to use direct or indirect addressing. When direct register addressing mode is used bits in the instruction directly provide the register address of one or more operands. When indirect register addressing mode is used, the register address of one or more operands may be computed based on an address register value and an address immediate field in the instruction.

In some embodiments instructions are grouped based on opcode 2912 bit-fields to simplify Opcode decode 2940. For an 8-bit opcode, bits 4, 5, and 6 allow the execution unit to determine the type of opcode. The precise opcode grouping shown is merely an example. In some embodiments, a move and logic opcode group 2942 includes data movement and logic instructions (e.g., move (mov), compare (cmp)). In some embodiments, move and logic group 2942 shares the five most significant bits (MSB), where move (mov) instructions are in the form of 0000xxxxb and logic instructions are in the form of 0001xxxxb. A flow control instruction group 2944 (e.g., call, jump (jmp)) includes instructions in the form of 0010xxxxb (e.g., 0x20). A miscellaneous instruction group 2946 includes a mix of instructions, including synchronization instructions (e.g., wait, send) in the form of 0011xxxxb (e.g., 0x30). A parallel math instruction group 2948 includes component-wise arithmetic instructions (e.g., add, multiply (mul)) in the form of 0100xxxxb (e.g., 0x40). The parallel math group 2948 performs the arithmetic operations in parallel across data channels. The vector math group 2950 includes arithmetic instructions (e.g., dp4) in the form of 0101xxxxb (e.g., 0x50). The vector math group performs arithmetic such as dot product calculations on vector operands.

Exemplary Additional Graphics Pipeline

FIG. 30 is a block diagram of another embodiment of a graphics processor 3000. Elements of FIG. 30 having the same reference numbers (or names) as the elements of any other figure herein can operate or function in any manner similar to that described elsewhere herein but are not limited to such.

In some embodiments, graphics processor 3000 includes a graphics pipeline 3020, a media pipeline 3030, a display engine 3040, thread execution logic 3050, and a render output pipeline 3070. In some embodiments, graphics processor 3000 is a graphics processor within a multi-core processing system that includes one or more general purpose processing cores. The graphics processor is controlled by register writes to one or more control registers (not shown) or via commands issued to graphics processor 3000 via a ring interconnect 3002. In some embodiments, ring interconnect 3002 couples graphics processor 3000 to other processing components, such as other graphics processors or general-purpose processors. Commands from ring interconnect 3002 are interpreted by a command streamer 3003, which supplies instructions to individual components of graphics pipeline 3020 or media pipeline 3030.

In some embodiments, command streamer 3003 directs the operation of a vertex fetcher 3005 that reads vertex data from memory and executes vertex-processing commands provided by command streamer 3003. In some embodiments, vertex fetcher 3005 provides vertex data to a vertex shader 3007, which performs coordinate space transformation and lighting operations to each vertex. In some embodiments, vertex fetcher 3005 and vertex shader 3007 execute vertex-processing instructions by dispatching execution threads to execution units 3052A-3052B via a thread dispatcher 3031.

In some embodiments, execution units 3052A-3052B are an array of vector processors having an instruction set for performing graphics and media operations. In some embodiments, execution units 3052A-3052B have an attached L1 cache 3051 that is specific for each array or shared between the arrays. The cache can be configured as a data cache, an instruction cache, or a single cache that is partitioned to contain data and instructions in different partitions.

In some embodiments, graphics pipeline 3020 includes tessellation components to perform hardware-accelerated tessellation of 3D objects. In some embodiments, a programmable hull shader 3011 configures the tessellation operations. A programmable domain shader 3017 provides back-end evaluation of tessellation output. A tessellator 3013 operates at the direction of hull shader 3011 and contains special purpose logic to generate a set of detailed geometric objects based on a coarse geometric model that is provided as input to graphics pipeline 3020. In some embodiments, if tessellation is not used, tessellation components (e.g., hull shader 3011, tessellator 3013, and domain shader 3017) can be bypassed.

In some embodiments, complete geometric objects can be processed by a geometry shader 3019 via one or more threads dispatched to execution units 3052A-3052B or can proceed directly to the clipper 3029. In some embodiments, the geometry shader operates on entire geometric objects, rather than vertices or patches of vertices as in previous stages of the graphics pipeline. If the tessellation is disabled the geometry shader 3019 receives input from the vertex shader 3007. In some embodiments, geometry shader 3019 is programmable by a geometry shader program to perform geometry tessellation if the tessellation units are disabled.

Before rasterization, a clipper 3029 processes vertex data. The clipper 3029 may be a fixed function clipper or a programmable clipper having clipping and geometry shader functions. In some embodiments, a rasterizer and depth test component 3073 in the render output pipeline 3070 dispatches pixel shaders to convert the geometric objects into their per pixel representations. In some embodiments, pixel shader logic is included in thread execution logic 3050. In some embodiments, an application can bypass the rasterizer and depth test component 3073 and access un-rasterized vertex data via a stream out unit 3023.

The graphics processor 3000 has an interconnect bus, interconnect fabric, or some other interconnect mechanism that allows data and message passing amongst the major components of the processor. In some embodiments, execution units 3052A-3052B and associated cache(s) 3051, texture and media sampler 3054, and texture/sampler cache 3058 interconnect via a data port 3056 to perform memory access and communicate with render output pipeline components of the processor. In some embodiments, sampler 3054, caches 3051, 3058 and execution units 3052A-3052B each have separate memory access paths.

In some embodiments, render output pipeline 3070 contains a rasterizer and depth test component 3073 that converts vertex-based objects into an associated pixel-based representation. In some embodiments, the rasterizer logic includes a windower/masker unit to perform fixed function triangle and line rasterization. An associated render cache 3078 and depth cache 3079 are also available in some embodiments. A pixel operations component 3077 performs pixel-based operations on the data, though in some instances, pixel operations associated with 2D operations (e.g. bit block image transfers with blending) are performed by the 2D engine 3041 or substituted at display time by the display controller 3043 using overlay display planes. In some embodiments, a shared L3 cache 3075 is available to all graphics components, allowing the sharing of data without the use of main system memory.

In some embodiments, graphics processor media pipeline 3030 includes a media engine 3037 and a video front-end 3034. In some embodiments, video front-end 3034 receives pipeline commands from the command streamer 3003. In some embodiments, media pipeline 3030 includes a separate command streamer. In some embodiments, video front-end 3034 processes media commands before sending the command to the media engine 3037. In some embodiments, media engine 3037 includes thread spawning functionality to spawn threads for dispatch to thread execution logic 3050 via thread dispatcher 3031.

In some embodiments, graphics processor 3000 includes a display engine 3040. In some embodiments, display engine 3040 is external to processor 3000 and couples with the graphics processor via the ring interconnect 3002, or some other interconnect bus or fabric. In some embodiments, display engine 3040 includes a 2D engine 3041 and a display controller 3043. In some embodiments, display engine 3040 contains special purpose logic capable of operating independently of the 3D pipeline. In some embodiments, display controller 3043 couples with a display device (not shown), which may be a system integrated display device, as in a laptop computer, or an external display device attached via a display device connector.

In some embodiments, graphics pipeline 3020 and media pipeline 3030 are configurable to perform operations based on multiple graphics and media programming interfaces and are not specific to any one application programming interface (API). In some embodiments, driver software for the graphics processor translates API calls that are specific to a particular graphics or media library into commands that can be processed by the graphics processor. In some embodiments, support is provided for the Open Graphics Library (OpenGL), Open Computing Language (OpenCL), and/or Vulkan graphics and compute API, all from the Khronos Group. In some embodiments, support may also be provided for the Direct3D library from the Microsoft Corporation. In some embodiments, a combination of these libraries may be supported. Support may also be provided for the Open Source Computer Vision Library (OpenCV). A future API with a compatible 3D pipeline would also be supported if a mapping can be made from the pipeline of the future API to the pipeline of the graphics processor.

Graphics Pipeline Programming

FIG. 31A is a block diagram illustrating a graphics processor command format 3100 according to some embodiments. FIG. 31B is a block diagram illustrating a graphics processor command sequence 3110 according to an embodiment. The solid lined boxes in FIG. 31A illustrate the components that are generally included in a graphics command while the dashed lines include components that are optional or that are only included in a sub-set of the graphics commands. The exemplary graphics processor command format 3100 of FIG. 31A includes data fields including a target client field 3102 to indicate a target client of the command, a command operation code (opcode) field 3104, and a data field 3106 that includes relevant data for the command. A sub-opcode field 3105 and a command size field 3108 are also included in some commands.

In some embodiments, the target client field 3102 specifies the client unit of the graphics device that processes the command data. In some embodiments, a graphics processor command parser examines the client field of each command to condition the further processing of the command and route the command data to the appropriate client unit. In some embodiments, the graphics processor client units include a memory interface unit, a render unit, a 2D unit, a 3D unit, and a media unit. Each client unit has a corresponding processing pipeline that processes the commands. Once the command is received by the client unit, the client unit reads the opcode field 3104 and, if present, sub-opcode field 3105 to determine the operation to perform. The client unit performs the command using information in data field 3106. For some commands the command size field 3108 is expected to be used to explicitly specify the size of the command. In some embodiments, the command parser automatically determines the size of at least some of the commands based on the command opcode. In some embodiments commands are aligned via multiples of a double word.

The flow diagram in FIG. 31B shows an exemplary graphics processor command sequence 3110. In some embodiments, software or firmware of a data processing system that features an embodiment of a graphics processor uses a version of the command sequence shown to set up, execute, and terminate a set of graphics operations. A sample command sequence is shown and described for purposes of example only as embodiments are not limited to these specific commands or to this command sequence. Moreover, the commands may be issued as batch of commands in a command sequence, such that the graphics processor will process the sequence of commands in at least partial concurrence.

In some embodiments, the graphics processor command sequence 3110 may begin with a pipeline flush command 3112 to cause any active graphics pipeline to complete the currently pending commands for the pipeline. In some embodiments, the 3D pipeline 3122 and the media pipeline 3124 do not operate concurrently. The pipeline flush is performed to cause the active graphics pipeline to complete any pending commands. In response to a pipeline flush, the command parser for the graphics processor will pause command processing until the active drawing engines complete pending operations and the relevant read caches are invalidated. Optionally, any data in the render cache that is marked ‘dirty’ can be flushed to memory. In some embodiments, pipeline flush command 3112 can be used for pipeline synchronization or before placing the graphics processor into a low power state.

In some embodiments, a pipeline select command 3113 is used when a command sequence requires the graphics processor to explicitly switch between pipelines. In some embodiments, a pipeline select command 3113 is required only once within an execution context before issuing pipeline commands unless the context is to issue commands for both pipelines. In some embodiments, a pipeline flush command 3112 is required immediately before a pipeline switch via the pipeline select command 3113.

In some embodiments, a pipeline control command 3114 configures a graphics pipeline for operation and is used to program the 3D pipeline 3122 and the media pipeline 3124. In some embodiments, pipeline control command 3114 configures the pipeline state for the active pipeline. In one embodiment, the pipeline control command 3114 is used for pipeline synchronization and to clear data from one or more cache memories within the active pipeline before processing a batch of commands.

In some embodiments, commands to configure a return buffer state 3116 are used to configure a set of return buffers for the respective pipelines to write data. Some pipeline operations require the allocation, selection, or configuration of one or more return buffers into which the operations write intermediate data during processing. In some embodiments, the graphics processor also uses one or more return buffers to store output data and to perform cross thread communication. In some embodiments, the return buffer state 3116 includes selecting the size and number of return buffers to use for a set of pipeline operations.

The remaining commands in the command sequence differ based on the active pipeline for operations. Based on a pipeline determination 3120, the command sequence is tailored to the 3D pipeline 3122 beginning with the 3D pipeline state 3130 or the media pipeline 3124 beginning at the media pipeline state 3140.

The commands to configure the 3D pipeline state 3130 include 3D state setting commands for vertex buffer state, vertex element state, constant color state, depth buffer state, and other state variables that are to be configured before 3D primitive commands are processed. The values of these commands are determined at least in part based on the particular 3D API in use. In some embodiments, 3D pipeline state 3130 commands are also able to selectively disable or bypass certain pipeline elements if those elements will not be used.

In some embodiments, 3D primitive 3132 command is used to submit 3D primitives to be processed by the 3D pipeline. Commands and associated parameters that are passed to the graphics processor via the 3D primitive 3132 command are forwarded to the vertex fetch function in the graphics pipeline. The vertex fetch function uses the 3D primitive 3132 command data to generate vertex data structures. The vertex data structures are stored in one or more return buffers. In some embodiments, 3D primitive 3132 command is used to perform vertex operations on 3D primitives via vertex shaders. To process vertex shaders, 3D pipeline 3122 dispatches shader execution threads to graphics processor execution units.

In some embodiments, 3D pipeline 3122 is triggered via an execute 3134 command or event. In some embodiments, a register writes triggers command execution. In some embodiments execution is triggered via a ‘go’ or ‘kick’ command in the command sequence. In one embodiment, command execution is triggered using a pipeline synchronization command to flush the command sequence through the graphics pipeline. The 3D pipeline will perform geometry processing for the 3D primitives. Once operations are complete, the resulting geometric objects are rasterized and the pixel engine colors the resulting pixels. Additional commands to control pixel shading and pixel back end operations may also be included for those operations.

In some embodiments, the graphics processor command sequence 3110 follows the media pipeline 3124 path when performing media operations. In general, the specific use and manner of programming for the media pipeline 3124 depends on the media or compute operations to be performed. Specific media decode operations may be offloaded to the media pipeline during media decode. In some embodiments, the media pipeline can also be bypassed and media decode can be performed in whole or in part using resources provided by one or more general purpose processing cores. In one embodiment, the media pipeline also includes elements for general-purpose graphics processor unit (GPGPU) operations, where the graphics processor is used to perform SIMD vector operations using computational shader programs that are not explicitly related to the rendering of graphics primitives.

In some embodiments, media pipeline 3124 is configured in a similar manner as the 3D pipeline 3122. A set of commands to configure the media pipeline state 3140 are dispatched or placed into a command queue before the media object commands 3142. In some embodiments, commands to configure the media pipeline state 3140 include data to configure the media pipeline elements that will be used to process the media objects. This includes data to configure the video decode and video encode logic within the media pipeline, such as encode or decode format. In some embodiments, commands to configure the media pipeline state 3140 also support the use of one or more pointers to “indirect” state elements that contain a batch of state settings.

In some embodiments, media object commands 3142 supply pointers to media objects for processing by the media pipeline. The media objects include memory buffers containing video data to be processed. In some embodiments, all media pipeline states must be valid before issuing a media object command 3142. Once the pipeline state is configured and media object commands 3142 are queued, the media pipeline 3124 is triggered via an execute command 3144 or an equivalent execute event (e.g., register write). Output from media pipeline 3124 may then be post processed by operations provided by the 3D pipeline 3122 or the media pipeline 3124. In some embodiments, GPGPU operations are configured and executed in a similar manner as media operations.

Graphics Software Architecture

FIG. 32 illustrates exemplary graphics software architecture for a data processing system 3200 according to some embodiments. In some embodiments, software architecture includes a 3D graphics application 3210, an operating system 3220, and at least one processor 3230. In some embodiments, processor 3230 includes a graphics processor 3232 and one or more general-purpose processor core(s) 3234. The graphics application 3210 and operating system 3220 each execute in the system memory 3250 of the data processing system.

In some embodiments, 3D graphics application 3210 contains one or more shader programs including shader instructions 3212. The shader language instructions may be in a high-level shader language, such as the High-Level Shader Language (HLSL) or the OpenGL Shader Language (GLSL). The application also includes executable instructions 3214 in a machine language suitable for execution by the general-purpose processor core 3234. The application also includes graphics objects 3216 defined by vertex data.

In some embodiments, operating system 3220 is a Microsoft® Windows® operating system from the Microsoft Corporation, a proprietary UNIX-like operating system, or an open source UNIX-like operating system using a variant of the Linux kernel. The operating system 3220 can support a graphics API 3222 such as the Direct3D API, the OpenGL API, or the Vulkan API. When the Direct3D API is in use, the operating system 3220 uses a front-end shader compiler 3224 to compile any shader instructions 3212 in HLSL into a lower-level shader language. The compilation may be a just-in-time (JIT) compilation or the application can perform shader pre-compilation. In some embodiments, high-level shaders are compiled into low-level shaders during the compilation of the 3D graphics application 3210. In some embodiments, the shader instructions 3212 are provided in an intermediate form, such as a version of the Standard Portable Intermediate Representation (SPIR) used by the Vulkan API.

In some embodiments, user mode graphics driver 3226 contains a back-end shader compiler 3227 to convert the shader instructions 3212 into a hardware specific representation. When the OpenGL API is in use, shader instructions 3212 in the GLSL high-level language are passed to a user mode graphics driver 3226 for compilation. In some embodiments, user mode graphics driver 3226 uses operating system kernel mode functions 3228 to communicate with a kernel mode graphics driver 3229. In some embodiments, kernel mode graphics driver 3229 communicates with graphics processor 3232 to dispatch commands and instructions.

IP Core Implementations

One or more aspects of at least one embodiment may be implemented by representative code stored on a machine-readable medium which represents and/or defines logic within an integrated circuit such as a processor. For example, the machine-readable medium may include instructions which represent various logic within the processor. When read by a machine, the instructions may cause the machine to fabricate the logic to perform 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 describes the structure of the integrated circuit. The hardware model may be supplied to various customers or manufacturing facilities, which load the hardware model on fabrication machines that manufacture the integrated circuit. The integrated circuit may be fabricated such that the circuit performs operations described in association with any of the embodiments described herein.

FIG. 33 is a block diagram illustrating an IP core development system 3300 that may be used to manufacture an integrated circuit to perform operations according to an embodiment. The IP core development system 3300 may be used to generate modular, re-usable designs that can be incorporated into a larger design or used to construct an entire integrated circuit (e.g., an SOC integrated circuit). A design facility 3330 can generate a software simulation 3310 of an IP core design in a high-level programming language (e.g., C/C++). The software simulation 3310 can be used to design, test, and verify the behavior of the IP core using a simulation model 3312. The simulation model 3312 may include functional, behavioral, and/or timing simulations. A register transfer level (RTL) design 3315 can then be created or synthesized from the simulation model 3312. The RTL design 3315 is an abstraction of the behavior of the integrated circuit that models the flow of digital signals between hardware registers, including the associated logic performed using the modeled digital signals. In addition to an RTL design 3315, lower-level designs at the logic level or transistor level may also be created, designed, or synthesized. Thus, the particular details of the initial design and simulation may vary.

The RTL design 3315 or equivalent may be further synthesized by the design facility into a hardware model 3320, which may be in a hardware description language (HDL), or some other representation of physical design data. The HDL may be further simulated or tested to verify the IP core design. The IP core design can be stored for delivery to a 3^rd party fabrication facility 3365 using non-volatile memory 3340 (e.g., hard disk, flash memory, or any non-volatile storage medium). Alternatively, the IP core design may be transmitted (e.g., via the Internet) over a wired connection 3350 or wireless connection 3360. The fabrication facility 3365 may then fabricate an integrated circuit that is based at least in part on the IP core design. The fabricated integrated circuit can be configured to perform operations in accordance with at least one embodiment described herein.

Exemplary System on a Chip Integrated Circuit

FIG. 34-36 illustrated exemplary integrated circuits and associated graphics processors that may be fabricated using one or more IP cores, according to various embodiments described herein. In addition to what is illustrated, other logic and circuits may be included, including additional graphics processors/cores, peripheral interface controllers, or general-purpose processor cores.

FIG. 34 is a block diagram illustrating an exemplary system on a chip integrated circuit 3400 that may be fabricated using one or more IP cores, according to an embodiment. Exemplary integrated circuit 3400 includes one or more application processor(s) 3405 (e.g., CPUs), at least one graphics processor 3410, and may additionally include an image processor 3415 and/or a video processor 3420, any of which may be a modular IP core from the same or multiple different design facilities. Integrated circuit 3400 includes peripheral or bus logic including a USB controller 3425, UART controller 3430, an SPI/SDIO controller 3435, and an I²S/I²C controller 3440. Additionally, the integrated circuit can include a display device 3445 coupled to one or more of a high-definition multimedia interface (HDMI) controller 3450 and a mobile industry processor interface (MIPI) display interface 3455. Storage may be provided by a flash memory subsystem 3460 including flash memory and a flash memory controller. Memory interface may be provided via a memory controller 3465 for access to SDRAM or SRAM memory devices. Some integrated circuits additionally include an embedded security engine 3470.

FIG. 35 is a block diagram illustrating an exemplary graphics processor 3510 of a system on a chip integrated circuit that may be fabricated using one or more IP cores, according to an embodiment. Graphics processor 3510 can be a variant of the graphics processor 3410 of FIG. 34. Graphics processor 3510 includes a vertex processor 3505 and one or more fragment processor(s) 3515A-3515N (e.g., 3515A, 3515B, 3515C, 3515D, through 3515N-1, and 3515N). Graphics processor 3510 can execute different shader programs via separate logic, such that the vertex processor 3505 is optimized to execute operations for vertex shader programs, while the one or more fragment processor(s) 3515A-3515N execute fragment (e.g., pixel) shading operations for fragment or pixel shader programs. The vertex processor 3505 performs the vertex processing stage of the 3D graphics pipeline and generates primitives and vertex data. The fragment processor(s) 3515A-3515N use the primitive and vertex data generated by the vertex processor 3505 to produce a framebuffer that is displayed on a display device. In one embodiment, the fragment processor(s) 3515A-3515N are optimized to execute fragment shader programs as provided for in the OpenGL API, which may be used to perform similar operations as a pixel shader program as provided for in the Direct 3D API.

Graphics processor 3510 additionally includes one or more memory management units (MMUs) 3520A-3520B, caches 3525A-3525B, and circuit interconnects 3530A-3530B. The one or more MMU(s) 3520A-3520B provide for virtual to physical address mapping for graphics processor 3510, including for the vertex processor 3505 and/or fragment processor(s) 3515A-3515N, which may reference vertex or image/texture data stored in memory, in addition to vertex or image/texture data stored in the one or more caches 3525A-3525B. In one embodiment the one or more MMU(s) 3520A-3520B may be synchronized with other MMUs within the system, including one or more MMUs associated with the one or more application processor(s) 3405, image processor 3415, and/or video processor 3420 of FIG. 34, such that each processor 3405-3420 can participate in a shared or unified virtual memory system. The one or more circuit interconnects 3530A-3530B enable graphics processor 3510 to interface with other IP cores within the SoC, either via an internal bus of the SoC or via a direct connection, according to embodiments.

FIG. 36 is a block diagram illustrating an additional exemplary graphics processor 3610 of a system on a chip integrated circuit that may be fabricated using one or more IP cores, according to an embodiment. Graphics processor 3610 can be a variant of the graphics processor 3410 of FIG. 34. Graphics processor 3610 includes the one or more MMU(s) 3520A-3520B, caches 3525A-3525B, and circuit interconnects 3530A-3530B of the graphics processor 3510 of FIG. 35.

Graphics processor 3610 includes one or more shader cores 3615A-3615N (e.g., 3615A, 3615B, 3615C, 3615D, 3615E, 3615F, through 3615N-1, and 3615N), which provides for a unified shader core architecture in which a single core or type or core can execute all types of programmable shader code, including shader program code to implement vertex shaders, fragment shaders, and/or compute shaders. The exact number of shader cores present can vary among embodiments and implementations. Additionally, graphics processor 3610 includes an inter-core task manager 3605, which acts as a thread dispatcher to dispatch execution threads to one or more shader cores 3615A-3615N and a tiling unit 3618 to accelerate tiling operations for tile-based rendering, in which rendering operations for a scene are subdivided in image space, for example to exploit local spatial coherence within a scene or to optimize use of internal caches.

The following clauses and/or examples pertain to specific embodiments or examples thereof. Specifics in the examples may be used anywhere in one or more embodiments. The various features of the different embodiments or examples may be variously combined with some features included and others excluded to suit a variety of different applications. Examples may include subject matter such as a method, means for performing acts of the method, at least one machine-readable medium including instructions that, when performed by a machine cause the machine to perform acts of the method, or of an apparatus or system according to embodiments and examples described herein. Various components can be a means for performing the operations or functions described.

One embodiment provides an accelerator module comprising a memory stack including multiple memory dies; a graphics processing unit (GPU) coupled with the memory stack via one or more memory controllers, the GPU including a plurality of multiprocessors having a single instruction, multiple thread (SIMT) architecture, the multiprocessors to execute at least one single instruction; the at least one single instruction to cause at least a portion of the GPU to perform a floating point operation on input having differing precisions; and the floating point operation is a two-dimensional matrix multiply and accumulate operation.

In one embodiment, the memory stack includes high bandwidth memory, which can be located on the same physical package as the GPU. Additionally, at least a portion of the plurality of multiprocessors can include a mixed precision core, the mixed precision core to execute at least one thread of the single instruction. The mixed precision core can perform a multiply having at least one 16-bit floating point input and an accumulate having a 32-bit floating point input. In one embodiment, the mixed precision core includes a floating point unit to perform a first operation at a first precision and a second operation at a second precision, where the first precision is lower than the second precision. In one embodiment, the single instruction can be used to accelerate a linear algebra subprogram associated with a machine learning framework.

One embodiment provides a method of accelerating a machine-learning operation, the method comprising decoding a single instruction on a graphics processing unit (GPU), the GPU having a single instruction, multiple thread (SIMT) architecture, the GPU coupled with a memory stack via one or more memory controllers; executing the single instruction via one or more multiprocessors within the GPU, the single instruction to cause at least a portion of the GPU to perform a floating point operation on input having differing precisions; and wherein the floating point operation is a two-dimensional matrix multiply and accumulate operation. In one embodiment the method additionally comprises executing at least one thread of the single instruction on a mixed precision core of the one or more multiprocessors, performing multiple operations on input using the mixed precision core to generate a two-dimensional output matrix, and storing the two-dimensional output matrix to the memory stack via the one or more memory controllers.

One embodiment provides for a data processing system comprising a non-transitory machine-readable medium to store instructions for execution; a memory stack including multiple memory dies, the memory stack including high bandwidth memory; a graphics processing unit (GPU) coupled with the memory stack via one or more memory controllers, the GPU including a plurality of multiprocessors having a single instruction, multiple thread (SIMT) architecture, the multiprocessors to execute the instructions; and wherein the instructions include at least one single instruction to perform a two-dimensional matrix multiply and accumulate operation, the at least one single instruction to cause at least a portion of the GPU to perform a floating point operation on input having differing precisions.

One embodiment provides for a graphics processor comprising a memory device, a level-two (L2) cache memory and a raster operations unit (ROP) coupled with the memory device, a compressor to perform lossless compression on data to be written to the memory device, and a streaming multiprocessor coupled with the memory device. The streaming multiprocessor can concurrently execute multiple thread groups and includes a single instruction, multiple thread (SIMT) architecture. The streaming multiprocessor can execute multiple threads for each of multiple instructions. In one embodiment the multiple instructions include a first instruction to cause at least a first portion of the streaming multiprocessor to perform a floating-point operation on multiple floating-point input operands and a second instruction to cause at least a second portion of the streaming multiprocessor to perform an integer operation on multiple integer operands, the first instruction to execute concurrently with the second instruction. In one embodiment the streaming multiprocessor includes a mixed precision core to perform operations for at least a third instruction of the multiple instructions, the mixed precision core to perform a first operation of the third instruction at a first precision and a second operation of the third instruction at a second precision, the first operation is a multiply having at least one 16-bit floating-point input, and the second operation is an accumulate having a 32-bit floating-point input.

In one embodiment the memory device is a memory stack including multiple memory dies. The memory stack can be a high-bandwidth memory (e.g., HBM) stack on the same package as the streaming multiprocessor. The graphics processor can also be included in a graphics processing system that includes the memory device. In such system the memory device can be a graphics memory device such as a GDDR (e.g., GDDR6) memory device.

The first portion of the streaming multiprocessor can include a set of logic units configured to perform floating-point operations, while the second portion of the streaming multiprocessor can include a set of logic units configured to perform integer operations. The second instruction can cause the second portion of the streaming multiprocessor to perform an 8-bit integer operation on multiple 8-bit floating-point input operands, while the first instruction can cause the first portion of the streaming multiprocessor to perform a 16-bit floating-point operation on multiple 32-bit floating-point input operands. While performing the floating-point operations, the set of logic units can track a loss of precision during execution of the first instruction. In one embodiment the third instruction is a matrix multiply and accumulate operation and the first operation of the third instruction has two or more 16-bit floating-point inputs. In one embodiment the graphics processor additionally includes a register file to store operands associated with the multiple instructions. The operands can be loaded into the register file from the L2 cache memory.

One embodiment provides a multi-chip module accelerator usable to execute tensor data processing operations a multi-chip module. The multi-chip module may include a memory stack including multiple memory dies and parallel processor circuitry communicatively coupled to the memory stack. The parallel processor circuitry may include multiprocessor cores to execute matrix multiplication and accumulate operations. The matrix multiplication and accumulate operations may include floating-point operations that are configurable to include two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions. The floating-point operations may include a first operation at a first precision and a second operation at a second precision. The first operation may include a multiply having at least one 16-bit floating-point input and the second operation may include an accumulate having a 32-bit floating-point input. In one embodiment, the memory stack comprises high bandwidth memory, which may be comprised in the same physical package as the parallel processor circuitry. The first operation may be at a 16-bit precision and the second operation may be at a 32-bit precision. The first operation, in one embodiment, involves two or more 16-bit floating-point inputs. In one embodiment, the multi-chip module accelerator is to be communicatively coupled to at least one accelerator cluster and the accelerator cluster comprises identical parallel processors communicatively coupled together. In one embodiment, the multiprocessor cores are to use a unified memory space associated with the memory stack.

Other embodiments include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the above described multi-chip module accelerator.

The embodiments described herein refer to specific configurations of hardware, such as application specific integrated circuits (ASICs), configured to perform certain operations or having a predetermined functionality. Such electronic devices typically include a set of one or more processors coupled to one or more other components, such as one or more storage devices (non-transitory machine-readable storage media), user input/output devices (e.g., a keyboard, a touchscreen, and/or a display), and network connections. The coupling of the set of processors and other components is typically through one or more busses and bridges (also termed as bus controllers). The storage device and signals carrying the network traffic respectively represent one or more machine-readable storage media and machine-readable communication media. Thus, the storage devices of a given electronic device typically store code and/or data for execution on the set of one or more processors of that electronic device.

Of course, one or more parts of an embodiment may be implemented using different combinations of software, firmware, and/or hardware. Throughout this detailed description, for the purposes of explanation, numerous specific details were set forth in order to provide a thorough understanding of the present invention. It will be apparent, however, to one skilled in the art that the embodiments may be practiced without some of these specific details. In certain instances, well-known structures and functions were not described in elaborate detail to avoid obscuring the inventive subject matter of the embodiments. Accordingly, the scope and spirit of the invention should be judged in terms of the claims that follow.

Claims

1. A multi-chip module accelerator usable to execute tensor data processing operations, the multi-chip module accelerator comprising: a multi-chip module comprising: a memory stack including multiple memory dies; andparallel processor circuitry communicatively coupled to the memory stack, the parallel processor circuitry comprising multiprocessor cores to execute matrix multiplication and accumulate operations;wherein: the matrix multiplication and accumulate operations comprise floating-point operations;the floating-point operations are configurable to comprise two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions;the floating-point operations comprise a first operation at a first precision and a second operation at a second precision; andthe first operation comprises a multiply having at least one 16-bit floating-point input and the second operation comprises an accumulate having a 32-bit floating-point input.

2. The multi-chip module accelerator of claim 1, wherein the memory stack comprises high bandwidth memory.

3. The multi-chip module accelerator of claim 1, wherein the memory stack is comprised in the same physical package as the parallel processor circuitry.

4. The multi-chip module accelerator of claim 1, wherein the first operation is at a 16-bit precision, and the second operation is at a 32-bit precision.

5. The multi-chip module accelerator of claim 1, wherein the first operation involves two or more 16-bit floating-point inputs.

6. The multi-chip module accelerator of claim 1, wherein: the multi-chip module accelerator is to be communicatively coupled to at least one accelerator cluster; andthe accelerator cluster comprises identical parallel processors communicatively coupled together.

7. The multi-chip module accelerator of claim 1, wherein: the multiprocessor cores are to use a unified memory space associated with the memory stack.

8. At least one non-transitory machine-readable storage medium storing instructions for being executed by at least one machine associated with a multi-chip module accelerator, the multi-chip module accelerator comprising a memory stack and parallel processor circuitry, the parallel processor circuitry being communicatively coupled to the memory stack, the memory stack including multiple memory dies, the parallel processor circuitry comprising multiprocessor cores, the instructions, when executed by the at least one machine, resulting in the multi-chip module accelerator being configured for performance of operations comprising: executing, by the multiprocessor cores, matrix multiplication and accumulate operations;wherein: the matrix multiplication and accumulate operations comprise floating-point operations;the floating-point operations are configurable to comprise two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions;the floating-point operations comprise a first operation at a first precision and a second operation at a second precision; andthe first operation comprises a multiply having at least one 16-bit floating-point input and the second operation comprises an accumulate having a 32-bit floating-point input.

9. The at least one non-transitory machine-readable storage medium of claim 8, wherein the memory stack comprises high bandwidth memory.

10. The at least one non-transitory machine-readable storage medium of claim 8, wherein the memory stack is comprised in the same physical package as the parallel processor circuitry.

11. The at least one non-transitory machine-readable storage medium of claim 8, wherein the first operation is at a 16-bit precision, and the second operation is at a 32-bit precision.

12. The at least one non-transitory machine-readable storage medium of claim 8, wherein the first operation involves two or more 16-bit floating-point inputs.

13. The at least one non-transitory machine-readable storage medium of claim 8, wherein: the multi-chip module accelerator is to be communicatively coupled to at least one accelerator cluster; andthe accelerator cluster comprises identical parallel processors communicatively coupled together.

14. The at least one non-transitory machine-readable storage medium of claim 8, wherein: the multiprocessor cores are to use a unified memory space associated with the memory stack.

15. A method implemented using a multi-chip module accelerator, the multi-chip module accelerator comprising a memory stack and parallel processor circuitry, the parallel processor circuitry being communicatively coupled to the memory stack, the memory stack including multiple memory dies, the parallel processor circuitry comprising multiprocessor cores, the method comprising: executing, by the multiprocessor cores, matrix multiplication and accumulate operations;wherein: the matrix multiplication and accumulate operations comprise floating-point operations;the floating-point operations are configurable to comprise two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions;the floating-point operations comprise a first operation at a first precision and a second operation at a second precision; andthe first operation comprises a multiply having at least one 16-bit floating-point input and the second operation comprises an accumulate having a 32-bit floating-point input.

16. The method of claim 15, wherein the memory stack comprises high bandwidth memory.

17. The method of claim 15, wherein the memory stack is comprised in the same physical package as the parallel processor circuitry.

18. The method of claim 15, wherein the first operation is at a 16-bit precision, and the second operation is at a 32-bit precision.

19. The method of claim 15, wherein the first operation involves two or more 16-bit floating-point inputs.

20. The method of claim 15, wherein: the multi-chip module accelerator is to be communicatively coupled to at least one accelerator cluster; andthe accelerator cluster comprises identical parallel processors communicatively coupled together.

21. The method of claim 15, wherein: the multiprocessor cores are to use a unified memory space associated with the memory stack.

22. An accelerator cluster comprising: identical parallel processors communicatively coupled together; anda multi-chip module accelerator communicatively coupled to the identical parallel processors;wherein: the multi-chip module accelerator is usable to execute tensor data processing operations;the multi-chip module accelerator comprises a multi-chip module;the multi-chip module comprises: a memory stack including multiple memory dies; andparallel processor circuitry communicatively coupled to the memory stack;the parallel processor circuitry comprises multiprocessor cores to execute matrix multiplication and accumulate operations;the matrix multiplication and accumulate operations comprise floating-point operations; the floating-point operations are configurable to comprise two-dimensional matrix multiply and accumulate operations involving inputs that have differing floating-point precisions;the floating-point operations comprise a first operation at a first precision and a second operation at a second precision; andthe first operation comprises a multiply having at least one 16-bit floating-point input and the second operation comprises an accumulate having a 32-bit floating-point input.

23. The accelerator cluster of claim 22, wherein the memory stack comprises high bandwidth memory.

24. The accelerator cluster of claim 22, wherein the memory stack is comprised in the same physical package as the parallel processor circuitry.

25. The accelerator cluster of claim 22, wherein the first operation is at a 16-bit precision, and the second operation is at a 32-bit precision.

26. The accelerator cluster of claim 22, wherein the first operation involves two or more 16-bit floating-point inputs.

27. The accelerator cluster of claim 22, wherein: the multiprocessor cores are to use a unified memory space associated with the memory stack.