The present disclosure pertains to the field of processing logic, microprocessors, and associated instruction set architecture that, when executed by the processor or other processing logic, perform logical, mathematical, or other functional operations.
Multiprocessor systems are becoming more and more common. Applications of multiprocessor systems include dynamic domain partitioning all the way down to desktop computing. In order to take advantage of multiprocessor systems, code to be executed may be separated into multiple threads for execution by various processing entities. Each thread may be executed in parallel with one another. Instructions as they are received on a processor may be decoded into terms or instruction words that are native, or more native, for execution on the processor. Processors may be implemented in a system on chip. Data structures that are organized in tuples of three to five elements may be used in media applications, High Performance Computing applications, and molecular dynamics applications.
Embodiments are illustrated by way of example and not limitation in the Figures of the accompanying drawings:
The following description describes embodiments of instructions and processing logic for performing permute sequences of operation on a processing apparatus. The permute sequences may be part of a striding operation, such as Stride-5. Such a processing apparatus may include an out-of-order processor. In the following description, numerous specific details such as processing logic, processor types, micro-architectural conditions, events, enablement mechanisms, and the like are set forth in order to provide a more thorough understanding of embodiments of the present disclosure. It will be appreciated, however, by one skilled in the art that the embodiments may be practiced without such specific details. Additionally, some well-known structures, circuits, and the like have not been shown in detail to avoid unnecessarily obscuring embodiments of the present disclosure.
Although the following embodiments are described with reference to a processor, other embodiments are applicable to other types of integrated circuits and logic devices. Similar techniques and teachings of embodiments of the present disclosure may be applied to other types of circuits or semiconductor devices that may benefit from higher pipeline throughput and improved performance. The teachings of embodiments of the present disclosure are applicable to any processor or machine that performs data manipulations. However, the embodiments are not limited to processors or machines that perform 512-bit, 256-bit, 128-bit, 64-bit, 32-bit, or 16-bit data operations and may be applied to any processor and machine in which manipulation or management of data may be performed. In addition, the following description provides examples, and the accompanying drawings show various examples for the purposes of illustration. However, these examples should not be construed in a limiting sense as they are merely intended to provide examples of embodiments of the present disclosure rather than to provide an exhaustive list of all possible implementations of embodiments of the present disclosure.
Although the below examples describe instruction handling and distribution in the context of execution units and logic circuits, other embodiments of the present disclosure may be accomplished by way of a data or instructions stored on a machine-readable, tangible medium, which when performed by a machine cause the machine to perform functions consistent with at least one embodiment of the disclosure. In one embodiment, functions associated with embodiments of the present disclosure are embodied in machine-executable instructions. The instructions may be used to cause a general-purpose or special-purpose processor that may be programmed with the instructions to perform the steps of the present disclosure. Embodiments of the present disclosure may be provided as a computer program product or software which may include a machine or computer-readable medium having stored thereon instructions which may be used to program a computer (or other electronic devices) to perform one or more operations according to embodiments of the present disclosure. Furthermore, steps of embodiments of the present disclosure might be performed by specific hardware components that contain fixed-function logic for performing the steps, or by any combination of programmed computer components and fixed-function hardware components.
Instructions used to program logic to perform embodiments of the present disclosure may be stored within a memory in the system, such as DRAM, cache, flash memory, or other storage. Furthermore, the instructions may be distributed via a network or by way of other computer-readable media. Thus a machine-readable medium may include any mechanism for storing or transmitting information in a form readable by a machine (e.g., a computer), but is not limited to, floppy diskettes, optical disks, Compact Disc, Read-Only Memory (CD-ROMs), and magneto-optical disks, Read-Only Memory (ROMs), Random Access Memory (RAM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), magnetic or optical cards, flash memory, or a tangible, machine-readable storage used in the transmission of information over the Internet via electrical, optical, acoustical or other forms of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.). Accordingly, the computer-readable medium may include any type of tangible machine-readable medium suitable for storing or transmitting electronic instructions or information in a form readable by a machine (e.g., a computer).
A design may go through various stages, from creation to simulation to fabrication. Data representing a design may represent the design in a number of manners. First, as may be useful in simulations, the hardware may be represented using a hardware description language or another functional description language. Additionally, a circuit level model with logic and/or transistor gates may be produced at some stages of the design process. Furthermore, designs, at some stage, may reach a level of data representing the physical placement of various devices in the hardware model. In cases wherein some semiconductor fabrication techniques are used, the data representing the hardware model may be the data specifying the presence or absence of various features on different mask layers for masks used to produce the integrated circuit. In any representation of the design, the data may be stored in any form of a machine-readable medium. A memory or a magnetic or optical storage such as a disc may be the machine-readable medium to store information transmitted via optical or electrical wave modulated or otherwise generated to transmit such information. When an electrical carrier wave indicating or carrying the code or design is transmitted, to the extent that copying, buffering, or retransmission of the electrical signal is performed, a new copy may be made. Thus, a communication provider or a network provider may store on a tangible, machine-readable medium, at least temporarily, an article, such as information encoded into a carrier wave, embodying techniques of embodiments of the present disclosure.
In modern processors, a number of different execution units may be used to process and execute a variety of code and instructions. Some instructions may be quicker to complete while others may take a number of clock cycles to complete. The faster the throughput of instructions, the better the overall performance of the processor. Thus it would be advantageous to have as many instructions execute as fast as possible. However, there may be certain instructions that have greater complexity and require more in terms of execution time and processor resources, such as floating point instructions, load/store operations, data moves, etc.
As more computer systems are used in internet, text, and multimedia applications, additional processor support has been introduced over time. In one embodiment, an instruction set may be associated with one or more computer architectures, including data types, instructions, register architecture, addressing modes, memory architecture, interrupt and exception handling, and external input and output (I/O).
In one embodiment, the instruction set architecture (ISA) may be implemented by one or more micro-architectures, which may include processor logic and circuits used to implement one or more instruction sets. Accordingly, processors with different micro-architectures may share at least a portion of a common instruction set. For example, Intel® Pentium 4 processors, Intel® Core™ processors, and processors from Advanced Micro Devices, Inc. of Sunnyvale Calif. implement nearly identical versions of the x86 instruction set (with some extensions that have been added with newer versions), but have different internal designs. Similarly, processors designed by other processor development companies, such as ARM Holdings, Ltd., MIPS, or their licensees or adopters, may share at least a portion of a common instruction set, but may include different processor designs. For example, the same register architecture of the ISA may be implemented in different ways in different micro-architectures using new or well-known techniques, including dedicated physical registers, one or more dynamically allocated physical registers using a register renaming mechanism (e.g., the use of a Register Alias Table (RAT), a Reorder Buffer (ROB) and a retirement register file. In one embodiment, registers may include one or more registers, register architectures, register files, or other register sets that may or may not be addressable by a software programmer.
An instruction may include one or more instruction formats. In one embodiment, an instruction format may indicate various fields (number of bits, location of bits, etc.) to specify, among other things, the operation to be performed and the operands on which that operation will be performed. In a further embodiment, some instruction formats may be further defined by instruction templates (or sub-formats). For example, the instruction templates of a given instruction format may be defined to have different subsets of the instruction format's fields and/or defined to have a given field interpreted differently. In one embodiment, an instruction may be expressed using an instruction format (and, if defined, in a given one of the instruction templates of that instruction format) and specifies or indicates the operation and the operands upon which the operation will operate.
Scientific, financial, auto-vectorized general purpose, RMS (recognition, mining, and synthesis), and visual and multimedia applications (e.g., 2D/3D graphics, image processing, video compression/decompression, voice recognition algorithms and audio manipulation) may require the same operation to be performed on a large number of data items. In one embodiment, Single Instruction Multiple Data (SIMD) refers to a type of instruction that causes a processor to perform an operation on multiple data elements. SIMD technology may be used in processors that may logically divide the bits in a register into a number of fixed-sized or variable-sized data elements, each of which represents a separate value. For example, in one embodiment, the bits in a 64-bit register may be organized as a source operand containing four separate 16-bit data elements, each of which represents a separate 16-bit value. This type of data may be referred to as ‘packed’ data type or ‘vector’ data type, and operands of this data type may be referred to as packed data operands or vector operands. In one embodiment, a packed data item or vector may be a sequence of packed data elements stored within a single register, and a packed data operand or a vector operand may a source or destination operand of a SIMD instruction (or ‘packed data instruction’ or a ‘vector instruction’). In one embodiment, a SIMD instruction specifies a single vector operation to be performed on two source vector operands to generate a destination vector operand (also referred to as a result vector operand) of the same or different size, with the same or different number of data elements, and in the same or different data element order.
SIMD technology, such as that employed by the Intel® Core™ processors having an instruction set including x86, MMX™, Streaming SIMD Extensions (SSE), SSE2, SSE3, SSE4.1, and SSE4.2 instructions, ARM processors, such as the ARM Cortex® family of processors having an instruction set including the Vector Floating Point (VFP) and/or NEON instructions, and MIPS processors, such as the Loongson family of processors developed by the Institute of Computing Technology (ICT) of the Chinese Academy of Sciences, has enabled a significant improvement in application performance (Core™ and MMX™ are registered trademarks or trademarks of Intel Corporation of Santa Clara, Calif.).
In one embodiment, destination and source registers/data may be generic terms to represent the source and destination of the corresponding data or operation. In some embodiments, they may be implemented by registers, memory, or other storage areas having other names or functions than those depicted. For example, in one embodiment, “DEST1” may be a temporary storage register or other storage area, whereas “SRC1” and “SRC2” may be a first and second source storage register or other storage area, and so forth. In other embodiments, two or more of the SRC and DEST storage areas may correspond to different data storage elements within the same storage area (e.g., a SIMD register). In one embodiment, one of the source registers may also act as a destination register by, for example, writing back the result of an operation performed on the first and second source data to one of the two source registers serving as a destination registers.
Embodiments are not limited to computer systems. Embodiments of the present disclosure may be used in other devices such as handheld devices and embedded applications. Some examples of handheld devices include cellular phones, Internet Protocol devices, digital cameras, personal digital assistants (PDAs), and handheld PCs. Embedded applications may include a micro controller, a digital signal processor (DSP), system on a chip, network computers (NetPC), set-top boxes, network hubs, wide area network (WAN) switches, or any other system that may perform one or more instructions in accordance with at least one embodiment.
Computer system 100 may include a processor 102 that may include one or more execution units 108 to perform an algorithm to perform at least one instruction in accordance with one embodiment of the present disclosure. One embodiment may be described in the context of a single processor desktop or server system, but other embodiments may be included in a multiprocessor system. System 100 may be an example of a ‘hub’ system architecture. System 100 may include a processor 102 for processing data signals. Processor 102 may include a complex instruction set computer (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor implementing a combination of instruction sets, or any other processor device, such as a digital signal processor, for example. In one embodiment, processor 102 may be coupled to a processor bus 110 that may transmit data signals between processor 102 and other components in system 100. The elements of system 100 may perform conventional functions that are well known to those familiar with the art.
In one embodiment, processor 102 may include a Level 1 (L1) internal cache memory 104. Depending on the architecture, the processor 102 may have a single internal cache or multiple levels of internal cache. In another embodiment, the cache memory may reside external to processor 102. Other embodiments may also include a combination of both internal and external caches depending on the particular implementation and needs. Register file 106 may store different types of data in various registers including integer registers, floating point registers, status registers, and instruction pointer register.
Execution unit 108, including logic to perform integer and floating point operations, also resides in processor 102. Processor 102 may also include a microcode (ucode) ROM that stores microcode for certain macroinstructions. In one embodiment, execution unit 108 may include logic to handle a packed instruction set 109. By including the packed instruction set 109 in the instruction set of a general-purpose processor 102, along with associated circuitry to execute the instructions, the operations used by many multimedia applications may be performed using packed data in a general-purpose processor 102. Thus, many multimedia applications may be accelerated and executed more efficiently by using the full width of a processor's data bus for performing operations on packed data. This may eliminate the need to transfer smaller units of data across the processor's data bus to perform one or more operations one data element at a time.
Embodiments of an execution unit 108 may also be used in micro controllers, embedded processors, graphics devices, DSPs, and other types of logic circuits. System 100 may include a memory 120. Memory 120 may be implemented as a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, flash memory device, or other memory device. Memory 120 may store instructions 119 and/or data 121 represented by data signals that may be executed by processor 102.
A system logic chip 116 may be coupled to processor bus 110 and memory 120. System logic chip 116 may include a memory controller hub (MCH). Processor 102 may communicate with MCH 116 via a processor bus 110. MCH 116 may provide a high bandwidth memory path 118 to memory 120 for storage of instructions 119 and data 121 and for storage of graphics commands, data and textures. MCH 116 may direct data signals between processor 102, memory 120, and other components in system 100 and to bridge the data signals between processor bus 110, memory 120, and system I/O 122. In some embodiments, the system logic chip 116 may provide a graphics port for coupling to a graphics controller 112. MCH 116 may be coupled to memory 120 through a memory interface 118. Graphics card 112 may be coupled to MCH 116 through an Accelerated Graphics Port (AGP) interconnect 114.
System 100 may use a proprietary hub interface bus 122 to couple MCH 116 to I/O controller hub (ICH) 130. In one embodiment, ICH 130 may provide direct connections to some I/O devices via a local I/O bus. The local I/O bus may include a high-speed I/O bus for connecting peripherals to memory 120, chipset, and processor 102. Examples may include the audio controller 129, firmware hub (flash BIOS) 128, wireless transceiver 126, data storage 124, legacy I/O controller 123 containing user input interface 125 (which may include a keyboard interface), a serial expansion port 127 such as Universal Serial Bus (USB), and a network controller 134. Data storage device 124 may comprise a hard disk drive, a floppy disk drive, a CD-ROM device, a flash memory device, or other mass storage device.
For another embodiment of a system, an instruction in accordance with one embodiment may be used with a system on a chip. One embodiment of a system on a chip comprises of a processor and a memory. The memory for one such system may include a flash memory. The flash memory may be located on the same die as the processor and other system components. Additionally, other logic blocks such as a memory controller or graphics controller may also be located on a system on a chip.
Computer system 140 comprises a processing core 159 for performing at least one instruction in accordance with one embodiment. In one embodiment, processing core 159 represents a processing unit of any type of architecture, including but not limited to a CISC, a RISC or a VLIW type architecture. Processing core 159 may also be suitable for manufacture in one or more process technologies and by being represented on a machine-readable media in sufficient detail, may be suitable to facilitate said manufacture.
Processing core 159 comprises an execution unit 142, a set of register files 145, and a decoder 144. Processing core 159 may also include additional circuitry (not shown) which may be unnecessary to the understanding of embodiments of the present disclosure. Execution unit 142 may execute instructions received by processing core 159. In addition to performing typical processor instructions, execution unit 142 may perform instructions in packed instruction set 143 for performing operations on packed data formats. Packed instruction set 143 may include instructions for performing embodiments of the disclosure and other packed instructions. Execution unit 142 may be coupled to register file 145 by an internal bus. Register file 145 may represent a storage area on processing core 159 for storing information, including data. As previously mentioned, it is understood that the storage area may store the packed data might not be critical. Execution unit 142 may be coupled to decoder 144. Decoder 144 may decode instructions received by processing core 159 into control signals and/or microcode entry points. In response to these control signals and/or microcode entry points, execution unit 142 performs the appropriate operations. In one embodiment, the decoder may interpret the opcode of the instruction, which will indicate what operation should be performed on the corresponding data indicated within the instruction.
Processing core 159 may be coupled with bus 141 for communicating with various other system devices, which may include but are not limited to, for example, synchronous dynamic random access memory (SDRAM) control 146, static random access memory (SRAM) control 147, burst flash memory interface 148, personal computer memory card international association (PCMCIA)/compact flash (CF) card control 149, liquid crystal display (LCD) control 150, direct memory access (DMA) controller 151, and alternative bus master interface 152. In one embodiment, data processing system 140 may also comprise an I/O bridge 154 for communicating with various I/O devices via an I/O bus 153. Such I/O devices may include but are not limited to, for example, universal asynchronous receiver/transmitter (UART) 155, universal serial bus (USB) 156, Bluetooth wireless UART 157 and I/O expansion interface 158.
One embodiment of data processing system 140 provides for mobile, network and/or wireless communications and a processing core 159 that may perform SIMD operations including a text string comparison operation. Processing core 159 may be programmed with various audio, video, imaging and communications algorithms including discrete transformations such as a Walsh-Hadamard transform, a fast Fourier transform (FFT), a discrete cosine transform (DCT), and their respective inverse transforms; compression/decompression techniques such as color space transformation, video encode motion estimation or video decode motion compensation; and modulation/demodulation (MODEM) functions such as pulse coded modulation (PCM).
In one embodiment, SIMD coprocessor 161 comprises an execution unit 162 and a set of register files 164. One embodiment of main processor 166 comprises a decoder 165 to recognize instructions of instruction set 163 including instructions in accordance with one embodiment for execution by execution unit 162. In other embodiments, SIMD coprocessor 161 also comprises at least part of decoder 165 (shown as 165B) to decode instructions of instruction set 163. Processing core 170 may also include additional circuitry (not shown) which may be unnecessary to the understanding of embodiments of the present disclosure.
In operation, main processor 166 executes a stream of data processing instructions that control data processing operations of a general type including interactions with cache memory 167, and input/output system 168. Embedded within the stream of data processing instructions may be SIMD coprocessor instructions. Decoder 165 of main processor 166 recognizes these SIMD coprocessor instructions as being of a type that should be executed by an attached SIMD coprocessor 161. Accordingly, main processor 166 issues these SIMD coprocessor instructions (or control signals representing SIMD coprocessor instructions) on the coprocessor bus 166. From coprocessor bus 171, these instructions may be received by any attached SIMD coprocessors. In this case, SIMD coprocessor 161 may accept and execute any received SIMD coprocessor instructions intended for it.
Data may be received via wireless interface 169 for processing by the SIMD coprocessor instructions. For one example, voice communication may be received in the form of a digital signal, which may be processed by the SIMD coprocessor instructions to regenerate digital audio samples representative of the voice communications. For another example, compressed audio and/or video may be received in the form of a digital bit stream, which may be processed by the SIMD coprocessor instructions to regenerate digital audio samples and/or motion video frames. In one embodiment of processing core 170, main processor 166, and a SIMD coprocessor 161 may be integrated into a single processing core 170 comprising an execution unit 162, a set of register files 164, and a decoder 165 to recognize instructions of instruction set 163 including instructions in accordance with one embodiment.
Some instructions may be converted into a single micro-op, whereas others need several micro-ops to complete the full operation. In one embodiment, if more than four micro-ops are needed to complete an instruction, decoder 228 may access microcode ROM 232 to perform the instruction. In one embodiment, an instruction may be decoded into a small number of micro ops for processing at instruction decoder 228. In another embodiment, an instruction may be stored within microcode ROM 232 should a number of micro-ops be needed to accomplish the operation. Trace cache 230 refers to an entry point programmable logic array (PLA) to determine a correct micro-instruction pointer for reading the micro-code sequences to complete one or more instructions in accordance with one embodiment from micro-code ROM 232. After microcode ROM 232 finishes sequencing micro-ops for an instruction, front end 201 of the machine may resume fetching micro-ops from trace cache 230.
Out-of-order execution engine 203 may prepare instructions for execution. The out-of-order execution logic has a number of buffers to smooth out and re-order the flow of instructions to optimize performance as they go down the pipeline and get scheduled for execution. The allocator logic in allocator/register renamer 215 allocates the machine buffers and resources that each uop needs in order to execute. The register renaming logic in allocator/register renamer 215 renames logic registers onto entries in a register file. The allocator 215 also allocates an entry for each uop in one of the two uop queues, one for memory operations (memory uop queue 207) and one for non-memory operations (integer/floating point uop queue 205), in front of the instruction schedulers: memory scheduler 209, fast scheduler 202, slow/general floating point scheduler 204, and simple floating point scheduler 206. Uop schedulers 202, 204, 206, determine when a uop is ready to execute based on the readiness of their dependent input register operand sources and the availability of the execution resources the uops need to complete their operation. Fast scheduler 202 of one embodiment may schedule on each half of the main clock cycle while the other schedulers may only schedule once per main processor clock cycle. The schedulers arbitrate for the dispatch ports to schedule uops for execution.
Register files 208, 210 may be arranged between schedulers 202, 204, 206, and execution units 212, 214, 216, 218, 220, 222, 224 in execution block 211. Each of register files 208, 210 perform integer and floating point operations, respectively. Each register file 208, 210, may include a bypass network that may bypass or forward just completed results that have not yet been written into the register file to new dependent uops. Integer register file 208 and floating point register file 210 may communicate data with the other. In one embodiment, integer register file 208 may be split into two separate register files, one register file for low-order thirty-two bits of data and a second register file for high order thirty-two bits of data. Floating point register file 210 may include 128-bit wide entries because floating point instructions typically have operands from 64 to 128 bits in width.
Execution block 211 may contain execution units 212, 214, 216, 218, 220, 222, 224. Execution units 212, 214, 216, 218, 220, 222, 224 may execute the instructions. Execution block 211 may include register files 208, 210 that store the integer and floating point data operand values that the micro-instructions need to execute. In one embodiment, processor 200 may comprise a number of execution units: address generation unit (AGU) 212, AGU 214, fast ALU 216, fast ALU 218, slow ALU 220, floating point ALU 222, floating point move unit 224. In another embodiment, floating point execution blocks 222, 224, may execute floating point, MMX, SIMD, and SSE, or other operations. In yet another embodiment, floating point ALU 222 may include a 64-bit by 64-bit floating point divider to execute divide, square root, and remainder micro-ops. In various embodiments, instructions involving a floating point value may be handled with the floating point hardware. In one embodiment, ALU operations may be passed to high-speed ALU execution units 216, 218. High-speed ALUs 216, 218 may execute fast operations with an effective latency of half a clock cycle. In one embodiment, most complex integer operations go to slow ALU 220 as slow ALU 220 may include integer execution hardware for long-latency type of operations, such as a multiplier, shifts, flag logic, and branch processing. Memory load/store operations may be executed by AGUs 212, 214. In one embodiment, integer ALUs 216, 218, 220 may perform integer operations on 64-bit data operands. In other embodiments, ALUs 216, 218, 220 may be implemented to support a variety of data bit sizes including sixteen, thirty-two, 128, 256, etc. Similarly, floating point units 222, 224 may be implemented to support a range of operands having bits of various widths. In one embodiment, floating point units 222, 224, may operate on 128-bit wide packed data operands in conjunction with SIMD and multimedia instructions.
In one embodiment, uops schedulers 202, 204, 206, dispatch dependent operations before the parent load has finished executing. As uops may be speculatively scheduled and executed in processor 200, processor 200 may also include logic to handle memory misses. If a data load misses in the data cache, there may be dependent operations in flight in the pipeline that have left the scheduler with temporarily incorrect data. A replay mechanism tracks and re-executes instructions that use incorrect data. Only the dependent operations might need to be replayed and the independent ones may be allowed to complete. The schedulers and replay mechanism of one embodiment of a processor may also be designed to catch instruction sequences for text string comparison operations.
The term “registers” may refer to the on-board processor storage locations that may be used as part of instructions to identify operands. In other words, registers may be those that may be usable from the outside of the processor (from a programmer's perspective). However, in some embodiments registers might not be limited to a particular type of circuit. Rather, a register may store data, provide data, and perform the functions described herein. The registers described herein may be implemented by circuitry within a processor using any number of different techniques, such as dedicated physical registers, dynamically allocated physical registers using register renaming, combinations of dedicated and dynamically allocated physical registers, etc. In one embodiment, integer registers store 32-bit integer data. A register file of one embodiment also contains eight multimedia SIMD registers for packed data. For the discussions below, the registers may be understood to be data registers designed to hold packed data, such as 64-bit wide MMX™ registers (also referred to as ‘mm’ registers in some instances) in microprocessors enabled with MMX technology from Intel Corporation of Santa Clara, Calif. These MMX registers, available in both integer and floating point forms, may operate with packed data elements that accompany SIMD and SSE instructions. Similarly, 128-bit wide XMM registers relating to SSE2, SSE3, SSE4, or beyond (referred to generically as “SSEx”) technology may hold such packed data operands. In one embodiment, in storing packed data and integer data, the registers do not need to differentiate between the two data types. In one embodiment, integer and floating point data may be contained in the same register file or different register files. Furthermore, in one embodiment, floating point and integer data may be stored in different registers or the same registers.
In the examples of the following figures, a number of data operands may be described.
Generally, a data element may include an individual piece of data that is stored in a single register or memory location with other data elements of the same length. In packed data sequences relating to SSEx technology, the number of data elements stored in a XMM register may be 128 bits divided by the length in bits of an individual data element. Similarly, in packed data sequences relating to MMX and SSE technology, the number of data elements stored in an MMX register may be 64 bits divided by the length in bits of an individual data element. Although the data types illustrated in
In
In
Core 490 may be a reduced instruction set computing (RISC) core, a complex instruction set computing (CISC) core, a very long instruction word (VLIW) core, or a hybrid or alternative core type. In one embodiment, core 490 may be a special-purpose core, such as, for example, a network or communication core, compression engine, graphics core, or the like.
Front end unit 430 may include a branch prediction unit 432 coupled to an instruction cache unit 434. Instruction cache unit 434 may be coupled to an instruction translation lookaside buffer (TLB) 436. TLB 436 may be coupled to an instruction fetch unit 438, which is coupled to a decode unit 440. Decode unit 440 may decode instructions, and generate as an output one or more micro-operations, micro-code entry points, microinstructions, other instructions, or other control signals, which may be decoded from, or which otherwise reflect, or may be derived from, the original instructions. The decoder may be implemented using various different mechanisms. Examples of suitable mechanisms include, but are not limited to, look-up tables, hardware implementations, programmable logic arrays (PLAs), microcode read-only memories (ROMs), etc. In one embodiment, instruction cache unit 434 may be further coupled to a level 2 (L2) cache unit 476 in memory unit 470. Decode unit 440 may be coupled to a rename/allocator unit 452 in execution engine unit 450.
Execution engine unit 450 may include rename/allocator unit 452 coupled to a retirement unit 454 and a set of one or more scheduler units 456. Scheduler units 456 represent any number of different schedulers, including reservations stations, central instruction window, etc. Scheduler units 456 may be coupled to physical register file units 458. Each of physical register file units 458 represents one or more physical register files, different ones of which store one or more different data types, such as scalar integer, scalar floating point, packed integer, packed floating point, vector integer, vector floating point, etc., status (e.g., an instruction pointer that is the address of the next instruction to be executed), etc. Physical register file units 458 may be overlapped by retirement unit 454 to illustrate various ways in which register renaming and out-of-order execution may be implemented (e.g., using one or more reorder buffers and one or more retirement register files, using one or more future files, one or more history buffers, and one or more retirement register files; using register maps and a pool of registers; etc.). Generally, the architectural registers may be visible from the outside of the processor or from a programmer's perspective. The registers might not be limited to any known particular type of circuit. Various different types of registers may be suitable as long as they store and provide data as described herein. Examples of suitable registers include, but might not be limited to, dedicated physical registers, dynamically allocated physical registers using register renaming, combinations of dedicated and dynamically allocated physical registers, etc. Retirement unit 454 and physical register file units 458 may be coupled to execution clusters 460. Execution clusters 460 may include a set of one or more execution units 462 and a set of one or more memory access units 464. Execution units 462 may perform various operations (e.g., shifts, addition, subtraction, multiplication) and on various types of data (e.g., scalar floating point, packed integer, packed floating point, vector integer, vector floating point). While some embodiments may include a number of execution units dedicated to specific functions or sets of functions, other embodiments may include only one execution unit or multiple execution units that all perform all functions. Scheduler units 456, physical register file units 458, and execution clusters 460 are shown as being possibly plural because certain embodiments create separate pipelines for certain types of data/operations (e.g., a scalar integer pipeline, a scalar floating point/packed integer/packed floating point/vector integer/vector floating point pipeline, and/or a memory access pipeline that each have their own scheduler unit, physical register file unit, and/or execution cluster—and in the case of a separate memory access pipeline, certain embodiments may be implemented in which only the execution cluster of this pipeline has memory access units 464). It should also be understood that where separate pipelines are used, one or more of these pipelines may be out-of-order issue/execution and the rest in-order.
The set of memory access units 464 may be coupled to memory unit 470, which may include a data TLB unit 472 coupled to a data cache unit 474 coupled to a level 2 (L2) cache unit 476. In one exemplary embodiment, memory access units 464 may include a load unit, a store address unit, and a store data unit, each of which may be coupled to data TLB unit 472 in memory unit 470. L2 cache unit 476 may be coupled to one or more other levels of cache and eventually to a main memory.
By way of example, the exemplary register renaming, out-of-order issue/execution core architecture may implement pipeline 400 as follows: 1) instruction fetch 438 may perform fetch and length decoding stages 402 and 404; 2) decode unit 440 may perform decode stage 406; 3) rename/allocator unit 452 may perform allocation stage 408 and renaming stage 410; 4) scheduler units 456 may perform schedule stage 412; 5) physical register file units 458 and memory unit 470 may perform register read/memory read stage 414; execution cluster 460 may perform execute stage 416; 6) memory unit 470 and physical register file units 458 may perform write-back/memory-write stage 418; 7) various units may be involved in the performance of exception handling stage 422; and 8) retirement unit 454 and physical register file units 458 may perform commit stage 424.
Core 490 may support one or more instructions sets (e.g., the x86 instruction set (with some extensions that have been added with newer versions); the MIPS instruction set of MIPS Technologies of Sunnyvale, Calif.; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, Calif.).
It should be understood that the core may support multithreading (executing two or more parallel sets of operations or threads) in a variety of manners. Multithreading support may be performed by, for example, including time sliced multithreading, simultaneous multithreading (where a single physical core provides a logical core for each of the threads that physical core is simultaneously multithreading), or a combination thereof. Such a combination may include, for example, time sliced fetching and decoding and simultaneous multithreading thereafter such as in the Intel® Hyperthreading technology.
While register renaming may be described in the context of out-of-order execution, it should be understood that register renaming may be used in an in-order architecture. While the illustrated embodiment of the processor may also include a separate instruction and data cache units 434/474 and a shared L2 cache unit 476, other embodiments may have a single internal cache for both instructions and data, such as, for example, a Level 1 (L1) internal cache, or multiple levels of internal cache. In some embodiments, the system may include a combination of an internal cache and an external cache that may be external to the core and/or the processor. In other embodiments, all of the caches may be external to the core and/or the processor.
Processor 500 may include any suitable mechanism for interconnecting cores 502, system agent 510, and caches 506, and graphics module 560. In one embodiment, processor 500 may include a ring-based interconnect unit 508 to interconnect cores 502, system agent 510, and caches 506, and graphics module 560. In other embodiments, processor 500 may include any number of well-known techniques for interconnecting such units. Ring-based interconnect unit 508 may utilize memory control units 552 to facilitate interconnections.
Processor 500 may include a memory hierarchy comprising one or more levels of caches within the cores, one or more shared cache units such as caches 506, or external memory (not shown) coupled to the set of integrated memory controller units 552. Caches 506 may include any suitable cache. In one embodiment, caches 506 may include one or more mid-level caches, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, a last level cache (LLC), and/or combinations thereof.
In various embodiments, one or more of cores 502 may perform multi-threading. System agent 510 may include components for coordinating and operating cores 502. System agent unit 510 may include for example a power control unit (PCU). The PCU may be or include logic and components needed for regulating the power state of cores 502. System agent 510 may include a display engine 512 for driving one or more externally connected displays or graphics module 560. System agent 510 may include an interface 514 for communications busses for graphics. In one embodiment, interface 514 may be implemented by PCI Express (PCIe). In a further embodiment, interface 514 may be implemented by PCI Express Graphics (PEG). System agent 510 may include a direct media interface (DMI) 516. DMI 516 may provide links between different bridges on a motherboard or other portion of a computer system. System agent 510 may include a PCIe bridge 518 for providing PCIe links to other elements of a computing system. PCIe bridge 518 may be implemented using a memory controller 520 and coherence logic 522.
Cores 502 may be implemented in any suitable manner. Cores 502 may be homogenous or heterogeneous in terms of architecture and/or instruction set. In one embodiment, some of cores 502 may be in-order while others may be out-of-order. In another embodiment, two or more of cores 502 may execute the same instruction set, while others may execute only a subset of that instruction set or a different instruction set.
Processor 500 may include a general-purpose processor, such as a Core™ i3, i5, i7, 2 Duo and Quad, Xeon™, Itanium™, XScale™ or StrongARM™ processor, which may be available from Intel Corporation, of Santa Clara, Calif. Processor 500 may be provided from another company, such as ARM Holdings, Ltd, MIPS, etc. Processor 500 may be a special-purpose processor, such as, for example, a network or communication processor, compression engine, graphics processor, co-processor, embedded processor, or the like. Processor 500 may be implemented on one or more chips. Processor 500 may be a part of and/or may be implemented on one or more substrates using any of a number of process technologies, such as, for example, BiCMOS, CMOS, or NMOS.
In one embodiment, a given one of caches 506 may be shared by multiple ones of cores 502. In another embodiment, a given one of caches 506 may be dedicated to one of cores 502. The assignment of caches 506 to cores 502 may be handled by a cache controller or other suitable mechanism. A given one of caches 506 may be shared by two or more cores 502 by implementing time-slices of a given cache 506.
Graphics module 560 may implement an integrated graphics processing subsystem. In one embodiment, graphics module 560 may include a graphics processor. Furthermore, graphics module 560 may include a media engine 565. Media engine 565 may provide media encoding and video decoding.
Front end 570 may be implemented in any suitable manner, such as fully or in part by front end 201 as described above. In one embodiment, front end 570 may communicate with other portions of processor 500 through cache hierarchy 503. In a further embodiment, front end 570 may fetch instructions from portions of processor 500 and prepare the instructions to be used later in the processor pipeline as they are passed to out-of-order execution engine 580.
Out-of-order execution engine 580 may be implemented in any suitable manner, such as fully or in part by out-of-order execution engine 203 as described above. Out-of-order execution engine 580 may prepare instructions received from front end 570 for execution. Out-of-order execution engine 580 may include an allocate module 582. In one embodiment, allocate module 582 may allocate resources of processor 500 or other resources, such as registers or buffers, to execute a given instruction. Allocate module 582 may make allocations in schedulers, such as a memory scheduler, fast scheduler, or floating point scheduler. Such schedulers may be represented in
Cache hierarchy 503 may be implemented in any suitable manner. For example, cache hierarchy 503 may include one or more lower or mid-level caches, such as caches 572, 574. In one embodiment, cache hierarchy 503 may include an LLC 595 communicatively coupled to caches 572, 574. In another embodiment, LLC 595 may be implemented in a module 590 accessible to all processing entities of processor 500. In a further embodiment, module 590 may be implemented in an uncore module of processors from Intel, Inc. Module 590 may include portions or subsystems of processor 500 necessary for the execution of core 502 but might not be implemented within core 502. Besides LLC 595, Module 590 may include, for example, hardware interfaces, memory coherency coordinators, interprocessor interconnects, instruction pipelines, or memory controllers. Access to RAM 599 available to processor 500 may be made through module 590 and, more specifically, LLC 595. Furthermore, other instances of core 502 may similarly access module 590. Coordination of the instances of core 502 may be facilitated in part through module 590.
Each processor 610,615 may be some version of processor 500. However, it should be noted that integrated graphics logic and integrated memory control units might not exist in processors 610,615.
GMCH 620 may be a chipset, or a portion of a chipset. GMCH 620 may communicate with processors 610, 615 and control interaction between processors 610, 615 and memory 640. GMCH 620 may also act as an accelerated bus interface between the processors 610, 615 and other elements of system 600. In one embodiment, GMCH 620 communicates with processors 610, 615 via a multi-drop bus, such as a frontside bus (FSB) 695.
Furthermore, GMCH 620 may be coupled to a display 645 (such as a flat panel display). In one embodiment, GMCH 620 may include an integrated graphics accelerator. GMCH 620 may be further coupled to an input/output (I/O) controller hub (ICH) 650, which may be used to couple various peripheral devices to system 600. External graphics device 660 may include a discrete graphics device coupled to ICH 650 along with another peripheral device 670.
In other embodiments, additional or different processors may also be present in system 600. For example, additional processors 610, 615 may include additional processors that may be the same as processor 610, additional processors that may be heterogeneous or asymmetric to processor 610, accelerators (such as, e.g., graphics accelerators or digital signal processing (DSP) units), field programmable gate arrays, or any other processor. There may be a variety of differences between the physical resources 610, 615 in terms of a spectrum of metrics of merit including architectural, micro-architectural, thermal, power consumption characteristics, and the like. These differences may effectively manifest themselves as asymmetry and heterogeneity amongst processors 610, 615. For at least one embodiment, various processors 610, 615 may reside in the same die package.
While
Processors 770 and 780 are shown including integrated memory controller units 772 and 782, respectively. Processor 770 may also include as part of its bus controller units point-to-point (P-P) interfaces 776 and 778; similarly, second processor 780 may include P-P interfaces 786 and 788. Processors 770, 780 may exchange information via a point-to-point (P-P) interface 750 using P-P interface circuits 778, 788. As shown in
Processors 770, 780 may each exchange information with a chipset 790 via individual P-P interfaces 752, 754 using point to point interface circuits 776, 794, 786, 798. In one embodiment, chipset 790 may also exchange information with a high-performance graphics circuit 738 via a high-performance graphics interface 739.
A shared cache (not shown) may be included in either processor or outside of both processors, yet connected with the processors via P-P interconnect, such that either or both processors' local cache information may be stored in the shared cache if a processor is placed into a low power mode.
Chipset 790 may be coupled to a first bus 716 via an interface 796. In one embodiment, first bus 716 may be a Peripheral Component Interconnect (PCI) bus, or a bus such as a PCI Express bus or another third generation I/O interconnect bus, although the scope of the present disclosure is not so limited.
As shown in
In some embodiments, instructions that benefit from highly parallel, throughput processors may be performed by the GPU, while instructions that benefit from the performance of processors that benefit from deeply pipelined architectures may be performed by the CPU. For example, graphics, scientific applications, financial applications and other parallel workloads may benefit from the performance of the GPU and be executed accordingly, whereas more sequential applications, such as operating system kernel or application code may be better suited for the CPU.
In
One or more aspects of at least one embodiment may be implemented by representative data stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores” may be stored on a tangible, machine-readable medium (“tape”) and supplied to various customers or manufacturing facilities to load into the fabrication machines that actually make the logic or processor. For example, IP cores, such as the Cortex™ family of processors developed by ARM Holdings, Ltd. and Loongson IP cores developed the Institute of Computing Technology (ICT) of the Chinese Academy of Sciences may be licensed or sold to various customers or licensees, such as Texas Instruments, Qualcomm, Apple, or Samsung and implemented in processors produced by these customers or licensees.
In some embodiments, one or more instructions may correspond to a first type or architecture (e.g., x86) and be translated or emulated on a processor of a different type or architecture (e.g., ARM). An instruction, according to one embodiment, may therefore be performed on any processor or processor type, including ARM, x86, MIPS, a GPU, or other processor type or architecture.
For example, instruction set architecture 1400 may include processing entities such as one or more cores 1406, 1407 and a graphics processing unit 1415. Cores 1406, 1407 may be communicatively coupled to the rest of instruction set architecture 1400 through any suitable mechanism, such as through a bus or cache. In one embodiment, cores 1406, 1407 may be communicatively coupled through an L2 cache control 1408, which may include a bus interface unit 1409 and an L2 cache 1411. Cores 1406, 1407 and graphics processing unit 1415 may be communicatively coupled to each other and to the remainder of instruction set architecture 1400 through interconnect 1410. In one embodiment, graphics processing unit 1415 may use a video code 1420 defining the manner in which particular video signals will be encoded and decoded for output.
Instruction set architecture 1400 may also include any number or kind of interfaces, controllers, or other mechanisms for interfacing or communicating with other portions of an electronic device or system. Such mechanisms may facilitate interaction with, for example, peripherals, communications devices, other processors, or memory. In the example of
Instruction architecture 1500 may include a memory system 1540 communicatively coupled to one or more execution entities 1565. Furthermore, instruction architecture 1500 may include a caching and bus interface unit such as unit 1510 communicatively coupled to execution entities 1565 and memory system 1540. In one embodiment, loading of instructions into execution entities 1565 may be performed by one or more stages of execution. Such stages may include, for example, instruction prefetch stage 1530, dual instruction decode stage 1550, register rename stage 1555, issue stage 1560, and writeback stage 1570.
In one embodiment, memory system 1540 may include an executed instruction pointer 1580. Executed instruction pointer 1580 may store a value identifying the oldest, undispatched instruction within a batch of instructions. The oldest instruction may correspond to the lowest Program Order (PO) value. A PO may include a unique number of an instruction. Such an instruction may be a single instruction within a thread represented by multiple strands. A PO may be used in ordering instructions to ensure correct execution semantics of code. A PO may be reconstructed by mechanisms such as evaluating increments to PO encoded in the instruction rather than an absolute value. Such a reconstructed PO may be known as an “RPO.” Although a PO may be referenced herein, such a PO may be used interchangeably with an RPO. A strand may include a sequence of instructions that are data dependent upon each other. The strand may be arranged by a binary translator at compilation time. Hardware executing a strand may execute the instructions of a given strand in order according to the PO of the various instructions. A thread may include multiple strands such that instructions of different strands may depend upon each other. A PO of a given strand may be the PO of the oldest instruction in the strand which has not yet been dispatched to execution from an issue stage. Accordingly, given a thread of multiple strands, each strand including instructions ordered by PO, executed instruction pointer 1580 may store the oldest—illustrated by the lowest number—PO in the thread.
In another embodiment, memory system 1540 may include a retirement pointer 1582. Retirement pointer 1582 may store a value identifying the PO of the last retired instruction. Retirement pointer 1582 may be set by, for example, retirement unit 454. If no instructions have yet been retired, retirement pointer 1582 may include a null value.
Execution entities 1565 may include any suitable number and kind of mechanisms by which a processor may execute instructions. In the example of
Unit 1510 may be implemented in any suitable manner. In one embodiment, unit 1510 may perform cache control. In such an embodiment, unit 1510 may thus include a cache 1525. Cache 1525 may be implemented, in a further embodiment, as an L2 unified cache with any suitable size, such as zero, 128 k, 256 k, 512 k, 1 M, or 2 M bytes of memory. In another, further embodiment, cache 1525 may be implemented in error-correcting code memory. In another embodiment, unit 1510 may perform bus interfacing to other portions of a processor or electronic device. In such an embodiment, unit 1510 may thus include a bus interface unit 1520 for communicating over an interconnect, intraprocessor bus, interprocessor bus, or other communication bus, port, or line. Bus interface unit 1520 may provide interfacing in order to perform, for example, generation of the memory and input/output addresses for the transfer of data between execution entities 1565 and the portions of a system external to instruction architecture 1500.
To further facilitate its functions, bus interface unit 1520 may include an interrupt control and distribution unit 1511 for generating interrupts and other communications to other portions of a processor or electronic device. In one embodiment, bus interface unit 1520 may include a snoop control unit 1512 that handles cache access and coherency for multiple processing cores. In a further embodiment, to provide such functionality, snoop control unit 1512 may include a cache-to-cache transfer unit that handles information exchanges between different caches. In another, further embodiment, snoop control unit 1512 may include one or more snoop filters 1514 that monitors the coherency of other caches (not shown) so that a cache controller, such as unit 1510, does not have to perform such monitoring directly. Unit 1510 may include any suitable number of timers 1515 for synchronizing the actions of instruction architecture 1500. Also, unit 1510 may include an AC port 1516.
Memory system 1540 may include any suitable number and kind of mechanisms for storing information for the processing needs of instruction architecture 1500. In one embodiment, memory system 1540 may include a load store unit 1546 for storing information such as buffers written to or read back from memory or registers. In another embodiment, memory system 1540 may include a translation lookaside buffer (TLB) 1545 that provides look-up of address values between physical and virtual addresses. In yet another embodiment, memory system 1540 may include a memory management unit (MMU) 1544 for facilitating access to virtual memory. In still yet another embodiment, memory system 1540 may include a prefetcher 1543 for requesting instructions from memory before such instructions are actually needed to be executed, in order to reduce latency.
The operation of instruction architecture 1500 to execute an instruction may be performed through different stages. For example, using unit 1510 instruction prefetch stage 1530 may access an instruction through prefetcher 1543. Instructions retrieved may be stored in instruction cache 1532. Prefetch stage 1530 may enable an option 1531 for fast-loop mode, wherein a series of instructions forming a loop that is small enough to fit within a given cache are executed. In one embodiment, such an execution may be performed without needing to access additional instructions from, for example, instruction cache 1532. Determination of what instructions to prefetch may be made by, for example, branch prediction unit 1535, which may access indications of execution in global history 1536, indications of target addresses 1537, or contents of a return stack 1538 to determine which of branches 1557 of code will be executed next. Such branches may be possibly prefetched as a result. Branches 1557 may be produced through other stages of operation as described below. Instruction prefetch stage 1530 may provide instructions as well as any predictions about future instructions to dual instruction decode stage 1550.
Dual instruction decode stage 1550 may translate a received instruction into microcode-based instructions that may be executed. Dual instruction decode stage 1550 may simultaneously decode two instructions per clock cycle. Furthermore, dual instruction decode stage 1550 may pass its results to register rename stage 1555. In addition, dual instruction decode stage 1550 may determine any resulting branches from its decoding and eventual execution of the microcode. Such results may be input into branches 1557.
Register rename stage 1555 may translate references to virtual registers or other resources into references to physical registers or resources. Register rename stage 1555 may include indications of such mapping in a register pool 1556. Register rename stage 1555 may alter the instructions as received and send the result to issue stage 1560.
Issue stage 1560 may issue or dispatch commands to execution entities 1565. Such issuance may be performed in an out-of-order fashion. In one embodiment, multiple instructions may be held at issue stage 1560 before being executed. Issue stage 1560 may include an instruction queue 1561 for holding such multiple commands. Instructions may be issued by issue stage 1560 to a particular processing entity 1565 based upon any acceptable criteria, such as availability or suitability of resources for execution of a given instruction. In one embodiment, issue stage 1560 may reorder the instructions within instruction queue 1561 such that the first instructions received might not be the first instructions executed. Based upon the ordering of instruction queue 1561, additional branching information may be provided to branches 1557. Issue stage 1560 may pass instructions to executing entities 1565 for execution.
Upon execution, writeback stage 1570 may write data into registers, queues, or other structures of instruction set architecture 1500 to communicate the completion of a given command. Depending upon the order of instructions arranged in issue stage 1560, the operation of writeback stage 1570 may enable additional instructions to be executed. Performance of instruction set architecture 1500 may be monitored or debugged by trace unit 1575.
Execution pipeline 1600 may include any suitable combination of steps or operations. In 1605, predictions of the branch that is to be executed next may be made. In one embodiment, such predictions may be based upon previous executions of instructions and the results thereof. In 1610, instructions corresponding to the predicted branch of execution may be loaded into an instruction cache. In 1615, one or more such instructions in the instruction cache may be fetched for execution. In 1620, the instructions that have been fetched may be decoded into microcode or more specific machine language. In one embodiment, multiple instructions may be simultaneously decoded. In 1625, references to registers or other resources within the decoded instructions may be reassigned. For example, references to virtual registers may be replaced with references to corresponding physical registers. In 1630, the instructions may be dispatched to queues for execution. In 1640, the instructions may be executed. Such execution may be performed in any suitable manner. In 1650, the instructions may be issued to a suitable execution entity. The manner in which the instruction is executed may depend upon the specific entity executing the instruction. For example, at 1655, an ALU may perform arithmetic functions. The ALU may utilize a single clock cycle for its operation, as well as two shifters. In one embodiment, two ALUs may be employed, and thus two instructions may be executed at 1655. At 1660, a determination of a resulting branch may be made. A program counter may be used to designate the destination to which the branch will be made. 1660 may be executed within a single clock cycle. At 1665, floating point arithmetic may be performed by one or more FPUs. The floating point operation may require multiple clock cycles to execute, such as two to ten cycles. At 1670, multiplication and division operations may be performed. Such operations may be performed in four clock cycles. At 1675, loading and storing operations to registers or other portions of pipeline 1600 may be performed. The operations may include loading and storing addresses. Such operations may be performed in four clock cycles. At 1680, write-back operations may be performed as required by the resulting operations of 1655-1675.
Electronic device 1700 may include processor 1710 communicatively coupled to any suitable number or kind of components, peripherals, modules, or devices. Such coupling may be accomplished by any suitable kind of bus or interface, such as I2C bus, system management bus (SMBus), low pin count (LPC) bus, SPI, high definition audio (HDA) bus, Serial Advance Technology Attachment (SATA) bus, USB bus (versions 1, 2, 3), or Universal Asynchronous Receiver/Transmitter (UART) bus.
Such components may include, for example, a display 1724, a touch screen 1725, a touch pad 1730, a near field communications (NFC) unit 1745, a sensor hub 1740, a thermal sensor 1746, an express chipset (EC) 1735, a trusted platform module (TPM) 1738, BIOS/firmware/flash memory 1722, a digital signal processor 1760, a drive 1720 such as a solid state disk (SSD) or a hard disk drive (HDD), a wireless local area network (WLAN) unit 1750, a Bluetooth unit 1752, a wireless wide area network (WWAN) unit 1756, a global positioning system (GPS) 1775, a camera 1754 such as a USB 3.0 camera, or a low power double data rate (LPDDR) memory unit 1715 implemented in, for example, the LPDDR3 standard. These components may each be implemented in any suitable manner.
Furthermore, in various embodiments other components may be communicatively coupled to processor 1710 through the components discussed above. For example, an accelerometer 1741, ambient light sensor (ALS) 1742, compass 1743, and gyroscope 1744 may be communicatively coupled to sensor hub 1740. A thermal sensor 1739, fan 1737, keyboard 1736, and touch pad 1730 may be communicatively coupled to EC 1735. Speakers 1763, headphones 1764, and a microphone 1765 may be communicatively coupled to an audio unit 1762, which may in turn be communicatively coupled to DSP 1760. Audio unit 1762 may include, for example, an audio codec and a class D amplifier. A SIM card 1757 may be communicatively coupled to WWAN unit 1756. Components such as WLAN unit 1750 and Bluetooth unit 1752, as well as WWAN unit 1756 may be implemented in a next generation form factor (NGFF).
The operations that cause data conversion performed by permuting may implement instruction striding, wherein multiple operations are applied to different elements of a structure simultaneously. For example, the operations may implement in part a Stride-5 operation, although the principles of the present disclosure may be applied to stride operations on a different number of elements. In one embodiment, the operations might be made on five elements of the same type. Each different structure within the array may be denoted by a different shading or color, and each element within a given structure may be shown by its number (0 . . . 4).
More specifically, the need to implement striding operations may arise when converting an array-of-structures (AOS) data format into a structure-of-arrays (SOA) data format. Such operations are shown briefly in
The process of converting AOS to SOA may be performed. System 1800 may perform such a conversion in an efficient manner.
As a result, a structure of arrays 2104 may result. Each array (Array1 . . . Array4) may be loaded into a different destination, such as a register or memory or cache location. Each array may include, for example, all the first elements from the structures, all the second elements from the structures, all the third elements from the structures, all the fourth elements from the structures, or all the fifth elements from the structure.
By arranging the structure of arrays 2104 into different registers, each with all of the particularly indexed elements from all of the structures of the array of structures 2102, additional operations may be performed on each register with increased efficiency. For example, in a loop of executing code, the first element of each structure might be added to a second element of each structure, or the third element of each structure might be analyzed. By isolating all such elements into a single register or other location, vector operations can be performed. Such vector operations, using SIMD techniques, could perform the addition, analysis, or other execution upon all elements of the array at a single time, in a clock cycle. Transformation of AOS to SOA format may allow vectorized operations such as these.
Returning to
The AOS-SOA conversion may be made upon any suitable trigger. In one embodiment, system 1800 may perform AOS-SOA conversion upon a specific instruction in instruction stream 1802 that such conversion is to be performed. In another embodiment, system 1800 may infer that AOS-SOA conversion should be performed based upon the proposed execution of another instruction from instruction stream 1802. For example, upon determination that a stride operation, a vector operation, or an operation upon strided data is to be performed, system 1800 may recognize that such execution will be more efficiently executed with data that is converted to strided data and perform AOS-SOA conversion. Any suitable portion of system 1800 may determine that AOS-SOA conversion is to be performed, such as a front end, a decoder, a dynamic translator, or other suitable portions, such as a just-in-time interpreter or compiler.
In some systems, an AOS-SOA conversion may be performed by gather instructions. In other systems, an AOS-SOA conversion may be performed by load, blend, and permute instructions. However, system 1800 may efficiently perform the conversion using permute instructions that reduce the total number of permute instructions that are needed.
System 1800 may include a processor, SoC, integrated circuit, or other mechanism. For example, system 1800 may include processor 1804. Although processor 1804 is shown and described as an example in
Instructions to be executed on processor 1804 may be included in instruction stream 1802. Instruction stream 1802 may be generated by, for example, a compiler, just-in-time interpreter, or other suitable mechanism (which might or might not be included in system 1800), or may be designated by a drafter of code resulting in instruction stream 1802. For example, a compiler may take application code and generate executable code in the form of instruction stream 1802. Instructions may be received by processor 1804 from instruction stream 1802. Instruction stream 1802 may be loaded to processor 1804 in any suitable manner. For example, instructions to be executed by processor 1804 may be loaded from storage, from other machines, or from other memory, such as memory system 1830. The instructions may arrive and be available in resident memory, such as RAM, wherein instructions are fetched from storage to be executed by processor 1804. The instructions may be fetched from resident memory by, for example. In one embodiment, instruction stream 1802 may include an instruction 1822 that will trigger AOS-SOA conversion.
Processor 1804 may include a front end 1806, which may include an instruction fetch pipeline stage and a decode pipeline stage. Front end 1806 may receive instructions with fetch unit 1808 and decode instructions from instruction stream 1802 using decode unit 1810. The decoded instructions may be dispatched, allocated, and scheduled for execution by an allocation stage of a pipeline (such as allocator 1814) and allocated to specific execution units 1816 for execution. One or more specific instructions to be executed by processor 1804 may be included in a library defined for execution by processor 1804. In another embodiment, specific instructions may be targeted by particular portions of processor 1804. For example, processor 1804 may recognize an attempt in instruction stream 1802 to execute a vector operation in software and may issue the instruction to a particular one of execution units 1816.
During execution, access to data or additional instructions (including data or instructions resident in memory system 1830) may be made through memory subsystem 1820. Moreover, results from execution may be stored in memory subsystem 1820 and may subsequently be flushed to other portions of memory. Memory subsystem 1820 may include, for example, memory, RAM, or a cache hierarchy, which may include one or more Level 1 (L1) caches or Level 2 (L2) caches, some of which may be shared by multiple cores 1812 or processors 1804. After execution by execution units 1816, instructions may be retired by a writeback stage or retirement stage in retirement unit 1818. Various portions of such execution pipelining may be performed by one or more cores 1812.
An execution unit 1816 that executes vector instructions may be implemented in any suitable manner. In one embodiment, an execution unit 1816 may include or may be communicatively coupled to memory elements to store information necessary to perform one or more vector operations. In one embodiment, an execution unit 1816 may include circuitry to perform strided operations upon stride5 or other data. For example, an execution unit 1816 may include circuitry to implement an instruction upon multiple elements of data simultaneously within a given clock cycle.
In embodiments of the present disclosure, the instruction set architecture of processor 1804 may implement one or more extended vector instructions that are defined as Intel® Advanced Vector Extensions 512 (Intel® AVX-512) instructions. Processor 1804 may recognize, either implicitly or through decoding and execution of specific instructions, that one of these extended vector operations is to be performed. In such cases, the extended vector operation may be directed to a particular one of the execution units 1816 for execution of the instruction. In one embodiment, the instruction set architecture may include support for 512-bit SIMD operations. For example, the instruction set architecture implemented by an execution unit 1816 may include 32 vector registers, each of which is 512 bits wide, and support for vectors that are up to 512 bits wide. The instruction set architecture implemented by an execution unit 1816 may include eight dedicated mask registers for conditional execution and efficient merging of destination operands. At least some extended vector instructions may include support for broadcasting. At least some extended vector instructions may include support for embedded masking to enable predication.
At least some extended vector instructions may apply the same operation to each element of a vector stored in a vector register at the same time. Other extended vector instructions may apply the same operation to corresponding elements in multiple source vector registers. For example, the same operation may be applied to each of the individual data elements of a packed data item stored in a vector register by an extended vector instruction. In another example, an extended vector instruction may specify a single vector operation to be performed on the respective data elements of two source vector operands to generate a destination vector operand.
In embodiments of the present disclosure, at least some extended vector instructions may be executed by a SIMD coprocessor within a processor core. For example, one or more of execution units 1816 within a core 1812 may implement the functionality of a SIMD coprocessor. The SIMD coprocessor may be implemented fully or in part by the elements described in
During execution, in response to an operation that may benefit from strided data, system 1800 may execute an instruction that causes AOS-SOA conversion 1830. Example operation of such conversion may be shown in the figures below.
Some aspects of AOS-SOA conversion may utilize permute instructions. Permute instructions may selectively identify any combination of the elements of two or more source vectors to be stored in a destination vector. Moreover, the combination of the elements may be stored in any desired order. In order to perform such an operation, an index vector may be specified, wherein each element of the index vector specifies, for an element of the destination vector, which element among the combined sources will be stored in the destination vector.
Several forms of permute instructions may be used. For example, a two-source permute instruction such as VPERMT2D may include a mask and three other operators or parameters. VPERMT2D may be called using, for example, VPERMT2D {mask} source1, index, source 2, although the order of parameters may be in any suitable arrangement. Source1, index, and source2 may all be vectors of the same size. The mask may be used to selective write to the destination. Thus, if mask is all 1's, all results will be written, but the binary mask may be set so as to selectively write a subset of the permutation. The permute operation will select values from the combination of source1 and source2 to write to the destination. Either source or the index may also serve as the destination of the permutation. For example, source1 may be used as the destination. In other examples, VPERMT2 may overwrite results on source registers, while VPERMI2 may overwrite results on index registers. The elements of the index may specify which elements of source1 and source2 are to be written to the destination. A given element of the index at a given position may specify which of source1 and source2 are to be written to the destination at a location in the destination at the given position. The element of the index may specify an offset within a combination of source1 and source2 that will be written to the destination.
For example, consider a call to VPERMT2D {mask=01111111} {source1=zmm0={a b c d e f g h} {index=zmm31={−1 11 6 1 15 10 5 0} {source2=zmm1=i j k l m n o p}. The first seven elements of source1 (zmm0) will be written according to the mask. Furthermore, index may specify offsets (from right to left) within the combination of source1 and source2 that will be written to the destination. The combination may include the concatenation of source2 to source1, or {i j k l m n o p a b c d e f g h}. Thus, index may specify that the zeroth element of the destination will be written with the zeroth element of the combination of source2 and source1, or “h”. The index may specify that the first element (of the destination will be written with the fifth element of the combination of source2 and source1, or “c”. The index may specify (zero-based numbering) that the second element of the destination will be written with the tenth element of the combination of source2 and source1, or “n”. The index may specify (zero-based numbering) that the third element of the destination will be written with the fifteenth element of the combination of source2 and source1, or “i”. The index may specify (zero-based numbering) that the fourth element of the destination will be written with the first element of the combination of source2 and source1, or “g”. The index may specify (zero-based numbering) that the fifth element of the destination will be written with the sixth element of the combination of source2 and source1, or “b”. The index may specify (zero-based numbering) that the sixth element of the destination will be written with the eleventh element of the combination of source2 and source1, or “m”. The index may specify (zero-based numbering) that the seventh element of the destination will not be written, as it is specified with a “−1”. Thus, as a result, the permute will yield {m b g i n c h} stored in source1, the zmm0 register.
Different permute operations provide significant flexibility. For example, different permute operations shown in
In the present disclosure, example pseudocode, instructions, and parameters may be shown. However, other pseudocode, instructions, and parameters may be substituted and used as appropriate. The instructions may include Intel® instructions that are used for example purposes.
In one embodiment, main processor 1920 may include a decoder 1922 to recognize instructions of extended SIMD instruction set 1916 for execution by SIMD coprocessor 1910. In other embodiments, SIMD coprocessor 1910 may include at least part of decoder (not shown) to decode instructions of extended SIMD instruction set 1916. Processor core 1900 may also include additional circuitry (not shown) which may be unnecessary to the understanding of embodiments of the present disclosure.
In embodiments of the present disclosure, main processor 1920 may execute a stream of data processing instructions that control data processing operations of a general type, including interactions with cache(s) 1924 and/or register file 1926. Embedded within the stream of data processing instructions may be SIMD coprocessor instructions of extended SIMD instruction set 1916. Decoder 1922 of main processor 1920 may recognize these SIMD coprocessor instructions as being of a type that should be executed by an attached SIMD coprocessor 1910. Accordingly, main processor 1920 may issue these SIMD coprocessor instructions (or control signals representing SIMD coprocessor instructions) on the coprocessor bus 1915. From coprocessor bus 1915, these instructions may be received by any attached SIMD coprocessor. In the example embodiment illustrated in
In one embodiment, main processor 1920 and SIMD coprocessor 1920 may be integrated into a single processor core 1900 that includes an execution unit, a set of register files, and a decoder to recognize instructions of extended SIMD instruction set 1916.
The example implementations depicted in
In one embodiment, extended vector instructions in extended SIMD instruction set 1916 may operate on any of the registers in extended vector register file 1914, including registers ZMM0-ZMM31, registers YMM0-YMM15, and registers XMM0-XMM7. In another embodiment, legacy SIMD instructions implemented prior to the development of the Intel® AVX-512 instruction set architecture may operate on a subset of the YMM or XMM registers in extended vector register file 1914. For example, access by some legacy SIMD instructions may be limited to registers YMM0-YMM15 or to registers XMM0-XMM7, in some embodiments.
In embodiments of the present disclosure, the instruction set architecture may support extended vector instructions that access up to four instruction operands. For example, in at least some embodiments, the extended vector instructions may access any of 32 extended vector registers ZMM0-ZMM31 shown in
In embodiments of the present disclosure, encodings of the extended vector instructions may include an opcode specifying a particular vector operation to be performed. Encodings of the extended vector instructions may include an encoding identifying any of eight dedicated mask registers, k0-k7. Each bit of the identified mask register may govern the behavior of a vector operation as it is applied to a respective source vector element or destination vector element. For example, in one embodiment, seven of these mask registers (k1-k7) may be used to conditionally govern the per-data-element computational operation of an extended vector instruction. In this example, the operation is not performed for a given vector element if the corresponding mask bit is not set. In another embodiment, mask registers k1-k7 may be used to conditionally govern the per-element updates to the destination operand of an extended vector instruction. In this example, a given destination element is not updated with the result of the operation if the corresponding mask bit is not set.
In one embodiment, encodings of the extended vector instructions may include an encoding specifying the type of masking to be applied to the destination (result) vector of an extended vector instruction. For example, this encoding may specify whether merging-masking or zero-masking is applied to the execution of a vector operation. If this encoding specifies merging-masking, the value of any destination vector element whose corresponding bit in the mask register is not set may be preserved in the destination vector. If this encoding specifies zero-masking, the value of any destination vector element whose corresponding bit in the mask register is not set may be replaced with a value of zero in the destination vector. In one example embodiment, mask register k0 is not used as a predicate operand for a vector operation. In this example, the encoding value that would otherwise select mask k0 may instead select an implicit mask value of all ones, thereby effectively disabling masking. In this example, mask register k0 may be used for any instruction that takes one or more mask registers as a source or destination operand.
One example of the use and syntax of an extended vector instruction is shown below:
VADDPS zmm1, zmm2, zmm3
In one embodiment, the instruction shown above would apply a vector addition operation to all of the elements of the source vector registers zmm2 and zmm3. In one embodiment, the instruction shown above would store the result vector in destination vector register zmm1. Alternatively, an instruction to conditionally apply a vector operation is shown below:
VADDPS zmm1 {k1} {z}, zmm2, zmm3
In this example, the instruction would apply a vector addition operation to the elements of the source vector registers zmm2 and zmm3 for which the corresponding bit in mask register k1 is set. In this example, if the {z} modifier is set, the values of the elements of the result vector stored in destination vector register zmm1 corresponding to bits in mask register k1 that are not set may be replaced with a value of zero. Otherwise, if the {z} modifier is not set, or if no {z} modifier is specified, the values of the elements of the result vector stored in destination vector register zmm1 corresponding to bits in mask register k1 that are not set may be preserved.
In one embodiment, encodings of some extended vector instructions may include an encoding to specify the use of embedded broadcast. If an encoding specifying the use of embedded broadcast is included for an instruction that loads data from memory and performs some computational or data movement operation, a single source element from memory may be broadcast across all elements of the effective source operand. For example, embedded broadcast may be specified for a vector instruction when the same scalar operand is to be used in a computation that is applied to all of the elements of a source vector. In one embodiment, encodings of the extended vector instructions may include an encoding specifying the size of the data elements that are packed into a source vector register or that are to be packed into a destination vector register. For example, the encoding may specify that each data element is a byte, word, doubleword, or quadword, etc. In another embodiment, encodings of the extended vector instructions may include an encoding specifying the data type of the data elements that are packed into a source vector register or that are to be packed into a destination vector register. For example, the encoding may specify that the data represents single or double precision integers, or any of multiple supported floating point data types.
In one embodiment, encodings of the extended vector instructions may include an encoding specifying a memory address or memory addressing mode with which to access a source or destination operand. In another embodiment, encodings of the extended vector instructions may include an encoding specifying a scalar integer or a scalar floating point number that is an operand of the instruction. While several specific extended vector instructions and their encodings are described herein, these are merely examples of the extended vector instructions that may be implemented in embodiments of the present disclosure. In other embodiments, more fewer, or different extended vector instructions may be implemented in the instruction set architecture and their encodings may include more, less, or different information to control their execution.
Data structures that are organized in tuples of three to five elements that can be accessed individually may be used in various applications. For examples, RGB (Red-Green-Blue) is a common format in many encoding schemes used in media applications. A data structure storing this type of information may consist of three data elements (an R component, a G component, and a B component), which are stored contiguously and are the same size (for example, they may all be 32-bit integers). A format that is common for encoding data in High Performance Computing applications includes two or more coordinate values that collectively represent a position within a multidimensional space. For example, a data structure may store X and Y coordinates representing a position within a 2D space or may store X, Y, and Z coordinates representing a position within a 3D space. Other common data structures having a higher number of elements may appear in these and other types of applications.
In some cases, these types of data structures may be organized as arrays. In embodiments of the present disclosure, multiple ones of these data structures may be stored in a single vector register, such as one of the XMM, YMM, or ZMM vector registers described above. In one embodiment, the individual data elements within such data structures may be re-organized into vectors of like elements that can then be used in SIMD loops, as these elements might not be stored next to each other in the data structures themselves. An application may include instructions to operate on all of the data elements of one type in the same way and instructions to operate on all of the data elements of a different type in a different way. In one example, for an array of data structures that each include an R component, a G components, and a B component in an RGB color space, a different computational operation may be applied to the R components in each of the rows of the array (each data structures) than a computational operation that is applied to the G components or the B components in each of the rows of the array.
In yet another example, many molecular dynamics applications operate on neighbor lists consisting of an array of XYZW data structures. In this example, each of the data structures may include an X component, a Y component, a Z component, and a W component. In embodiments of the present disclosure, in order to operate on individual ones of these types of components, one or more even or odd vector GET instructions may be used to extract the X values, Y values, Z values, and W values from the array of XYZW data structures into separate vectors that contain elements of the same type. As a result, one of the vectors may include all of the X values, one may include all of the Y values, one may include all of the Z values, and one may include all of the W values. In some cases, after operating on at least some of the data elements within these separate vectors, an application may include instructions that operate on the XYZW data structures as a whole. For example, after updating at least some of the X, Y, Z, or W values in the separate vectors, the application may include instructions that access one of the data structures to retrieve or operate on an XYZW data structure as a whole. In this case, one or more other instructions may be called in order to store the XYZW values back in their original format.
In embodiments of the present disclosure, the instructions that may cause AOS to SOA conversion may be implemented by a processor core (such as core 1812 in system 1800) or by a SIMD coprocessor (such as SIMD coprocessor 1910) may include an instruction to perform an even vector GET operation or an odd vector GET operation. The instructions may store the extracted data elements into respective vectors containing the different data elements of a data structure in memory. In one embodiment, these instructions may be used to extract data elements from data structures whose data elements are stored together in contiguous locations within one or more source vector registers. In one embodiment, each of the multiple-element data structures may represent a row of an array.
In embodiments of the present disclosure, different “lanes” within a vector register may be used to hold data elements of different types. In one embodiment, each lane may hold multiple data elements of a single type. In another embodiment, the data elements held in a single lane may not be of the same type, but they may be operated on by an application in the same way. For example, one lane may hold X values, one lane may hold Y values, and so on. In this context, the term “lane” may refer to a portion of the vector register that holds multiple data elements that are to be treated in the same way, rather than to a portion of the vector register that holds a single data element. In another embodiment, different “lanes” within a vector register may be used to hold the data elements of different data structures. In this context, the term “lane” may refer to a portion of the vector register that holds multiple data elements of a single data structure. In this example, the data elements stored in each lane may be of two or more different types. In one embodiment in which the vector registers are 512 bits wide, there may be four 128-bit lanes. For example, the lowest-order 128 bits within a 512-bit vector register may be referred as the first lane, the next 128 bits may be referred to as the second lane, and so on. In this example, each of the 128-bit lanes may store two 64-bit data elements, four 32-bit data elements, eight 16-bit data elements, or four 8-bit data elements. In another embodiment in which the vector registers are 512 bits wide, there may be two 256-bit lanes, each of which stores data elements of a respective data structure. In this example, each of the 256-bit lanes may store multiple data elements of up to 128 bits each.
The process of converting AOS to SOA may be performed. System 1800 may perform such a conversion in an efficient manner.
As a result, a structure of arrays 2104 may result. Each array (Array1 . . . Array4) may be loaded into a different destination, such as a register or memory or cache location. Each array may include, for example, all the first elements from the structures, all the second elements from the structures, all the third elements from the structures, all the fourth elements from the structures, or all the fifth elements from the structure.
By arranging the structure of arrays 2104 into different registers, each with all of the particularly indexed elements from all of the structures of the array of structures 2102, additional operations may be performed on each register with increased efficiency. For example, in a loop of executing code, the first element of each structure might be added to a second element of each structure, or the third element of each structure might be analyzed. By isolating all such elements into a single register or other location, vector operations can be performed. Such vector operations, using SIMD techniques, could perform the addition, analysis, or other execution upon all elements of the array at a single time, in a clock cycle. Transformation of AOS to SOA format may allow vectorized operations such as these.
For example, given sources zmm1 and zmm0, each with register elements identified as x-, y-, z-, and w-coordinate elements, a permute instruction may be used to permute the x-coordinate and y-coordinate elements into a destination register. The destination register may include the source zmm0. As only seven x-coordinate and y-coordinate elements exist in the sources, a write to the last element of the destination may be masked off (mask=0x7F). An index (stored in zmm31) may define which of the elements from the combination of zmm1 and zmm0 are to be stored in zmm0, and in what order. For example, the index vector may include corresponding positions for the x-coordinate elements, to be stored in the least significant positions of the destination register, and the y-coordinate elements, to be stored in the next significant portions of the destination register. As a result VPERMT2D {0x7F} zmm0, zmm31 zmm1 may be called, resulting in zmm0 storing the results as shown in
In another example, given sources zmm1 and zmm0, each with register elements identified as x-, y-, z-, and w-coordinate elements, a permute instruction may be used to permute elements into a destination register. However, the order of the elements might not be arbitrarily selectable. For each relative position in the sources, an element from the source must be chosen to be written to the destination. The mask may define, for a given relative position in the sources, which source will be written to the destination. As a result VBLENDMPD {0x9c} zmm2, zmm0, zmm1 may be called, resulting in zmm2 storing the results as shown in
Permute operations may be used to perform portions or all of the AOS-SOA conversion. These are described in more complete detail in subsequent figures.
Suppose it is a goal to obtain the x-coordinates stored in the registers zmm0, zmm1, zmm2, and zmm3. Each register might include contents loaded from memory and may contain more than one x-coordinate, as each register includes contents from more than one structure. The contents of each register may include an x-coordinate (albeit an x-coordinate from various structures) in the same relative position in each register. These positions may be, for example, the zeroth and fifth locations in a given index. Accordingly, given the flexibility of different permute functions, a single index vector (stored in zmm4) may be used to perform various permute operations. The index vector may define that x values are located, for a combination of any two of the sources, in the same locations (indices 0, 5, 8, 13). The index vector may repeat these values and rely upon selective usage of permute operation (through masking) to arrive at the correct composition of the destination vector.
For example, VPERMT2D may be called to permute zmm2 and zmm3 into zmm2 using the index zmm4. Furthermore, as these two source registers are the left-half of the source, their results may be stored in the left-half of the eventual destination. Accordingly, the permute operation may be masked with {0xF0} so that the left-half of zmm2 is filled with the x-coordinates from zmm2 and zmm3. VPERMI2D may be called to permute zmm0 and zmm1 into zmm4 using the index zmm4. As these two source registers are the right-half of the source, their results may be stored in the right-half of the eventual destination. Accordingly, the permute operation may be masked with {0x0F} so that the right-half of zmm4 is filled with the x-coordinates from zmm0 and zmm1. Notably, each of the results in zmm2 and zmm4 include x-coordinates from their respective sources in-order. Two results in zmm2 and zmm4 may be blended. A blend operation such as VLENDMPD may be called to blend zmm4 and zmm2 into zmm5. The blend may use a mask of {0xF0} to indicate that, for the right-half, zmm4 values should be used, and for the left-half, zmm2 values should be used. The result may be a collection of the x-coordinates from the sources ordered in zmm5.
In one embodiment, operation of permute instructions to perform aspects of AOS to SOA conversion may rely upon a feature of permute instructions to reuse the index vector to store results. By selectively storing results in only part of the index vector and preserving the remainder of the index vector, an operation may be saved. As discussed above, as the same relative position of a given coordinate (such as the x-coordinate) may exist across multiple sources, reflecting portions of an AOS to convert, an index vector might repeat part of itself (such as {13 8 5 0 13 8 5 0}) and the permute operation may be masked (such as with 0x0F or 0xF0} to arrive a destination vector with all x-coordinates. In such cases, the part of the index vector that repeats may be eliminated, and a permute operation masked for the remaining portion may be used. Conversely, data elements that are not needed may be overwritten with index values using a mask. The same write mask may be used with the permute instruction, which overwrites the index register as a destination, preserving some data values and overwriting unneeded index values with data combine from the other source registers. Consequently, the particular variant of permute instructions denoted by the “i” in VPERMI instructions may allow merging of writes that depositing of data values mixed with index control values, converting the two-source instruction effectively into a three-source permute instruction.
For example, given the same source vectors zmm0-zmm3 of
The resulting zmm4 register will serve as the index vector source for another call to VPERM2I. The zmm4 register will also be the destination of the permute operation. The other sources, zmm2 and zmm3, may be permuted according to the values of the left-half of zmm4, as the permute operation is masked with 0xF0. Thus, the lowest significant four locations in zmm4, which store the x-coordinates from zmm0 and zmm4, will be preserved. The additional elements (the x-coordinates) from zmm2 and zmm3 will be stored as the index values in the most significant four locations in zmm4 are overwritten. As a result, zmm4 will include the x-coordinates from all four sources, in-order. This result may be the same as that in
The principles of this operation may be applied in the operations discussed further below.
As shown in
The conversion shown in
The AOS may be loaded into the registers through the use of five gather instructions. Five KNORB operations may be used to set masks.
First, gather indices may be created. They may be created with the pseudocode:
_declspec (align(32)) const_int32 gather0_index[8]={0, 5, 10, 15, 20, 25, 30, 35};
_declspec (align(32)) const_int32 gather1_index[8]={1, 6, 11, 16, 21, 26, 31, 36};
_declspec (align(32)) const_int32 gather2_index[8]={2, 7, 12, 17, 22, 27, 32, 37};
_declspec (align(32)) const_int32 gather3_index[8]={3, 8, 13, 18, 23, 28, 33, 38};
_declspec (align(32)) const_int32 gather4_index[8]={4, 9, 14, 19, 24, 29, 34, 39};
The index for gather0 may identify, in the AOS, the relative location of each “0” element. The index for gather1 may identify, in the AOS, the relative location of each “1” element. The index for gather2 may identify, in the AOS, the relative location of each “2” element. The index for gather3 may identify, in the AOS, the relative location of each “3” element. The index for gather5 may identify, in the AOS, the relative location of each “4” element.
Given these, KNORW may be called to generate masks, followed by five calls to VGATHERDPD. Each call to VGATHERDPD may gather packed values (in this case, of doubles) based upon the indices supplied to each call. The indices provided (r8+[ymm5->ymm9]*8) may be used to identify particular locations in memory (from a base address r8, scaled by the size of the doubles) from where the values will be gathered and loaded into respective registers. The calls may be expressed in the following pseudocode:
kxnorw k1, k0, k0
kxnorw k2, k0, k0
kxnorw k3, k0, k0
kxnorw k4, k0, k0
kxnorw k5, k0, k0
vgatherdpd zmm4{k1}, zmmword ptr [r8+ymm9*8]
vgatherdpd zmm3 {k2}, zmmword ptr [r8+ymm8*8]
vgatherdpd zmm2 {k3}, zmmword ptr [r8+ymm7*8]
vgatherdpd zmm1 {k4}, zmmword ptr [r8+ymm6*8]
vgatherdpd zmm0{k5}, zmmword ptr [r8+ymm5*8]
Given the AOS of eight doubles in memory, five load operations may be made to load data into registers. While each structure might include five elements, a load operation may be made in multiples of eight. Consequently, rather than load the eight structures into five registers wherein each register includes unused space, the eight structures may be loaded into five registers. Some structures may be broken up across multiple registers. The AOS to SOA conversion may then attempt to sort the contents of these eight registers so that all (eight) of the first elements of the structures are in a common register, all (eight) of the second elements of the structures are in a common register, and so on. In other examples, where structures with another number of elements (such as four) will be processed, four registers might be needed to be to store the results.
Five additional loads may be performed to load data from the memory into the registers. However, these loads may be performed with masks so that only some of the contents of a given memory section are loaded into the respective registers. The specific masks may be selected according to those that are needed to filter the correct element (such as the first, second, third, fourth, or fifth) from a given segment into the register. As a given register will only contain the same indexed element (that is, all first elements, all second elements, etc.), the mask is selected to filter only that element into a corresponding register. In some cases, such as in the present figure, the same mask might be used in all of these load operations. For example, it may be observed that for these particular structures, a mask of {01000010} may uniquely identify a different indexed element (first elements, second elements, etc.) for different memory segments. Thus, applying this same mask to the original memory segments that were loaded from memory will yield the application of indexed elements. Applying the mask, then, to the appropriate register may copy the required elements (that is, the first, second, or other elements).
The same process may be repeated for different masks and combination of sources, until the registers are each filled with respective elements (first elements, or second elements, and so on). The process may be repeated with five loads with a second mask, five loads with a third mask, and five loads with a fourth mask to accomplish the correct loading combinations. The result may be that each register is filled only with respective ones of first elements, second elements, third elements, fourth elements, or fifth elements of the original array of structures. However, the elements within a given register might not be ordered in the same way that they were ordered in the original array.
Accordingly, a number of permute operations may be performed to reorder the contents of the registers to match the original order of the array of structures. For example, five permute operations may be performed. Interim registers may be used as needed. A separate index vector may be needed for each permute to provide the order of the original array. As a result, the contents of each register may be reordered according to the order of the original array. The result may be the converted AOS resulting in a SOA. The arrays may be represented in each respective register. The structure may be the combination of the arrays.
In total, the operations of
vmovups zmm5, zmmword ptr [r8]
vmovups zmm11, zmmword ptr [r8+0x40]
vmovups zmm7, zmmword ptr [r8+0x80]
vmovups zmm13, zmmword ptr [r8+0xc0]
vmovups zmm9, zmmword ptr [r8+0x100]
vmovapd zmm5{k4}, zmmword ptr [r8+0xc0]
vmovapd zmm11{k4}, zmmword ptr [r8+0x100]
vmovapd zmm7{k4}, zmmword ptr [r8]
vmovapd zmm13{k4}, zmmword ptr [r8+0x40]
vmovapd zmm9{k4}, zmmword ptr [r8+0x80]
vmovapd zmm5{k3}, zmmword ptr [r8+0x40]
vmovapd zmm11{k3}, zmmword ptr [r8+0x80]
vmovapd zmm7{k3}, zmmword ptr [r8+0xc0]
vmovapd zmm13{k3}, zmmword ptr [r8+0x100]
vmovapd zmm9{k3}, zmmword ptr [r8]
vmovapd zmm5{k2}, zmmword ptr [r8+0x100]
vmovapd zmm11{k2}, zmmword ptr [r8]
vmovapd zmm7{k2}, zmmword ptr [r8+0x40]
vmovapd zmm13{k2}, zmmword ptr [r8+0x80]
vmovapd zmm9{k2}, zmmword ptr [r8+0xc0]
vmovapd zmm5{k1}, zmmword ptr [r8+0x80]
vmovapd zmm11{k1}, zmmword ptr [r8+0xc0]
vmovapd zmm7{k1}, zmmword ptr [r8+0x100]
vmovapd zmm13{k1}, zmmword ptr [r8]
vmovapd zmm9{k1}, zmmword ptr [r8+0x40]
vpermpd zmm6, zmm4, zmm5
vpermpd zmm8, zmm3, zmm7
vpermpd zmm10, zmm2, zmm9
vpermpd zmm12, zmm1, zmm11
vpermpd zmm14, zmm0, zmm13
First, the eight structures of the array may be loaded, unaligned, into five registers as previously shown. The registers may include mm0 . . . mm4. This process may take five load operations. Some of the data to be permuted may be loaded into another register. That register is then partially overwritten with an index vector. The index vector may use half of the available space. The permute operation that results will be performed with a mask, so that the half with the original data elements are not overwritten, but are instead preserved. This may performed with a VPERMI instruction and may use its index vector parameter as a destination vector. Then, the same mask used to load the indices to the index vector register as the write mask so that only index values in the index vector register are overwritten.
Using this technique on data that is loaded from memory with five loads into each register, with the original order preserved across the registers, a total of fourteen permute operations may be needed to perform the AOS-SOA conversion. To perform these fourteen permute operations, a total of thirteen different index vectors and three different masks may be needed.
The operation of system 1800 in
Index vectors may be initialized as:
mm0 {0,2,4,6,8,9,14,12}
mm1 {9,11,13,15,3,2,7,5}
mm3 {0,2,4,5,8,10,12,14}
mm4 {9,11,13,15,1,3,5,7}
mm5 {3,4,8,9,13,14,-,-}
mm6 {2,3,7,8,12,13,-,-}
mm7 {2,3,7,8,12,13,-,-}
mm8 {0,1,5,6,8,9,10,11}
mm9 {2,3,4,5,9,10,14,15}
mm10 {0,2,4,6,8,10,12,14}
mm11 {1,3,5,7,9,11,13,15}
mm12 {0,2,4,6,8,9,12,14}
mm13 {1,3,5,7,10,11,13,15}
mm14 {2-,12,7,8,3,-,13}
mm15 {4,-,-,5,11,12,-,-}
mm16 {0,3,2,1,6,5,4,7}
For example, mm7 may be created as a permute of mm3 into mm2 using the mm7 index vector. As a result, mm7 may consolidate the “w” and “v” elements from these registers.
The register mm2 may be permuted with mm1 using the vector index mm6, storing the results into mm6. As a result, mm6 may consolidate the “x” and “y” elements from these registers.
As the register mm2 has had its “x”, “y” “w”, and “v” elements permuted into other locations, it only needs to retain its “z” elements. Accordingly, register mm2 may serve both as a source of “z” elements and be loaded with other index values and serve as an index vector for a subsequent permute. In particular, it may serve as an index vector for a permute operation wherein the “z” elements will be consolidated. Efficiency may be gained wherein register mm2 does not need to serve as a typical source in a permute, but may be added on as a de-facto third source for another permute operation to consolidate “z” elements from another two vectors. For example, mm2 may be loaded with offset values that identify the “z” element locations in mm3 and mm4. The register mm2 may be loaded with index elements in its locations that are not otherwise holding “z” elements. Subsequently, mm2 may be used as an index vector to permute the “z” elements from mm3 and mm4. The permute may have a write mask that matches the index vector elements stored in mm2, such as {0xB0}. Then, “z” elements from mm4 and mm3 may be stored into mm2, overwriting index elements but preserving the “z” elements already within mm2.
The registers mm0 and mm1 may be permuted with an index vector in mm5 to consolidate the “v” and “w” elements therein into mm5. The resulting register mm5 may itself be permuted with mm7, which contained the consolidation of “v” and “w” from mm2 and mm3. This permutation may be performed with a new index vector, mm13. However, mm13 might not be big enough to hold all the “v” and “w” elements from all four original source registers. Accordingly, the “v” and “w” set that bridged the original mm2-mm3 may be dropped, but consolidated in other permute operations. The result may be performed with a permute instruction that stores the result back into mm5.
The registers mm7 and mm4 may be permuted with a new index vector in mm9 to consolidate the “v” and “w” elements therein into mm9. This register mm9 with “v” and “w” elements may include the “v” and “w” element combination that bridged the original mm2-mm3 that is missing from mm5. Furthermore, mm9 and mm5 may each include the “v” and “w elements that are missing from the other register. Accordingly, these registers may be permuted twice according to different index vectors to return registers with all “v” elements or all “w” elements. For example, mm9 and mm5 may be permuted by index vector mm11, storing all “v” elements in mm11. In another example, mm9 and mm5 may be permuted by index vector mm10, storing all “w” elements into mm10. These may be copied back to original ones of mm0 . . . mm4 as needed upon completion of the conversion.
The registers mm3 and mm4 may be permuted to obtain the “z” elements. These may be permuted according to the contents of mm2, which, as shown above, may itself have been permuted to preserve “z” elements. Furthermore, mm2 may have been populated, in indices not containing “z” elements, with index values to reference “z” elements from mm3 and mm4.
Accordingly, mm3 and mm4 may be permuted with mm2 as its index and store the results back in to mm2. Moreover, the permute may be performed with a mask, wherein the mask (0xB0) protects the already-existing “z” elements in mm2. Furthermore, the mask may also protect index elements not used in mm2 to obtain “z’ elements from mm3 or mm4. In fact, these index elements, as Thus, at the end of the permute, mm2 may include the “z” elements consolidated from the original mm2, mm3, and mm4. Furthermore, mm2 may still retain two index elements to indicate the positions in subsequent permutes with mm1 and mm0 to obtain their “z” elements.
The resulting mm2 may include the “z” elements consolidated from permute operations upon the original mm2, mm3, and mm4. Furthermore, mm2 may include indices for identifying the position of “z” elements in mm1 and mm0. Thus, mm2 may be used as vector index for a permute of mm1 and mm0 to consolidate the “z” elements from these additional registers. The permute may apply the mask (0xBD) based upon the location of “z” elements and indices within mm2. The result of the mask may be that the existing “z” elements are preserved while the indices indicating “z” element locations in mm1 and mm0 are overwritten with such “z” elements. The result may be mm2, filled with “z” elements from the original array. However, the order of the “z” elements might not match the order as presented in the original array. A permute operation may be called on mm2 with a vector index to reorder the “z” elements therein. The resulting mm2 may be the “z” array. These may be copied back to original ones of mm0 . . . mm4 as needed upon completion of the conversion.
As discussed above, mm6 may include “x” and “y” elements permuted from mm1 and the original mm2. Furthermore, “x” and “y” elements may be permuted from mm0 and mm6 using a new vector index in mm8. The result may be stored in mm8. The results may omit the “x” and “y” elements from the second half of the original mm2, as mm8 does not have room to store all “x” and “y” elements from the original mm1, mm2, and mm0. However, these may be recovered from mm6 in a separate permute function as described below.
The register mm3 may be converted to an index vector for use with mm4 and mm6 “x” and “y” element permute operation. However, mm3 may still retain its own “x” and “y” elements, using the other positions for the index vector values. A load or move function may be masked (0x39) to only edit the non-“x” and non-“y” elements in mm3. The index vector values may otherwise be loaded from a new index vector, mm15. The result may still be referenced as mm3.
The resulting mm3 may be used as an index vector and source for permute of mm4 and mm6 with respect to “x” and “y” elements. The same mask (0x39) may be used to perform writes of the permute back in to mm3, such that the “x” and “y” elements from mm4 and mm6 may be consolidated into mm3 at the locations that previously served as index values. This version of mm3 may include “x” and “y” elements from the original mm4, original mm3, and original second half of mm2.
Meanwhile, mm8 may include “x” and “y” elements from the other original register contents. Accordingly, mm3 and mm8 may be permuted with two different permute operations, each with its own index, to yield an array of “x” elements and an array of “y” elements. Register contents may be copied back to original ones of mm0 . . . mm4 as needed.
Accordingly, the AOS-SOA conversion may be complete.
Pseudocode to perform this conversion may be specified as:
vmovups zmm10, zmmword ptr [r8+0x40]
vmovups zmm13, zmmword ptr [r8+0x80]
vmovups zmm17, zmmword ptr [r8]
vmovups zmm16, zmmword ptr [r8+0xc0]
vmovups zmm20, zmmword ptr [r8+0x100]
vmovaps zmm11, zmm10
vpermt2pd zmm11, zmm8, zmm13
vmovaps zmm19, zmm13
vmovapd zmm13{k3}, zmmword ptr [rip+0x76f2]
vpermt2pd zmm19, zmm8, zmm16
vpermi2pd zmm13 {k3}, zmm17, zmm10
vmovapd zmm13{k2}, zmmword ptr [rip+0x775c]
vpermi2pd zmm13{k2}, zmm16, zmm20
vmovapd zmm16{k1}, zmmword ptr [rip+0x77cc]
vpermpd zmm14, zmm4, zmm13
vpermi2pd zmm16{k1}, zmm11, zmm20
vpermt2pd zmm20, zmm6, zmm19
vmovaps zmm12, zmm17
vpermt2pd zmm12, zmm9, zmm10
vpermt2pd zmm17, zmm7, zmm11
vpermt2pd zmm19, zmm5, zmm12
vmovaps zmm15, zmm16
vmovaps zmm18, zmm19
vpermt2pd zmm15, zmm3, zmm17
vpermt2pd zmm17, zmm2, zmm16
vpermt2pd zmm18, zmm1, zmm20
vpermt2pd zmm20, zmm0, zmm19
The operation of system 1800 in
For example, the array of structures may include eight structures, each with five elements denoted in
As a result, mm0 through mm3 may have identical relative indices. Other loading schemes may be used depending upon the particular size of the structures and arrays. However, each may be performed according to the teachings of
To perform the loading, four loads without masking may be executed. The first eight elements of the array may be loaded to mm0 using a load operation. Thus, mm0 may include elements of different structures including “z y x v w z y x”. An unaligned load may be called to load the first five elements of the third structure of the array and the first three elements of the fourth structure. Another load may be called to load the first five elements of the fifth structure of the array and the first three elements of the sixth array. Yet another load may be called to load the first five elements of the seventh structure of the array and the first three elements of the eighth structure. Each of these, mm0 . . . mm3, may include elements of different structures including “z y x v w z y x”.
The loading may also include loading the elements that were skipped in the OOO loading described above. These include elements “w” and “v” of every even structure in the array. These may be loaded with four load operations, wherein each load operation uses a mask to identify the portion of the array segment that includes the missing “w” and “v” elements. The load operations may be made to mm4.
The number of permutes may be simplified because mm0, mm1, mm2, and mm3 each have the same elements arranged at the same relative locations therein. Accordingly, index vector, such as mm9 defined as “12 8 5 0 12 8 5 0” may define the respective locations of “x” elements within any pair of mm0, mm1, mm2, and mm3. Moreover, this index vector may be selectively overwritten during permute to allow it to be a source for a subsequent permute.
For example, mm0 and mm1 may be permuted so as to consolidate the “x” elements therein into the right-half of mm9. The selective write may be made through use of a mask such as (0x0F). The left-half of mm9 may maintain vector index values for “x” elements, which might be used in any combination of mm0, mm1, mm2, and mm3. Thus, the resulting mm9 may be used again as a vector index and a de-facto source for a permute to consolidate “x” elements from mm2 and mm3 back into mm9. The permute may selectively write to the left-half of mm9 using a mask (0xF0), thus preserving the previously-written elements of “x” from the previous permute operation. The result may be that mm9 includes an array entirely of “x” elements. This was accomplished with two permute operations, a vector index, and two masks.
The process performed on mm0, mm1, mm2, and mm3 for the “x” elements may be repeated on mm0, mm1, mm2, and mm3 for the “y” elements and the “z” elements, yielding arrays entirely of “y” elements and “z” elements. Each such process may require two permute operations and a vector index. The vector index for each process may be unique, wherein each vector index identifies the respective locations of “y” and “z” elements within the registers. While each such process may also require two masks, the same masks that were used for “x” permute operations may be reused for “y” and “z” permute operations.
The process performed on mm0, mm1, mm2, and mm3 for the “x”, “y”, and “z” elements may be repeated, but to consolidate “v” and “w” values into a register. The vector index for the permute functions may identify the locations of “v” and “w” (4 and 5, respectively). As a result, mm4 may include “v” and “w” components from four structures, while the result of the permute functions performed on mm0 . . . mm3 (mm5, for example) may include the “v” and “w’ components from the structures within these registers. Accordingly, mm4 and mm5 may be permuted with two separate VPERM instructions and two indices, each identifying the location of “v” and “w” within the combination of the registers. One such permute may yield an array of “v” elements, and the other permute may yield an array of “w” elements.
The data conversion may thus be complete.
Pseudocode to perform this conversion may be specified as:
vmovups zmm10, zmmword ptr [r8+0x40]
vmovups zmm6, zmmword ptr [r8+0x50]
vmovups zmm7, zmmword ptr [r8+0xa0]
vmovups zmm8, zmmword ptr [r8]
vmovups zmm9, zmmword ptr [r8+0xf0]
vmovapd zmm10{k7}, zmmword ptr [r8+0x80]
vmovapd zmm10{k6}, zmmword ptr [r8+0xc0]
vmovaps zmm15, zmm2
vpermi2pd zmm15{k3}, zmm6, zmm7
vmovapd zmm10{k5}, zmmword ptr [r8+0x100]
vpermi2pd zmm15{k1}, zmm8, zmm9
vmovaps zmm11, zmm5
vmovaps zmm12, zmm4
vmovaps zmm13, zmm3
vpermi2pd zmm11{k4}, zmm8, zmm6
vpermi2pd zmm12{k4}, zmm8, zmm6
vpermi2pd zmm13 {k4}, zmm8, zmm6
vpermi2pd zmm11{k2}, zmm7, zmm9
vpermi2pd zmm12{k2}, zmm7, zmm9
vpermi2pd zmm13{k2}, zmm7, zmm9
vmovaps zmm14, zmm15
vpermt2pd zmm14, zmm1, zmm10
vpermt2pd zmm15, zmm0, zmm10
In one embodiment, data may be initially loaded into registers for data conversion with gaps that align with the vector position of certain elements in their final place. This may be performed using six move or load operations (VMOVUPS—from memory or cache, not counting moves between registers, as these have significantly less latency). These may use masks to accomplish the gaps and offset. This may be fewer than the load operations needed in
As shown in
After mm2 is permuted with other registers to consolidate the “0”, “1”, “3”, and “4” elements therein to the other registers, mm2 may be available to serve as both a vector index and a de-facto source for permute operations to consolidate “2” elements from mm0, mm1, mm3, and mm4. The register mm2 may be loaded with vector index values identifying the location of “2” elements in these other registers. The already-set “2” element in mm2 may be preserved through masking, while during consolidation vector index elements may be reclaimed with written “2” elements from the other registers.
As shown in
The vector permute operations shown in the previous figures may be applied to consolidate the respective identified elements within individual registers, resulting in arrays.
Pseudocode to perform this conversion may be specified as:
At 3105, in one embodiment, an instruction may be loaded and at 3110 the instruction may be decoded.
At 3115, it may be determined that the instruction requires AOS-SOA conversion of data. Such data may include strided data. In one embodiment, the stride data may include Stride5 data. The instruction may be determined to require such data because vector operations on the data are to be performed. The data conversion may result in the data being in an appropriate format so that a vectorized operation may be applied simultaneously, in a clock cycle, to each element of a bank of data. The instruction may specifically identify that the AOS-SOA conversion is to be performed or it may be inferred from the desire to execute an instruction that the AOS-SOA is needed.
At 3120, an array to be converted may be loaded into registers. In one embodiment, structures in the array may be loaded into registers such that as many registers as possible have the same element layout. For example, “1” elements are all in the same relative positions, “2” elements are all in the same relative positions, etc. The load operations may be performed with masks. The load operations may cut off certain elements from every other register that would have otherwise been loaded. These may be referenced as excess elements. The excess elements may be the same for every other register.
At 3125, the excess elements may be loaded into a common register using mask load operations. A larger number of load operations may be performed as a consequence. This common register may have a different element layout than the registers with the common element layout.
At 3130, index vectors may be generated for the common element layouts. An index vector may be created identifying relative positions in the common element layouts for a given element. The index vector may be used as an index vector and a partial source for a permute function to consolidate given elements. At 3135, permutes may be performed on registers with the common layout using these index vectors. 3135 may be repeated as necessary to generate arrays of elements within the common layout other than those among the excess element. These generated arrays may represent a partial output of the data conversion.
At 3140, index vectors for the elements among the excess elements and the common register may be generated. The index vectors may also serve as de-facto sources. At 3145, permute may be performed on a combination of the common register and various appropriate results from 3135. The elements among the excess elements may be consolidated to arrays. These generated arrays may represent the remainder output of the data conversion.
At 3150, the execution upon the different registers may be performed. As a given register is to be used with the vector instruction for execution, each element may be executed-upon in parallel. Results may be stored as necessary. At 3155, it may be determined if subsequent vector execution is to be performed on the same converted data. If so, method 3100 may return to 3150. Otherwise, method 3100 may proceed to 3160.
At 3160, it may be determined whether additional execution is needed for other stride5 data. If so, method 3100 may proceed to 3120. Otherwise, at 3165 the instruction may be retired. Method 3100 may optionally repeat or terminate.
At 3205, in one embodiment, an instruction may be loaded and at 3210 the instruction may be decoded.
At 3215, it may be determined that the instruction requires AOS-SOA conversion of data. Such data may include strided data. In one embodiment, the stride data may include Stride5 data. The instruction may be determined to require such data because vector operations on the data are to be performed. The data conversion may result in the data being in an appropriate format so that a vectorized operation may be applied simultaneously, in a clock cycle, to each element of a bank of data. The instruction may specifically identify that the AOS-SOA conversion is to be performed or it may be inferred from the desire to execute an instruction that the AOS-SOA is needed.
At 3220, an array to be converted may be prepared to be loaded into registers. The mapping of the array to the registers may be evaluated in view of the final conversion of data. One or more elements may be identified that can be initially loaded into a given vector register at a given location that matches the same position and vector register that is to contain the element after data conversion. At 3225, load operations may be performed to load the array into the registers such that the identified element is loaded to the designated register and position. Such load operations may require shifting of data or leaving gaps in various registers such that the alignment occurs. At 3230, permute operations may be performed to consolidate given elements from each of the registers into a single register. These arrays of elements may be generated and used for vector execution. However, the aligned element might not require a permute operation.
At 3250, the execution upon the different registers may be performed. As a given register is to be used with the vector instruction for execution, each element may be executed-upon in parallel. Results may be stored as necessary. At 3255, it may be determined if subsequent vector execution is to be performed on the same converted data. If so, method 3200 may return to 3250. Otherwise, method 3200 may proceed to 3260.
At 3260, it may be determined whether additional execution is needed for other stride5 data. If so, method 3200 may proceed to 3220. Otherwise, at 3265 the instruction may be retired. Method 3200 may optionally repeat or terminate.
Embodiments of the mechanisms disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Embodiments of the disclosure may be implemented as computer programs or program code executing on programmable systems comprising at least one processor, a storage system (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device.
Program code may be applied to input instructions to perform the functions described herein and generate output information. The output information may be applied to one or more output devices, in known fashion. For purposes of this application, a processing system may include any system that has a processor, such as, for example; a digital signal processor (DSP), a microcontroller, an application specific integrated circuit (ASIC), or a microprocessor.
The program code may be implemented in a high level procedural or object oriented programming language to communicate with a processing system. The program code may also be implemented in assembly or machine language, if desired. In fact, the mechanisms described herein are not limited in scope to any particular programming language. In any case, the language may be a compiled or interpreted language.
One or more aspects of at least one embodiment may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores” may be stored on a tangible, machine-readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that actually make the logic or processor.
Such machine-readable storage media may include, without limitation, non-transitory, tangible arrangements of articles manufactured or formed by a machine or device, including storage media such as hard disks, any other type of disk including floppy disks, optical disks, compact disk read-only memories (CD-ROMs), compact disk rewritables (CD-RWs), and magneto-optical disks, semiconductor devices such as read-only memories (ROMs), random access memories (RAMs) such as dynamic random access memories (DRAMs), static random access memories (SRAMs), erasable programmable read-only memories (EPROMs), flash memories, electrically erasable programmable read-only memories (EEPROMs), magnetic or optical cards, or any other type of media suitable for storing electronic instructions.
Accordingly, embodiments of the disclosure may also include non-transitory, tangible machine-readable media containing instructions or containing design data, such as Hardware Description Language (HDL), which defines structures, circuits, apparatuses, processors and/or system features described herein. Such embodiments may also be referred to as program products.
In some cases, an instruction converter may be used to convert an instruction from a source instruction set to a target instruction set. For example, the instruction converter may translate (e.g., using static binary translation, dynamic binary translation including dynamic compilation), morph, emulate, or otherwise convert an instruction to one or more other instructions to be processed by the core. The instruction converter may be implemented in software, hardware, firmware, or a combination thereof. The instruction converter may be on processor, off processor, or part-on and part-off processor.
Thus, techniques for performing one or more instructions according to at least one embodiment are disclosed. While certain exemplary embodiments have been described and shown in the accompanying drawings, it is to be understood that such embodiments are merely illustrative of and not restrictive on other embodiments, and that such embodiments not be limited to the specific constructions and arrangements shown and described, since various other modifications may occur to those ordinarily skilled in the art upon studying this disclosure. In an area of technology such as this, where growth is fast and further advancements are not easily foreseen, the disclosed embodiments may be readily modifiable in arrangement and detail as facilitated by enabling technological advancements without departing from the principles of the present disclosure or the scope of the accompanying claims.
Some embodiments of the present disclosure include a processor. The processor may include a front end to receive an instruction, a decoder to decode the instruction, a core to execute the instruction, and a retirement unit to retire the instruction. In combination with any of the above embodiments, the core includes logic to determine that the instruction will require strided data converted from source data in memory. The strided data is to include corresponding indexed elements from a plurality of structures in the source data to be loaded into a final register to be used to execute the instruction. In combination with any of the above embodiments, the core includes logic to load source data into a plurality of preliminary vector registers to align a defined element of one of the preliminary vector registers in a position that corresponds to a required position in the final register for execution. In combination with any of the above embodiments, the core includes logic to apply a plurality of permute instructions to contents of the preliminary vector registers to cause corresponding indexed elements from the plurality of structures to be loaded into respective source vector registers. In combination with any of the above embodiments, the core includes logic to execute the instruction upon one or more source vector registers upon completion of conversion of source data to strided data. In combination with any of the above embodiments, the core includes logic to omit permute instruction execution for the defined element. In combination with any of the above embodiments, the core includes logic to load source data into the plurality of preliminary vector registers with a plurality of gaps to align the defined element to the required position. In combination with any of the above embodiments, the core includes logic to load source data into a number of preliminary vector registers that is greater than a number of the structures. In combination with any of the above embodiments, the strided data is to include eight registers of vectors, each vector to include five elements that correspond with the other vectors. In combination with any of the above embodiments, ten permute operations are to be applied to contents of the preliminary vector registers to yield contents of the respective source vector registers. In combination with any of the above embodiments, the core further includes logic to create ten index vectors to be used with permute instructions yield contents of the source vector registers.
Some embodiments of the present disclosure include a system. The system may include a front end to receive an instruction, a decoder to decode the instruction, a core to execute the instruction, and a retirement unit to retire the instruction. In combination with any of the above embodiments, the core includes logic to determine that the instruction will require strided data converted from source data in memory. The strided data is to include corresponding indexed elements from a plurality of structures in the source data to be loaded into a final register to be used to execute the instruction. In combination with any of the above embodiments, the core includes logic to load source data into a plurality of preliminary vector registers to align a defined element of one of the preliminary vector registers in a position that corresponds to a required position in the final register for execution. In combination with any of the above embodiments, the core includes logic to apply a plurality of permute instructions to contents of the preliminary vector registers to cause corresponding indexed elements from the plurality of structures to be loaded into respective source vector registers. In combination with any of the above embodiments, the core includes logic to execute the instruction upon one or more source vector registers upon completion of conversion of source data to strided data. In combination with any of the above embodiments, the core includes logic to omit permute instruction execution for the defined element. In combination with any of the above embodiments, the core includes logic to load source data into the plurality of preliminary vector registers with a plurality of gaps to align the defined element to the required position. In combination with any of the above embodiments, the core includes logic to load source data into a number of preliminary vector registers that is greater than a number of the structures. In combination with any of the above embodiments, the strided data is to include eight registers of vectors, each vector to include five elements that correspond with the other vectors. In combination with any of the above embodiments, ten permute operations are to be applied to contents of the preliminary vector registers to yield contents of the respective source vector registers. In combination with any of the above embodiments, the core further includes logic to create ten index vectors to be used with permute instructions yield contents of the source vector registers.
Embodiments of the present disclosure may include an apparatus. The apparatus may include means for receiving an instruction, decoding the instruction, executing the instruction, and retiring the instruction. In combination with any of the above embodiments, the apparatus may include means for determining that the instruction will require strided data converted from source data in memory. The strided data is to means for corresponding indexed elements from a plurality of structures in the source data to be loaded into a final register to be used to execute the instruction. In combination with any of the above embodiments, the apparatus may include means for loading source data into a plurality of preliminary vector registers to align a defined element of one of the preliminary vector registers in a position that corresponds to a required position in the final register for execution. In combination with any of the above embodiments, the apparatus may include means for applying a plurality of permute instructions to contents of the preliminary vector registers to cause corresponding indexed elements from the plurality of structures to be loaded into respective source vector registers. In combination with any of the above embodiments, the apparatus may include means for executing the instruction upon one or more source vector registers upon completion of conversion of source data to strided data. In combination with any of the above embodiments, the apparatus may include means for omitting permute instruction execution for the defined element. In combination with any of the above embodiments, the apparatus may include means for loading source data into the plurality of preliminary vector registers with a plurality of gaps to align the defined element to the required position. In combination with any of the above embodiments, the apparatus may include means for loading source data into a number of preliminary vector registers that is greater than a number of the structures. In combination with any of the above embodiments, the strided data is to means for eight registers of vectors, each vector to means for five elements that correspond with the other vectors. In combination with any of the above embodiments, ten permute operations are to be applied to contents of the preliminary vector registers to yield contents of the respective source vector registers. In combination with any of the above embodiments, the apparatus may include means for creating ten index vectors to be used with permute instructions yield contents of the source vector registers.
Embodiments of the present disclosure may include a method. The method may include receiving an instruction, decoding the instruction, executing the instruction, and retiring the instruction. In combination with any of the above embodiments, the method may include determining that the instruction will require strided data converted from source data in memory. The strided data is to include corresponding indexed elements from a plurality of structures in the source data to be loaded into a final register to be used to execute the instruction. In combination with any of the above embodiments, the method may include loading source data into a plurality of preliminary vector registers to align a defined element of one of the preliminary vector registers in a position that corresponds to a required position in the final register for execution. In combination with any of the above embodiments, the method may include applying a plurality of permute instructions to contents of the preliminary vector registers to cause corresponding indexed elements from the plurality of structures to be loaded into respective source vector registers. In combination with any of the above embodiments, the method may include executing the instruction upon one or more source vector registers upon completion of conversion of source data to strided data. In combination with any of the above embodiments, the method may include omitting permute instruction execution for the defined element. In combination with any of the above embodiments, the method may include loading source data into the plurality of preliminary vector registers with a plurality of gaps to align the defined element to the required position. In combination with any of the above embodiments, the method may include loading source data into a number of preliminary vector registers that is greater than a number of the structures. In combination with any of the above embodiments, the strided data is to include eight registers of vectors, each vector to include five elements that correspond with the other vectors. In combination with any of the above embodiments, ten permute operations are to be applied to contents of the preliminary vector registers to yield contents of the respective source vector registers. In combination with any of the above embodiments, t the method may include creating ten index vectors to be used with permute instructions yield contents of the source vector registers.