[公众留言]关于兰阿公路

"SIMD" redirects here. For the cryptographic hash function, see SIMD (hash function). For the Scottish statistical tool, see Scottish index of multiple deprivation.

See also: SIMD within a register and Single instruction, multiple threads

Single instruction, multiple data

百度 去年的11月30日，“二十四节气”被列入联合国教科文组织人类非物质文化遗产代表作名录。

Single instruction, multiple data (SIMD) is a type of parallel computing (processing) in Flynn's taxonomy. SIMD describes computers with multiple processing elements that perform the same operation on multiple data points simultaneously. SIMD can be internal (part of the hardware design) and it can be directly accessible through an instruction set architecture (ISA), but it should not be confused with an ISA.

Such machines exploit data level parallelism, but not concurrency: there are simultaneous (parallel) computations, but each unit performs exactly the same instruction at any given moment (just with different data). A simple example is to add many pairs of numbers together, all of the SIMD units are performing an addition, but each one has different pairs of values to add. SIMD is especially applicable to common tasks such as adjusting the contrast in a digital image or adjusting the volume of digital audio. Most modern central processing unit (CPU) designs include SIMD instructions to improve the performance of multimedia use. In recent CPUs, SIMD units are tightly coupled with cache hierarchies and prefetch mechanisms, which minimize latency during large block operations. For instance, AVX-512-enabled processors can prefetch entire cache lines and apply fused multiply-add operations (FMA) in a single SIMD cycle.

Confusion between SIMT and SIMD

See also: SIMD within a register; Single instruction, multiple threads; and Vector processor

ILLIAC IV Array overview, from ARPA-funded Introductory description by Steward Denenberg, July 15 1971.

SIMD has three different subcategories in Flynn's 1972 Taxonomy, one of which is single instruction, multiple threads (SIMT). SIMT should not be confused with software threads or hardware threads, both of which are task time-sharing (time-slicing). SIMT is true simultaneous parallel hardware-level execution, such as in the ILLIAC IV.

SIMD should not be confused with Vector processing, characterized by the Cray 1 and clarified in Duncan's taxonomy. The difference between SIMD and vector processors is primarily the presence of a Cray-style SET VECTOR LENGTH instruction.

One key distinction between SIMT and SIMD is that the SIMD unit will not have its own memory. Another key distinction in SIMT is the presence of control flow mechanisms like warps (Nvidia terminology) or wavefronts (Advanced Micro Devices (AMD) terminology). ILLIAC IV simply called them "Control Signals". These signals ensure that each Processing Element in the entire parallel array is synchronized in its simultaneous execution of the (one, current) broadcast instruction.

Each hardware element (PU, or PE in ILLIAC IV terminology) working on individual data item sometimes also referred to as a SIMD lane or channel, although the ILLIAC IV PE was a scalar 64-bit unit. Modern graphics processing units (GPUs) are invariably wide SIMD within a register (SWAR) and typically have more that 16 data lanes or channels of such Processing Elements. Some newer GPUs integrate mixed-precision SWAR pipelines, which performs concurrent sub-word 8-bit, 16-bit, and 32-bit operations. This is critical for applications like AI inference, where mixed precision boosts throughput.

History

The first known operational use to date of SIMD within a register was the TX-2, in 1958. It was capable of 36-bit operations and two 18-bit or four 9-bit sub-word operations.

The first commercial use of SIMD instructions was in the ILLIAC IV, which was completed in 1972. This included 64 (of an original design of 256) processors that had local memory to hold different values while performing the same instruction. Separate hardware quickly sent out the values to be processed and gathered up the results.

Vector supercomputers of the early 1970s such as the CDC Star-100 and the Texas Instruments ASC could operate on a "vector" of data with a single instruction. Vector processing was especially popularized by Cray in the 1970s and 1980s. Vector processing architectures are now considered separate from SIMD computers: Duncan's Taxonomy includes them whereas Flynn's Taxonomy does not, due to Flynn's work (1966, 1972) pre-dating the Cray-1 (1977). The complexity of Vector processors however inspired a simpler arrangement known as SIMD within a register.

The first era of modern SIMD computers was characterized by massively parallel processing-style supercomputers such as the Thinking Machines Connection Machine CM-1 and CM-2. These computers had many limited-functionality processors that would work in parallel. For example, each of 65,536 single-bit processors in a Thinking Machines CM-2 would execute the same instruction at the same time, allowing, for instance, to logically combine 65,536 pairs of bits at a time, using a hypercube-connected network or processor-dedicated RAM to find its operands. Supercomputing moved away from the SIMD approach when inexpensive scalar multiple instruction, multiple data (MIMD) approaches based on commodity processors such as the Intel i860 XP became more powerful, and interest in SIMD waned.

The current era of SIMD processors grew out of the desktop-computer market rather than the supercomputer market. As desktop processors became powerful enough to support real-time gaming and audio/video processing during the 1990s, demand grew for this type of computing power, and microprocessor vendors turned to SIMD to meet the demand. This resurgence also coincided with the rise of DirectX and OpenGL shader models, which heavily leveraged SIMD under the hood. The graphics APIs encouraged programmers to adopt data-parallel programming styles, indirectly accelerating SIMD adoption in desktop software. Hewlett-Packard introduced Multimedia Acceleration eXtensions (MAX) instructions into PA-RISC 1.1 desktops in 1994 to accelerate MPEG decoding. Sun Microsystems introduced SIMD integer instructions in its "VIS" instruction set extensions in 1995, in its UltraSPARC I microprocessor. MIPS followed suit with their similar MDMX system.

The first widely deployed desktop SIMD was with Intel's MMX extensions to the x86 architecture in 1996. This sparked the introduction of the much more powerful AltiVec system in the Motorola PowerPC and IBM's POWER systems. Intel responded in 1999 by introducing the all-new SSE system. Since then, there have been several extensions to the SIMD instruction sets for both architectures. Advanced vector extensions AVX, AVX2 and AVX-512 are developed by Intel. AMD supports AVX, AVX2, and AVX-512 in their current products.

All of these developments have been oriented toward support for real-time graphics, and are therefore oriented toward processing in two, three, or four dimensions, usually with vector lengths of between two and sixteen words, depending on data type and architecture. When new SIMD architectures need to be distinguished from older ones, the newer architectures are then considered "short-vector" architectures, as earlier SIMD and vector supercomputers had vector lengths from 64 to 64,000. A modern supercomputer is almost always a cluster of MIMD computers, each of which implements (short-vector) SIMD instructions.

Advantages

An application that may take advantage of SIMD is one where the same value is being added to (or subtracted from) a large number of data points, a common operation in many multimedia applications. One example would be changing the brightness of an image. Each pixel of an image consists of three values for the brightness of the red (R), green (G) and blue (B) portions of the color. To change the brightness, the R, G and B values are read from memory, a value is added to (or subtracted from) them, and the resulting values are written back out to memory. Audio digital signal processors (DSPs) would likewise, for volume control, multiply both Left and Right channels simultaneously.

With a SIMD processor there are two improvements to this process. For one the data is understood to be in blocks, and a number of values can be loaded all at once. Instead of a series of instructions saying "retrieve this pixel, now retrieve the next pixel", a SIMD processor will have a single instruction that effectively says "retrieve n pixels" (where n is a number that varies from design to design). For a variety of reasons, this can take much less time than retrieving each pixel individually, as with a traditional CPU design. Moreover, SIMD instructions can exploit data reuse, where the same operand is used across multiple calculations, via broadcasting features. For example, multiplying several pixels by a constant scalar value can be done more efficiently by loading the scalar once and broadcasting it across a SIMD register.

Another advantage is that the instruction operates on all loaded data in a single operation. In other words, if the SIMD system works by loading up eight data points at once, the add operation being applied to the data will happen to all eight values at the same time. This parallelism is separate from the parallelism provided by a superscalar processor; the eight values are processed in parallel even on a non-superscalar processor, and a superscalar processor may be able to perform multiple SIMD operations in parallel.

Disadvantages

• Not all algorithms can be vectorized easily. For example, a flow-control-heavy task like code parsing may not easily benefit from SIMD; however, it is theoretically possible to vectorize comparisons and "batch flow" to target maximal cache optimality, though this technique will require more intermediate state. Note: Batch-pipeline systems (example: GPUs or software rasterization pipelines) are most advantageous for cache control when implemented with SIMD intrinsics, but they are not exclusive to SIMD features. Further complexity may be apparent to avoid dependence within series such as code strings; while independence is required for vectorization. Additionally, divergent control flow—where different data lanes would follow different execution paths—can lead to underutilization of SIMD hardware. To handle such divergence, techniques like masking and predication are often employed, but they introduce performance overhead and complexity.
• Large register files which increases power consumption and required chip area.
• Currently, implementing an algorithm with SIMD instructions usually requires human labor; most compilers do not generate SIMD instructions from a typical C program, for instance. Automatic vectorization in compilers is an active area of computer science research. (Compare Vector processor.)
• Programming with given SIMD instruction sets can involve many low-level challenges.
  1. SIMD may have restrictions on data alignment; programmers familiar with a given architecture may not expect this. Worse: the alignment may change from one revision or "compatible" processor to another.
  2. Gathering data into SIMD registers and scattering it to the correct destination locations is tricky (sometimes requiring permute instructions (operations) and can be inefficient.
  3. Specific instructions like rotations or three-operand addition are not available in some SIMD instruction sets.
  4. Instruction sets are architecture-specific: some processors lack SIMD instructions entirely, so programmers must provide non-vectorized implementations (or different vectorized implementations) for them.
  5. Different architectures provide different register sizes (e.g. 64, 128, 256 and 512 bits) and instruction sets, meaning that programmers must provide multiple implementations of vectorized code to operate optimally on any given CPU. In addition, the possible set of SIMD instructions grows with each new register size. Unfortunately, for legacy support reasons, the older versions cannot be retired.
  6. The early MMX instruction set shared a register file with the floating-point stack, which caused inefficiencies when mixing floating-point and MMX code. However, SSE2 corrects this.

To remedy problems 1 and 5, Cray-style Vector processors use an alternative approach: instead of exposing the sub-register-level details directly to the programmer, the instruction set abstracts out at least the length (number of elements) into a runtime control register, usually named "VL" (Vector Length). The hardware then handles all alignment issues and "strip-mining" of loops. Machines with different vector sizes would be able to run the same code. LLVM calls this vector type "vscale".

With SIMD, an order of magnitude increase in code size is not uncommon, when compared to equivalent scalar or equivalent vector code, and an order of magnitude or greater effectiveness (work done per instruction) is achievable with Vector ISAs.

ARM's Scalable Vector Extension takes another approach, known in Flynn's Taxonomy more commonly known today as "Predicated" (masked) SIMD. This approach is not as compact as vector processing but is still far better than non-predicated SIMD. Detailed comparative examples are given at Vector processor § Vector instruction example. In addition, all versions of the ARM architecture have offered Load and Store multiple instructions, to Load or Store a block of data from a continuous block of memory, into a range or non-continuous set of registers.

Chronology

SIMD supercomputer examples excluding vector processors

Year	Example
1974	ILLIAC IV - an Array Processor comprising scalar 64-bit PEs
1974	ICL Distributed Array Processor (DAP)
1976	Burroughs Scientific Processor
1981	Geometric-Arithmetic Parallel Processor from Martin Marietta (continued at Lockheed Martin, then at Teranex and Silicon Optix)
1983–1991	Massively Parallel Processor (MPP), from NASA/Goddard Space Flight Center
1985	Connection Machine, models 1 and 2 (CM-1 and CM-2), from Thinking Machines Corporation
1987–1996	MasPar MP-1 and MP-2
1991	Zephyr DC from Wavetracer
2001	Xplor from Pyxsys, Inc.

Hardware

Small-scale (64 or 128 bits) SIMD became popular on general-purpose CPUs in the early 1990s and continued through 1997 and later with Motion Video Instructions (MVI) for Alpha. SIMD instructions can be found, to one degree or another, on most CPUs, including IBM's AltiVec and Signal Processing Engine (SPE) for PowerPC, Hewlett-Packard's (HP) PA-RISC Multimedia Acceleration eXtensions (MAX), Intel's MMX and iwMMXt, Streaming SIMD Extensions (SSE), SSE2, SSE3 SSSE3 and SSE4.x, AMD's 3DNow!, ARC's ARC Video subsystem, SPARC's VIS and VIS2, Sun's MAJC, ARM's Neon technology, MIPS' MDMX (MaDMaX) and MIPS-3D. The IBM, Sony, Toshiba co-developed Cell processor's Synergistic Processing Element's (SPE's) instruction set is heavily SIMD based. Philips, now NXP, developed several SIMD processors named Xetal. The Xetal has 320 16-bit processor elements especially designed for vision tasks. Apple's M1 and M2 chips also incorporate SIMD units deeply integrated with their GPU and Neural Engine, using Apple-designed SIMD pipelines optimized for image filtering, convolution, and matrix multiplication. This unified memory architecture helps SIMD instructions operate on shared memory pools more efficiently.

Intel's AVX-512 SIMD instructions process 512 bits of data at once.

Software

The ordinary tripling of four 8-bit numbers. The CPU loads one 8-bit number into R1, multiplies it with R2, and then saves the answer from R3 back to RAM. This process is repeated for each number.

The SIMD tripling of four 8-bit numbers. The CPU loads 4 numbers at once, multiplies them all in one SIMD-multiplication, and saves them all at once back to RAM. In theory, the speed can be multiplied by 4.

SIMD instructions are widely used to process 3D graphics, although modern graphics cards with embedded SIMD have largely taken over this task from the CPU. Some systems also include permute functions that re-pack elements inside vectors, making them especially useful for data processing and compression. They are also used in cryptography. The trend of general-purpose computing on GPUs (GPGPU) may lead to wider use of SIMD in the future. Recent compilers such as LLVM, GNU Compiler Collection (GCC), and Intel's ICC offer aggressive auto-vectoring options. Developers can often enable these with flags like -O3 or -ftree-vectorize, which guide the compiler to restructure loops for SIMD compatibility.

Adoption of SIMD systems in personal computer software was at first slow, due to a number of problems. One was that many of the early SIMD instruction sets tended to slow overall performance of the system due to the re-use of existing floating point registers. Other systems, like MMX and 3DNow!, offered support for data types that were not interesting to a wide audience and had expensive context switching instructions to switch between using the FPU and MMX registers. Compilers also often lacked support, requiring programmers to resort to assembly language coding.

SIMD on x86 had a slow start. The introduction of 3DNow! by AMD and SSE by Intel confused matters somewhat, but today the system seems to have settled down (after AMD adopted SSE) and newer compilers should result in more SIMD-enabled software. Intel and AMD now both provide optimized math libraries that use SIMD instructions, and open source alternatives like libSIMD, SIMDx86 and SLEEF have started to appear (see also libm).

Apple Computer had somewhat more success, even though they entered the SIMD market later than the rest. AltiVec offered a rich system and can be programmed using increasingly sophisticated compilers from Motorola, IBM and GNU, therefore assembly language programming is rarely needed. Additionally, many of the systems that would benefit from SIMD were supplied by Apple itself, for example iTunes and QuickTime. However, in 2006, Apple computers moved to Intel x86 processors. Apple's APIs and development tools (XCode) were modified to support SSE2 and SSE3 as well as AltiVec. Apple was the dominant purchaser of PowerPC chips from IBM and Freescale Semiconductor. Even though Apple has stopped using PowerPC processors in their products, further development of AltiVec is continued in several PowerPC and Power ISA designs from Freescale and IBM.

SIMD within a register, or SWAR, is a range of techniques and tricks used for performing SIMD in general-purpose registers on hardware that does not provide any direct support for SIMD instructions. This can be used to exploit parallelism in certain algorithms even on hardware that does not support SIMD directly.

Programmer interface

It is common for publishers of the SIMD instruction sets to make their own C and C++ language extensions with intrinsic functions or special datatypes (with operator overloading) guaranteeing the generation of vector code. Intel, AltiVec, and ARM NEON provide extensions widely adopted by the compilers targeting their CPUs. (More complex operations are the task of vector math libraries.)

The GNU C Compiler takes the extensions a step further by abstracting them into a universal interface that can be used on any platform by providing a way of defining SIMD datatypes. The LLVM Clang compiler also implements the feature, with an analogous interface defined in the IR. Rust's packed_simd crate (and the experimental std::simd) uses this interface, and so does Swift 2.0+.

C++ has an experimental interface std::experimental::simd that works similarly to the GCC extension. LLVM's libcxx seems to implement it. For GCC and libstdc++, a wrapper library that builds on top of the GCC extension is available.

Microsoft added SIMD to .NET in RyuJIT. The System.Numerics.Vector package, available on NuGet, implements SIMD datatypes. Java also has a new proposed API for SIMD instructions available in OpenJDK 17 in an incubator module. It also has a safe fallback mechanism on unsupported CPUs to simple loops.

Instead of providing an SIMD datatype, compilers can also be hinted to auto-vectorize some loops, potentially taking some assertions about the lack of data dependency. This is not as flexible as manipulating SIMD variables directly, but is easier to use. OpenMP 4.0+ has a #pragma omp simd hint. This OpenMP interface has replaced a wide set of nonstandard extensions, including Cilk's #pragma simd, GCC's #pragma GCC ivdep, and many more.

SIMD multi-versioning

Consumer software is typically expected to work on a range of CPUs covering multiple generations, which could limit the programmer's ability to use new SIMD instructions to improve the computational performance of a program. The solution is to include multiple versions of the same code that uses either older or newer SIMD technologies, and pick one that best fits the user's CPU at run-time (dynamic dispatch). There are two main camps of solutions:

• Function multi-versioning (FMV): a subroutine in the program or a library is duplicated and compiled for many instruction set extensions, and the program decides which one to use at run-time.
• Library multi-versioning (LMV): the entire programming library is duplicated for many instruction set extensions, and the operating system or the program decides which one to load at run-time.

FMV, manually coded in assembly language, is quite commonly used in a number of performance-critical libraries such as glibc and libjpeg-turbo. Intel C++ Compiler, GNU Compiler Collection since GCC 6, and Clang since clang 7 allow for a simplified approach, with the compiler taking care of function duplication and selection. GCC and clang requires explicit target_clones labels in the code to "clone" functions, while ICC does so automatically (under the command-line option /Qax). The Rust programming language also supports FMV. The setup is similar to GCC and Clang in that the code defines what instruction sets to compile for, but cloning is manually done via inlining.

As using FMV requires code modification on GCC and Clang, vendors more commonly use library multi-versioning: this is easier to achieve as only compiler switches need to be changed. Glibc supports LMV and this functionality is adopted by the Intel-backed Clear Linux project.

SIMD on the web

In 2013 John McCutchan announced that he had created a high-performance interface to SIMD instruction sets for the Dart programming language, bringing the benefits of SIMD to web programs for the first time. The interface consists of two types:

• Float32x4, 4 single precision floating point values.
• Int32x4, 4 32-bit integer values.

Instances of these types are immutable and in optimized code are mapped directly to SIMD registers. Operations expressed in Dart typically are compiled into a single instruction without any overhead. This is similar to C and C++ intrinsics. Benchmarks for 4×4 matrix multiplication, 3D vertex transformation, and Mandelbrot set visualization show near 400% speedup compared to scalar code written in Dart.

Intel announced at IDF 2013 that they were implementing McCutchan's specification for both V8 and SpiderMonkey. However, by 2017, SIMD.js was taken out of the ECMAScript standard queue in favor of pursuing a similar interface in WebAssembly. Support for SIMD was added to the WebAssembly 2.0 specification, which was finished on 2022 and became official on December 2024. LLVM's autovectoring, when compiling C or C++ to WebAssembly, can target WebAssembly SIMD to automatically make use of SIMD, while SIMD intrinsic are also available.

Commercial applications

It has generally proven difficult to find sustainable commercial applications for SIMD-only processors.

One that has had some measure of success is the GAPP, which was developed by Lockheed Martin and taken to the commercial sector by their spin-off Teranex. The GAPP's recent incarnations have become a powerful tool in real-time video processing applications like conversion between various video standards and frame rates (NTSC to/from PAL, NTSC to/from high-definition television (HDTV) formats, etc.), deinterlacing, image noise reduction, adaptive video compression, and image enhancement.

A more ubiquitous application for SIMD is found in video games: nearly every modern video game console since 1998 has incorporated a SIMD processor somewhere in its architecture. The PlayStation 2 was unusual in that one of its vector-float units could function as an autonomous digital signal processor (DSP) executing its own instruction stream, or as a coprocessor driven by ordinary CPU instructions. 3D graphics applications tend to lend themselves well to SIMD processing as they rely heavily on operations with 4-dimensional vectors. Microsoft's Direct3D 9.0 now chooses at runtime processor-specific implementations of its own math operations, including the use of SIMD-capable instructions.

A later processor that used vector processing is the Cell processor used in the Playstation 3, which was developed by IBM in cooperation with Toshiba and Sony. It uses a number of SIMD processors (a non-uniform memory access (NUMA) architecture, each with independent local store and controlled by a general purpose CPU) and is geared towards the huge datasets required by 3D and video processing applications. It differs from traditional ISAs by being SIMD from the ground up with no separate scalar registers.

Ziilabs produced an SIMD type processor for use on mobile devices, such as media players and mobile phones.

Larger scale commercial SIMD processors are available from ClearSpeed Technology, Ltd. and Stream Processors, Inc. ClearSpeed's CSX600 (2004) has 96 cores each with two double-precision floating point units while the CSX700 (2008) has 192. Stream Processors is headed by computer architect Bill Dally. Their Storm-1 processor (2007) contains 80 SIMD cores controlled by a MIPS CPU.

See also

• Streaming SIMD Extensions, MMX, SSE2, SSE3, Advanced Vector Extensions, AVX-512
• Instruction set architecture
• Flynn's taxonomy
• SIMD within a register (SWAR)
• Single program, multiple data (SPMD)
• OpenCL

External links

• SIMD architectures (2000)
• Cracking Open The Pentium 3 (1999)
• Short Vector Extensions in Commercial Microprocessor
• Article about Optimizing the Rendering Pipeline of Animated Models Using the Intel Streaming SIMD Extensions
• "Yeppp!": cross-platform, open-source SIMD library from Georgia Tech
• Introduction to Parallel Computing from LLNL Lawrence Livermore National Laboratory Archived 2025-08-06 at the Wayback Machine
• simde on GitHub: A portable implementation of platform-specific intrinsics for other platforms (e.g. SSE intrinsics for ARM NEON), using C/C++ headers