SIMD

SIMD (pronounced "seem-dee") is short for Single Instruction/Multiple Data which is one classification of computer architectures . SIMD allows one same operation to be performed on multiple data points resulting in data level parallelism and thus performance gains, for example for 3D graphics and video processing, physics simulations or cryptography, and other domains.

Description

The JavaScript SIMD API consists of several new types and operations. Browsers provide highly optimized implementations of this API depending on the underlying hardware of the user. Currently, SIMD is especially modeled for ARMv7 platforms with NEON and x86 platforms with SSE.

The SIMD API types are installed on a SIMD module. Unlike the other global objects, SIMD is not a constructor. You can not use it with a new operator or invoke the SIMD object as a function. All properties and methods of SIMD are static (like it is the case with the Math object).

Overview

A SIMD value has multiple lanes. For a vector of length 4, the lanes are named x, y, z, and w. Now, instead of having to perform 4 separate operations on each of these lanes, SIMD allows you to perform the operation on all 4 lanes simultaneously. This requires fewer operations, which leads to performance improvements and better energy efficiency compared to scalar operations (SISD). Note that SIMD operations cannot be used to process multiple data in different ways. In the following figure, there is only a single instruction (addition) and thus it could be operated with SIMD:

Figures 1 and 2: SISD and SIMD compared.

Simple addition arithmetic

The JavaScript code for a simple SIMD operation like in figure 2 looks like this:

var a = new SIMD.float32x4(1, 2, 3, 4);
var b = new SIMD.float32x4(5, 6, 7, 8);
var c = SIMD.float32x4.add(a,b); // float32x4[6,8,10,12]

Data types

All SIMD data types are immutable. You can not alter them directly. Instead, you perform operations that create new immutable SIMD data types. The following figure shows the different SIMD data types in a 128-bit SIMD register. The current SIMD JavaScript API has 5 different types with lane lengths of either 2, 4, 8 or 16.

Figure 3: Lanes per type in a 128-bit SIMD register.

SIMD.int8x16

128-bits divided into 16 lanes storing 8-bit signed integer values.

SIMD.int16x8

128-bits divided into 8 lanes storing 16-bit signed integer values.

SIMD.int32x4

128-bits divided into 4 lanes storing 32-bit signed integer values.

SIMD.float32x4

128-bits divided into 4 lanes storing single precision floating point values.

SIMD.float64x2

128-bits divided into 2 lanes storing double precision floating point values.

Constructors

In addition to the simple constructors (e.g. new SIMD.int32x4(1,2,3,4)), the SIMD API provides the following constructors. Note that you can also convert from one SIMD data type to another.

SIMD.%type%.splat()

Creates SIMD data type with all lanes set to a given value.

SIMD.%type%.bool()

Creates a SIMD data type with boolean parameters, allowing you to create an explicit selection mask.

Operations

To actually do something with SIMD types, SIMD operations are needed that work on SIMD data types.

Checking SIMD types

SIMD.%type%.check()

Returns a new instance if the parameter is a valid SIMD data type and the same as %type%. Throws a TypeError otherwise.

Accessing and mutating lanes

SIMD.%type%.extractLane()

Returns the value of the given lane.

SIMD.%type%.replaceLane()

Returns a new instance with the given lane value replaced.

Loading from and storing into typed arrays

SIMD.%type%.load()

Returns a new instance with the lane values loaded from a typed array.

SIMD.%type%.store()

Store a SIMD data type into a typed array.

Arithmetic operations

SIMD.%type%.abs()

Returns a new instance with the absolute lane values.

SIMD.%type%.add()

Returns a new instance with the lane values added (a + b).

SIMD.%type%.div()

Returns a new instance with the lane values divided (a / b).

SIMD.%type%.mul()

Returns a new instance with the lane values multiplied (a * b).

SIMD.%type%.neg()

Returns a new instance with the negated lane values.

SIMD.%type%.reciprocalApproximation()

Returns a new instance with an approximation of the reciprocal lane values.

SIMD.%type%.reciprocalSqrtApproximation()

Returns a new instance with an approximation of the reciprocal square root lane values.

SIMD.%type%.sub()

Returns a new instance with the lane values subtracted (a - b).

SIMD.%type%.sqrt()

Returns a new instance with the square root of the lane values.

Shuffling and swizzling

SIMD.%type%.shuffle()

Returns a new instance with the lane values shuffled.

SIMD.%type%.swizzle()

Returns a new instance with the lane values swizzled.

Min/max and clamping

SIMD.%type%.clamp()

Returns a new instance with the lane values clamped between a lower limit and an upper limit.

SIMD.%type%.max()

Returns a new instance with the maximum of the lane values.

SIMD.%type%.maxNum()

Returns a new instance with the maximum of the lane values, preferring numbers over NaN.

SIMD.%type%.min()

Returns a new instance with the minimum of the lane values.

SIMD.%type%.minNum()

Returns a new instance with the minimum of the lane values, preferring numbers over NaN.

Selections

SIMD.%type%.select()

Returns a new instance with the lane values being a mix of the lanes depending on the selector mask.

SIMD.%type%.selectBits()

Returns a new instance with the lane values being a mix of bits depending on the selector mask.

Comparisons

SIMD.%type%.equal()

Returns a selection mask depending on a == b.

SIMD.%type%.notEqual()

Returns a selection mask depending on a != b.

SIMD.%type%.lessThan()

Returns a selection mask depending on a < b.

SIMD.%type%.lessThanOrEqual()

Returns selection mask depending on a <= b.

SIMD.%type%.greaterThan()

Returns a selection mask depending on a > b.

SIMD.%type%.greaterThanOrEqual()

Returns a selection mask depending on a >= b.

Bitwise logical operations

SIMD.%type%.and()

Returns a new instance with the logical AND of the lane values (a & b).

SIMD.%type%.or()

Returns a new instance with the logical OR of the lane values (a | b).

SIMD.%type%.xor()

Returns a new instance with the logical XOR of the lane values (a ^ b).

SIMD.%type%.not()

Returns a new instance with the logical NOT of the lane values (~a).

Bitwise shift operations

SIMD.%type%.shiftLeftByScalar()

Returns a new instance with the lane values shifted left by a given bit count (a << bits).

SIMD.%type%.shiftRightArithmeticByScalar()

Returns a new instance with the lane values shifted right (arithmetic) by a given bit count (a >> bits).

SIMD.%type%.shiftRightLogicalByScalar()

Returns a new instance with the lane values shifted right (logical) by a given bit count (a >>> bits).

Data conversions (packing / unpacking)

SIMD.%type%.fromFloat32x4()

Creates a new SIMD data type with a float conversion from a float32x4.

SIMD.%type%.fromFloat32x4Bits()

Creates a new SIMD data type with a bit-wise copy from a float32x4.

SIMD.%type%.fromFloat64x2()

Creates a new SIMD data type with a float conversion from a float64x2.

SIMD.%type%.fromFloat64x2Bits()

Creates a new SIMD data type with a bit-wise copy from a float64x2.

SIMD.%type%.fromInt32x4()

Creates a new SIMD data type with an integer conversion from a in32x4.

SIMD.%type%.fromInt32x4Bits()

Creates a new SIMD data type with a bit-wise copy from an int32x4.

SIMD.%type%.fromInt16x8Bits()

Creates a new SIMD data type with a bit-wise copy from an int16x8.

SIMD.%type%.fromInt8x16Bits()

Creates a new SIMD data type with a bit-wise copy from an int8x16.

Specifications

SIMD is not yet part of an official standards document or draft. For standardization work and a Polyfill implementation based on typed arrays, see the ecmascript_simd GitHub repository.

Browser compatibility

Status notes

• "In development" in Microsoft Edge
• "Intent to implement" in Blink/Chromium

See also

• Glossary: SIMD
• Data types and data structures
• JavaScript typed arrays
• SIMD Programming in JavaScript, talk by Peter Jensen, Intel.