Gensimd

Gensimd is a collection of packages and a command, gensimd, for using SIMD in Go.

Write functions in a subset of Go, using the gensimd/simd, gensimd/simd/sse2 packages and convert them to Go assembly with the gensimd command.

CODEGEN REWRITE - Code generation is being rewritten to use the new Go compiler SSA backend, the rewrite is in progress at https://github.com/bjwbell/ssa and https://github.com/bjwbell/gir.

Install

export GO15VENDOREXPERIMENT=1
go get github.com/bjwbell/gensimd
go install github.com/bjwbell/gensimd
go get github.com/bjwbell/gensimd/simd

Vendoring golang.org/x/tools/go/ssa and enabling GO15VENDOREXPERIMENT is a temporary hack.

Optional - SSE2

The SSE2 intrinsics package is github.com/bjwbell/gensimd/simd/sse

Examples

• Distance calculation
• Cross platform - SIMD
• Platform specific - SSE2

To build and run all examples execute ./run_examples.sh.

Tests

To build and run the reference tests execute ./run_tests.sh.

Gensimd Command

[bjwbell]$ gensimd --help
  -debug
    	include debug comments in assembly
  -f string
    	input file with function definitions
  -fn string
    	comma separated list of function names
  -goprotofile string
    	output file for SIMD function prototype(s)
  -o string
    	Go assembly output file
  -outfn string
    	comma separated list of output function names
  -spills
    	print each register spill
  -ssa
    	dump ssa representation

Go Language Subset

For functions gensimd translates from Go to assembly it supports only a small subset of Go.

Go - Supported

• Integers and floats - uint8/int8, uint16/int16, uint32/int32, uint64/int64, float32/float64
• if statements, for loops (except with range)
• Arrays and slices

Go - Unsupported

• Heap allocated local variables
• Multiple and named return values
• Builtins except len
• Function calls except to simd.*
• Method calls
• All struct types except simd.*
• Keywords range, map, select, chan, defer
• Slice creation e.g. newslice := slice[1:len(slice) - 2]

TODO

• Slice access bounds checking

SIMD

SIMD intrinsics are availabe if simd.Available() returns true. The common vs platform specific functionality is inspired from Huon Wilson's SIMD work.

Integer Overflow

For unsigned integer values, the SIMD functions Add*, Sub*, Mul*, and Shl* are computed modulo 2^n, where n is the bit width of the unsigned integer's type. These unsigned integer operations discard high bits upon overflow.

For signed integers, the SIMD functions Add*, Sub*, Mul*, and Shl* may overflow and the resulting value is defined by the signed integer representation, the operation, and its operands. The behavior is guaranteed to be identical to the Go versions in gensimd/simd/simd.go.

For both unsigned and signed integer values, the SIMD function Shr* is guaranteed to have the same behavior as the Go version in gensimd/simd/simd.go

Floating Point

The behavior of the floating point SIMD functions Add*, Sub*, Mul*, and Div* is guaranteed to be identical to the Go versions in gensimd/simd/simd.go.

SIMD types

type I8x16 [16]int8
type I16x8 [8]int16
type I32x4 [4]int32
type I64x2 [2]int64
type U8x16 [16]uint8
type U16x8 [8]uint16
type U32x4 [4]uint32
type U64x2 [2]uint64
type F32x4 [4]float32
type F64x2 [2]float64

SIMD functions

func AddI8x16(x, y I8x16) I8x16
func SubI8x16(x, y I8x16) I8x16
func AddU8x16(x, y U8x16) U8x16
func SubU8x16(x, y U8x16) U8x16

func AddI16x8(x, y I16x8) I16x8
func SubI16x8(x, y I16x8) I16x8
func MulI16x8(x, y I16x8) I16x8
func ShlI16x8(x I16x8, shift uint8) I16x8
func ShrI16x8(x I16x8, shift uint8) I16x8
func AddU16x8(x, y U16x8) U16x8
func SubU16x8(x, y U16x8) U16x8
func MulU16x8(x, y U16x8) U16x8
func ShlU16x8(x U16x8, shift uint8) U16x8
func ShrU16x8(x U16x8, shift uint8) U16x8

func AddI32x4(x, y I32x4) I32x4
func SubI32x4(x, y I32x4) I32x4
func MulI32x4(x, y I32x4) I32x4
func ShlI32x4(x I32x4, shift uint8) I32x4
func ShrI32x4(x I32x4, shift uint8) I32x4
func ShuffleI32x4(x I32x4, order uint8) I32x4
func AddU32x4(x, y U32x4) U32x4
func SubU32x4(x, y U32x4) U32x4
func MulU32x4(x, y U32x4) U32x4
func ShlU32x4(x U32x4, shift uint8) U32x4
func ShrU32x4(x U32x4, shift uint8) U32x4
func ShuffleU32x4(x U32x4, order uint8) U32x4

func AddI64x2(x, y I64x2) I64x2
func SubI64x2(x, y I64x2) I64x2
func AddU64x2(x, y U64x2) U64x2
func SubU64x2(x, y U64x2) U64x2

func AddF32x4(x, y F32x4) F32x4
func SubF32x4(x, y F32x4) F32x4
func MulF32x4(x, y F32x4) F32x4
func DivF32x4(x, y F32x4) F32x4

func AddF64x2(x, y F64x2) F64x2
func SubF64x2(x, y F64x2) F64x2
func MulF64x2(x, y F64x2) F64x2
func DivF64x2(x, y F64x2) F64x2

Gotchas

There are no SIMD functions for 64 bit integer multiplication because there's no equivalent SSE2 instruction.

Until Go 1.5.2, ShrU16x8 is slow because of golang/go#13010 "cmd/asm: x86, incorrect Optab entry - PSRLW".

MulI32x4 is slow because the instruction "PMULLD" wasn't added until SSE4.1. It's emulated using SSE2 instructions, SSE multiplication of 4 32-bit integers.

The shuffle order operand in ShuffleI32x4/ShuffleU32x4 MUST be a constant not a variable. Example:

const order uint8 = 8
simd.SuffleU32x4(x, order)

Misc

The below functions aren't implemented because they have no directly equivalent SSE2 instructions.

func ShlI64x2(x, shift uint8) I64x2
func ShrI64x2(x, shift uint8) I64x2
func ShlU64x2(x, shift uint8) U64x2
func ShrU64x2(x, shift uint8) U64x2

Platform Specific - SSE2

SSE2 intrinsics are availabe if simd.SSE2() returns true.

SSE2 types

type M128i [16]byte
type M128 [4]float32
type M128d [2]float64

SSE2 intrinsics

func AddEpi64(x, y simd.M128i) simd.M128i
func SubEpi64(x, y simd.M128i) simd.M128i
func MulEpu32(x, y simd.M128i) simd.M128i
func ShufflehiEpi16(x, y simd.M128i) simd.M128i
func ShuffleloEpi16(x, y simd.M128i) simd.M128i
func ShuffleEpi32(x, y simd.M128i) simd.M128i
func SlliSi128(x, y simd.M128i) simd.M128i
func SrliSi128(x, y simd.M128i) simd.M128i
func UnpackhiEpi64(x, y simd.M128i) simd.M128i
func UnpackloEpi64(x, y simd.M128i) simd.M128i
func AddPd(x, y simd.M128d) simd.M128d
func AddSd(x, y simd.M128d) simd.M128d
func AndnotPd(x, y simd.M128d) simd.M128d
func CmpeqPd(x, y simd.M128d) simd.M128d
func CmpeqSd(x, y simd.M128d) simd.M128d