ref

package

v1.0.5 Latest Latest Go to latest Published: Jul 31, 2026 License: MIT Imports: 8 Imported by: 0

Details

Valid go.mod file
Redistributable license
Tagged version
Stable version
Learn more about best practices

Repository

github.com/sebishogun/simd

Links

Open Source Insights

Documentation ¶

Overview ¶

Package ref is the portable Go reference implementation of every kernel.

It defines the semantics. Every generated backend is differential-tested against it, and it is also the live fallback: the dispatcher runs these functions on architectures with no backend, in builds made with the purego tag, and below the per-kernel element threshold where crossing into assembly costs more than it saves.

It is therefore not throwaway code. The loops are written for bounds-check elimination because they run in the small-n hot path.

The numerical contract these functions define is documented on package kernel. The parts that are easy to get wrong:

Floating-point reductions use exactly kernel.SumLanes accumulators and kernel.CombineTree, so every vector width reproduces them bit for bit.
Dot multiplies and adds with separate roundings; it does not fuse.
Minimum, Maximum, Min and Max implement IEEE-754-2019 minimum/maximum: NaN propagates, and +0 compares greater than -0.
Integer Abs and Neg wrap, so Abs(MinInt32) is MinInt32, matching PABSD.

Index ¶

func AbsFloat[T Float](dst, a []T)
func AbsInt[T Integer](dst, a []T)
func Acos[T float](dst, a []T)
func Acosh[T float](dst, a []T)
func Add[T Number](dst, a, b []T)
func Add3[T Number](dst, a, b, c []T)
func Add4[T Number](dst, a, b, c, d []T)
func AddScalar[T Number](dst, a []T, s T)
func AddScaled[T Number](dst, a, b []T, s T)
func ArgMaxFloat[T Float](a []T) int
func ArgMaxInt[T Integer](a []T) int
func ArgMinFloat[T Float](a []T) int
func ArgMinInt[T Integer](a []T) int
func Asin[T float](dst, a []T)
func Asinh[T float](dst, a []T)
func Atan[T float](dst, a []T)
func Atan2[T float](dst, a, b []T)
func Atanh[T float](dst, a []T)
func B64Decode(dst, src []byte) int
func B64Encode(dst, src []byte) int
func BF16ToF32(dst []float32, a []uint16)
func BitAnd(dst, a, b []byte)
func BitAndNot(dst, a, b []byte)
func BitNot(dst, a []byte)
func BitOr(dst, a, b []byte)
func BitPackU32(dst, a []uint32, bits int32)
func BitUnpackU32(dst, a []uint32, bits int32)
func BitXor(dst, a, b []byte)
func ByteSwap[T Integer](dst, a []T)
func CAbs64(dst []float32, a []complex64)
func CAbs128(dst []float64, a []complex128)
func CAdd[C complexT](dst, a, b []C)
func CConj64(dst, a []complex64)
func CConj128(dst, a []complex128)
func CDiv64(dst, a, b []complex64)
func CDiv128(dst, a, b []complex128)
func CDot64(a, b []complex64) complex64
func CDot128(a, b []complex128) complex128
func CDotConj64(a, b []complex64) complex64
func CDotConj128(a, b []complex128) complex128
func CFromParts64(dst []complex64, re, im []float32)
func CFromParts128(dst []complex128, re, im []float64)
func CImag64(dst []float32, a []complex64)
func CImag128(dst []float64, a []complex128)
func CMul64(dst, a, b []complex64)
func CMul128(dst, a, b []complex128)
func CNeg[C complexT](dst, a []C)
func CReal64(dst []float32, a []complex64)
func CReal128(dst []float64, a []complex128)
func CScale64(dst, a []complex64, s float32)
func CScale128(dst, a []complex128, s float64)
func CSub[C complexT](dst, a, b []C)
func CSum64(a []complex64) complex64
func CSum128(a []complex128) complex128
func Cbrt[T float](dst, a []T)
func Ceil[T Float](dst, a []T)
func ClampFloat[T Float](dst, a []T, lo, hi T)
func ClampInt[T Integer](dst, a []T, lo, hi T)
func CommonPrefix(a, b []byte) int
func CompareBytes(a, b []byte) int
func Compress[T any](dst, src []T, keep []bool) int
func CompressFloat32(dst, src []float32, keep []bool) int
func CompressFloat64(dst, src []float64, keep []bool) int
func CompressInt32(dst, src []int32, keep []bool) int
func CompressInt64(dst, src []int64, keep []bool) int
func Convolve[T Number](dst, sig, ker []T)
func Correlate[T Number](dst, sig, ker []T)
func Cos[T float](dst, a []T)
func Cosh[T float](dst, a []T)
func CountAny(b, chars []byte) int
func CountByte(b []byte, c byte) int
func CountSeq(haystack, needle []byte) int
func CumMaxFloat[T Float](dst, a []T)
func CumMaxInt[T Integer](dst, a []T)
func CumMinFloat[T Float](dst, a []T)
func CumMinInt[T Integer](dst, a []T)
func CumProdInt[T Integer](dst, a []T)
func DequantizeI8(dst []float32, a []int8, scale float32, zeroPoint int32)
func DequantizePerChannelI8(dst []float32, a []int8, scale []float32, zeroPoint []int32, ...)
func DequantizePerChannelU8(dst []float32, a []uint8, scale []float32, zeroPoint []int32, ...)
func DequantizeU8(dst []float32, a []uint8, scale float32, zeroPoint int32)
func Diff[T Number](dst, a []T)
func DifferenceInt[T Integer](dst, a, b []T) int
func Div[T Float](dst, a, b []T)
func DivScalar[T Number](dst, a []T, s T)
func DotFloat[T Float](a, b []T) T
func DotInt[T Integer](a, b []T) T
func EqualBytes(a, b []byte) bool
func EqualFoldASCII(a, b []byte) bool
func EqualMask[T number](dst []bool, a, b []T)
func EqualScalarMask[T number](dst []bool, a []T, v T)
func Erf[T float](dst, a []T)
func Erfc[T float](dst, a []T)
func Exp[T float](dst, a []T)
func Exp2[T float](dst, a []T)
func Expand[T any](dst, src []T, keep []bool) int
func Expm1[T float](dst, a []T)
func F8E4M3ToF32(dst []float32, a []byte)
func F8E5M2ToF32(dst []float32, a []byte)
func F16ToF32(dst []float32, a []uint16)
func F32ToBF16(dst []uint16, a []float32)
func F32ToF8E4M3(dst []byte, a []float32)
func F32ToF8E5M2(dst []byte, a []float32)
func F32ToF16(dst []uint16, a []float32)
func FastCumProdFloat[T Float](dst, a []T)
func FastCumSumFloat[T Float](dst, a []T)
func Fill[T Number](dst []T, v T)
func FillBytes(dst []byte, v byte)
func FillFastFallbacks(s *kernel.Set)
func Floor[T Float](dst, a []T)
func FormatInts(dst []byte, vals []int64, sep byte) int
func Gather[T number](dst, src []T, idx []int32)
func GemmPackB[T Number](bp, b []T, k, n int)
func GemvFloat[T float](dst, a, x []T, m, k int)
func GemvInt[T integer](dst, a, x []T, m, k int)
func Grayscale(dst, r, g, b []byte)
func GreaterEqualMask[T number](dst []bool, a, b []T)
func GreaterEqualScalarMask[T number](dst []bool, a []T, v T)
func GreaterMask[T number](dst []bool, a, b []T)
func GreaterScalarMask[T number](dst []bool, a []T, v T)
func Hamming(a, b []byte) int
func HammingWords(a, b []uint64) int
func HexDecode(dst, src []byte) (int, bool)
func HexEncode(dst, src []byte) int
func Hypot[T float](dst, a, b []T)
func Index(haystack, needle []byte) int
func IndexAll(dst []int32, b []byte, c byte) int
func IndexAny(b, chars []byte) int
func IndexByte(b []byte, c byte) int
func IndexNonASCII(b []byte) int
func IndexNonASCII16(b []uint16) int
func IndexNotAny(b, chars []byte) int
func IntersectInt[T Integer](dst, a, b []T) int
func IsASCII(b []byte) bool
func L1DiffFloat[T Float](a, b []T) T
func L1DiffInt[T Integer](a, b []T) T
func L1NormFloat[T Float](a []T) T
func L1NormInt[T Integer](a []T) T
func LastIndex(haystack, needle []byte) int
func LastIndexByte(b []byte, c byte) int
func LastIndexNotAny(b, chars []byte) int
func LayerNorm[T float](dst, a, gamma, beta []T, shift, denom T)
func LeadingZeros[T Integer](dst, a []T)
func Lerp[T Number](dst, a, b []T, t T)
func LessEqualMask[T number](dst []bool, a, b []T)
func LessEqualScalarMask[T number](dst []bool, a []T, v T)
func LessMask[T number](dst []bool, a, b []T)
func LessScalarMask[T number](dst []bool, a []T, v T)
func Log[T float](dst, a []T)
func Log1p[T float](dst, a []T)
func Log2[T float](dst, a []T)
func Log10[T float](dst, a []T)
func LowerBoundFloat[T Float](dst []int32, a, q []T)
func LowerBoundInt[T Integer](dst []int32, a, q []T)
func MaskAll(m []bool) bool
func MaskAnd(dst, a, b []bool)
func MaskAny(m []bool) bool
func MaskCount(m []bool) int
func MaskNot(dst, a []bool)
func MaskOr(dst, a, b []bool)
func MaskXor(dst, a, b []bool)
func MatMul[T number](dst, a, b []T, m, k, n int)
func MatMulPk[T Number](dst, a, bp []T, m, k, n int)
func MaxReduceFloat[T Float](a []T) T
func MaxReduceInt[T Integer](a []T) T
func MaximumFloat[T Float](dst, a, b []T)
func MaximumInt[T Integer](dst, a, b []T)
func MinMaxFloat[T Float](a []T) (T, T)
func MinMaxInt[T Integer](a []T) (T, T)
func MinReduceFloat[T Float](a []T) T
func MinReduceInt[T Integer](a []T) T
func MinimumFloat[T Float](dst, a, b []T)
func MinimumInt[T Integer](dst, a, b []T)
func MovingAverage[T Number](dst, a []T, width int)
func Mul[T Number](dst, a, b []T)
func Mul3[T Number](dst, a, b, c []T)
func Mul4[T Number](dst, a, b, c, d []T)
func NarrowU16U8(dst []byte, s []uint16)
func NarrowU32U8(dst []byte, s []uint32)
func NegFloat[T Float](dst, a []T)
func NegInt[T Integer](dst, a []T)
func NormFloat[T Float](a []T) T
func NotEqualMask[T number](dst []bool, a, b []T)
func NotEqualScalarMask[T number](dst []bool, a []T, v T)
func OnesCount[T Integer](dst, a []T)
func ParseInts(dst []int64, src []byte, idx []int32) (int, bool)
func ParseUints(dst []uint64, src []byte, idx []int32) (int, bool)
func Partition[T Number](dst, src []T, pivot T) int
func PolyEval[T Number](dst, x, coeffs []T)
func PopCount(b []byte) int
func Pow[T float](dst, a, b []T)
func ProdInt[T Integer](a []T) T
func QMatMulI8(dst []int32, a, b []int8, m, k, n int)
func QuantizeI8(dst []int8, a []float32, scale float32, zeroPoint int32)
func QuantizePerChannelI8(dst []int8, a []float32, scale []float32, zeroPoint []int32, ...)
func QuantizePerChannelU8(dst []uint8, a []float32, scale []float32, zeroPoint []int32, ...)
func QuantizeU8(dst []uint8, a []float32, scale float32, zeroPoint int32)
func RGBToUV(u, v, r, g, b []byte)
func Ramp[T Number](dst []T, start, step T)
func RandomF32(dst []float32, seed uint64)
func RandomF64(dst []float64, seed uint64)
func RandomU64(dst []uint64, seed uint64)
func Reciprocal[T Float](dst, a []T)
func ReplaceByte(dst, b []byte, old, with byte)
func RequantizeI8(dst []int8, a []int32, scale float32, zeroPoint int32)
func Reverse[T Number](dst, a []T)
func ReverseBits[T Integer](dst, a []T)
func RollingMaxFloat[T Float](dst, a []T, window int)
func RollingMaxInt[T Integer](dst, a []T, window int)
func RollingMinFloat[T Float](dst, a []T, window int)
func RollingMinInt[T Integer](dst, a []T, window int)
func Rotl[T Integer](dst, a []T, s uint64)
func Rotr[T Integer](dst, a []T, s uint64)
func Round[T Float](dst, a []T)
func RoundToEven[T Float](dst, a []T)
func RunStartsI32(dst []bool, a []int32)
func RunStartsI64(dst []bool, a []int64)
func RunStartsU8(dst []bool, a []byte)
func SatAdd[T Saturating](dst, a, b []T)
func SatSub[T Saturating](dst, a, b []T)
func Scale[T Number](dst, a []T, s T)
func Scatter[T number](dst []T, idx []int32, src []T)
func Select[T number](dst []T, mask []bool, yes, no []T)
func Set() kernel.Set
func ShiftDiv[T float](dst, a []T, shift, denom T)
func Shl[T Integer](dst, a []T, s uint64)
func Shr[T Integer](dst, a []T, s uint64)
func Sigmoid[T float](dst, a []T)
func Sin[T float](dst, a []T)
func Sinh[T float](dst, a []T)
func SortOrdered[T ordered](a []T)
func SparseDotFloat[T Float](v []T, idx []int32, x []T) T
func Sqrt[T Float](dst, a []T)
func Sub[T Number](dst, a, b []T)
func SubScalar[T Number](dst, a []T, s T)
func SumFloat[T Float](a []T) T
func SumInt[T Integer](a []T) T
func SumLanesFloat[T float](dst, a []T)
func SumSqDevFloat[T Float](a []T, c T) T
func SumSqDevInt[T Integer](a []T, c T) T
func SumSqDiffFloat[T Float](a, b []T) T
func SumSqDiffInt[T Integer](a, b []T) T
func SumSquaresFloat[T Float](a []T) T
func SumSquaresInt[T Integer](a []T) T
func Tan[T float](dst, a []T)
func Tanh[T float](dst, a []T)
func Tile[T number](dst, pattern []T)
func ToLowerASCII(dst, b []byte)
func ToUpperASCII(dst, b []byte)
func TrailingZeros[T Integer](dst, a []T)
func Transpose[T Number](dst, a []T, m, n int)
func Trunc[T Float](dst, a []T)
func ValidUTF8(b []byte) bool
func VarintLenU32(dst []int32, a []uint32)
func VarintLenU64(dst []int32, a []uint64)
func VarintSizeU32(a []uint32) int
func VarintSizeU64(a []uint64) int
func WidenU8U16(dst []uint16, s []byte)
func WidenU8U32(dst []uint32, s []byte)
func ZigzagDecodeI8(dst []int8, a []byte)
func ZigzagDecodeI16(dst []int16, a []uint16)
func ZigzagDecodeI32(dst []int32, a []uint32)
func ZigzagDecodeI64(dst []int64, a []uint64)
func ZigzagEncodeI8(dst []byte, a []int8)
func ZigzagEncodeI16(dst []uint16, a []int16)
func ZigzagEncodeI32(dst []uint32, a []int32)
func ZigzagEncodeI64(dst []uint64, a []int64)
type Float
type Integer
type Number
type Saturating

Constants ¶

This section is empty.

Variables ¶

This section is empty.

Functions ¶

func AbsFloat ¶

func AbsFloat[T Float](dst, a []T)

func AbsInt ¶

func AbsInt[T Integer](dst, a []T)

func Acos ¶

func Acos[T float](dst, a []T)

func Acosh ¶

func Acosh[T float](dst, a []T)

func Add ¶

func Add[T Number](dst, a, b []T)

func Add3 ¶

func Add3[T Number](dst, a, b, c []T)

func Add4 ¶

func Add4[T Number](dst, a, b, c, d []T)

func AddScalar ¶

func AddScalar[T Number](dst, a []T, s T)

func AddScaled ¶

func AddScaled[T Number](dst, a, b []T, s T)

func ArgMaxFloat ¶

func ArgMaxFloat[T Float](a []T) int

func ArgMaxInt ¶

func ArgMaxInt[T Integer](a []T) int

func ArgMinFloat ¶

func ArgMinFloat[T Float](a []T) int

func ArgMinInt ¶

func ArgMinInt[T Integer](a []T) int

func Asin ¶

func Asin[T float](dst, a []T)

func Asinh ¶

func Asinh[T float](dst, a []T)

func Atan ¶

func Atan[T float](dst, a []T)

func Atan2 ¶

func Atan2[T float](dst, a, b []T)

func Atanh ¶

func Atanh[T float](dst, a []T)

func B64Decode ¶

func B64Decode(dst, src []byte) int

func B64Encode ¶

func B64Encode(dst, src []byte) int

func BF16ToF32 ¶

func BF16ToF32(dst []float32, a []uint16)

func BitAnd ¶

func BitAnd(dst, a, b []byte)

func BitAndNot ¶

func BitAndNot(dst, a, b []byte)

func BitNot ¶

func BitNot(dst, a []byte)

func BitOr ¶

func BitOr(dst, a, b []byte)

func BitPackU32 ¶

func BitPackU32(dst, a []uint32, bits int32)

func BitUnpackU32 ¶

func BitUnpackU32(dst, a []uint32, bits int32)

func BitXor ¶

func BitXor(dst, a, b []byte)

func ByteSwap ¶

func ByteSwap[T Integer](dst, a []T)

func CAbs64 ¶

func CAbs64(dst []float32, a []complex64)

func CAbs128 ¶

func CAbs128(dst []float64, a []complex128)

func CAdd ¶

func CAdd[C complexT](dst, a, b []C)

func CConj64 ¶

func CConj64(dst, a []complex64)

func CConj128 ¶

func CConj128(dst, a []complex128)

func CDiv64 ¶

func CDiv64(dst, a, b []complex64)

func CDiv128 ¶

func CDiv128(dst, a, b []complex128)

func CDot64 ¶

func CDot64(a, b []complex64) complex64

func CDot128 ¶

func CDot128(a, b []complex128) complex128

func CDotConj64 ¶

func CDotConj64(a, b []complex64) complex64

func CDotConj128 ¶

func CDotConj128(a, b []complex128) complex128

func CFromParts64 ¶

func CFromParts64(dst []complex64, re, im []float32)

func CFromParts128 ¶

func CFromParts128(dst []complex128, re, im []float64)

func CImag64 ¶

func CImag64(dst []float32, a []complex64)

func CImag128 ¶

func CImag128(dst []float64, a []complex128)

func CMul64 ¶

func CMul64(dst, a, b []complex64)

func CMul128 ¶

func CMul128(dst, a, b []complex128)

func CNeg ¶

func CNeg[C complexT](dst, a []C)

func CReal64 ¶

func CReal64(dst []float32, a []complex64)

func CReal128 ¶

func CReal128(dst []float64, a []complex128)

func CScale64 ¶

func CScale64(dst, a []complex64, s float32)

func CScale128 ¶

func CScale128(dst, a []complex128, s float64)

func CSub ¶

func CSub[C complexT](dst, a, b []C)

func CSum64 ¶

func CSum64(a []complex64) complex64

func CSum128 ¶

func CSum128(a []complex128) complex128

func Cbrt ¶

func Cbrt[T float](dst, a []T)

func Ceil ¶

func Ceil[T Float](dst, a []T)

func ClampFloat ¶

func ClampFloat[T Float](dst, a []T, lo, hi T)

func ClampInt ¶

func ClampInt[T Integer](dst, a []T, lo, hi T)

func CommonPrefix ¶

func CommonPrefix(a, b []byte) int

func CompareBytes ¶

func CompareBytes(a, b []byte) int

func Compress ¶

func Compress[T any](dst, src []T, keep []bool) int

func CompressFloat32 ¶

func CompressFloat32(dst, src []float32, keep []bool) int

func CompressFloat64 ¶

func CompressFloat64(dst, src []float64, keep []bool) int

func CompressInt32 ¶

func CompressInt32(dst, src []int32, keep []bool) int

func CompressInt64 ¶

func CompressInt64(dst, src []int64, keep []bool) int

func Convolve ¶

func Convolve[T Number](dst, sig, ker []T)

func Correlate ¶

func Correlate[T Number](dst, sig, ker []T)

func Cos ¶

func Cos[T float](dst, a []T)

func Cosh ¶

func Cosh[T float](dst, a []T)

func CountAny ¶

func CountAny(b, chars []byte) int

func CountByte ¶

func CountByte(b []byte, c byte) int

func CountSeq ¶

func CountSeq(haystack, needle []byte) int

func CumMaxFloat ¶

func CumMaxFloat[T Float](dst, a []T)

func CumMaxInt ¶

func CumMaxInt[T Integer](dst, a []T)

func CumMinFloat ¶

func CumMinFloat[T Float](dst, a []T)

func CumMinInt ¶

func CumMinInt[T Integer](dst, a []T)

func CumProdInt ¶

func CumProdInt[T Integer](dst, a []T)

CumProdInt is the EXACT integer product scan. Two's-complement multiplication is associative, so the blocked grouping is bit-identical to the serial loop for every input including ones that overflow — verified over four million deliberately overflowing values. It therefore needs no Fast prefix, and this is the reference the int32 kernel falls back to.

It is written as the blocked scan rather than as the obvious serial loop so that reference and kernel have the same shape, and a change to one is visibly a change to the other.

func DequantizeI8 ¶

func DequantizeI8(dst []float32, a []int8, scale float32, zeroPoint int32)

func DequantizePerChannelI8 ¶

func DequantizePerChannelI8(dst []float32, a []int8, scale []float32, zeroPoint []int32, channels, inner int)

func DequantizePerChannelU8 ¶

func DequantizePerChannelU8(dst []float32, a []uint8, scale []float32, zeroPoint []int32, channels, inner int)

func DequantizeU8 ¶

func DequantizeU8(dst []float32, a []uint8, scale float32, zeroPoint int32)

func Diff ¶

func Diff[T Number](dst, a []T)

func DifferenceInt ¶

func DifferenceInt[T Integer](dst, a, b []T) int

func Div ¶

func Div[T Float](dst, a, b []T)

func DivScalar ¶

func DivScalar[T Number](dst, a []T, s T)

func DotFloat ¶

func DotFloat[T Float](a, b []T) T

func DotInt ¶

func DotInt[T Integer](a, b []T) T

func EqualBytes ¶

func EqualBytes(a, b []byte) bool

func EqualFoldASCII ¶

func EqualFoldASCII(a, b []byte) bool

func EqualMask ¶

func EqualMask[T number](dst []bool, a, b []T)

func EqualScalarMask ¶

func EqualScalarMask[T number](dst []bool, a []T, v T)

func Erf ¶

func Erf[T float](dst, a []T)

func Erfc ¶

func Erfc[T float](dst, a []T)

func Exp ¶

func Exp[T float](dst, a []T)

func Exp2 ¶

func Exp2[T float](dst, a []T)

func Expand ¶

func Expand[T any](dst, src []T, keep []bool) int

func Expm1 ¶

func Expm1[T float](dst, a []T)

func F8E4M3ToF32 ¶

func F8E4M3ToF32(dst []float32, a []byte)

func F8E5M2ToF32 ¶

func F8E5M2ToF32(dst []float32, a []byte)

func F16ToF32 ¶

func F16ToF32(dst []float32, a []uint16)

func F32ToBF16 ¶

func F32ToBF16(dst []uint16, a []float32)

func F32ToF8E4M3 ¶

func F32ToF8E4M3(dst []byte, a []float32)

func F32ToF8E5M2 ¶

func F32ToF8E5M2(dst []byte, a []float32)

func F32ToF16 ¶

func F32ToF16(dst []uint16, a []float32)

func FastCumProdFloat ¶

func FastCumProdFloat[T Float](dst, a []T)

func FastCumSumFloat ¶

func FastCumSumFloat[T Float](dst, a []T)

FastCumSum and FastCumProd are the exported references the generated backends fall back to and the differential tests compare against.

FastCumSumFloat is the blocked scan for float32 and the SERIAL loop for float64, and that asymmetry is measured rather than arbitrary. Eight doubles fill one AVX-512 register, so the shift steps become cross-lane permutes, and against a serial chain of four-cycle adds the blocked form measured 0.91x — slower. float32 has sixteen lanes to hide the same latency and wins. Since no kernel is generated for float64, this branch is what every tier runs there, so they still agree with each other.

func Fill ¶

func Fill[T Number](dst []T, v T)

func FillBytes ¶

func FillBytes(dst []byte, v byte)

func FillFastFallbacks ¶

func FillFastFallbacks(s *kernel.Set)

FillFastFallbacks points any Fast slot with no generated kernel at the accurate one, for both float groups of a backend.

It runs after a backend is fully assembled, which is the only moment the two cases can be distinguished, and it is what lets a caller use FastExp on every architecture without asking whether that architecture has it. A target where the Fast tier did not measure faster simply computes a more accurate answer, which an upper bound on error permits.

func Floor ¶

func Floor[T Float](dst, a []T)

func FormatInts ¶

func FormatInts(dst []byte, vals []int64, sep byte) int

FormatInts is the reference formatter: exact-fit, so it succeeds wherever success is possible, which is what makes it the safe fallback for short destinations.

func Gather ¶

func Gather[T number](dst, src []T, idx []int32)

func GemmPackB ¶

func GemmPackB[T Number](bp, b []T, k, n int)

func GemvFloat ¶

func GemvFloat[T float](dst, a, x []T, m, k int)

func GemvInt ¶

func GemvInt[T integer](dst, a, x []T, m, k int)

func Grayscale ¶

func Grayscale(dst, r, g, b []byte)

func GreaterEqualMask ¶

func GreaterEqualMask[T number](dst []bool, a, b []T)

func GreaterEqualScalarMask ¶

func GreaterEqualScalarMask[T number](dst []bool, a []T, v T)

func GreaterMask ¶

func GreaterMask[T number](dst []bool, a, b []T)

func GreaterScalarMask ¶

func GreaterScalarMask[T number](dst []bool, a []T, v T)

func Hamming ¶

func Hamming(a, b []byte) int

func HammingWords ¶

func HammingWords(a, b []uint64) int

func HexDecode ¶

func HexDecode(dst, src []byte) (int, bool)

HexDecode is exported for the generated dispatch tables, which name the reference by its exported identifier. The two results are what kept this portable before the generator could return a pair.

func HexEncode ¶

func HexEncode(dst, src []byte) int

func Hypot ¶

func Hypot[T float](dst, a, b []T)

func Index ¶

func Index(haystack, needle []byte) int

func IndexAll ¶

func IndexAll(dst []int32, b []byte, c byte) int

func IndexAny ¶

func IndexAny(b, chars []byte) int

func IndexByte ¶

func IndexByte(b []byte, c byte) int

func IndexNonASCII ¶

func IndexNonASCII(b []byte) int

func IndexNonASCII16 ¶

func IndexNonASCII16(b []uint16) int

func IndexNotAny ¶

func IndexNotAny(b, chars []byte) int

func IntersectInt ¶

func IntersectInt[T Integer](dst, a, b []T) int

func IsASCII ¶

func IsASCII(b []byte) bool

func L1DiffFloat ¶

func L1DiffFloat[T Float](a, b []T) T

func L1DiffInt ¶

func L1DiffInt[T Integer](a, b []T) T

func L1NormFloat ¶

func L1NormFloat[T Float](a []T) T

func L1NormInt ¶

func L1NormInt[T Integer](a []T) T

func LastIndex ¶

func LastIndex(haystack, needle []byte) int

func LastIndexByte ¶

func LastIndexByte(b []byte, c byte) int

func LastIndexNotAny ¶

func LastIndexNotAny(b, chars []byte) int

func LayerNorm ¶

func LayerNorm[T float](dst, a, gamma, beta []T, shift, denom T)

func LeadingZeros ¶

func LeadingZeros[T Integer](dst, a []T)

func Lerp ¶

func Lerp[T Number](dst, a, b []T, t T)

func LessEqualMask ¶

func LessEqualMask[T number](dst []bool, a, b []T)

func LessEqualScalarMask ¶

func LessEqualScalarMask[T number](dst []bool, a []T, v T)

func LessMask ¶

func LessMask[T number](dst []bool, a, b []T)

func LessScalarMask ¶

func LessScalarMask[T number](dst []bool, a []T, v T)

func Log ¶

func Log[T float](dst, a []T)

func Log1p ¶

func Log1p[T float](dst, a []T)

func Log2 ¶

func Log2[T float](dst, a []T)

func Log10 ¶

func Log10[T float](dst, a []T)

func LowerBoundFloat ¶

func LowerBoundFloat[T Float](dst []int32, a, q []T)

func LowerBoundInt ¶

func LowerBoundInt[T Integer](dst []int32, a, q []T)

func MaskAll ¶

func MaskAll(m []bool) bool

func MaskAnd ¶

func MaskAnd(dst, a, b []bool)

func MaskAny ¶

func MaskAny(m []bool) bool

func MaskCount ¶

func MaskCount(m []bool) int

func MaskNot ¶

func MaskNot(dst, a []bool)

func MaskOr ¶

func MaskOr(dst, a, b []bool)

func MaskXor ¶

func MaskXor(dst, a, b []bool)

func MatMul ¶

func MatMul[T number](dst, a, b []T, m, k, n int)

func MatMulPk ¶

func MatMulPk[T Number](dst, a, bp []T, m, k, n int)

func MaxReduceFloat ¶

func MaxReduceFloat[T Float](a []T) T

func MaxReduceInt ¶

func MaxReduceInt[T Integer](a []T) T

func MaximumFloat ¶

func MaximumFloat[T Float](dst, a, b []T)

func MaximumInt ¶

func MaximumInt[T Integer](dst, a, b []T)

func MinMaxFloat ¶

func MinMaxFloat[T Float](a []T) (T, T)

func MinMaxInt ¶

func MinMaxInt[T Integer](a []T) (T, T)

func MinReduceFloat ¶

func MinReduceFloat[T Float](a []T) T

func MinReduceInt ¶

func MinReduceInt[T Integer](a []T) T

func MinimumFloat ¶

func MinimumFloat[T Float](dst, a, b []T)

func MinimumInt ¶

func MinimumInt[T Integer](dst, a, b []T)

func MovingAverage ¶

func MovingAverage[T Number](dst, a []T, width int)

func Mul ¶

func Mul[T Number](dst, a, b []T)

func Mul3 ¶

func Mul3[T Number](dst, a, b, c []T)

func Mul4 ¶

func Mul4[T Number](dst, a, b, c, d []T)

func NarrowU16U8 ¶

func NarrowU16U8(dst []byte, s []uint16)

func NarrowU32U8 ¶

func NarrowU32U8(dst []byte, s []uint32)

func NegFloat ¶

func NegFloat[T Float](dst, a []T)

func NegInt ¶

func NegInt[T Integer](dst, a []T)

func NormFloat ¶

func NormFloat[T Float](a []T) T

func NotEqualMask ¶

func NotEqualMask[T number](dst []bool, a, b []T)

func NotEqualScalarMask ¶

func NotEqualScalarMask[T number](dst []bool, a []T, v T)

func OnesCount ¶

func OnesCount[T Integer](dst, a []T)

func ParseInts ¶

func ParseInts(dst []int64, src []byte, idx []int32) (int, bool)

ParseInts is the reference for the integer field parser. Fields are src[start:idx[k]] with start one past the previous separator, which is the shape IndexAll produces.

func ParseUints ¶

func ParseUints(dst []uint64, src []byte, idx []int32) (int, bool)

ParseUints is ParseInts over the full uint64 range and with no sign.

A leading '+' is rejected rather than skipped, matching strconv.ParseUint with bitSize 64, which accepts no sign at all.

func Partition ¶

func Partition[T Number](dst, src []T, pivot T) int

Partition is the exported entry point the generated guards call.

func PolyEval ¶

func PolyEval[T Number](dst, x, coeffs []T)

func PopCount ¶

func PopCount(b []byte) int

func Pow ¶

func Pow[T float](dst, a, b []T)

func ProdInt ¶

func ProdInt[T Integer](a []T) T

func QMatMulI8 ¶

func QMatMulI8(dst []int32, a, b []int8, m, k, n int)

func QuantizeI8 ¶

func QuantizeI8(dst []int8, a []float32, scale float32, zeroPoint int32)

func QuantizePerChannelI8 ¶

func QuantizePerChannelI8(dst []int8, a []float32, scale []float32, zeroPoint []int32, channels, inner int)

func QuantizePerChannelU8 ¶

func QuantizePerChannelU8(dst []uint8, a []float32, scale []float32, zeroPoint []int32, channels, inner int)

func QuantizeU8 ¶

func QuantizeU8(dst []uint8, a []float32, scale float32, zeroPoint int32)

func RGBToUV ¶

func RGBToUV(u, v, r, g, b []byte)

func Ramp ¶

func Ramp[T Number](dst []T, start, step T)

func RandomF32 ¶

func RandomF32(dst []float32, seed uint64)

func RandomF64 ¶

func RandomF64(dst []float64, seed uint64)

RandomF64 takes the top 53 bits and scales by 2^-53, the construction that cannot produce 1.0 — the largest result is 1 - 2^-53.

func RandomU64 ¶

func RandomU64(dst []uint64, seed uint64)

func Reciprocal ¶

func Reciprocal[T Float](dst, a []T)

func ReplaceByte ¶

func ReplaceByte(dst, b []byte, old, with byte)

func RequantizeI8 ¶

func RequantizeI8(dst []int8, a []int32, scale float32, zeroPoint int32)

func Reverse ¶

func Reverse[T Number](dst, a []T)

func ReverseBits ¶

func ReverseBits[T Integer](dst, a []T)

func RollingMaxFloat ¶

func RollingMaxFloat[T Float](dst, a []T, window int)

func RollingMaxInt ¶

func RollingMaxInt[T Integer](dst, a []T, window int)

func RollingMinFloat ¶

func RollingMinFloat[T Float](dst, a []T, window int)

func RollingMinInt ¶

func RollingMinInt[T Integer](dst, a []T, window int)

func Rotl ¶

func Rotl[T Integer](dst, a []T, s uint64)

func Rotr ¶

func Rotr[T Integer](dst, a []T, s uint64)

func Round ¶

func Round[T Float](dst, a []T)

func RoundToEven ¶

func RoundToEven[T Float](dst, a []T)

func RunStartsI32 ¶

func RunStartsI32(dst []bool, a []int32)

func RunStartsI64 ¶

func RunStartsI64(dst []bool, a []int64)

func RunStartsU8 ¶

func RunStartsU8(dst []bool, a []byte)

func SatAdd ¶

func SatAdd[T Saturating](dst, a, b []T)

func SatSub ¶

func SatSub[T Saturating](dst, a, b []T)

func Scale ¶

func Scale[T Number](dst, a []T, s T)

func Scatter ¶

func Scatter[T number](dst []T, idx []int32, src []T)

func Select ¶

func Select[T number](dst []T, mask []bool, yes, no []T)

func Set ¶

func Set() kernel.Set

Set returns the reference backend: every kernel, portable Go.

func ShiftDiv ¶

func ShiftDiv[T float](dst, a []T, shift, denom T)

func Shl ¶

func Shl[T Integer](dst, a []T, s uint64)

func Shr ¶

func Shr[T Integer](dst, a []T, s uint64)

func Sigmoid ¶

func Sigmoid[T float](dst, a []T)

func Sin ¶

func Sin[T float](dst, a []T)

func Sinh ¶

func Sinh[T float](dst, a []T)

func SortOrdered ¶

func SortOrdered[T ordered](a []T)

SortOrdered sorts in place with the standard library's pdqsort.

It is here as the reference and as the small-case path, and on integers it is also the shipped implementation: see the note on the exported Sort.

func SparseDotFloat ¶

func SparseDotFloat[T Float](v []T, idx []int32, x []T) T

func Sqrt ¶

func Sqrt[T Float](dst, a []T)

func Sub ¶

func Sub[T Number](dst, a, b []T)

func SubScalar ¶

func SubScalar[T Number](dst, a []T, s T)

func SumFloat ¶

func SumFloat[T Float](a []T) T

func SumInt ¶

func SumInt[T Integer](a []T) T

func SumLanesFloat ¶

func SumLanesFloat[T float](dst, a []T)

func SumSqDevFloat ¶

func SumSqDevFloat[T Float](a []T, c T) T

func SumSqDevInt ¶

func SumSqDevInt[T Integer](a []T, c T) T

func SumSqDiffFloat ¶

func SumSqDiffFloat[T Float](a, b []T) T

func SumSqDiffInt ¶

func SumSqDiffInt[T Integer](a, b []T) T

func SumSquaresFloat ¶

func SumSquaresFloat[T Float](a []T) T

func SumSquaresInt ¶

func SumSquaresInt[T Integer](a []T) T

func Tan ¶

func Tan[T float](dst, a []T)

func Tanh ¶

func Tanh[T float](dst, a []T)

func Tile ¶

func Tile[T number](dst, pattern []T)

func ToLowerASCII ¶

func ToLowerASCII(dst, b []byte)

func ToUpperASCII ¶

func ToUpperASCII(dst, b []byte)

func TrailingZeros ¶

func TrailingZeros[T Integer](dst, a []T)

func Transpose ¶

func Transpose[T Number](dst, a []T, m, n int)

func Trunc ¶

func Trunc[T Float](dst, a []T)

func ValidUTF8 ¶

func ValidUTF8(b []byte) bool

func VarintLenU32 ¶

func VarintLenU32(dst []int32, a []uint32)

func VarintLenU64 ¶

func VarintLenU64(dst []int32, a []uint64)

func VarintSizeU32 ¶

func VarintSizeU32(a []uint32) int

func VarintSizeU64 ¶

func VarintSizeU64(a []uint64) int

func WidenU8U16 ¶

func WidenU8U16(dst []uint16, s []byte)

func WidenU8U32 ¶

func WidenU8U32(dst []uint32, s []byte)

func ZigzagDecodeI8 ¶

func ZigzagDecodeI8(dst []int8, a []byte)

func ZigzagDecodeI16 ¶

func ZigzagDecodeI16(dst []int16, a []uint16)

func ZigzagDecodeI32 ¶

func ZigzagDecodeI32(dst []int32, a []uint32)

func ZigzagDecodeI64 ¶

func ZigzagDecodeI64(dst []int64, a []uint64)

func ZigzagEncodeI8 ¶

func ZigzagEncodeI8(dst []byte, a []int8)

func ZigzagEncodeI16 ¶

func ZigzagEncodeI16(dst []uint16, a []int16)

func ZigzagEncodeI32 ¶

func ZigzagEncodeI32(dst []uint32, a []int32)

func ZigzagEncodeI64 ¶

func ZigzagEncodeI64(dst []uint64, a []int64)

Types ¶

type Float ¶

type Float interface{ ~float32 | ~float64 }

Float is the element types with a fixed-tree reduction.

type Integer ¶

type Integer = integer

Integer is the integer half of it, for the operations that are not the same on floats: integer minimum is not IEEE minimum, and integer Abs wraps where float Abs clears a sign bit.

type Number ¶

type Number interface {
	~float32 | ~float64 |
		~int8 | ~int16 | ~int32 | ~int64 |
		~uint8 | ~uint16 | ~uint32 | ~uint64
}

Number is any element type the kernels handle.

type Saturating ¶

type Saturating = satInteger

Saturating is the integer types that have saturating add and subtract. The 64-bit ones are absent for the reason kernels.saturating gives: there is nothing wider to detect the overflow in.

Source Files ¶

View all Source files

?	: This menu
/	: Search site
f or F	: Jump to
y or Y	: Canonical URL