Skip to content

Latest commit

 

History

History
283 lines (227 loc) · 6.7 KB

File metadata and controls

283 lines (227 loc) · 6.7 KB

Warning

SpinalHDL fixed-point support is only partially used/tested, if you find any bugs with it, or you think that some functionality is missing, please create a Github issue. Also, please do not use undocumented features in your code.

UFix/SFix

Description

The UFix and SFix types correspond to a vector of bits that can be used for fixed-point arithmetic.

Declaration

The syntax to declare a fixed-point number is as follows:

Unsigned Fixed-Point

Syntax bit width resolution max min
UFix(peak: ExpNumber, resolution: ExpNumber) peak-resolution 2^resolution 2^peak-2^resolution 0
UFix(peak: ExpNumber, width: BitCount) width 2^(peak-width) 2^peak-2^(peak-width) 0

Signed Fixed-Point

Syntax bit width resolution max min
SFix(peak: ExpNumber, resolution: ExpNumber) peak-resolution+1 2^resolution 2^peak-2^resolution -(2^peak)
SFix(peak: ExpNumber, width: BitCount) width 2^(peak-width-1) 2^peak-2^(peak-width-1) -(2^peak)

Format

The chosen format follows the usual way of defining fixed-point number format using Q notation. More information can be found on the Wikipedia page about the Q number format.

For example Q8.2 will mean a fixed-point number of 8+2 bits, where 8 bits are used for the natural part and 2 bits for the fractional part. If the fixed-point number is signed, one more bit is used for the sign.

The resolution is defined as being the smallest power of two that can be represented in this number.

Note

To make representing power-of-two numbers less error prone, there is a numeric type in spinal.core called ExpNumber, which is used for the fixed-point type constructors. A convenience wrapper exists for this type, in the form of the exp function (used in the code samples on this page).

Examples

// Unsigned Fixed-Point
val UQ_8_2 = UFix(peak = 8 exp, resolution = -2 exp) // bit width = 8 - (-2) = 10 bits
val UQ_8_2 = UFix(8 exp, -2 exp)

val UQ_8_2 = UFix(peak = 8 exp, width = 10 bits)
val UQ_8_2 = UFix(8 exp, 10 bits)

// Signed Fixed-Point
val Q_8_2 = SFix(peak = 8 exp, resolution = -2 exp) // bit width = 8 - (-2) + 1 = 11 bits
val Q_8_2 = SFix(8 exp, -2 exp)

val Q_8_2 = SFix(peak = 8 exp, width = 11 bits)
val Q_8_2 = SFix(8 exp, 11 bits)

Assignments

Valid Assignments

An assignment to a fixed-point value is valid when there is no bit loss. Any bit loss will result in an error.

If the source fixed-point value is too big, the truncated function will allow you to resize the source number to match the destination size.

Example

val i16_m2 = SFix(16 exp, -2 exp)
val i16_0  = SFix(16 exp,  0 exp)
val i8_m2  = SFix( 8 exp, -2 exp)
val o16_m2 = SFix(16 exp, -2 exp)
val o16_m0 = SFix(16 exp,  0 exp)
val o14_m2 = SFix(14 exp, -2 exp)

o16_m2 := i16_m2            // OK
o16_m0 := i16_m2            // Not OK, Bit loss
o14_m2 := i16_m2            // Not OK, Bit loss
o16_m0 := i16_m2.truncated  // OK, as it is resized to match assignment target
o14_m2 := i16_m2.truncated  // OK, as it is resized to match assignment target
val o18_m2 = i16_m2.truncated(18 exp, -2 exp)
val o18_22b = i16_m2.truncated(18 exp, 22 bit)

From a Scala constant

Scala BigInt or Double types can be used as constants when assigning to UFix or SFix signals.

Example

val i4_m2 = SFix(4 exp, -2 exp)
i4_m2 := 1.25    // Will load 5 in i4_m2.raw
i4_m2 := 4       // Will load 16 in i4_m2.raw

Raw value

The integer representation of the fixed-point number can be read or written by using the raw property.

Example

val UQ_8_2 = UFix(8 exp, 10 bits)
UQ_8_2.raw := 4        // Assign the value corresponding to 1.0
UQ_8_2.raw := U(17)    // Assign the value corresponding to 4.25

Operators

The following operators are available for the UFix type:

Arithmetic

Operator Description Returned resolution Returned amplitude
x + y Addition Min(x.resolution, y.resolution) Max(x.amplitude, y.amplitude)
x - y Subtraction Min(x.resolution, y.resolution) Max(x.amplitude, y.amplitude)
x * y Multiplication x.resolution * y.resolution) x.amplitude * y.amplitude
x >> y Arithmetic shift right, y : Int x.amplitude >> y x.resolution >> y
x << y Arithmetic shift left, y : Int x.amplitude << y x.resolution << y
x >>| y Arithmetic shift right, y : Int x.amplitude >> y x.resolution
x <<| y Arithmetic shift left, y : Int x.amplitude << y x.resolution

Comparison

Operator Description Return type
x === y Equality Bool
x =/= y Inequality Bool
x > y Greater than Bool
x >= y Greater than or equal Bool
x < y Less than Bool
x >= y Less than or equal Bool

Type cast

Operator Description Return
x.asBits Binary cast to Bits Bits(w(x) bits)
x.asUInt Binary cast to UInt UInt(w(x) bits)
x.asSInt Binary cast to SInt SInt(w(x) bits)
x.asBools Cast into a array of Bool Vec(Bool(),width(x))
x.toUInt Return the corresponding UInt (with truncation) UInt
x.toSInt Return the corresponding SInt (with truncation) SInt
x.toUFix Return the corresponding UFix UFix
x.toSFix Return the corresponding SFix SFix

Misc

Name Return Description
x.maxValue Return the maximum value storable Double
x.minValue Return the minimum value storable Double
x.resolution x.amplitude * y.amplitude Double