remove NV

Author: AntonOresten

Project.toml

@@ -1,7 +1,7 @@
 name = "Microfloats"
 uuid = "31c70f10-a750-4521-b13c-797315ae2933"
 authors = ["Anton Oresten <[email protected]> and contributors"]
-version = "0.0.4+1"
+version = "0.0.5"
 
 [deps]
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

README.md

@@ -23,20 +23,11 @@ const Float8_4 = Microfloat(1, 4, 3)
 const Float8_5 = Microfloat(1, 5, 2)
 ```
 
-### MX format
+### Microscaling (MX) Element Types
 
-Microfloats additionally implements the `E4M3`, `E5M2`, `E2M3`, `E3M2`, `E2M1`, and `E8M0` formats from the [Open Compute Project Microscaling Formats (MX) Specification](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf), with most of these using saturated arithmetic (no infinities), and different bit layouts for NaNs. These can be constructed by passing an additional `:MX` argument to the `Microfloat` constructor:
+Microfloats implements the E4M3, E5M2, E2M3, E3M2, E2M1, and E8M0 types from the [Open Compute Project Microscaling Formats (MX) Specification](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf), with most of these using saturated arithmetic (no infinities), and different bit layouts for NaNs. These are exported as `MX_E4M3`, `MX_E5M2`, `MX_E2M3`, `MX_E3M2`, `MX_E2M1`, and `MX_E8M0`, respectively.
 
-```julia
-const E4M3 = Microfloat(1, 4, 3, :MX)
-const E5M2 = Microfloat(1, 5, 2, :MX)
-const E2M3 = Microfloat(1, 2, 3, :MX)
-const E3M2 = Microfloat(1, 3, 2, :MX)
-const E2M1 = Microfloat(1, 2, 1, :MX)
-const E8M0 = Microfloat(0, 8, 0, :MX)
-```
-
-For `INT8`, see `FixedPointNumbers.Q1f6`.
+For INT8, see `FixedPointNumbers.Q1f6`.
 
 ## Installation
 

src/Microfloats.jl

@@ -4,12 +4,15 @@ include("float-bits.jl")
 
 include("Microfloat.jl")
 export Microfloat
-export IEEE
-export Float8_E3M4, Float8_E4M3, Float8_E5M2, Float6_E2M3, Float6_E3M2, Float4_E2M1
+export Float8_E3M4, Float8_E4M3, Float8_E5M2
+export Float6_E2M3, Float6_E3M2
+export Float4_E2M1
 
 include("microscaled/microscaled.jl")
-export MX, NV
-export MX_E5M2, MX_E4M3, MX_E3M2, MX_E2M3, MX_E2M1, MX_E8M0, NV_E2M1
+export MX_E5M2, MX_E4M3
+export MX_E3M2, MX_E2M3
+export MX_E2M1
+export MX_E8M0
 
 include("conversion/conversion.jl")
 

src/microscaled/microscaled.jl

@@ -1,22 +1,17 @@
 abstract type MX <: Variant end
-abstract type NV <: Variant end
-const Microscaled = Union{MX, NV}
 
-const MXMicrofloat{S,E,M} = Microfloat{S,E,M,MX}
-const NVMicrofloat{S,E,M} = Microfloat{S,E,M,NV}
-const MicroscaledMicrofloat{S,E,M} = Microfloat{S,E,M,<:Microscaled}
+const MX_Microfloat{S,E,M} = Microfloat{S,E,M,MX}
 
-const MX_E5M2 = MXMicrofloat{1,5,2}
-const MX_E4M3 = MXMicrofloat{1,4,3}
-const MX_E3M2 = MXMicrofloat{1,3,2}
-const MX_E2M3 = MXMicrofloat{1,2,3}
-const MX_E2M1 = MXMicrofloat{1,2,1}
-const MX_E8M0 = MXMicrofloat{0,8,0}
-const NV_E2M1 = NVMicrofloat{1,2,1}
+const MX_E5M2 = MX_Microfloat{1,5,2}
+const MX_E4M3 = MX_Microfloat{1,4,3}
+const MX_E3M2 = MX_Microfloat{1,3,2}
+const MX_E2M3 = MX_Microfloat{1,2,3}
+const MX_E2M1 = MX_Microfloat{1,2,1}
+const MX_E8M0 = MX_Microfloat{0,8,0}
 
-const NO_INF = Union{MX_E4M3, MX_E3M2, MX_E2M3, MX_E2M1, MX_E8M0, NV_E2M1}
-const NO_NAN = Union{MX_E3M2, MX_E2M3, MX_E2M1, NV_E2M1}
-const NO_NAN_OR_INF = Union{MX_E3M2, MX_E2M3, MX_E2M1, NV_E2M1}
+const NO_INF = Union{MX_E4M3, MX_E3M2, MX_E2M3, MX_E2M1, MX_E8M0}
+const NO_NAN = Union{MX_E3M2, MX_E2M3, MX_E2M1}
+const NO_NAN_OR_INF = Union{MX_E3M2, MX_E2M3, MX_E2M1}
 
 Base.isinf(::NO_INF) = false
 Base.isnan(::NO_NAN) = false
@@ -38,7 +33,7 @@ nan(::Type{MX_E8M0}) = reinterpret(MX_E8M0, 0xff)
 # Float32 conversion for MX variants:
 # - exp=all-ones is "normal" except for the MX NaN sentinel(s)
 # - otherwise identical mapping as IEEE
-function _float32(x::T) where {T<:MicroscaledMicrofloat}
+function _float32(x::T) where {T<:MX_Microfloat}
     T isa MX_E8M0 && reinterpret(UInt8, x) == 0xff && return NaN32
 
     sgn = UInt32(right_aligned_sign(x))
@@ -78,7 +73,7 @@ function _float32(x::T) where {T<:MicroscaledMicrofloat}
 end
 
 # Saturating to_microfloat tables for MX (no Infs; overflow -> ±floatmax)
-function create_base_shifttable(::Type{T}) where {T<:MicroscaledMicrofloat}
+function create_base_shifttable(::Type{T}) where {T<:MX_Microfloat}
     basetable = Vector{T}(undef, 512)
     shifttable = Vector{UInt8}(undef, 512)
 
@@ -115,5 +110,5 @@ function create_base_shifttable(::Type{T}) where {T<:MicroscaledMicrofloat}
 end
 
 # Saturating bounds for MX: use finite extrema
-Base.typemax(::Type{T}) where {S,E,M,T<:MicroscaledMicrofloat{S,E,M}} = floatmax(T)
-Base.typemin(::Type{T}) where {S,E,M,T<:MicroscaledMicrofloat{S,E,M}} = ifelse(n_sign_bits(T) == 0, zero(T), -floatmax(T))
+Base.typemax(::Type{T}) where {S,E,M,T<:MX_Microfloat{S,E,M}} = floatmax(T)
+Base.typemin(::Type{T}) where {S,E,M,T<:MX_Microfloat{S,E,M}} = ifelse(n_sign_bits(T) == 0, zero(T), -floatmax(T))