native_qr

Author: Jutho

src/MatrixAlgebraKit.jl

@@ -72,6 +72,7 @@ export notrunc, truncrank, trunctol, truncerror, truncfilter
 end
 
 include("common/defaults.jl")
+include("common/householder.jl")
 include("common/initialization.jl")
 include("common/pullbacks.jl")
 include("common/safemethods.jl")

src/common/householder.jl

@@ -0,0 +1,142 @@
+const IndexRange{T <: Integer} = Base.AbstractRange{T}
+
+# Elementary Householder reflection
+struct Householder{T, V <: AbstractVector, R <: IndexRange}
+    β::T
+    v::V
+    r::R
+end
+Base.adjoint(H::Householder) = Householder(conj(H.β), H.v, H.r)
+
+function householder(x::AbstractVector, r::IndexRange = axes(x, 1), k = first(r))
+    i = findfirst(equalto(k), r)
+    i == nothing && error("k = $k should be in the range r = $r")
+    β, v, ν = _householder!(x[r], i)
+    return Householder(β, v, r), ν
+end
+# Householder reflector h that zeros the elements A[r,col] (except for A[k,col]) upon lmul!(h,A)
+function householder(A::AbstractMatrix, r::IndexRange, col::Int, k = first(r))
+    i = findfirst(equalto(k), r)
+    i == nothing && error("k = $k should be in the range r = $r")
+    β, v, ν = _householder!(A[r, col], i)
+    return Householder(β, v, r), ν
+end
+# Householder reflector that zeros the elements A[row,r] (except for A[row,k]) upon rmul!(A,h')
+function householder(A::AbstractMatrix, row::Int, r::IndexRange, k = first(r))
+    i = findfirst(equalto(k), r)
+    i == nothing && error("k = $k should be in the range r = $r")
+    β, v, ν = _householder!(conj!(A[row, r]), i)
+    return Householder(β, v, r), ν
+end
+
+# generate Householder vector based on vector v, such that applying the reflection
+# to v yields a vector with single non-zero element on position i, whose value is
+# positive and thus equal to norm(v)
+function _householder!(v::AbstractVector{T}, i::Int = 1) where {T}
+    β::T = zero(T)
+    @inbounds begin
+        σ = abs2(zero(T))
+        @simd for k in 1:(i - 1)
+            σ += abs2(v[k])
+        end
+        @simd for k in (i + 1):length(v)
+            σ += abs2(v[k])
+        end
+        vi = v[i]
+        ν = sqrt(abs2(vi) + σ)
+
+        if σ == 0 && vi == ν
+            β = zero(vi)
+        else
+            if real(vi) < 0
+                vi = vi - ν
+            else
+                vi = ((vi - conj(vi)) * ν - σ) / (conj(vi) + ν)
+            end
+            @simd for k in 1:(i - 1)
+                v[k] /= vi
+            end
+            v[i] = 1
+            @simd for k in (i + 1):length(v)
+                v[k] /= vi
+            end
+            β = -conj(vi) / (ν)
+        end
+    end
+    return β, v, ν
+end
+
+function LinearAlgebra.lmul!(H::Householder, x::AbstractVector)
+    v = H.v
+    r = H.r
+    β = H.β
+    β == 0 && return x
+    @inbounds begin
+        μ = conj(zero(v[1])) * zero(x[r[1]])
+        i = 1
+        @simd for j in r
+            μ += conj(v[i]) * x[j]
+            i += 1
+        end
+        μ *= β
+        i = 1
+        @simd for j in H.r
+            x[j] -= μ * v[i]
+            i += 1
+        end
+    end
+    return x
+end
+function LinearAlgebra.lmul!(H::Householder, A::AbstractMatrix; cols = axes(A, 2))
+    v = H.v
+    r = H.r
+    β = H.β
+    β == 0 && return A
+    @inbounds begin
+        for k in cols
+            μ = conj(zero(v[1])) * zero(A[r[1], k])
+            i = 1
+            @simd for j in r
+                μ += conj(v[i]) * A[j, k]
+                i += 1
+            end
+            μ *= β
+            i = 1
+            @simd for j in H.r
+                A[j, k] -= μ * v[i]
+                i += 1
+            end
+        end
+    end
+    return A
+end
+function LinearAlgebra.rmul!(A::AbstractMatrix, H::Householder; rows = axes(A, 1))
+    v = H.v
+    r = H.r
+    β = H.β
+    β == 0 && return A
+    w = similar(A, length(rows))
+    fill!(w, 0)
+    all(in(axes(A, 2)), r) || error("Householder range r = $r not compatible with matrix A of size $(size(A))")
+    @inbounds begin
+        l = 1
+        for k in r
+            j = 1
+            @simd for i in rows
+                w[j] += A[i, k] * v[l]
+                j += 1
+            end
+            l += 1
+        end
+        l = 1
+        for k in r
+            j = 1
+            @simd for i in rows
+                A[i, k] -= β * w[j] * conj(v[l])
+                j += 1
+            end
+            l += 1
+        end
+    end
+    return A
+end

src/implementations/qr.jl

@@ -233,10 +233,9 @@ end
 
 _diagonal_qr_null!(A::AbstractMatrix, N; positive::Bool = false) = N
 
-### GPU logic
-# placed here to avoid code duplication since much of the logic is replicable across
-# CUDA and AMDGPU
-###
+# GPU logic
+# --------------
+# placed here to avoid code duplication since much of the logic is replicable across CUDA and AMDGPU
 function MatrixAlgebraKit.qr_full!(
         A::AbstractMatrix, QR, alg::Union{CUSOLVER_HouseholderQR, ROCSOLVER_HouseholderQR}
     )
@@ -325,3 +324,83 @@ function _gpu_qr_null!(
     N = _gpu_unmqr!('L', 'N', A, τ, N)
     return N
 end
+
+# Native logic
+# --------------
+function qr_full!(A::AbstractMatrix, QR, alg::Native_HouseholderQR)
+    check_input(qr_full!, A, QR, alg)
+    Q, R = QR
+    A === Q &&
+        throw(ArgumentError("inplace Q not supported with native QR implementation"))
+    _native_qr!(A, Q, R; alg.kwargs...)
+    return Q, R
+end
+function qr_compact!(A::AbstractMatrix, QR, alg::Native_HouseholderQR)
+    check_input(qr_compact!, A, QR, alg)
+    Q, R = QR
+    A === Q &&
+        throw(ArgumentError("inplace Q not supported with native QR implementation"))
+    _native_qr!(A, Q, R; alg.kwargs...)
+    return Q, R
+end
+function qr_null!(A::AbstractMatrix, N, alg::Native_HouseholderQR)
+    check_input(qr_null!, A, N, alg)
+    _native_qr_null!(A, N; alg.kwargs...)
+    return N
+end
+
+function _native_qr!(
+        A::AbstractMatrix, Q::AbstractMatrix, R::AbstractMatrix
+    )
+    m, n = size(A)
+    minmn = min(m, n)
+    @inbounds for j in 1:minmn
+        for i in 1:(j - 1)
+            R[i, j] = A[i, j]
+        end
+        β, v, R[j, j] = _householder!(view(A, j:m, j), 1)
+        for i in (j + 1):size(R, 1)
+            R[i, j] = 0
+        end
+        H = Householder(β, v, j:m)
+        lmul!(H, A; cols = (j + 1):n)
+        # A[j,j] == 1; store β instead
+        A[j, j] = β
+    end
+    @inbounds for j in (minmn + 1):n
+        for i in 1:size(R, 1)
+            R[i, j] = A[i, j]
+        end
+    end
+    # build Q
+    one!(Q)
+    for j in minmn:-1:1
+        β = A[j, j]
+        A[j, j] = 1
+        Hᴴ = Householder(conj(β), view(A, j:m, j), j:m)
+        lmul!(Hᴴ, Q)
+    end
+    return Q, R
+end
+
+function _native_qr_null!(A::AbstractMatrix, N::AbstractMatrix)
+    m, n = size(A)
+    minmn = min(m, n)
+    @inbounds for j in 1:minmn
+        β, v, ν = _householder!(view(A, j:m, j), 1)
+        H = Householder(β, v, j:m)
+        lmul!(H, A; cols = (j + 1):n)
+        # A[j,j] == 1; store β instead
+        A[j, j] = β
+    end
+    # build Q
+    fill!(N, zero(eltype(N)))
+    one!(view(N, (minmn + 1):m, 1:(m - minmn)))
+    for j in minmn:-1:1
+        β = A[j, j]
+        A[j, j] = 1
+        Hᴴ = Householder(conj(β), view(A, j:m, j), j:m)
+        lmul!(Hᴴ, N)
+    end
+    return N
+end

src/interface/decompositions.jl

@@ -9,6 +9,15 @@
 
 # QR, LQ, QL, RQ Decomposition
 # ----------------------------
+"""
+    Native_HouseholderQR(; blocksize, positive = false, pivoted = false)
+
+Algorithm type to denote a native implementation for computing the QR decomposition of
+a matrix using Householder reflectors, .The diagonal elements of `R` will be non-negative
+by construction.
+"""
+@algdef Native_HouseholderQR
+
 """
     LAPACK_HouseholderQR(; blocksize, positive = false, pivoted = false)