DArray/QR: Pick better ib and p defaults

docs/src/darray.md

@@ -705,7 +705,7 @@ From `LinearAlgebra`:
 - `mul!` (In-place Matrix-Matrix and Matrix-Vector multiply)
 - `cholesky`/`cholesky!` (In-place/Out-of-place Cholesky factorization)
 - `lu`/`lu!` (In-place/Out-of-place LU factorization (`NoPivot` and `RowMaximum`))
-- `qr`/`qr!` (In-place/Out-of-place QR factorization)
+- `qr`/`qr!` (In-place/Out-of-place tiled Compact-WY QR; keywords `ib` and `p` for inner block size and CAQR domains)
 - `\`/`ldiv!` (In-place/Out-of-place Linear solving with LU, Cholesky, QR, and SVD factorizations)
 - `inv` (Out-of-place matrix inversion, including via SVD)
 - `svd`/`svd!`/`svdvals!` (In-place/Out-of-place Singular Value Decomposition)

src/array/qr.jl

@@ -774,14 +774,38 @@ function qr_measure_workspace(A::DMatrix{T}, ib::Int) where {T<: Number}
     lm, ln
 end
 
-function _qr_impl!(A::DMatrix{T}; ib::Int=1, p::Int=1) where {T<:Number}
+"""
+    _default_qr_ib(A; p=1) -> Int
+
+Choose a Compact-WY inner block size for tiled QR.
+
+`ib == 1` is numerically fine but runs NextLA's unblocked kernels and scales
+poorly. A modest inner block (capped at the tile width) restores BLAS-3
+intensity inside `geqrt!`/`tsmqr!` without lengthening the panel critical path.
+
+CAQR (`p > 1`) currently only produces correct results for `ib == 1`, so that
+path keeps the unblocked default until the tree kernels are fixed for `ib > 1`.
+"""
+function _default_qr_ib(A::DMatrix; p::Int=1)
+    p > 1 && return 1
+    nb = Int(A.partitioning.blocksize[2])
+    nb <= 0 && return 1
+    return min(32, nb)
+end
+
+function _qr_impl!(A::DMatrix{T}; ib::Union{Int,Nothing}=nothing, p::Int=1) where {T<:Number}
     p >= 1 || throw(ArgumentError("p must be >= 1, got $p"))
-    lm, ln = qr_measure_workspace(A, ib)
+    ib_eff = something(ib, _default_qr_ib(A; p=p))
+    ib_eff >= 1 || throw(ArgumentError("ib must be >= 1, got $ib_eff"))
+    if p > 1 && ib_eff > 1
+        throw(ArgumentError("CAQR (p > 1) currently requires ib=1, got ib=$ib_eff"))
+    end
+    lm, ln = qr_measure_workspace(A, ib_eff)
     nb = A.partitioning.blocksize[2]
     irregular = _use_irregular_qr_tiling(A)
 
     Tm_cols = p > 1 ? 2 * ln : ln
-    Tm = DArray{T}(undef, Blocks(ib, nb), (lm, Tm_cols))
+    Tm = DArray{T}(undef, Blocks(ib_eff, nb), (lm, Tm_cols))
 
     if irregular
         # Hierarchical irregular path (no CAQR tree metadata).
@@ -795,8 +819,22 @@ function _qr_impl!(A::DMatrix{T}; ib::Int=1, p::Int=1) where {T<:Number}
     return QRCompactWY(A, Tm)
 end
 
-function LinearAlgebra.qr!(A::DMatrix{T}; ib::Int=1, p::Int=1) where {T<:Number}
+"""
+    qr!(A::DMatrix; ib=nothing, p=1) -> QRCompactWY
+
+In-place tiled Compact-WY QR factorization of a distributed matrix.
+
+# Keywords
+- `ib`: Compact-WY inner block size. For flat tile QR (`p=1`), defaults to
+  `min(32, tilewidth)` — much faster than `ib=1` for multi-threaded runs.
+  CAQR (`p>1`) currently requires `ib=1`.
+- `p`: CAQR domain count (`p=1` is flat tile QR). `p>1` needs uniform square
+  tiling with `mt % p == 0`; on shared-memory multi-threading it is rarely faster
+  than `p=1`.
+"""
+function LinearAlgebra.qr!(A::DMatrix{T}; ib::Union{Int,Nothing}=nothing, p::Int=1) where {T<:Number}
     p >= 1 || throw(ArgumentError("p must be >= 1, got $p"))
+    ib === nothing || ib >= 1 || throw(ArgumentError("ib must be >= 1, got $ib"))
 
     if !_use_irregular_qr_tiling(A)
         return _qr_impl!(A; ib=ib, p=p)

test/array/linalg/qr.jl

@@ -20,7 +20,7 @@ function check_qr_residuals(A_col, DQ, DR; tol=2.0)
 end
 
 # ======================================================================
-# 1. Basic qr(DA) — varying shapes and block sizes (p=1, ib=1)
+# 1. Basic qr(DA) — varying shapes and block sizes (p=1, default ib)
 # ======================================================================
 @testset "Tile QR:  $T" for T in (Float64, ComplexF64)
     @testset "Square matrices" begin
@@ -510,6 +510,15 @@ end
     @test_throws ArgumentError cageqrf!(DA_copy, Tm; p=2)
 end
 
+# ======================================================================
+# 18b. CAQR currently requires ib=1
+# ======================================================================
+@testset "CAQR rejects ib>1" begin
+    A = rand(Float64, 128, 64)
+    DA = distribute(A, Blocks(32, 32))
+    @test_throws ArgumentError qr!(copy(DA); ib=32, p=2)
+end
+
 # ======================================================================
 # 19. Edge case — single tile
 # ======================================================================