Use BLAS.trsm! instead of LAPACK.trtrs! in left-triangular solves

Co-authored-by: Alexis Montoison <[email protected]>

Author: dkarrasch

src/triangular.jl

@@ -1223,11 +1223,13 @@ function generic_mattrimul!(C::StridedMatrix{T}, uploc, isunitc, tfun::Function,
     end
 end
 # division
-function generic_trimatdiv!(C::StridedVecOrMat{T}, uploc, isunitc, tfun::Function, A::StridedMatrix{T}, B::AbstractVecOrMat{T}) where {T<:BlasFloat}
+generic_trimatdiv!(C::StridedVector{T}, uploc, isunitc, tfun::Function, A::StridedMatrix{T}, B::AbstractVector{T}) where {T<:BlasFloat} =
+    BLAS.trsv!(uploc, tfun === identity ? 'N' : tfun === transpose ? 'T' : 'C', isunitc, A, C === B ? C : copyto!(C, B))
+function generic_trimatdiv!(C::StridedVecOrMat{T}, uploc, isunitc, tfun::Function, A::StridedMatrix{T}, B::AbstractMatrix{T}) where {T<:BlasFloat}
     if stride(C,1) == stride(A,1) == 1
-        LAPACK.trtrs!(uploc, tfun === identity ? 'N' : tfun === transpose ? 'T' : 'C', isunitc, A, C === B ? C : copyto!(C, B))
+        BLAS.trsm!('L', uploc, tfun === identity ? 'N' : tfun === transpose ? 'T' : 'C', isunitc, one(T), A, C === B ? C : copyto!(C, B))
     else # incompatible with LAPACK
-        @invoke generic_trimatdiv!(C::AbstractVecOrMat, uploc, isunitc, tfun::Function, A::AbstractMatrix, B::AbstractVecOrMat)
+        @invoke generic_trimatdiv!(C::AbstractVecOrMat, uploc, isunitc, tfun::Function, A::AbstractMatrix, B::AbstractMatrix)
     end
 end
 function generic_mattridiv!(C::StridedMatrix{T}, uploc, isunitc, tfun::Function, A::AbstractMatrix{T}, B::StridedMatrix{T}) where {T<:BlasFloat}

test/triangular.jl

@@ -886,8 +886,13 @@ end
     end
 end
 
-@testset "(l/r)mul! and (l/r)div! for non-contiguous matrices" begin
+@testset "(l/r)mul! and (l/r)div! for non-contiguous arrays" begin
     U = UpperTriangular(reshape(collect(3:27.0),5,5))
+    b = float.(1:10)
+    b2 = copy(b); b2v = view(b2, 1:2:9); b2vc = copy(b2v)
+    @test lmul!(U, b2v) == lmul!(U, b2vc)
+    b2 = copy(b); b2v = view(b2, 1:2:9); b2vc = copy(b2v)
+    @test ldiv!(U, b2v) ≈ ldiv!(U, b2vc)
     B = float.(collect(reshape(1:100, 10,10)))
     B2 = copy(B); B2v = view(B2, 1:2:9, 1:5); B2vc = copy(B2v)
     @test lmul!(U, B2v) == lmul!(U, B2vc)