Add storage_type for adjoint, tranpose and Diagonal s.t. they work for GPU arrays

Author: nHacke

src/abstract.jl

@@ -174,6 +174,11 @@ storage_type(op::AbstractLinearOperator) = error("please implement storage_type
 storage_type(op::LinearOperator) = typeof(op.Mv5)
 storage_type(M::AbstractMatrix{T}) where {T} = Vector{T}
 
+# Lazy wrappers
+storage_type(op::Adjoint) = storage_type(parent(op))
+storage_type(op::Transpose) = storage_type(parent(op))
+storage_type(op::Diagonal) = typeof(parent(op))
+
 """
   reset!(op)
 

test/gpu/amdgpu.jl

@@ -6,10 +6,18 @@ using LinearOperators, AMDGPU
   B = ROCArray(rand(Float32, 10, 10))
   C = ROCArray(rand(Float32, 20, 20))
   M = BlockDiagonalOperator(A, B, C)
+  v = ROCArray(rand(5))
+
 
   v = ROCArray(rand(Float32, 35))
   y = M * v
   @test y isa ROCArray{Float32}
 
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(adjoint(A))
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(transpose(A))
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(adjoint(A))
+  @test LinearOperators.storage_type(Diagonal(v)) == typeof(v)
+
+
   @testset "AMDGPU S kwarg" test_S_kwarg(arrayType = ROCArray)
 end

test/gpu/nvidia.jl

@@ -9,6 +9,13 @@ using LinearOperators, CUDA, CUDA.CUSPARSE, CUDA.CUSOLVER
   B = CUDA.rand(10, 10)
   C = CUDA.rand(20, 20)
   M = BlockDiagonalOperator(A, B, C)
+  v = CUDA.rand(5)
+
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(adjoint(A))
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(transpose(A))
+  @test LinearOperators.storage_type(A) == LinearOperators.storage_type(adjoint(A))
+  @test LinearOperators.storage_type(Diagonal(v)) == typeof(v)
+  
 
   v = CUDA.rand(35)
   y = M * v