Support for adjoint and transpose of strided inputs (non-contiguous views) (#452)

Co-authored-by: Tim Besard <tim.besard@gmail.com>

Author: evelyne-ringoot

lib/GPUArraysCore/src/GPUArraysCore.jl

@@ -6,7 +6,8 @@ using Adapt
 ## essential types
 
 export AbstractGPUArray, AbstractGPUVector, AbstractGPUMatrix, AbstractGPUVecOrMat,
-       WrappedGPUArray, AnyGPUArray, AbstractGPUArrayStyle
+       WrappedGPUArray, AnyGPUArray, AbstractGPUArrayStyle,
+       AnyGPUArray, AnyGPUVector, AnyGPUMatrix
 
 """
     AbstractGPUArray{T, N} <: DenseArray{T, N}
@@ -24,6 +25,8 @@ const AbstractGPUVecOrMat{T} = Union{AbstractGPUArray{T, 1}, AbstractGPUArray{T,
 # convenience aliases for working with wrapped arrays
 const WrappedGPUArray{T,N} = WrappedArray{T,N,AbstractGPUArray,AbstractGPUArray{T,N}}
 const AnyGPUArray{T,N} = Union{AbstractGPUArray{T,N}, WrappedGPUArray{T,N}}
+const AnyGPUVector{T} = AnyGPUArray{T, 1}
+const AnyGPUMatrix{T} = AnyGPUArray{T, 2}
 
 ## broadcasting
 

src/host/linalg.jl

@@ -29,9 +29,9 @@ function LinearAlgebra.adjoint!(B::AbstractGPUMatrix, A::AbstractGPUVector)
     B
 end
 
-LinearAlgebra.transpose!(B::AbstractGPUArray, A::AbstractGPUArray) = transpose_f!(transpose, B, A)
-LinearAlgebra.adjoint!(B::AbstractGPUArray, A::AbstractGPUArray) = transpose_f!(adjoint, B, A)
-function transpose_f!(f, B::AbstractGPUMatrix{T}, A::AbstractGPUMatrix{T}) where T
+LinearAlgebra.transpose!(B::AnyGPUArray, A::AnyGPUArray) = transpose_f!(transpose, B, A)
+LinearAlgebra.adjoint!(B::AnyGPUArray, A::AnyGPUArray) = transpose_f!(adjoint, B, A)
+function transpose_f!(f, B::AnyGPUMatrix{T}, A::AnyGPUMatrix{T}) where T
     axes(B,1) == axes(A,2) && axes(B,2) == axes(A,1) || throw(DimensionMismatch(string(f)))
     gpu_call(B, A) do ctx, B, A
         idx = @cartesianidx A