Backports for 5.8.5

.buildkite/pipeline.yml

@@ -25,7 +25,7 @@ steps:
           build.message !~ /\[only/ &&
             build.message !~ /\[skip tests\]/ &&
             build.message !~ /\[skip julia\]/
-        timeout_in_minutes: 60
+        timeout_in_minutes: 90
         matrix:
           setup:
             julia:
@@ -123,16 +123,12 @@ steps:
             println("--- :julia: Instantiating project")
             withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0) do
               Pkg.activate(joinpath("lib", lowercase("{{matrix.package}}")))
-              try
-                Pkg.instantiate()
-              catch
-                # if we fail to instantiate, assume that we need newer dependencies
-                deps = [PackageSpec(path=".")]
-                if "{{matrix.package}}" == "cuTensorNet"
-                  push!(deps, PackageSpec(path="lib/cutensor"))
-                end
-                Pkg.develop(deps)
+
+              deps = [PackageSpec(path=".")]
+              if "{{matrix.package}}" == "cuTensorNet"
+                push!(deps, PackageSpec(path="lib/cutensor"))
               end
+              Pkg.develop(deps)
 
               Pkg.add("CUDA_Runtime_jll")
               write(joinpath("lib", lowercase("{{matrix.package}}"), "LocalPreferences.toml"),

Project.toml

@@ -1,6 +1,6 @@
 name = "CUDA"
 uuid = "052768ef-5323-5732-b1bb-66c8b64840ba"
-version = "5.8.4"
+version = "5.8.5"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

lib/cusolver/linalg.jl

@@ -110,6 +110,13 @@ function Base.:\(F::Union{LinearAlgebra.LAPACKFactorizations{<:Any,<:CuArray},
     return LinearAlgebra._cut_B(BB, 1:n)
 end
 
+# make copyto! for Hermitian and Symmetric dispatch to the Base implementation
+# instead of being overridden by GPUArrays' dense copy (because of AnyGPUArray)
+Base.copyto!(dst::Symmetric{<:Any,<:CuMatrix}, src::Symmetric{<:Any,<:CuMatrix}) =
+    @invoke copyto!(dst::Symmetric, src::Symmetric)
+Base.copyto!(dst::Hermitian{<:Any,<:CuMatrix}, src::Hermitian{<:Any,<:CuMatrix}) =
+    @invoke copyto!(dst::Hermitian, src::Hermitian)
+
 # eigenvalues
 
 function LinearAlgebra.eigen(A::Symmetric{T,<:CuMatrix}) where {T<:BlasReal}

lib/cutensornet/src/types.jl

@@ -110,12 +110,12 @@ mutable struct CuTensorNetwork{T}
     desc::CuTensorNetworkDescriptor
     input_modes::Vector{Vector{Int32}}
     input_extents::Vector{Vector{Int32}}
-    input_strides::Vector{<:Union{Ptr{Nothing}, Vector{Int32}}}
+    input_strides::Vector{<:Union{Ptr{Nothing}, Vector{Int64}}}
     input_qualifiers::Vector{cutensornetTensorQualifiers_t}
     input_arrs::Vector{CuArray{T}}
     output_modes::Vector{Int32}
     output_extents::Vector{Int32}
-    output_strides::Union{Ptr{Nothing}, Vector{Int32}}
+    output_strides::Union{Ptr{Nothing}, Vector{Int64}}
     output_arr::CuArray{T}
 end
 function CuTensorNetwork(T::DataType, input_modes, input_extents, input_strides, input_qualifiers, output_modes, output_extents, output_strides)

src/mapreduce.jl

@@ -276,7 +276,7 @@ function GPUArrays.mapreducedim!(f::F, op::OP, R::AnyCuArray{T},
         partial_shmem = partial_reduce_shmem
         partial_blocks = partial_reduce_blocks*other_blocks
 
-        partial = similar(R, (size(R)..., partial_blocks))
+        partial = similar(R, (size(R)..., partial_reduce_blocks))
         if init === nothing
             # without an explicit initializer we need to copy from the output container
             partial .= R