Allow Kokkos::create_mirror* to be used with WithoutInitializing

[masterleinad]

Fixes #4259. For the moment, I only considered the overloads for Kokkos::View and not yet the other View-like containers.

[masterleinad]

With

int main(int argc, char* argv[]) {
  Kokkos::initialize(argc, argv);
  {
    Kokkos::View<int*, Kokkos::Experimental::SYCLDeviceUSMSpace> device_view("device view", 10);
    Kokkos::View<int*, Kokkos::HostSpace> host_view("host view", 10);

    auto mirror_device = Kokkos::create_mirror(device_view);
    auto mirror_host = Kokkos::create_mirror(Kokkos::Experimental::SYCLDeviceUSMSpace{}, host_view);
    auto mirror_device_view = Kokkos::create_mirror_view(device_view);
    auto mirror_host_view = Kokkos::create_mirror_view(Kokkos::Experimental::SYCLDeviceUSMSpace{}, host_view);
  }
  {
    Kokkos::View<int*, Kokkos::Experimental::SYCLDeviceUSMSpace> device_view("device view", 10);
    Kokkos::View<int*, Kokkos::HostSpace> host_view("host view", 10);

    auto mirror_device_wi = Kokkos::create_mirror(Kokkos::WithoutInitializing, device_view);
    auto mirror_host_wi = Kokkos::create_mirror(Kokkos::WithoutInitializing, Kokkos::Experimental::SYCLDeviceUSMSpace{}, host_view);
    auto mirror_device_view_wi = Kokkos::create_mirror(Kokkos::WithoutInitializing, device_view);
    auto mirror_host_view_wi = Kokkos::create_mirror(Kokkos::WithoutInitializing, Kokkos::Experimental::SYCLDeviceUSMSpace{}, host_view);
  }
  Kokkos::finalize();
}

I am getting

KokkosP: Example Library Initialized (sequence is 0, version: 20210623)
KokkosP: Allocate<SYCLDeviceUSM> name: device view pointer: 0x2970a80 size: 40
KokkosP: Executing parallel-for kernel on device 117440513 with unique execution identifier 0
KokkosP: Kokkos::View::initialization [device view] via memset
KokkosP: Execution of kernel 0 is completed.
KokkosP: Allocate<Host> name: host view pointer: 0x2926180 size: 40
KokkosP: Executing parallel-for kernel on device 1 with unique execution identifier 1
KokkosP: Kokkos::View::initialization [host view] via memset
KokkosP: Execution of kernel 1 is completed.
KokkosP: Allocate<Host> name: device view_mirror pointer: 0x2972780 size: 40
KokkosP: Executing parallel-for kernel on device 1 with unique execution identifier 2
KokkosP: Kokkos::View::initialization [device view_mirror] via memset
KokkosP: Execution of kernel 2 is completed.
KokkosP: Allocate<SYCLDeviceUSM> name: host view pointer: 0x2972f00 size: 40
KokkosP: Executing parallel-for kernel on device 117440513 with unique execution identifier 3
KokkosP: Kokkos::View::initialization [host view] via memset
KokkosP: Execution of kernel 3 is completed.
KokkosP: Allocate<Host> name: device view_mirror pointer: 0x2973e40 size: 40
KokkosP: Executing parallel-for kernel on device 1 with unique execution identifier 4
KokkosP: Kokkos::View::initialization [device view_mirror] via memset
KokkosP: Execution of kernel 4 is completed.
KokkosP: Allocate<SYCLDeviceUSM> name: host view pointer: 0x2974600 size: 40
KokkosP: Executing parallel-for kernel on device 117440513 with unique execution identifier 5
KokkosP: Kokkos::View::initialization [host view] via memset
KokkosP: Execution of kernel 5 is completed.
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x2974600 size: 40
KokkosP: Deallocate<Host> name: device view_mirror pointer: 0x2973e40 size: 40
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x2972f00 size: 40
KokkosP: Deallocate<Host> name: device view_mirror pointer: 0x2972780 size: 40
KokkosP: Deallocate<Host> name: host view pointer: 0x2926180 size: 40
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x2970a80 size: 40
KokkosP: Kokkos library finalization called.

for the overloads without WithoutInitializing and

KokkosP: Example Library Initialized (sequence is 0, version: 20210623)
KokkosP: Allocate<SYCLDeviceUSM> name: device view pointer: 0x1ef2180 size: 40
KokkosP: Executing parallel-for kernel on device 117440513 with unique execution identifier 0
KokkosP: Kokkos::View::initialization [device view] via memset
KokkosP: Execution of kernel 0 is completed.
KokkosP: Allocate<Host> name: host view pointer: 0x1e9bb40 size: 40
KokkosP: Executing parallel-for kernel on device 1 with unique execution identifier 1
KokkosP: Kokkos::View::initialization [host view] via memset
KokkosP: Execution of kernel 1 is completed.
KokkosP: Allocate<Host> name: device view_mirror pointer: 0x1ef3bc0 size: 40
KokkosP: Allocate<SYCLDeviceUSM> name: host view pointer: 0x1ef4300 size: 40
KokkosP: Allocate<Host> name: device view_mirror pointer: 0x1ef4cc0 size: 40
KokkosP: Allocate<SYCLDeviceUSM> name: host view pointer: 0x1ef5400 size: 40
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x1ef5400 size: 40
KokkosP: Deallocate<Host> name: device view_mirror pointer: 0x1ef4cc0 size: 40
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x1ef4300 size: 40
KokkosP: Deallocate<Host> name: device view_mirror pointer: 0x1ef3bc0 size: 40
KokkosP: Deallocate<Host> name: host view pointer: 0x1e9bb40 size: 40
KokkosP: Deallocate<SYCLDeviceUSM> name: ceUSM pointer: 0x1ef2180 size: 40
KokkosP: Kokkos library finalization called.

with WithoutInitializing. We should add a test like this once we have #4331.

[masterleinad]

I added a test after we merged #4331. This only fixes some issues when including the basic testing include files.

[masterleinad]

Retest this please.

[janciesko]

LGTM

[crtrott]

Are we introducing all kinds of overloads with unconstrained additional arguments here? If so shouldn't those be functions in the Impl namespace?

[masterleinad]

Are we introducing all kinds of overloads with unconstrained additional arguments here? If so shouldn't those be functions in the Impl namespace?

All the overloads with unconstrained arguments (for WithouutInitializing and other view properties) are in namespace Impl already.

[crtrott]

Ah saw the namespace stuff now.