Reenable stdexec GCC CI configurations

[msimberg]

Fixes #1184. Based on #1180, since that disables the stdexec CI.

This is a workaround to avoid compilation errors like:

error: exception specification of 'decltype(auto) stdexec::dependent_domain::transform_sender(_Sender&&, const _Env&) const [with _Sender = stdexec::__sexpr<<lambda closure object>stdexec::{anonymous}::<lambda()>(), stdexec::{anonymous}::__anon>; _Env = stdexec::__env::env<>]' depends on itself
  179 |     noexcept(__is_nothrow_transform_sender<_Sender, _Env>()) -> decltype(auto) {
      |              ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~

This is triggered by the two nested withTemporaryTiles used in all_reduce.h (

DLA-Future/include/dlaf/communication/kernels/internal/all_reduce.h

Lines 51 to 99 in a64c822

	template <Device DCommIn, Device DCommOut, class T, Device DIn, Device DOut>
	[[nodiscard]] auto scheduleAllReduce(
	pika::execution::experimental::unique_any_sender<CommunicatorPipelineExclusiveWrapper> pcomm,
	MPI_Op reduce_op, dlaf::matrix::ReadOnlyTileSender<T, DIn> tile_in,
	dlaf::matrix::ReadWriteTileSender<T, DOut> tile_out) {
	using dlaf::comm::CommunicationDevice_v;
	using dlaf::comm::internal::allReduce_o;
	using dlaf::comm::internal::transformMPI;
	using dlaf::internal::CopyFromDestination;
	using dlaf::internal::CopyToDestination;
	using dlaf::internal::RequireContiguous;
	using dlaf::internal::SenderSingleValueType;
	using dlaf::internal::whenAllLift;
	using dlaf::internal::withTemporaryTile;
	// We create two nested scopes for the input and output tiles with
	// withTemporaryTile. The output tile is in the outer scope as the output tile
	// will be returned by the returned sender.
	auto all_reduce_final = [reduce_op, pcomm = std::move(pcomm),
	tile_in = std::move(tile_in)](const auto& tile_out_comm) mutable {
	auto all_reduce = [reduce_op, pcomm = std::move(pcomm),
	&tile_out_comm](const auto& tile_in_comm) mutable {
	return whenAllLift(std::move(pcomm), reduce_op, std::cref(tile_in_comm),
	std::cref(tile_out_comm)) |
	transformMPI(allReduce_o);
	};
	// The input tile must be copied to the temporary tile used for the
	// reduction, but the temporary tile does not need to be copied back to the
	// input since the data is not changed by the reduction (the result is
	// written into the output tile instead). The reduction is explicitly done
	// on CPU memory so that we can manage potential asynchronous copies between
	// CPU and GPU. A reduction requires contiguous memory.
	return withTemporaryTile<DCommIn, CopyToDestination::Yes, CopyFromDestination::No,
	RequireContiguous::Yes>(std::move(tile_in), std::move(all_reduce));
	};
	#if !defined(DLAF_WITH_MPI_GPU_AWARE)
	static_assert(DCommIn == Device::CPU, "DLAF_WITH_MPI_GPU_AWARE=off, MPI accepts only CPU memory.");
	static_assert(DCommOut == Device::CPU, "DLAF_WITH_MPI_GPU_AWARE=off, MPI accepts only CPU memory.");
	#endif
	// The output tile does not need to be copied to the temporary tile since it
	// is only written to. The written data is copied back from the temporary tile
	// to the output tile. The reduction is explicitly done on CPU memory so that
	// we can manage potential asynchronous copies between CPU and GPU. A
	// reduction requires contiguous memory.
	return withTemporaryTile<DCommOut, CopyToDestination::No, CopyFromDestination::Yes,
	RequireContiguous::Yes>(std::move(tile_out), std::move(all_reduce_final));
	}

). The smallest reproducer I've found so far involves both pika and stdexec, so I can't yet conclude if it's a bug in either or if we should just avoid this pattern in general.

[msimberg]

cscs-ci run

[msimberg]

I've marked this ready for review. The workaround is the same as before. The smallest reproducer I've found so far is:

#include <pika/execution.hpp>

auto helper = [](auto f) {
  return stdexec::let_value(stdexec::just(f), [](auto f) noexcept {
           return f() |
               pika::execution::experimental::drop_value() |
               stdexec::then([]{});
         });
};

auto algo() {
  return helper([]() noexcept {
    return helper([]() noexcept { return stdexec::just(); });
  });
}

i.e. involves both pika and stdexec. Further investigation is going to take quite a bit longer so I would prefer going with the workaround right now.

[aurianer]

LGTM thanks a lot!

[rasolca]

cscs-ci run