Update to use sockets instead of numa_domains with pika's resource partitioner

ci/.gitlab-ci.yml

@@ -7,8 +7,9 @@ include:
   - local: 'ci/cpu/clang16_release.yml'
   - local: 'ci/cpu/gcc11_codecov.yml'
   - local: 'ci/cpu/gcc11_release.yml'
-  - local: 'ci/cpu/gcc11_release_stdexec.yml'
-  - local: 'ci/cpu/gcc11_debug_stdexec.yml'
+  # https://github.com/eth-cscs/DLA-Future/issues/1184
+  # - local: 'ci/cpu/gcc11_release_stdexec.yml'
+  # - local: 'ci/cpu/gcc11_debug_stdexec.yml'
   - local: 'ci/cpu/gcc12_release_cxx20.yml'
   - local: 'ci/cuda/gcc11_release.yml'
   - local: 'ci/cuda/gcc11_release_scalapack.yml'

ci/common-ci.yml

@@ -34,7 +34,7 @@ stages:
     reports:
       dotenv: build.env
   variables:
-    SPACK_SHA: 1184de8352281ce1b34a8f3f6b54d7fadc2b216a
+    SPACK_SHA: 0905edf592752742eb4ddab3a528d3aee8f92930
     SPACK_DLAF_REPO: ./spack
     DOCKER_BUILD_ARGS: '[
         "BASE_IMAGE",

ci/docker/debug-cpu-stdexec.yaml

@@ -31,5 +31,5 @@ spack:
         - 'malloc=system'
     stdexec:
       require:
-        - '@git.nvhpc-23.09.rc4=main'
+        - '@git.8bc7c7f06fe39831dea6852407ebe7f6be8fa9fd=main'
         - 'build_type=Debug'

ci/docker/debug-cuda-scalapack.yaml

@@ -35,5 +35,6 @@ spack:
         - '+fortran'
     pika:
       require:
+        - '@0.19.1'
         - 'build_type=Debug'
         - 'malloc=system'

ci/docker/release-cpu-stdexec.yaml

@@ -29,4 +29,4 @@ spack:
         - '+stdexec'
     stdexec:
       require:
-        - '@git.48c52df0f81c6151eecf4f39fa5eed2dc0216204=main'
+        - '@git.8bc7c7f06fe39831dea6852407ebe7f6be8fa9fd=main'

ci/docker/release-rocm533-stdexec.yaml

@@ -31,7 +31,7 @@ spack:
         - '+stdexec'
     stdexec:
       require:
-        - '@git.48c52df0f81c6151eecf4f39fa5eed2dc0216204=main'
+        - '@git.8bc7c7f06fe39831dea6852407ebe7f6be8fa9fd=main'
     blas:
       require:: openblas
     lapack:

include/dlaf/tune.h

@@ -13,8 +13,11 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <exception>
 #include <iosfwd>
+#include <iostream>
 
+#include <pika/init.hpp>
 #include <pika/runtime.hpp>
 
 #include <dlaf/types.h>
@@ -97,10 +100,14 @@ struct TuneParameters {
     // Some parameters require the pika runtime to be initialized since they depend on the number of
     // threads used by the runtime. We initialize them separately in the constructor after checking that
     // pika is initialized.
+#if PIKA_VERSION_FULL >= 0x001600  // >= 0.22.0
+    if (!pika::is_runtime_initialized()) {
+      std::cerr
+          << "[ERROR] Trying to initialize DLA-Future tune parameters but the pika runtime is not initialized. Make sure pika is initialized first.\n";
+      std::terminate();
+    }
+#endif
 
-    // TODO: Check upfront whether pika has been initialized to be able to give a better error message.
-    // If pika has not been initialized getting the default thread pool will either trigger an assertion
-    // or a segfault.
     const auto default_pool_thread_count =
         pika::resource::get_thread_pool("default").get_os_thread_count();
     red2band_panel_nworkers = std::max<std::size_t>(1, default_pool_thread_count / 2);

spack/packages/dla-future/package.py

@@ -92,6 +92,7 @@ class DlaFuture(CMakePackage, CudaPackage, ROCmPackage):
     depends_on("pika@0.17:", when="@0.2.1")
     depends_on("pika@0.18:", when="@0.3")
     depends_on("pika@0.19.1:", when="@0.4.0:")
+    conflicts("^pika@0.28:", when="@:0.6")
     depends_on("pika-algorithms@0.1:", when="@:0.2")
     depends_on("pika +mpi")
     depends_on("pika +cuda", when="+cuda")

src/init.cpp

@@ -398,6 +398,10 @@ void initResourcePartitionerHandler(pika::resource::partitioner& rp,
   // Create a thread pool with a single core that we will use for all
   // communication related tasks
   rp.create_thread_pool("mpi", pika::resource::scheduling_policy::static_priority, mode);
+#if PIKA_VERSION_FULL >= 0x001C00  // >= 0.28.0
+  rp.add_resource(rp.sockets()[0].cores()[0].pus()[0], "mpi");
+#else
   rp.add_resource(rp.numa_domains()[0].cores()[0].pus()[0], "mpi");
+#endif
 }
 }