Rfa

.github/workflows/macos_debug_fetch_hwloc.yml

@@ -19,7 +19,6 @@ jobs:
       run: |
           brew install --overwrite python-tk && \
           brew install --overwrite boost gperftools ninja autoconf automake && \
-          autoreconf -f -i \
           brew upgrade cmake
     - name: Configure
       shell: bash

libs/core/algorithms/CMakeLists.txt

@@ -37,7 +37,9 @@ set(algorithms_headers
     hpx/parallel/algorithms/detail/parallel_stable_sort.hpp
     hpx/parallel/algorithms/detail/pivot.hpp
     hpx/parallel/algorithms/detail/reduce.hpp
+    hpx/parallel/algorithms/detail/reduce_deterministic.hpp
     hpx/parallel/algorithms/detail/replace.hpp
+    hpx/parallel/algorithms/detail/rfa.hpp
     hpx/parallel/algorithms/detail/rotate.hpp
     hpx/parallel/algorithms/detail/sample_sort.hpp
     hpx/parallel/algorithms/detail/search.hpp
@@ -72,6 +74,7 @@ set(algorithms_headers
     hpx/parallel/algorithms/partition.hpp
     hpx/parallel/algorithms/reduce_by_key.hpp
     hpx/parallel/algorithms/reduce.hpp
+    hpx/parallel/algorithms/reduce_deterministic.hpp
     hpx/parallel/algorithms/remove_copy.hpp
     hpx/parallel/algorithms/remove.hpp
     hpx/parallel/algorithms/replace.hpp

libs/core/algorithms/include/hpx/parallel/algorithms/detail/reduce_deterministic.hpp

@@ -0,0 +1,83 @@
+//  Copyright (c) 2024 Shreyas Atre
+//
+//  SPDX-License-Identifier: BSL-1.0
+//  Distributed under the Boost Software License, Version 1.0. (See accompanying
+//  file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#pragma once
+
+#include <hpx/config.hpp>
+#include <hpx/functional/detail/tag_fallback_invoke.hpp>
+#include <hpx/functional/invoke.hpp>
+#include <hpx/parallel/algorithms/detail/rfa.hpp>
+#include <hpx/parallel/util/loop.hpp>
+
+#include <cstddef>
+#include <cstring>
+#include <limits>
+#include <type_traits>
+#include <utility>
+#include "rfa.hpp"
+
+namespace hpx::parallel::detail {
+
+    template <typename ExPolicy>
+    struct sequential_reduce_deterministic_t final
+      : hpx::functional::detail::tag_fallback<
+            sequential_reduce_deterministic_t<ExPolicy>>
+    {
+    private:
+        template <typename InIterB, typename InIterE, typename T,
+            typename Reduce>
+        friend constexpr T tag_fallback_invoke(
+            sequential_reduce_deterministic_t, ExPolicy&&, InIterB first,
+            InIterE last, T init, Reduce&& r)
+        {
+            /// TODO: Put constraint on Reduce to be a binary plus operator
+            (void) r;
+            hpx::parallel::detail::rfa::RFA_bins<T> bins;
+            bins.initialize_bins();
+            std::memcpy(rfa::__rfa_bin_host_buffer__, &bins, sizeof(bins));
+
+            hpx::parallel::detail::rfa::ReproducibleFloatingAccumulator<T> rfa;
+            rfa.set_max_abs_val(init);
+            rfa.unsafe_add(init);
+            rfa.renorm();
+            size_t count = 0;
+            T max_val = std::abs(*first);
+            for (auto e = first; e != last; ++e)
+            {
+                T temp_max_val = std::abs(static_cast<T>(*e));
+                if (max_val < temp_max_val)
+                {
+                    rfa.set_max_abs_val(temp_max_val);
+                    max_val = temp_max_val;
+                }
+                rfa.unsafe_add(*e);
+                count++;
+                if (count == rfa.endurance())
+                {
+                    rfa.renorm();
+                    count = 0;
+                }
+            }
+            return rfa.conv();
+        }
+    };
+
+#if !defined(HPX_COMPUTE_DEVICE_CODE)
+    template <typename ExPolicy>
+    inline constexpr sequential_reduce_deterministic_t<ExPolicy>
+        sequential_reduce_deterministic =
+            sequential_reduce_deterministic_t<ExPolicy>{};
+#else
+    template <typename ExPolicy, typename... Args>
+    HPX_HOST_DEVICE HPX_FORCEINLINE auto sequential_reduce_deterministic(
+        Args&&... args)
+    {
+        return sequential_reduce_deterministic_t<ExPolicy>{}(
+            std::forward<Args>(args)...);
+    }
+#endif
+
+}    // namespace hpx::parallel::detail