Add multidimensional memset vectorization benchmark.

benchmark/CMakeLists.txt

@@ -6,4 +6,7 @@ add_executable(internal-iteration-benchmark internal_iteration_benchmark.cpp)
 target_link_libraries(internal-iteration-benchmark PUBLIC nanobench flux)
 
 add_executable(sort-benchmark sort_benchmark.cpp)
-target_link_libraries(sort-benchmark PUBLIC nanobench flux)
+target_link_libraries(sort-benchmark PUBLIC nanobench flux)
+
+add_executable(multidimensional-memset-benchmark multidimensional_memset_benchmark.cpp multidimensional_memset_benchmark_kernels.cpp)
+target_link_libraries(multidimensional-memset-benchmark PUBLIC nanobench flux)

benchmark/multidimensional_memset_benchmark.cpp

@@ -0,0 +1,46 @@
+
+// Copyright (c) 2021 Barry Revzin
+// Copyright (c) 2023 NVIDIA Corporation
+// 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)
+
+#include <nanobench.h>
+
+#include <flux.hpp>
+
+#include <numeric>
+
+namespace an = ankerl::nanobench;
+
+extern void memset_2d_reference(double* A, std::size_t N, std::size_t M);
+
+extern void memset_2d_flux_cartesian_product_iota(double* A, std::size_t N, std::size_t M);
+
+int main(int argc, char** argv)
+{
+    int const n_iters = argc > 1 ? std::atoi(argv[1]) : 40;
+
+    constexpr std::size_t N = 1024;
+    constexpr std::size_t M = 2048;
+    std::vector<double> A(N * M);
+
+    {
+        auto bench = an::Bench().minEpochIterations(n_iters).relative(true);
+
+        std::iota(A.begin(), A.end(), 0);
+
+        bench.run("memset_2d_handwritten",
+            [&] { memset_2d_reference(A.data(), N, M); });
+
+        if (auto it = std::ranges::find_if_not(A, [&] (auto e) { return e == 0; }); it != A.end())
+            throw false;
+
+        std::iota(A.begin(), A.end(), 0);
+
+        bench.run("memset_2d_flux_cartesian_product_iota",
+            [&] { memset_2d_flux_cartesian_product_iota(A.data(), N, M); });
+
+        if (auto it = std::ranges::find_if_not(A, [&] (auto e) { return e == 0; }); it != A.end())
+            throw false;
+    }
+}

benchmark/multidimensional_memset_benchmark_kernels.cpp

@@ -0,0 +1,23 @@
+
+// Copyright (c) 2023 NVIDIA Corporation
+// 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)
+
+#include <flux/op/cartesian_product.hpp>
+#include <flux/source/iota.hpp>
+#include <flux/op/for_each.hpp>
+
+void memset_2d_reference(double* A, std::size_t N, std::size_t M)
+{
+    for (std::size_t j = 0; j != M; ++j)
+        for (std::size_t i = 0; i != N; ++i)
+            A[i + j * N] = 0.0;
+}
+
+void memset_2d_flux_cartesian_product_iota(double* A, std::size_t N, std::size_t M)
+{
+    flux::cartesian_product(flux::iota(0LU, N), flux::iota(0LU, M))
+        .for_each(flux::unpack([&] (auto i, auto j) {
+            A[i + j * N] = 0.0;
+        }));
+}