add executor singleton and remove tbb

vlm/backends/avx2/src/vlm_backend_avx2.cpp

@@ -5,18 +5,13 @@
 #include "vlm_mesh.hpp"
 #include "vlm_data.hpp"
 #include "vlm_utils.hpp"
+#include "vlm_executor.hpp" // includes taskflow/taskflow.hpp
 
 #include <algorithm>
 #include <cstdio>
 #include <fstream>
 #include <immintrin.h>
 
-// TODO: evaluate possible replacement of TBB with TaskFlow or OMP
-#include <oneapi/tbb/global_control.h>
-#include <oneapi/tbb/blocked_range.h>
-#include <oneapi/tbb/parallel_for.h>
-
-#include <taskflow/taskflow.hpp>
 #include <taskflow/algorithm/for_each.hpp>
 
 #include <lapacke.h>
@@ -27,7 +22,6 @@ using namespace vlm;
 BackendAVX2::~BackendAVX2() = default; // Destructor definition
 
 BackendAVX2::BackendAVX2(Mesh& mesh) : Backend(mesh) {
-    //tbb::global_control global_limit(oneapi::tbb::global_control::max_allowed_parallelism, 1);
     lhs.resize((u64)mesh.nb_panels_wing() * (u64)mesh.nb_panels_wing());
     rhs.resize(mesh.nb_panels_wing());
     ipiv.resize(mesh.nb_panels_wing());
@@ -283,34 +277,13 @@ void BackendAVX2::compute_lhs(const FlowData& flow) {
     SimpleTimer timer("LHS");
     Mesh& m = mesh;
     const f32 sigma_p4 = pow<4>(flow.sigma_vatistas); // Vatistas coeffcient (^2n with n=2)
-    tbb::affinity_partitioner ap;
 
     const u32 start_wing = 0;
     const u32 end_wing = (m.nc - 1) * m.ns;
-    // tbb::parallel_for(tbb::blocked_range<u32>(start_wing, end_wing),[&](const tbb::blocked_range<u32> &r) {
-    // for (u32 i = r.begin(); i < r.end(); i++) {
-    //     macro_kernel_avx2<true>(m, lhs, i, i, sigma_p4);
-    //     macro_kernel_remainder_scalar<true>(m, lhs, i, i);
-    // }
-    // }, ap);
-
-    // for (u32 i = m.nc - 1; i < m.nc + m.nw; i++) {
-    //     tbb::parallel_for(tbb::blocked_range<u32>(0, m.ns),[&](const tbb::blocked_range<u32> &r) {
-    //     for (u32 j = r.begin(); j < r.end(); j++) {
-    //         const u32 ia = (m.nc - 1) * m.ns + j;
-    //         const u32 lidx = i * m.ns + j;
-    //         macro_kernel_avx2<false>(m, lhs, ia, lidx, sigma_p4);
-    //         macro_kernel_remainder_scalar<false>(m, lhs, i, i);
-    //     }
-    //     }, ap);
-    // }
 
-    tf::Executor executor{};
     tf::Taskflow taskflow;
 
     auto init = taskflow.placeholder();
-    auto sync = taskflow.placeholder();
-
     auto wing_pass = taskflow.for_each_index(start_wing, end_wing, [&] (u32 i) {
         macro_kernel_avx2<true>(m, lhs, i, i, sigma_p4);
         macro_kernel_remainder_scalar<true>(m, lhs, i, i);
@@ -330,14 +303,15 @@ void BackendAVX2::compute_lhs(const FlowData& flow) {
         idx++;
         return 0; // 0 means continue
     });
+    auto sync = taskflow.placeholder();
 
     init.precede(wing_pass, cond);
     wing_pass.precede(sync);
     cond.precede(wake_pass, sync);
     wake_pass.precede(back);
     back.precede(cond);
 
-    executor.run(taskflow).wait();
+    Executor::get().run(taskflow).wait();
 }
 
 void BackendAVX2::compute_rhs(const FlowData& flow) {

vlm/backends/avx2/xmake.lua

@@ -1,15 +1,10 @@
-includes("../../../packages/taskflow.lua")
-
-add_requires("tbb")
-
 add_requires("openblas")
 add_requires("taskflow_custom")
 
 target("backend-avx2")
     set_kind("static")
     set_default(false)
     add_vectorexts("avx2", "fma")
-    add_packages("tbb")
     add_packages("taskflow_custom")
     add_defines("HAVE_LAPACK_CONFIG_H")
     add_packages("openblas", { public = true })

vlm/dev/main.cpp

@@ -1,5 +1,6 @@
 #include "vlm.hpp"
 #include "parser.hpp"
+#include "vlm_executor.hpp"
 #include "vlm_types.hpp"
 #include <iostream>
 #include <cstdio>
@@ -53,6 +54,7 @@ int main(int argc, char **argv) {
     });
 
     try {
+        //vlm::Executor::instance(1); // 1 thread
         LinearVLM solver(cfg);
         std::vector<f32> alphas = cfg().section("solver").get_vector<f32>("alphas", {0.0f});
         std::transform(alphas.begin(), alphas.end(), alphas.begin(), 

vlm/include/vlm_executor.hpp

@@ -0,0 +1,27 @@
+// This header should not be included by another header file, only by source files.
+#pragma once
+
+#include <mutex>
+#include <taskflow/taskflow.hpp> // includes <thread>, <mutex>, <memory>
+
+// Taskflow executor singleton wrapper (not thread-safe by design)
+namespace vlm {
+    class Executor {
+    public:
+        Executor(const Executor&) = delete;
+        Executor& operator=(const Executor&) = delete;
+        static tf::Executor& instance(size_t num_threads) {
+            if (!_instance) _instance = std::make_unique<tf::Executor>(num_threads ? num_threads : std::thread::hardware_concurrency());
+            return *_instance;
+        }
+        static tf::Executor& get() {
+            if (!_instance) return instance(0);
+            return *_instance;
+        }
+    private:
+        static std::unique_ptr<tf::Executor> _instance;
+
+        Executor() = default;
+        ~Executor() = default;
+    };
+}

vlm/include/vlm_fwd.hpp

@@ -4,5 +4,8 @@ namespace vlm {
 
 class FlowData;
 class Mesh;
+class Backend;
+template<typename Interpolator> class Database2D;
+template<typename Interpolator> class WingProfile;
 
 }

vlm/src/vlm_executor.cpp

@@ -0,0 +1,5 @@
+#include "vlm_executor.hpp"
+
+using namespace vlm;
+
+std::unique_ptr<tf::Executor> Executor::_instance = nullptr;

vlm/xmake.lua

@@ -1,7 +1,9 @@
--- add_requires("openmp")
+add_requires("taskflow_custom")
 
 target("libvlm")
     set_kind("static")
+    add_packages("taskflow_custom", {public = true})
+
     for _,name in ipairs(backends) do
         if has_config(backend_option(name)) then
             add_defines(backend_defines(name))
@@ -15,16 +17,9 @@ target("libvlm")
 target("vlm")
     set_kind("binary")
     set_default(true)
-    add_rpathdirs("$ORIGIN") -- tbb dll must be in same dir as exe
-    -- add_packages("openmp") -- needed for gcc linker (for eigen)
+    add_rpathdirs("$ORIGIN")
     add_deps("libvlm") -- core library
     set_runargs({"-i"}, {"../../../../config/elliptic.vlm"}, {"-m"}, {"../../../../mesh/elliptic_64x64.x"}, {"-o"}, {"../../../../results/elliptic.vtu"})
     add_files("dev/main.cpp")
 
 -- xmake run vlm -i ../../../../config/elliptic.vlm -m ../../../../mesh/elliptic_128x128.x -o ../../../../results/rec.vtu
-
--- target("vlm-tests")
---     set_kind("binary")
---     set_default(false)
---     add_deps("libvlm")
---     add_files("dev/tests.cpp")

xmake.lua

@@ -17,11 +17,6 @@ set_warnings("all")
 set_languages("c++17", "c99")
 set_runtimes("MD") -- msvc runtime library (MD/MT/MDd/MTd)
 
--- TBB macro for profiling parallel objects
-if is_mode("debug", "releasedbg") then
-    add_defines("TBB_USE_THREADING_TOOLS")
-end
-
 -- Define backends and helper functions
 backends = {"cuda", "avx2"}
 backend_includes = function(name) return "vlm/backends/" .. name .. "/xmake.lua" end
@@ -45,8 +40,7 @@ for _,name in ipairs(backends) do
     end
 end
 
--- add_requires("openmp")
-
+includes("packages/taskflow.lua")
 includes("vlm/xmake.lua") -- library and main driver
 
 -- Create tests
@@ -55,7 +49,6 @@ for _, file in ipairs(os.files("tests/test_*.cpp")) do
     target(name)
         set_kind("binary")
         set_default(false)
-        -- add_packages("openmp") -- needed for gcc linker (for eigen)
         add_deps("libvlm")
         add_files("tests/" .. name .. ".cpp")
         add_tests("default")