add custom taskflow fork

packages/taskflow.lua

@@ -0,0 +1,30 @@
+package("taskflow_custom", function()
+    set_urls("https://github.com/samayala22/taskflow.git")
+    add_urls("https://github.com/samayala22/taskflow/archive/$(version).tar.gz")
+    add_versions("v3.6.2", "0a1d306f90e8e17cb98b95eaae9e8b9455beeeca0f0a72afee7719b27706c68c")
+
+    if is_plat("linux") then
+        add_syslinks("pthread")
+    end
+
+    on_install("linux", "macosx", "windows", "iphoneos", "android", "cross", "mingw", "bsd", function (package)
+        os.cp("taskflow", package:installdir("include"))
+    end)
+
+    on_test(function (package)
+        assert(package:check_cxxsnippets({test = [[
+            #include <taskflow/taskflow.hpp>
+            #include <taskflow/algorithm/for_each.hpp>
+            static void test() {
+                tf::Executor executor;
+                tf::Taskflow taskflow;
+                std::vector<int> range(10);
+                std::iota(range.begin(), range.end(), 0);
+                taskflow.for_each(range.begin(), range.end(), [&] (int i) {
+                    std::printf("for_each on container item: %d\n", i);
+                });
+                executor.run(taskflow).wait();
+            }
+        ]]}, {configs = {languages = "c++17"}}))
+    end)
+end)

vlm/backends/avx2/src/vlm_backend_avx2.cpp

@@ -311,7 +311,7 @@ void BackendAVX2::compute_lhs(const FlowData& flow) {
     auto init = taskflow.placeholder();
     auto sync = taskflow.placeholder();
 
-    auto wing_pass = taskflow.for_each_index(start_wing, end_wing, (u32)1, [&] (u32 i) {
+    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);
     }, tf::GuidedPartitioner());
@@ -320,7 +320,7 @@ void BackendAVX2::compute_lhs(const FlowData& flow) {
     auto cond = taskflow.emplace([&]{
         return idx < m.nc + m.nw ? 0 : 1; // 0 means continue, 1 means break
     });
-    auto wake_pass = taskflow.for_each_index(0u, m.ns, (u32)1, [&] (u32 j) {
+    auto wake_pass = taskflow.for_each_index(0u, m.ns, [&] (u32 j) {
         const u32 ia = (m.nc - 1) * m.ns + j;
         const u32 lidx = idx * m.ns + j;
         macro_kernel_avx2<false>(m, lhs, ia, lidx, sigma_p4);

vlm/backends/avx2/xmake.lua

@@ -1,18 +1,19 @@
+includes("../../../packages/taskflow.lua")
+
 add_requires("tbb")
 
 add_requires("openblas")
-add_requires("taskflow")
+add_requires("taskflow_custom")
 
 target("backend-avx2")
     set_kind("static")
     set_default(false)
     add_vectorexts("avx2", "fma")
     add_packages("tbb")
-    add_packages("taskflow")
+    add_packages("taskflow_custom")
     add_defines("HAVE_LAPACK_CONFIG_H")
     add_packages("openblas", { public = true })
 
     add_includedirs("../../include")
     add_files("src/*.cpp")
     add_includedirs("include", {public = true})
-