Application that have MKL linking work correct Intermittently in Windows

[Kohn-Sham]

Hi.

Following source code runs well in Ubuntu, oneAPI version: 2024.0, but doesn't in Windows, oneAPI version: 2024.1 or 2024.2

Compiled app sometimes runs well or sometimes runs without flow_graph or just do nothing. (Running app several times, app somtimes runs well)

Changing cmake build option "--config" to Debug or Release mitigated (but not completely) described bug but still occur.

Setting compiler optimization flag to "/Od" also mitigated but not completely resolved bug.

<bug_tbb_mkl_problem_windows.cpp>

#define _USE_MATH_DEFINES
#include <mkl.h>
#include <oneapi/tbb/flow_graph.h>
#include <fstream>
#include <cmath>
#include <cstring>
#include <iostream>
#include <vector>


class src_body{
    int m_f,m_N,m_F;
    MKL_Complex8* m_arr_in;

public:
    src_body(MKL_LONG N,MKL_LONG F,MKL_Complex8* arr_in):m_N(N),m_F(F),m_arr_in(arr_in){};
    MKL_Complex8* operator()(oneapi::tbb::flow_control& fc){
        if(m_f<m_F){
            MKL_Complex8*  arr{(MKL_Complex8*)mkl_malloc(m_N*8,64)};
            memcpy(arr,m_arr_in+m_N*m_f++,m_N*8);
            std::cout<<"run src_body"<<'\n';
            return arr;
        }else{
            fc.stop();
            return nullptr;
        }
    }
};

int main(){
    MKL_LONG            N{1048576},F{500};
    MKL_Complex8*       arr_in{(MKL_Complex8*)mkl_malloc(N*F*8,64)};
    MKL_Complex8*       arr_out{(MKL_Complex8*)mkl_malloc(N*F*8,64)}; 
    oneapi::tbb::flow::graph g;
    std::atomic<int>    f_fft_node{},f_out_node{};
    
    // init test dataset
    for(MKL_LONG f=0;f<F;f++){
        for(MKL_LONG n=0;n<N;n++){
            arr_in[f*N+n].real=sin(2*M_PI*10*f*n*1e-6);
            arr_in[f*N+n].imag=0;
        }
    }

    // define node
    oneapi::tbb::flow::input_node<MKL_Complex8*> input_node(g,src_body(N,F,arr_in));
    oneapi::tbb::flow::function_node<MKL_Complex8*,MKL_Complex8*> compute_node(g,2,[&](MKL_Complex8* arr){
        for(MKL_LONG n=0;n<N;n++){
            arr[n].real++;
            arr[n].imag++;
        }
        return arr;
    });
    oneapi::tbb::flow::function_node<MKL_Complex8*,uint64_t> output_node(g,1,[&](MKL_Complex8* arr){
        memcpy(arr_out+N*f_out_node,arr,N*8);
        mkl_free(arr);
        return f_out_node++;
    });

    // connect edge
    oneapi::tbb::flow::make_edge(input_node,compute_node);
    oneapi::tbb::flow::make_edge(compute_node,output_node);

    // flow data
    std::chrono::high_resolution_clock::time_point t0{std::chrono::high_resolution_clock::now()};
    input_node.activate();
    g.wait_for_all();
    std::chrono::duration<float,std::milli> d{std::chrono::high_resolution_clock::now()-t0};
    std::cout<<"process time : "<<d.count()<<" ms"<<std::endl;

    mkl_free(arr_out);
    mkl_free(arr_in);
    return 0;
}

<CMakeLists.txt>

set(CMAKE_C_COMPILER icx)
set(CMAKE_CXX_COMPILER icx)
set(CMAKE_CXX_FLAGS_DEBUG "/Od /arch:AVX2")
# set(CMAKE_CXX_FLAGS_RELEASE "/O2 /arch:AVX2")
set(CMAKE_CXX_FLAGS_RELEASE "/Od /arch:AVX2")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "/O2 /arch:AVX2")
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
    set(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDebug)
    # set(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDebugDLL) # not working
elseif(CMAKE_BUILD_TYPE STREQUAL "Release") # not working
    set(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDebug)
    # set(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreaded)
elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") # not working
    set(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDebug)
endif()

cmake_minimum_required(VERSION 3.28.0)
project(test VERSION 0.0.0 LANGUAGES C CXX)
set(MKL_LINK dynamic)
set(MKL_THREADING sequential)
set(MKL_INTERFACE ilp64)
find_package(MKL CONFIG REQUIRED PATHS $ENV{MKLROOT}) # flag setting reference : MKLConfig.cmake
find_package(TBB REQUIRED COMPONENTS tbb tbbmalloc)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/bin)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/bin)


add_executable(test     
    bug_tbb_mkl_problem_windows.cpp
)
set_target_properties(test PROPERTIES 
    OUTPUT_NAME ${PROJECT_NAME}
)
target_link_libraries(test
    PRIVATE MKL::MKL
    PRIVATE TBB::tbb
)

[Kohn-Sham]

I tested compiler optimization flags yesterday.

I found /Ot flag makes this bug. Maybe /Os so do.

/O2 flag is same as /Og /Oi /Ot /Oy /Ob2 /GF /Gy. everything except /Ot works code well.

It looks like compiler related bug.

[Mark_L_Intel]

Hello @Kohn-Sham, the issue may need to be moved to Compiler realted Forum if it is compiler related bug.