diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4cd7b13472961881d145be297834d22768ccdaa1..d1f4c7a64e63a4cdf51e8b11275bf763c2fb4792 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -99,7 +99,7 @@ if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
     endif()
 elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
     # 去掉 -DMKL_LP64 -DEIGEN_USE_MKL_ALL -m64
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE} -pthread -fvisibility=hidden -fvisibility-inlines-hidden")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE} -I ../submods/plink-ng/2.0/simde/ -pthread -fvisibility=hidden -fvisibility-inlines-hidden")
     set(CMAKE_CXX_FLAGS_RELEASE  "${CMAKE_CXX_FLAGS_RELEASE} -O3 -DNDEBUG") #-flto 
     set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} -O0 -fno-inline -fno-implicit-inline-templates -g3")
     add_compile_options(-fopenmp)
@@ -148,20 +148,14 @@ if(APPLE)
 else()
     if (CMAKE_CXX_COMPILER_ID MATCHES "Clang")
         # error keeps only
-        #target_link_libraries(gcta64 ${libs_list} mainV1 z ${KML_LIB
-    } iomp5 pthread m dl )
-        #target_link_libraries(gcta64 mainV1 ${libs_list} z ${KML_LIB
-    } pthread m dl)
+        #target_link_libraries(gcta64 ${libs_list} mainV1 z ${KML_LIB} iomp5 pthread m dl )
+        #target_link_libraries(gcta64 mainV1 ${libs_list} z ${KML_LIB} pthread m dl)
         # Dynamic link works
-        #target_link_libraries(gcta64 mainV1 ${libs_list} -Wl,-Bstatic z -Wl,--start-group ${KML_LIB
-    } -Wl,--end-group iomp5 -Wl,-Bdynamic pthread m dl)
-        target_link_libraries(gcta64 mainV1 ${libs_list} Pgenlib -static z sqlite3 zstd -Wl,--start-group ${KML_LIB
-    } -Wl,--end-group iomp5 -Wl,--whole-archive -lpthread -Wl,--no-whole-archive m dl)
+        #target_link_libraries(gcta64 mainV1 ${libs_list} -Wl,-Bstatic z -Wl,--start-group ${KML_LIB} -Wl,--end-group iomp5 -Wl,-Bdynamic pthread m dl)
+        target_link_libraries(gcta64 mainV1 ${libs_list} Pgenlib -static z sqlite3 zstd -Wl,--start-group ${KML_LIB} -Wl,--end-group iomp5 -Wl,--whole-archive -lpthread -Wl,--no-whole-archive m dl)
     else()
-        #target_link_libraries(gcta64 mainV1 ${libs_list} -Wl,-Bstatic z -Wl,--start-group ${KML_LIB
-    } -Wl,--end-group gomp -Wl,-Bdynamic pthread m dl)
-        target_link_libraries(gcta64 mainV1 ${libs_list} Pgenlib -static z sqlite3 zstd -Wl,--start-group ${KML_LIB
-    } -Wl,--end-group gomp -Wl,--whole-archive -lpthread -Wl,--no-whole-archive m dl)
+        #target_link_libraries(gcta64 mainV1 ${libs_list} -Wl,-Bstatic z -Wl,--start-group ${KML_LIB} -Wl,--end-group gomp -Wl,-Bdynamic pthread m dl)
+        target_link_libraries(gcta64 mainV1 ${libs_list} Pgenlib -static z sqlite3 zstd -Wl,--start-group ${KML_LIB} -Wl,--end-group gomp -Wl,--whole-archive -lpthread -Wl,--no-whole-archive m dl)
     endif()
 endif()
 
diff --git a/include/Matrix.hpp b/include/Matrix.hpp
index f31aa83369efd5033ad5c172cd46522aa6a1f96c..8e4b9c205bfcdd77c6e45ff71482ebffc6a6fe01 100644
--- a/include/Matrix.hpp
+++ b/include/Matrix.hpp
@@ -2,7 +2,7 @@
 #define GCTA2_MATRIX_H
 
 #include <Eigen/Eigen>
-#include <mkl.h>
+#include <kfft.h>
 #include <iostream>
 #include <Logger.h>
 
diff --git a/main/gcta.h b/main/gcta.h
index a6049e8453dbf07b7ff122755faa211edc62f6d7..b870df5c0724dfe956a9d2a118e82c106e620a9f 100644
--- a/main/gcta.h
+++ b/main/gcta.h
@@ -37,8 +37,8 @@
 #include <unsupported/Eigen/SparseExtra>
 #include <unsupported/Eigen/IterativeSolvers>
 #include <omp.h>
-#include <mkl_cblas.h>
-#include <mkl_lapack.h>
+#include <kblas.h>
+#include <klapack.h>
 #include "Logger.h"
 #include "Matrix.hpp"
 
diff --git a/src/GRM.cpp b/src/GRM.cpp
index 15cfedc03e99a7f8e82476db8f6497d1fc225d5d..78c7357ef7f1aa2b795741bb487401a2ea56a104 100644
--- a/src/GRM.cpp
+++ b/src/GRM.cpp
@@ -33,7 +33,7 @@
 #include <boost/algorithm/string.hpp>
 #include <boost/algorithm/string/join.hpp>
 #include <sstream>
-#include <mkl.h>
+#include <kfft.h>
 #include <csignal>
 
 using std::to_string;
diff --git a/src/LD.cpp b/src/LD.cpp
index d34fa24c6547708fcfe677139bfc3fde3fef68d9..f83134188ace8f9298351020eff5c630a24532fe 100644
--- a/src/LD.cpp
+++ b/src/LD.cpp
@@ -7,7 +7,7 @@
 #include <string>
 #include <functional>
 #include <omp.h>
-#include <mkl.h>
+#include <kfft.h>
 #include <cstdio>
 #include <algorithm>
 
diff --git a/src/StatLib.cpp b/src/StatLib.cpp
index 4656ed2b02760a1815a8eb127bb5f0172f1c383c..fb2606ade5823ccd55e8f5321153dd5a79b85fbb 100644
--- a/src/StatLib.cpp
+++ b/src/StatLib.cpp
@@ -18,7 +18,7 @@
 #include <cstdio>
 #include <limits>
 #include <boost/math/distributions/chi_squared.hpp>
-#include <mkl.h>
+#include <kfft.h>
 
 using namespace boost::math;
 
diff --git a/test/tests/CMakeLists.txt b/test/tests/CMakeLists.txt
index 8637b164c5ba17b422830d58668bc830c06cb15a..5bac532a5716f3599828e3c93e7e2e54c3ac0306 100644
--- a/test/tests/CMakeLists.txt
+++ b/test/tests/CMakeLists.txt
@@ -15,7 +15,7 @@ function(addTestItem test_name source lib_name run_param)
         ${lib_name} 
         libgtest
         libgmock
-        -Wl,-Bstatic z -Wl,--start-group ${MKL_LIB} -Wl,--end-group gomp -Wl,-Bdynamic pthread m dl
+        -Wl,-Bstatic z -Wl,--start-group ${KML_LIB} -Wl,--end-group gomp -Wl,-Bdynamic pthread m dl
         
     )
     add_test(${test_name} ${test_name} "${run_param}")