diff --git a/bench/CMakeLists.txt b/bench/CMakeLists.txt
index 291a0fc93..9bcf227f3 100644
--- a/bench/CMakeLists.txt
+++ b/bench/CMakeLists.txt
@@ -1,7 +1,7 @@
 set(CMAKE_BUILD_TYPE Release)
 
 set(BLAS_DEF)
-set(BLAS_LIBS atidlas)
+set(BLAS_LIBS atidlas ${OPENCL_LIBRARIES})
 
 #CUBLAS
 find_package(CUDA)
@@ -28,7 +28,13 @@ endif()
 string(REPLACE ";" " " BLAS_DEF_STR "${BLAS_DEF}")
 
 foreach(PROG blas overhead)
-     add_executable(${PROG}-bench  ${PROG}.cpp)
-     target_link_libraries(${PROG}-bench ${BLAS_LIBS} ${OPENCL_LIBRARIES})
+    if(CUDA_FOUND)
+        set(CUPROG ${CMAKE_CURRENT_SOURCE_DIR}/${PROG}.cu)
+        file(COPY ${PROG}.cpp ${CUPROG})
+        cuda_add_executable(${PROG}-bench  ${CUPROG})
+    else()
+        add_executable(${PROG}-bench  ${PROG}.cpp)
+    endif()
+     target_link_libraries(${PROG}-bench ${BLAS_LIBS})
      set_target_properties(${PROG}-bench PROPERTIES COMPILE_FLAGS "-Wall -Wextra ${BLAS_DEF_STR}")
 endforeach(PROG)
diff --git a/bench/blas.cpp b/bench/blas.cpp
index e4127113f..de7585cb5 100644
--- a/bench/blas.cpp
+++ b/bench/blas.cpp
@@ -1,9 +1,15 @@
 #include "atidlas/array.h"
 #include "atidlas/tools/timer.hpp"
-#include "clAmdBlas.h"
 #include "common.hpp"
-#include "cblas.h"
-
+#ifdef BENCH_CLAMDBLAS
+  #include "clAmdBlas.h"
+#endif
+#ifdef BENCH_CBLAS
+  #include "cblas.h"
+#endif
+#ifdef BENCH_CUBLAS
+  #include <cublas.h>
+#endif
 #include <iomanip>
 #include <stdlib.h>
 #include <cmath>
@@ -12,6 +18,7 @@
 namespace ad = atidlas;
 typedef atidlas::int_t int_t;
 
+template<class T>
 void bench(ad::numeric_type dtype)
 {
   unsigned int dtsize = ad::size_of(dtype);
@@ -57,6 +64,15 @@ void bench(ad::numeric_type dtype)
     atidlas::copy(x, cx);
     atidlas::copy(y, cy);
     BENCHMARK(cblas_saxpy(N, 1, cx.data(), 1, cy.data(), 1), bandwidth(3*N, tres, dtsize));
+#endif
+    /* CuBLAS */
+#ifdef BENCH_CUBLAS
+    T *cux, *cuy;
+    cudaMalloc((void**) &cux, N * sizeof(T));
+    cudaMalloc((void**) &cuy, N * sizeof(T));
+    BENCHMARK(cublasSaxpy(N, 2, x, 1, y, 1), bandwidth(3*N, tres, dtsize))
+    cudaFree(cux);
+    cudaFree(cuy);
 #endif
     std::cout << std::endl;
   }
@@ -175,7 +191,7 @@ int main(int argc, char* argv[])
   atidlas::cl::default_context_idx = device_idx;
   std::cout << "#Benchmark : BLAS" << std::endl;
   std::cout << "#----------------" << std::endl;
-  bench(ad::FLOAT_TYPE);
+  bench<float>(ad::FLOAT_TYPE);
 
 #ifdef BENCH_CLAMDBLAS
   clAmdBlasTeardown();