diff --git a/include/isaac/driver/buffer.h b/include/isaac/driver/buffer.h
index 3e8bfb853..2782b2ba4 100644
--- a/include/isaac/driver/buffer.h
+++ b/include/isaac/driver/buffer.h
@@ -23,6 +23,7 @@ public:
   bool operator<(Buffer const &) const;
   bool operator==(Buffer const &) const;
   HANDLE_TYPE(cl::Buffer, CUdeviceptr)&  handle();
+  HANDLE_TYPE(cl::Buffer, CUdeviceptr) const &  handle() const;
 private:
   backend_type backend_;
   Context context_;
diff --git a/lib/driver/buffer.cpp b/lib/driver/buffer.cpp
index fde34364d..8743e1246 100644
--- a/lib/driver/buffer.cpp
+++ b/lib/driver/buffer.cpp
@@ -36,6 +36,9 @@ bool Buffer::operator<(Buffer const & other) const
 HANDLE_TYPE(cl::Buffer, CUdeviceptr) & Buffer::handle()
 { return h_; }
 
+HANDLE_TYPE(cl::Buffer, CUdeviceptr) const & Buffer::handle() const
+{ return h_; }
+
 }
 
 }
diff --git a/lib/driver/program.cpp b/lib/driver/program.cpp
index a381b1526..977eff3b5 100644
--- a/lib/driver/program.cpp
+++ b/lib/driver/program.cpp
@@ -17,13 +17,13 @@ namespace driver
 
 Program::Program(Context const & context, std::string const & source) : backend_(context.backend_), context_(context), source_(source), h_(backend_)
 {
+//  std::cout << source << std::endl;
   std::string cache_path = context.cache_path_;
   switch(backend_)
   {
 #ifdef ISAAC_WITH_CUDA
     case CUDA:
     {
-//      std::cout << source << std::endl;
 
       std::string prefix = context_.device_.name() + "cuda";
       std::string sha1 = tools::sha1(prefix + source);
diff --git a/lib/wrap/clBLAS.cpp b/lib/wrap/clBLAS.cpp
index 8c86c06c9..ce1915191 100644
--- a/lib/wrap/clBLAS.cpp
+++ b/lib/wrap/clBLAS.cpp
@@ -161,7 +161,7 @@ extern "C"
         is::array x(N, TYPE_ISAAC, cl::Buffer(mx), offx, incx);\
         clRetainMemObject(mx);\
         \
-        is::array y(N, TYPE_ISAAC, cl::Buffer(my), offy, incy);\
+        is::array y(M, TYPE_ISAAC, cl::Buffer(my), offy, incy);\
         clRetainMemObject(my);\
         \
         is::driver::Context const & context = A.context();\
diff --git a/tests/linalg/common.hpp b/tests/linalg/common.hpp
index 3f00ff301..42978dc22 100644
--- a/tests/linalg/common.hpp
+++ b/tests/linalg/common.hpp
@@ -10,6 +10,7 @@ template<class T> struct BLAS;
 template<> struct BLAS<float> { template<class FT, class DT> static FT F(FT SAXPY, DT ) { return SAXPY; } };
 template<> struct BLAS<double> { template<class FT, class DT> static DT F(FT , DT DAXPY) { return DAXPY; } };
 
+#define CHANDLE(X) (*X.data().handle().cl)()
 /*------ Simple Vector ---------*/
 template<class T>
 class simple_vector_base
diff --git a/tests/linalg/mreduction.cpp b/tests/linalg/mreduction.cpp
index 621fc6b29..cfacce2a1 100644
--- a/tests/linalg/mreduction.cpp
+++ b/tests/linalg/mreduction.cpp
@@ -1,6 +1,7 @@
 #include <cmath>
 #include "common.hpp"
 #include "isaac/array.h"
+#include "isaac/wrap/clBLAS.h"
 
 namespace ad = isaac;
 
@@ -17,6 +18,9 @@ void test_row_wise_reduction(T epsilon, simple_vector_base<T> & cy, simple_matri
   simple_vector<T> bufy(M);
   simple_vector<T> bufx(N);
 
+  ad::driver::CommandQueue queue = ad::driver::queues[y.context()][0];
+  cl_command_queue clqueue = (*queue.handle().cl)();
+
   T yi = 0, xi = 0;
 #define TEST_OPERATION(NAME, SIZE1, SIZE2, REDUCTION, ASSIGNMENT, GPU_REDUCTION, RES, BUF, CRES)\
   std::cout << NAME "..." << std::flush;\
@@ -38,6 +42,14 @@ void test_row_wise_reduction(T epsilon, simple_vector_base<T> & cy, simple_matri
   else\
     std::cout << std::endl;
 
+  ad::int_t offA = A.start()[0] + A.start()[1]*A.ld();
+
+//  TEST_OPERATION("GEMV(COL, NoTrans)", M, N, yi+=cA(i,j)*cx[j], cy[i] = yi,
+//                 BLAS<T>::F(clblasSgemv, clblasDgemv)(clblasColumnMajor, clblasNoTrans, M, N, 1, CHANDLE(A), offA, A.ld(),
+//                            CHANDLE(x), x.start()[0], x.stride()[0], 0, CHANDLE(y), y.start()[0], y.start()[1],
+//                            1, &clqueue, 0, NULL, NULL), y, bufy, cy);
+
+
   TEST_OPERATION("y = A.x", M, N, yi+=cA(i,j)*cx[j], cy[i] = yi, y = dot(A,x), y, bufy, cy);
   TEST_OPERATION("x = A'.y", N, M, xi+=cA(j,i)*cy[j], cx[i] = xi, x = dot(trans(A),y), x, bufx, cx);