Converting all sycl buffers to uninitialised device only buffers; adding memcpyHostToDevice and memcpyDeviceToHost on syclDevice; modifying all examples to obey the new rules; moving sycl queue creating to the device based on Benoit suggestion; removing the sycl specefic condition for returning m_result in TensorReduction.h according to Benoit suggestion.

5 files changed, 167 insertions, 190 deletions

diff --git a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
index ecebf7d68..7201bfe37 100644
--- a/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_broadcast_sycl.cpp
@@ -25,55 +25,50 @@ using Eigen::SyclDevice;
 using Eigen::Tensor;
 using Eigen::TensorMap;
 
-// Types used in tests:
-using TestTensor = Tensor<float, 3>;
-using TestTensorMap = TensorMap<Tensor<float, 3>>;
-static void test_broadcast_sycl(){
+static void test_broadcast_sycl(const Eigen::SyclDevice &sycl_device){
 
-	cl::sycl::gpu_selector s;
-  cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-    for (const auto& e : l) {
-      try {
-        std::rethrow_exception(e);
-      } catch (cl::sycl::exception e) {
-        std::cout << e.what() << std::endl;
-      }
-    }
-  });
-	SyclDevice sycl_device(q);
-	// BROADCAST test:
-	array<int, 4> in_range   = {{2, 3, 5, 7}};
-	array<int, in_range.size()> broadcasts = {{2, 3, 1, 4}};
-	array<int, in_range.size()> out_range;  // = in_range * broadcasts
-	for (size_t i = 0; i < out_range.size(); ++i)
-		out_range[i] = in_range[i] * broadcasts[i];
+  // BROADCAST test:
+  array<int, 4> in_range   = {{2, 3, 5, 7}};
+  array<int, 4> broadcasts = {{2, 3, 1, 4}};
+  array<int, 4> out_range;  // = in_range * broadcasts
+  for (size_t i = 0; i < out_range.size(); ++i)
+    out_range[i] = in_range[i] * broadcasts[i];
+
+  Tensor<float, 4>  input(in_range);
+  Tensor<float, 4> out(out_range);
 
-	Tensor<float, in_range.size()>  input(in_range);
-	Tensor<float, out_range.size()> output(out_range);
+  for (size_t i = 0; i < in_range.size(); ++i)
+    VERIFY_IS_EQUAL(out.dimension(i), out_range[i]);
 
-	for (int i = 0; i < input.size(); ++i)
-		input(i) = static_cast<float>(i);
 
-	TensorMap<decltype(input)>  gpu_in(input.data(), in_range);
-	TensorMap<decltype(output)> gpu_out(output.data(), out_range);
-		gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts);
-		sycl_device.deallocate(output.data());
+  for (int i = 0; i < input.size(); ++i)
+    input(i) = static_cast<float>(i);
 
-	for (size_t i = 0; i < in_range.size(); ++i)
-		VERIFY_IS_EQUAL(output.dimension(i), out_range[i]);
+  float * gpu_in_data  = static_cast<float*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_out_data  = static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float)));
 
-	for (int i = 0; i < 4; ++i) {
-		for (int j = 0; j < 9; ++j) {
-			for (int k = 0; k < 5; ++k) {
-				for (int l = 0; l < 28; ++l) {
-					VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), output(i,j,k,l));
-				}
-			}
-		}
-	}
-	printf("Broadcast Test Passed\n");
+  TensorMap<Tensor<float, 4>>  gpu_in(gpu_in_data, in_range);
+  TensorMap<Tensor<float, 4>> gpu_out(gpu_out_data, out_range);
+  sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(float));
+  gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts);
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        for (int l = 0; l < 28; ++l) {
+          VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), out(i,j,k,l));
+        }
+      }
+    }
+  }
+  printf("Broadcast Test Passed\n");
+  sycl_device.deallocate(gpu_in_data);
+  sycl_device.deallocate(gpu_out_data);
 }
 
 void test_cxx11_tensor_broadcast_sycl() {
-  CALL_SUBTEST(test_broadcast_sycl());
+  cl::sycl::gpu_selector s;
+  Eigen::SyclDevice sycl_device(s);
+  CALL_SUBTEST(test_broadcast_sycl(sycl_device));
 }
diff --git a/unsupported/test/cxx11_tensor_device_sycl.cpp b/unsupported/test/cxx11_tensor_device_sycl.cpp
index f54fc8786..7f79753c5 100644
--- a/unsupported/test/cxx11_tensor_device_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_device_sycl.cpp
@@ -20,20 +20,12 @@
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
 
-void test_device_sycl() {
-  cl::sycl::gpu_selector s;
-  cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-    for (const auto& e : l) {
-      try {
-        std::rethrow_exception(e);
-      } catch (cl::sycl::exception e) {
-        std::cout << e.what() << std::endl;
-      }
-    }
-  });
-  Eigen::SyclDevice sycl_device(q);
-	printf("Helo from ComputeCpp: Device Exists\n");
+void test_device_sycl(const Eigen::SyclDevice &sycl_device) {
+  std::cout <<"Helo from ComputeCpp: the requested device exists and the device name is : "
+    << sycl_device.m_queue.get_device(). template get_info<cl::sycl::info::device::name>() <<std::endl;;
 }
 void test_cxx11_tensor_device_sycl() {
-  CALL_SUBTEST(test_device_sycl());
+  cl::sycl::gpu_selector s;
+  Eigen::SyclDevice sycl_device(s);
+  CALL_SUBTEST(test_device_sycl(sycl_device));
 }
diff --git a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
index 182ec7fa8..5690da723 100644
--- a/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_forced_eval_sycl.cpp
@@ -22,18 +22,7 @@
 
 using Eigen::Tensor;
 
-void test_forced_eval_sycl() {
-	cl::sycl::gpu_selector s;
-  cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-    for (const auto& e : l) {
-      try {
-        std::rethrow_exception(e);
-      } catch (cl::sycl::exception e) {
-        std::cout << e.what() << std::endl;
-      }
-    }
-  });
-	SyclDevice sycl_device(q);
+void test_forced_eval_sycl(const Eigen::SyclDevice &sycl_device) {
 
   int sizeDim1 = 100;
   int sizeDim2 = 200;
@@ -43,17 +32,22 @@ void test_forced_eval_sycl() {
   Eigen::Tensor<float, 3> in2(tensorRange);
   Eigen::Tensor<float, 3> out(tensorRange);
 
+  float * gpu_in1_data  = static_cast<float*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_in2_data  = static_cast<float*>(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_out_data =  static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float)));
+
   in1 = in1.random() + in1.constant(10.0f);
   in2 = in2.random() + in2.constant(10.0f);
 
-	// creating TensorMap from tensor
-  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in1(in1.data(), tensorRange);
-  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in2(in2.data(), tensorRange);
-  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_out(out.data(), tensorRange);
-
+  // creating TensorMap from tensor
+  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange);
+  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange);
+  Eigen::TensorMap<Eigen::Tensor<float, 3>> gpu_out(gpu_out_data, tensorRange);
+  sycl_device.memcpyHostToDevice(gpu_in1_data, in1.data(),(in1.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in1.dimensions().TotalSize())*sizeof(float));
   /// c=(a+b)*b
-	gpu_out.device(sycl_device) =(gpu_in1 + gpu_in2).eval() * gpu_in2;
-	sycl_device.deallocate(out.data());
+  gpu_out.device(sycl_device) =(gpu_in1 + gpu_in2).eval() * gpu_in2;
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -62,7 +56,15 @@ void test_forced_eval_sycl() {
       }
     }
   }
-	printf("(a+b)*b Test Passed\n");
+  printf("(a+b)*b Test Passed\n");
+  sycl_device.deallocate(gpu_in1_data);
+  sycl_device.deallocate(gpu_in2_data);
+  sycl_device.deallocate(gpu_out_data);
+
 }
 
-void test_cxx11_tensor_forced_eval_sycl() { CALL_SUBTEST(test_forced_eval_sycl()); }
+void test_cxx11_tensor_forced_eval_sycl() {
+  cl::sycl::gpu_selector s;
+  Eigen::SyclDevice sycl_device(s);
+  CALL_SUBTEST(test_forced_eval_sycl(sycl_device));
+}
diff --git a/unsupported/test/cxx11_tensor_reduction_sycl.cpp b/unsupported/test/cxx11_tensor_reduction_sycl.cpp
index bd09744a6..a9ef82907 100644
--- a/unsupported/test/cxx11_tensor_reduction_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_reduction_sycl.cpp
@@ -22,126 +22,117 @@
 
 
 
-static void test_full_reductions_sycl() {
-
-
-  cl::sycl::gpu_selector s;
-    cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-      for (const auto& e : l) {
-        try {
-          std::rethrow_exception(e);
-        } catch (cl::sycl::exception e) {
-          std::cout << e.what() << std::endl;
-        }
-      }
-    });
-  Eigen::SyclDevice sycl_device(q);
+static void test_full_reductions_sycl(const Eigen::SyclDevice&  sycl_device) {
 
   const int num_rows = 452;
   const int num_cols = 765;
   array<int, 2> tensorRange = {{num_rows, num_cols}};
 
   Tensor<float, 2> in(tensorRange);
+  Tensor<float, 0> full_redux;
+  Tensor<float, 0> full_redux_gpu;
+
   in.setRandom();
 
-  Tensor<float, 0> full_redux;
-  Tensor<float, 0> full_redux_g;
   full_redux = in.sum();
-  float* out_data = (float*)sycl_device.allocate(sizeof(float));
-  TensorMap<Tensor<float, 2> >  in_gpu(in.data(), tensorRange);
-  TensorMap<Tensor<float, 0> >  full_redux_gpu(out_data);
-  full_redux_gpu.device(sycl_device) = in_gpu.sum();
-  sycl_device.deallocate(out_data);
-  // Check that the CPU and GPU reductions return the same result.
-  VERIFY_IS_APPROX(full_redux_gpu(), full_redux());
 
-}
+  float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float)));
+  float* gpu_out_data =(float*)sycl_device.allocate(sizeof(float));
 
+  TensorMap<Tensor<float, 2> >  in_gpu(gpu_in_data, tensorRange);
+  TensorMap<Tensor<float, 0> >  out_gpu(gpu_out_data);
 
-static void test_first_dim_reductions_sycl() {
+  sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float));
+  out_gpu.device(sycl_device) = in_gpu.sum();
+  sycl_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_data, sizeof(float));
+  // Check that the CPU and GPU reductions return the same result.
+  VERIFY_IS_APPROX(full_redux_gpu(), full_redux());
 
+  sycl_device.deallocate(gpu_in_data);
+  sycl_device.deallocate(gpu_out_data);
+}
 
-  cl::sycl::gpu_selector s;
-    cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-      for (const auto& e : l) {
-        try {
-          std::rethrow_exception(e);
-        } catch (cl::sycl::exception e) {
-          std::cout << e.what() << std::endl;
-        }
-      }
-    });
-  Eigen::SyclDevice sycl_device(q);
+static void test_first_dim_reductions_sycl(const Eigen::SyclDevice& sycl_device) {
 
   int dim_x = 145;
   int dim_y = 1;
   int dim_z = 67;
 
   array<int, 3> tensorRange = {{dim_x, dim_y, dim_z}};
-
-  Tensor<float, 3> in(tensorRange);
-  in.setRandom();
   Eigen::array<int, 1> red_axis;
   red_axis[0] = 0;
-  Tensor<float, 2> redux = in.sum(red_axis);
   array<int, 2> reduced_tensorRange = {{dim_y, dim_z}};
-  Tensor<float, 2> redux_g(reduced_tensorRange);
-  TensorMap<Tensor<float, 3> >  in_gpu(in.data(), tensorRange);
-  float* out_data = (float*)sycl_device.allocate(dim_y*dim_z*sizeof(float));
-  TensorMap<Tensor<float, 2> >  redux_gpu(out_data, dim_y, dim_z );
-  redux_gpu.device(sycl_device) = in_gpu.sum(red_axis);
 
-  sycl_device.deallocate(out_data);
-  // Check that the CPU and GPU reductions return the same result.
-  for(int j=0; j<dim_y; j++ )
-    for(int k=0; k<dim_z; k++ )
-      VERIFY_IS_APPROX(redux_gpu(j,k), redux(j,k));
-}
+  Tensor<float, 3> in(tensorRange);
+  Tensor<float, 2> redux(reduced_tensorRange);
+  Tensor<float, 2> redux_gpu(reduced_tensorRange);
+
+  in.setRandom();
 
+  redux= in.sum(red_axis);
 
-static void test_last_dim_reductions_sycl() {
+  float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float)));
+  float* gpu_out_data = static_cast<float*>(sycl_device.allocate(redux_gpu.dimensions().TotalSize()*sizeof(float)));
 
+  TensorMap<Tensor<float, 3> >  in_gpu(gpu_in_data, tensorRange);
+  TensorMap<Tensor<float, 2> >  out_gpu(gpu_out_data, reduced_tensorRange);
 
-  cl::sycl::gpu_selector s;
-    cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-      for (const auto& e : l) {
-        try {
-          std::rethrow_exception(e);
-        } catch (cl::sycl::exception e) {
-          std::cout << e.what() << std::endl;
-        }
-      }
-    });
-  Eigen::SyclDevice sycl_device(q);
+  sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float));
+  out_gpu.device(sycl_device) = in_gpu.sum(red_axis);
+  sycl_device.memcpyDeviceToHost(redux_gpu.data(), gpu_out_data, redux_gpu.dimensions().TotalSize()*sizeof(float));
+
+  // Check that the CPU and GPU reductions return the same result.
+  for(int j=0; j<reduced_tensorRange[0]; j++ )
+    for(int k=0; k<reduced_tensorRange[1]; k++ )
+      VERIFY_IS_APPROX(redux_gpu(j,k), redux(j,k));
+
+  sycl_device.deallocate(gpu_in_data);
+  sycl_device.deallocate(gpu_out_data);
+}
+
+static void test_last_dim_reductions_sycl(const Eigen::SyclDevice &sycl_device) {
 
   int dim_x = 567;
   int dim_y = 1;
   int dim_z = 47;
 
   array<int, 3> tensorRange = {{dim_x, dim_y, dim_z}};
-
-  Tensor<float, 3> in(tensorRange);
-  in.setRandom();
   Eigen::array<int, 1> red_axis;
   red_axis[0] = 2;
-  Tensor<float, 2> redux = in.sum(red_axis);
   array<int, 2> reduced_tensorRange = {{dim_x, dim_y}};
-  Tensor<float, 2> redux_g(reduced_tensorRange);
-  TensorMap<Tensor<float, 3> >  in_gpu(in.data(), tensorRange);
-  float* out_data = (float*)sycl_device.allocate(dim_x*dim_y*sizeof(float));
-  TensorMap<Tensor<float, 2> >  redux_gpu(out_data, dim_x, dim_y );
-  redux_gpu.device(sycl_device) = in_gpu.sum(red_axis);
 
-  sycl_device.deallocate(out_data);
+  Tensor<float, 3> in(tensorRange);
+  Tensor<float, 2> redux(reduced_tensorRange);
+  Tensor<float, 2> redux_gpu(reduced_tensorRange);
+
+  in.setRandom();
+
+  redux= in.sum(red_axis);
+
+  float* gpu_in_data = static_cast<float*>(sycl_device.allocate(in.dimensions().TotalSize()*sizeof(float)));
+  float* gpu_out_data = static_cast<float*>(sycl_device.allocate(redux_gpu.dimensions().TotalSize()*sizeof(float)));
+
+  TensorMap<Tensor<float, 3> >  in_gpu(gpu_in_data, tensorRange);
+  TensorMap<Tensor<float, 2> >  out_gpu(gpu_out_data, reduced_tensorRange);
+
+  sycl_device.memcpyHostToDevice(gpu_in_data, in.data(),(in.dimensions().TotalSize())*sizeof(float));
+  out_gpu.device(sycl_device) = in_gpu.sum(red_axis);
+  sycl_device.memcpyDeviceToHost(redux_gpu.data(), gpu_out_data, redux_gpu.dimensions().TotalSize()*sizeof(float));
   // Check that the CPU and GPU reductions return the same result.
-  for(int j=0; j<dim_x; j++ )
-    for(int k=0; k<dim_y; k++ )
+  for(int j=0; j<reduced_tensorRange[0]; j++ )
+    for(int k=0; k<reduced_tensorRange[1]; k++ )
       VERIFY_IS_APPROX(redux_gpu(j,k), redux(j,k));
+
+  sycl_device.deallocate(gpu_in_data);
+  sycl_device.deallocate(gpu_out_data);
+
 }
 
 void test_cxx11_tensor_reduction_sycl() {
-  CALL_SUBTEST((test_full_reductions_sycl()));
-  CALL_SUBTEST((test_first_dim_reductions_sycl()));
-  CALL_SUBTEST((test_last_dim_reductions_sycl()));
+  cl::sycl::gpu_selector s;
+  Eigen::SyclDevice sycl_device(s);
+  CALL_SUBTEST((test_full_reductions_sycl(sycl_device)));
+  CALL_SUBTEST((test_first_dim_reductions_sycl(sycl_device)));
+  CALL_SUBTEST((test_last_dim_reductions_sycl(sycl_device)));
 
 }
diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp
index 0f66cd8f0..6a9c33422 100644
--- a/unsupported/test/cxx11_tensor_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_sycl.cpp
@@ -27,42 +27,33 @@ using Eigen::SyclDevice;
 using Eigen::Tensor;
 using Eigen::TensorMap;
 
-// Types used in tests:
-using TestTensor = Tensor<float, 3>;
-using TestTensorMap = TensorMap<Tensor<float, 3>>;
-
-void test_sycl_cpu() {
-  cl::sycl::gpu_selector s;
-  cl::sycl::queue q(s, [=](cl::sycl::exception_list l) {
-    for (const auto& e : l) {
-      try {
-        std::rethrow_exception(e);
-      } catch (cl::sycl::exception e) {
-        std::cout << e.what() << std::endl;
-      }
-    }
-  });
-  SyclDevice sycl_device(q);
+void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) {
 
   int sizeDim1 = 100;
   int sizeDim2 = 100;
   int sizeDim3 = 100;
   array<int, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
-  TestTensor in1(tensorRange);
-  TestTensor in2(tensorRange);
-  TestTensor in3(tensorRange);
-  TestTensor out(tensorRange);
-  in1 = in1.random();
+  Tensor<float, 3> in1(tensorRange);
+  Tensor<float, 3> in2(tensorRange);
+  Tensor<float, 3> in3(tensorRange);
+  Tensor<float, 3> out(tensorRange);
+
   in2 = in2.random();
   in3 = in3.random();
-  TestTensorMap gpu_in1(in1.data(), tensorRange);
-  TestTensorMap gpu_in2(in2.data(), tensorRange);
-  TestTensorMap gpu_in3(in3.data(), tensorRange);
-  TestTensorMap gpu_out(out.data(), tensorRange);
+
+  float * gpu_in1_data  = static_cast<float*>(sycl_device.allocate(in1.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_in2_data  = static_cast<float*>(sycl_device.allocate(in2.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_in3_data  = static_cast<float*>(sycl_device.allocate(in3.dimensions().TotalSize()*sizeof(float)));
+  float * gpu_out_data =  static_cast<float*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(float)));
+
+  TensorMap<Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange);
+  TensorMap<Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange);
+  TensorMap<Tensor<float, 3>> gpu_in3(gpu_in3_data, tensorRange);
+  TensorMap<Tensor<float, 3>> gpu_out(gpu_out_data, tensorRange);
 
   /// a=1.2f
   gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f);
-  sycl_device.deallocate(in1.data());
+  sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -74,7 +65,7 @@ void test_sycl_cpu() {
 
   /// a=b*1.2f
   gpu_out.device(sycl_device) = gpu_in1 * 1.2f;
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -86,8 +77,9 @@ void test_sycl_cpu() {
   printf("a=b*1.2f Test Passed\n");
 
   /// c=a*b
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.dimensions().TotalSize())*sizeof(float));
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in2;
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -101,7 +93,7 @@ void test_sycl_cpu() {
 
   /// c=a+b
   gpu_out.device(sycl_device) = gpu_in1 + gpu_in2;
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -115,7 +107,7 @@ void test_sycl_cpu() {
 
   /// c=a*a
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in1;
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -125,12 +117,11 @@ void test_sycl_cpu() {
       }
     }
   }
-
   printf("c= a*a Test Passed\n");
 
   //a*3.14f + b*2.7f
   gpu_out.device(sycl_device) =  gpu_in1 * gpu_in1.constant(3.14f) + gpu_in2 * gpu_in2.constant(2.7f);
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -143,8 +134,9 @@ void test_sycl_cpu() {
   printf("a*3.14f + b*2.7f Test Passed\n");
 
   ///d= (a>0.5? b:c)
+  sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.dimensions().TotalSize())*sizeof(float));
   gpu_out.device(sycl_device) =(gpu_in1 > gpu_in1.constant(0.5f)).select(gpu_in2, gpu_in3);
-  sycl_device.deallocate(out.data());
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -155,8 +147,13 @@ void test_sycl_cpu() {
     }
   }
   printf("d= (a>0.5? b:c) Test Passed\n");
-
+  sycl_device.deallocate(gpu_in1_data);
+  sycl_device.deallocate(gpu_in2_data);
+  sycl_device.deallocate(gpu_in3_data);
+  sycl_device.deallocate(gpu_out_data);
 }
 void test_cxx11_tensor_sycl() {
-  CALL_SUBTEST(test_sycl_cpu());
+  cl::sycl::gpu_selector s;
+  Eigen::SyclDevice sycl_device(s);
+  CALL_SUBTEST(test_sycl_cpu(sycl_device));
 }