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.

9 files changed, 241 insertions, 271 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 4231a11ff..8333301ea 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -16,95 +16,93 @@
 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
 
 namespace Eigen {
-/// \struct BufferT is used to specialise add_sycl_buffer function for
-//  two types of buffer we have. When the MapAllocator is true, we create the
-//  sycl buffer with MapAllocator.
-/// We have to const_cast the input pointer in order to work around the fact
-/// that sycl does not accept map allocator for const pointer.
-template <typename T, bool MapAllocator>
-struct BufferT {
-  using Type = cl::sycl::buffer<T, 1, cl::sycl::map_allocator<T>>;
-  static inline void add_sycl_buffer(const T *ptr, size_t num_bytes,std::map<const void *, std::shared_ptr<void>> &buffer_map) {
-    buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>(ptr, std::shared_ptr<void>(std::make_shared<Type>(Type(const_cast<T *>(ptr), cl::sycl::range<1>(num_bytes))))));
-  }
-};
-
-/// specialisation of the \ref BufferT when the MapAllocator is false. In this
-/// case we only create the device-only buffer.
-template <typename T>
-struct BufferT<T, false> {
-  using Type = cl::sycl::buffer<T, 1>;
-  static inline void add_sycl_buffer(const T *ptr, size_t num_bytes, std::map<const void *, std::shared_ptr<void>> &buffer_map) {
-    buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>(ptr, std::shared_ptr<void>(std::make_shared<Type>(Type(cl::sycl::range<1>(num_bytes))))));
-  }
-};
-
 struct SyclDevice {
   /// class members
   /// sycl queue
-  cl::sycl::queue &m_queue;
+  mutable cl::sycl::queue m_queue;
   /// std::map is the container used to make sure that we create only one buffer
-  /// per pointer. The lifespan of the buffer
-  /// now depends on the lifespan of SyclDevice. If a non-read-only pointer is
-  /// needed to be accessed on the host we should manually deallocate it.
+  /// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice.
+  /// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it.
   mutable std::map<const void *, std::shared_ptr<void>> buffer_map;
-
-  SyclDevice(cl::sycl::queue &q) : m_queue(q) {}
+  /// creating device by using selector
+  template<typename dev_Selector> SyclDevice(dev_Selector s)
+  :m_queue(cl::sycl::queue(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;
+        }
+    }
+  })) {}
   // destructor
   ~SyclDevice() { deallocate_all(); }
 
-  template <typename T>
-  void deallocate(const T *p) const {
+  template <typename T> void deallocate(T *p) const {
     auto it = buffer_map.find(p);
     if (it != buffer_map.end()) {
       buffer_map.erase(it);
+      internal::aligned_free(p);
+    }
+  }
+  void deallocate_all() const {
+    std::map<const void *, std::shared_ptr<void>>::iterator it=buffer_map.begin();
+    while (it!=buffer_map.end()) {
+      auto p=it->first;
+      buffer_map.erase(it);
+      internal::aligned_free(const_cast<void*>(p));
+      it=buffer_map.begin();
     }
+    buffer_map.clear();
   }
-  void deallocate_all() const { buffer_map.clear(); }
 
   /// creation of sycl accessor for a buffer. This function first tries to find
-  /// the buffer in the buffer_map.
-  /// If found it gets the accessor from it, if not, the function then adds an
-  /// entry by creating a sycl buffer
-  /// for that particular pointer.
-  template <cl::sycl::access::mode AcMd, bool MapAllocator, typename T>
-  inline cl::sycl::accessor<T, 1, AcMd, cl::sycl::access::target::global_buffer>
+  /// the buffer in the buffer_map. If found it gets the accessor from it, if not,
+  ///the function then adds an entry by creating a sycl buffer for that particular pointer.
+  template <cl::sycl::access::mode AcMd, typename T> inline cl::sycl::accessor<T, 1, AcMd, cl::sycl::access::target::global_buffer>
   get_sycl_accessor(size_t num_bytes, cl::sycl::handler &cgh, const T * ptr) const {
-    return (get_sycl_buffer<MapAllocator,T>(num_bytes, ptr).template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
+    return (get_sycl_buffer<T>(num_bytes, ptr)->template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
   }
 
-template <bool MapAllocator, typename T>
- inline typename BufferT<T, MapAllocator>::Type
-  get_sycl_buffer(size_t num_bytes,const T * ptr) const {
-    if(MapAllocator && !ptr){
-      eigen_assert("pointer with map_Allocator cannot be null. Please initialise the input pointer"); }
-    auto it = buffer_map.find(ptr);
-    if (it == buffer_map.end()) {
-      BufferT<T, MapAllocator>::add_sycl_buffer(ptr, num_bytes, buffer_map);
-    }
-    return (*((typename BufferT<T, MapAllocator>::Type*)((buffer_map.at(ptr).get()))));
+  template<typename T> inline  std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> add_sycl_buffer(const T *ptr, size_t num_bytes) const {
+    using Type = cl::sycl::buffer<T, 1>;
+    std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> ret = buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>(ptr, std::shared_ptr<void>(new Type(cl::sycl::range<1>(num_bytes)),
+      [](void *dataMem) { delete static_cast<Type*>(dataMem); })));
+    (static_cast<Type*>(buffer_map.at(ptr).get()))->set_final_data(nullptr);
+    return ret;
+  }
+
+  template <typename T> inline cl::sycl::buffer<T, 1>* get_sycl_buffer(size_t num_bytes,const T * ptr) const {
+    return static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(ptr, num_bytes).first->second.get());
   }
 
   /// allocating memory on the cpu
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
-    return internal::aligned_malloc(num_bytes);
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t) const {
+    return internal::aligned_malloc(8);
   }
 
   // some runtime conditions that can be applied here
   bool isDeviceSuitable() const { return true; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void *buffer) const {
-    internal::aligned_free(buffer);
-  }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void *dst, const void *src, size_t n) const {
     ::memcpy(dst, src, n);
   }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(void *dst, const void *src, size_t n) const {
-    memcpy(dst, src, n);
+
+  template<typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(T *dst, const T *src, size_t n) const {
+    auto host_acc= (static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(dst, n).first->second.get()))-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>();
+    memcpy(host_acc.get_pointer(), src, n);
   }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const void *src, size_t n) const {
-    memcpy(dst, src, n);
+ /// whith the current implementation of sycl, the data is copied twice from device to host. This will be fixed soon.
+  template<typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(T *dst, const T *src, size_t n) const {
+    auto it = buffer_map.find(src);
+    if (it != buffer_map.end()) {
+      auto host_acc= (static_cast<cl::sycl::buffer<T, 1>*>(it->second.get()))-> template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::host_buffer>();
+      memcpy(dst,host_acc.get_pointer(),  n);
+    } else{
+      eigen_assert("no device memory found. The memory might be destroyed before creation");
+    }
   }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void *buffer, int c, size_t n) const {
     ::memset(buffer, c, n);
   }
@@ -112,6 +110,7 @@ template <bool MapAllocator, typename T>
   return 1;
   }
 };
+
 }  // end namespace Eigen
 
 #endif  // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 367bccf63..f731bf17e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -662,13 +662,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
     }
   }
 
-  /// required by sycl in order to extract the output accessor
-#ifndef EIGEN_USE_SYCL
-  EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const { return NULL; }
-#else
-  EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const {
-    return m_result; }
-#endif
+  EIGEN_DEVICE_FUNC typename MakePointer_<Scalar>::Type data() const { return m_result; }
   /// required by sycl in order to extract the accessor
   const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
   /// added for sycl in order to construct the buffer from the sycl device
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index 1c89132db..3daecb045 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -27,9 +27,9 @@ namespace internal {
 
 template<typename CoeffReturnType, typename KernelName> struct syclGenericBufferReducer{
 template<typename BufferTOut, typename BufferTIn>
-static void run(BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
+static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
   do {
-          auto f = [length, local, &bufOut, &bufI](cl::sycl::handler& h) mutable {
+          auto f = [length, local, bufOut, &bufI](cl::sycl::handler& h) mutable {
             cl::sycl::nd_range<1> r{cl::sycl::range<1>{std::max(length, local)},
                                     cl::sycl::range<1>{std::min(length, local)}};
             /* Two accessors are used: one to the buffer that is being reduced,
@@ -37,7 +37,7 @@ static void run(BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDevice& de
             auto aI =
                 bufI.template get_access<cl::sycl::access::mode::read_write>(h);
             auto aOut =
-                bufOut.template get_access<cl::sycl::access::mode::discard_write>(h);
+                bufOut->template get_access<cl::sycl::access::mode::discard_write>(h);
             cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write,
                                cl::sycl::access::target::local>
                 scratch(cl::sycl::range<1>(local), h);
@@ -134,7 +134,7 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
     /// if the shared memory is less than the GRange, we set shared_mem size to the TotalSize and in this case one kernel would be created for recursion to reduce all to one.
     if (GRange < outTileSize) outTileSize=GRange;
     // getting final out buffer at the moment the created buffer is true because there is no need for assign
-    auto out_buffer =dev.template get_sycl_buffer<true, typename Eigen::internal::remove_all<CoeffReturnType>::type>(self.dimensions().TotalSize(), output);
+    auto out_buffer =dev.template get_sycl_buffer<typename Eigen::internal::remove_all<CoeffReturnType>::type>(self.dimensions().TotalSize(), output);
     /// creating the shared memory for calculating reduction.
     /// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can
     /// recursively apply reduction on it in order to reduce the whole.
@@ -208,7 +208,7 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
     dev.m_queue.submit([&](cl::sycl::handler &cgh) {
       // create a tuple of accessors from Evaluator
       auto tuple_of_accessors =  TensorSycl::internal::createTupleOfAccessors(cgh, self.impl());
-      auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write, true>(num_coeffs_to_preserve,cgh, output);
+      auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write>(num_coeffs_to_preserve,cgh, output);
 
       cgh.parallel_for<Self>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
         typedef typename TensorSycl::internal::ConvertToDeviceExpression<const HostExpr>::Type DevExpr;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
index 3af5f8cfc..b1da6858e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
@@ -56,10 +56,10 @@ struct AccessorConstructor{
   -> decltype(utility::tuple::append(ExtractAccessor<Arg1>::getTuple(cgh, eval1),utility::tuple::append(ExtractAccessor<Arg2>::getTuple(cgh, eval2), ExtractAccessor<Arg3>::getTuple(cgh, eval3)))) {
     return utility::tuple::append(ExtractAccessor<Arg1>::getTuple(cgh, eval1),utility::tuple::append(ExtractAccessor<Arg2>::getTuple(cgh, eval2), ExtractAccessor<Arg3>::getTuple(cgh, eval3)));
   }
-  template< cl::sycl::access::mode AcM,  bool MapAllocator, typename Arg> static inline auto getAccessor(cl::sycl::handler& cgh, Arg eval)
-  -> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM, MapAllocator,
+  template< cl::sycl::access::mode AcM, typename Arg> static inline auto getAccessor(cl::sycl::handler& cgh, Arg eval)
+  -> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM,
   typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()))){
-    return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM, MapAllocator, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()));
+    return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()));
   }
 };
 
@@ -141,8 +141,8 @@ struct ExtractAccessor<TensorEvaluator<TensorAssignOp<LHSExpr, RHSExpr>, Dev> >
 template <typename PlainObjectType, int Options_, typename Dev>\
 struct ExtractAccessor<TensorEvaluator<CVQual TensorMap<PlainObjectType, Options_>, Dev> > {\
   static inline auto getTuple(cl::sycl::handler& cgh,const TensorEvaluator<CVQual TensorMap<PlainObjectType, Options_>, Dev> eval)\
-  -> decltype(AccessorConstructor::template getAccessor<ACCType, true>(cgh, eval)){\
-    return AccessorConstructor::template getAccessor<ACCType, true>(cgh, eval);\
+  -> decltype(AccessorConstructor::template getAccessor<ACCType>(cgh, eval)){\
+    return AccessorConstructor::template getAccessor<ACCType>(cgh, eval);\
   }\
 };
 TENSORMAPEXPR(const, cl::sycl::access::mode::read)
@@ -153,8 +153,8 @@ TENSORMAPEXPR(, cl::sycl::access::mode::read_write)
 template <typename Expr, typename Dev>
 struct ExtractAccessor<TensorEvaluator<const TensorForcedEvalOp<Expr>, Dev> > {
   static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const TensorForcedEvalOp<Expr>, Dev> eval)
-  -> decltype(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read, false>(cgh, eval)){
-    return AccessorConstructor::template getAccessor<cl::sycl::access::mode::read, false>(cgh, eval);
+  -> decltype(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval)){
+    return AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval);
   }
 };
 
@@ -167,8 +167,8 @@ struct ExtractAccessor<TensorEvaluator<TensorForcedEvalOp<Expr>, Dev> >
 template <typename Expr, typename Dev>
 struct ExtractAccessor<TensorEvaluator<const TensorEvalToOp<Expr>, Dev> > {
   static inline auto getTuple(cl::sycl::handler& cgh,const TensorEvaluator<const TensorEvalToOp<Expr>, Dev> eval)
-  -> decltype(utility::tuple::append(AccessorConstructor::template getAccessor<cl::sycl::access::mode::write, false>(cgh, eval), AccessorConstructor::getTuple(cgh, eval.impl()))){
-    return utility::tuple::append(AccessorConstructor::template getAccessor<cl::sycl::access::mode::write, false>(cgh, eval), AccessorConstructor::getTuple(cgh, eval.impl()));
+  -> decltype(utility::tuple::append(AccessorConstructor::template getAccessor<cl::sycl::access::mode::write>(cgh, eval), AccessorConstructor::getTuple(cgh, eval.impl()))){
+    return utility::tuple::append(AccessorConstructor::template getAccessor<cl::sycl::access::mode::write>(cgh, eval), AccessorConstructor::getTuple(cgh, eval.impl()));
   }
 };
 
@@ -181,8 +181,8 @@ struct ExtractAccessor<TensorEvaluator<TensorEvalToOp<Expr>, Dev> >
 template <typename OP, typename Dim, typename Expr, typename Dev>
 struct ExtractAccessor<TensorEvaluator<const TensorReductionOp<OP, Dim, Expr>, Dev> > {
   static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<const TensorReductionOp<OP, Dim, Expr>, Dev> eval)
-  -> decltype(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read, false>(cgh, eval)){
-    return AccessorConstructor::template getAccessor<cl::sycl::access::mode::read, false>(cgh, eval);
+  -> decltype(AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval)){
+    return AccessorConstructor::template getAccessor<cl::sycl::access::mode::read>(cgh, eval);
   }
 };
 
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));
 }