Merged in benoitsteiner/opencl (pull request PR-247)
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 7c03989..844cec1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -44,14 +44,14 @@
   // destructor
   ~SyclDevice() { deallocate_all(); }
 
-  template <typename T> void deallocate(T *p) const {
+  template <typename T> EIGEN_STRONG_INLINE 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 {
+  EIGEN_STRONG_INLINE 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;
@@ -72,9 +72,14 @@
 
   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);
+    std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> ret;
+    if(ptr!=nullptr){
+       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*>(ret.first->second.get()))->set_final_data(nullptr);
+    } else {
+      eigen_assert("The device memory is not allocated. Please call allocate on the device!!");
+    }
     return ret;
   }
 
@@ -83,36 +88,77 @@
   }
 
   /// allocating memory on the cpu
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t) const {
+  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_STRONG_INLINE bool isDeviceSuitable() const { return true; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void *dst, const void *src, size_t n) const {
+  EIGEN_STRONG_INLINE void memcpy(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 {
+  template<typename T> 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);
   }
- /// 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 {
+
+  EIGEN_STRONG_INLINE void parallel_for_setup(size_t n, size_t &tileSize, size_t &rng, size_t &GRange)  const {
+      tileSize =m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
+      rng = n;
+      if (rng==0) rng=1;
+       GRange=rng;
+      if (tileSize>GRange) tileSize=GRange;
+      else if(GRange>tileSize){
+        size_t xMode = GRange % tileSize;
+        if (xMode != 0) GRange += (tileSize - xMode);
+      }
+    }
+
+  template<typename T> 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);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
+
+    auto dest_buf = cl::sycl::buffer<T, 1, cl::sycl::map_allocator<T>>(dst, cl::sycl::range<1>(rng));
+    typedef decltype(dest_buf) SYCLDTOH;
+    m_queue.submit([&](cl::sycl::handler &cgh) {
+      auto src_acc= (static_cast<cl::sycl::buffer<T, 1>*>(it->second.get()))-> template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+      auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
+      cgh.parallel_for<SYCLDTOH>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
+      auto globalid=itemID.get_global_linear_id();
+      if (globalid< dst_acc.get_size()) {
+          dst_acc[globalid] = src_acc[globalid];
+      }
+      });
+    });
+    m_queue.throw_asynchronous();
+
     } 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);
+  template<typename T>  EIGEN_STRONG_INLINE void memset(T *buff, int c, size_t n) const {
+
+      size_t rng, GRange, tileSize;
+      parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
+      m_queue.submit([&](cl::sycl::handler &cgh) {
+        auto buf_acc =(static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(buff, n).first->second.get()))-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
+        cgh.parallel_for<SyclDevice>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
+        auto globalid=itemID.get_global_linear_id();
+        auto buf_ptr= reinterpret_cast<typename cl::sycl::global_ptr<unsigned char>::pointer_t>((&(*buf_acc.get_pointer())));
+        if (globalid< buf_acc.get_size()) {
+          for(size_t i=0; i<sizeof(T); i++)
+            buf_ptr[globalid*sizeof(T) + i] = c;
+        }
+        });
+      });
+      m_queue.throw_asynchronous();
   }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const {
   return 1;
   }
 };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index 3daecb0..db23bd7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -188,15 +188,8 @@
     typedef const typename Self::ChildType HostExpr; /// this is the child of reduction
     typedef  typename TensorSycl::internal::createPlaceHolderExpression<HostExpr>::Type PlaceHolderExpr;
     auto functors = TensorSycl::internal::extractFunctors(self.impl());
-
-    size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
-
-    size_t GRange=num_coeffs_to_preserve;
-    if (tileSize>GRange) tileSize=GRange;
-    else if(GRange>tileSize){
-      size_t xMode = GRange % tileSize;
-      if (xMode != 0) GRange += (tileSize - xMode);
-    }
+    size_t range, GRange, tileSize;
+    dev.parallel_for_setup(num_coeffs_to_preserve, tileSize, range, GRange);
     // getting final out buffer at the moment the created buffer is true because there is no need for assign
     /// 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
@@ -223,7 +216,7 @@
         auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
         /// const cast added as a naive solution to solve the qualifier drop error
         auto globalid=itemID.get_global_linear_id();
-        if (globalid< static_cast<size_t>(num_coeffs_to_preserve)) {
+        if (globalid< range) {
           typename DeiceSelf::CoeffReturnType accum = functor.initialize();
           GenericDimReducer<DeiceSelf::NumReducedDims-1, DeiceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(globalid),const_cast<Op&>(functor), &accum);
           functor.finalize(accum);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
index 7914b6f..724eebd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
@@ -37,18 +37,12 @@
     typedef  typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr;
     auto functors = internal::extractFunctors(evaluator);
 
-    size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
     dev.m_queue.submit([&](cl::sycl::handler &cgh) {
-
       // create a tuple of accessors from Evaluator
       auto tuple_of_accessors = internal::createTupleOfAccessors<decltype(evaluator)>(cgh, evaluator);
-      const auto range = utility::tuple::get<0>(tuple_of_accessors).get_range()[0];
-      size_t GRange=range;
-      if (tileSize>GRange) tileSize=GRange;
-      else if(GRange>tileSize){
-        size_t xMode = GRange % tileSize;
-        if (xMode != 0) GRange += (tileSize - xMode);
-      }
+      size_t range, GRange, tileSize;
+      dev.parallel_for_setup(utility::tuple::get<0>(tuple_of_accessors).get_range()[0], tileSize, range, GRange);
+
       // run the kernel
       cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
         typedef  typename internal::ConvertToDeviceExpression<Expr>::Type DevExpr;
diff --git a/unsupported/test/cxx11_tensor_device_sycl.cpp b/unsupported/test/cxx11_tensor_device_sycl.cpp
index 7f79753..584fa80 100644
--- a/unsupported/test/cxx11_tensor_device_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_device_sycl.cpp
@@ -19,10 +19,23 @@
 
 #include "main.h"
 #include <unsupported/Eigen/CXX11/Tensor>
+#include<stdint.h>
 
 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;;
+  int sizeDim1 = 100;
+
+  array<int, 1> tensorRange = {{sizeDim1}};
+  Tensor<int, 1> in(tensorRange);
+  Tensor<int, 1> in1(tensorRange);
+  memset(in1.data(), 1,in1.size()*sizeof(int));
+  int * gpu_in_data  = static_cast<int*>(sycl_device.allocate(in.size()*sizeof(int)));
+  sycl_device.memset(gpu_in_data, 1,in.size()*sizeof(int) );
+  sycl_device.memcpyDeviceToHost(in.data(), gpu_in_data, in.size()*sizeof(int) );
+  for (int i=0; i<in.size(); i++)
+    VERIFY_IS_APPROX(in(i), in1(i));
+  sycl_device.deallocate(gpu_in_data);
 }
 void test_cxx11_tensor_device_sycl() {
   cl::sycl::gpu_selector s;
diff --git a/unsupported/test/cxx11_tensor_sycl.cpp b/unsupported/test/cxx11_tensor_sycl.cpp
index 6a9c334..05fbf9e 100644
--- a/unsupported/test/cxx11_tensor_sycl.cpp
+++ b/unsupported/test/cxx11_tensor_sycl.cpp
@@ -27,7 +27,46 @@
 using Eigen::Tensor;
 using Eigen::TensorMap;
 
-void test_sycl_cpu(const Eigen::SyclDevice &sycl_device) {
+void test_sycl_mem_transfers(const Eigen::SyclDevice &sycl_device) {
+  int sizeDim1 = 100;
+  int sizeDim2 = 100;
+  int sizeDim3 = 100;
+  array<int, 3> tensorRange = {{sizeDim1, sizeDim2, sizeDim3}};
+  Tensor<float, 3> in1(tensorRange);
+  Tensor<float, 3> out1(tensorRange);
+  Tensor<float, 3> out2(tensorRange);
+  Tensor<float, 3> out3(tensorRange);
+
+  in1 = in1.random();
+
+  float* gpu_data1  = static_cast<float*>(sycl_device.allocate(in1.size()*sizeof(float)));
+  float* gpu_data2  = static_cast<float*>(sycl_device.allocate(out1.size()*sizeof(float)));
+  //float* gpu_data =  static_cast<float*>(sycl_device.allocate(out2.size()*sizeof(float)));
+
+  TensorMap<Tensor<float, 3>> gpu1(gpu_data1, tensorRange);
+  TensorMap<Tensor<float, 3>> gpu2(gpu_data2, tensorRange);
+  //TensorMap<Tensor<float, 3>> gpu_out2(gpu_out2_data, tensorRange);
+  
+  sycl_device.memcpyHostToDevice(gpu_data1, in1.data(),(in1.size())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_data2, in1.data(),(in1.size())*sizeof(float));
+  gpu1.device(sycl_device) = gpu1 * 3.14f;
+  gpu2.device(sycl_device) = gpu2 * 2.7f;
+  sycl_device.memcpyDeviceToHost(out1.data(), gpu_data1,(out1.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out2.data(), gpu_data1,(out2.size())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out3.data(), gpu_data2,(out3.size())*sizeof(float));
+  //  sycl_device.Synchronize();
+
+  for (int i = 0; i < in1.size(); ++i) {
+    VERIFY_IS_APPROX(out1(i), in1(i) * 3.14f);
+    VERIFY_IS_APPROX(out2(i), in1(i) * 3.14f);
+    VERIFY_IS_APPROX(out3(i), in1(i) * 2.7f);
+  }
+
+  sycl_device.deallocate(gpu_data1);
+  sycl_device.deallocate(gpu_data2);
+}
+
+void test_sycl_computations(const Eigen::SyclDevice &sycl_device) {
 
   int sizeDim1 = 100;
   int sizeDim2 = 100;
@@ -41,10 +80,10 @@
   in2 = in2.random();
   in3 = in3.random();
 
-  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)));
+  float * gpu_in1_data  = static_cast<float*>(sycl_device.allocate(in1.size()*sizeof(float)));
+  float * gpu_in2_data  = static_cast<float*>(sycl_device.allocate(in2.size()*sizeof(float)));
+  float * gpu_in3_data  = static_cast<float*>(sycl_device.allocate(in3.size()*sizeof(float)));
+  float * gpu_out_data =  static_cast<float*>(sycl_device.allocate(out.size()*sizeof(float)));
 
   TensorMap<Tensor<float, 3>> gpu_in1(gpu_in1_data, tensorRange);
   TensorMap<Tensor<float, 3>> gpu_in2(gpu_in2_data, tensorRange);
@@ -53,7 +92,7 @@
 
   /// a=1.2f
   gpu_in1.device(sycl_device) = gpu_in1.constant(1.2f);
-  sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(in1.data(), gpu_in1_data ,(in1.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -65,7 +104,7 @@
 
   /// a=b*1.2f
   gpu_out.device(sycl_device) = gpu_in1 * 1.2f;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data ,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -77,9 +116,9 @@
   printf("a=b*1.2f Test Passed\n");
 
   /// c=a*b
-  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyHostToDevice(gpu_in2_data, in2.data(),(in2.size())*sizeof(float));
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in2;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -93,7 +132,7 @@
 
   /// c=a+b
   gpu_out.device(sycl_device) = gpu_in1 + gpu_in2;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -107,7 +146,7 @@
 
   /// c=a*a
   gpu_out.device(sycl_device) = gpu_in1 * gpu_in1;
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -121,7 +160,7 @@
 
   //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.memcpyDeviceToHost(out.data(),gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(),gpu_out_data,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -134,9 +173,9 @@
   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));
+  sycl_device.memcpyHostToDevice(gpu_in3_data, in3.data(),(in3.size())*sizeof(float));
   gpu_out.device(sycl_device) =(gpu_in1 > gpu_in1.constant(0.5f)).select(gpu_in2, gpu_in3);
-  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(float));
+  sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.size())*sizeof(float));
   for (int i = 0; i < sizeDim1; ++i) {
     for (int j = 0; j < sizeDim2; ++j) {
       for (int k = 0; k < sizeDim3; ++k) {
@@ -152,8 +191,10 @@
   sycl_device.deallocate(gpu_in3_data);
   sycl_device.deallocate(gpu_out_data);
 }
+
 void test_cxx11_tensor_sycl() {
   cl::sycl::gpu_selector s;
   Eigen::SyclDevice sycl_device(s);
-  CALL_SUBTEST(test_sycl_cpu(sycl_device));
+  CALL_SUBTEST(test_sycl_mem_transfers(sycl_device));
+  CALL_SUBTEST(test_sycl_computations(sycl_device));
 }