Merged in benoitsteiner/opencl (pull request PR-309)

OpenCL improvements

1 files changed, 266 insertions, 145 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index e209799bb..c5142b7c9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -15,9 +15,22 @@
 #if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H)
 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
 
+template <typename Scalar, size_t Align = EIGEN_MAX_ALIGN_BYTES, class Allocator = std::allocator<Scalar>>
+struct SyclAllocator {
+  typedef Scalar value_type;
+  typedef typename std::allocator_traits<Allocator>::pointer pointer;
+  typedef typename std::allocator_traits<Allocator>::size_type size_type;
+
+  SyclAllocator( ){};
+  Scalar* allocate(std::size_t elements) { return static_cast<Scalar*>(aligned_alloc(Align, elements)); }
+  void deallocate(Scalar * p, std::size_t size) { EIGEN_UNUSED_VARIABLE(size); free(p); }
+};
+
 namespace Eigen {
 
   #define ConvertToActualTypeSycl(Scalar, buf_acc) reinterpret_cast<typename cl::sycl::global_ptr<Scalar>::pointer_t>((&(*buf_acc.get_pointer())))
+  #define ConvertToActualSyclOffset(Scalar, offset) offset/sizeof(Scalar)
+
 
   template <typename Scalar, typename read_accessor, typename write_accessor> class MemCopyFunctor {
   public:
@@ -40,27 +53,50 @@ namespace Eigen {
     size_t m_offset;
   };
 
+template<typename AccType>
   struct memsetkernelFunctor{
-   typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> AccType;
    AccType m_acc;
+   const ptrdiff_t buff_offset;
    const size_t m_rng, m_c;
-   memsetkernelFunctor(AccType acc, const size_t rng, const size_t c):m_acc(acc), m_rng(rng), m_c(c){}
+   memsetkernelFunctor(AccType acc, const ptrdiff_t buff_offset_, const size_t rng, const size_t c):m_acc(acc),  buff_offset(buff_offset_), m_rng(rng), m_c(c){}
    void operator()(cl::sycl::nd_item<1> itemID) {
      auto globalid=itemID.get_global_linear_id();
-     if (globalid< m_rng) m_acc[globalid] = m_c;
+     if (globalid< m_rng) m_acc[globalid + buff_offset] = m_c;
    }
 
   };
 
+struct memsetCghFunctor{
+  cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& m_buf;
+  const ptrdiff_t& buff_offset;
+  const size_t& rng , GRange, tileSize;
+  const int  &c;
+  memsetCghFunctor(cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& buff,  const ptrdiff_t& buff_offset_, const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
+  :m_buf(buff), buff_offset(buff_offset_), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
+
+  void operator()(cl::sycl::handler &cgh) const {
+    auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+    typedef decltype(buf_acc) AccType;
+    cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor<AccType>(buf_acc, buff_offset, rng, c));
+  }
+};
+
+  //get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU  and intel GPU)
 EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device::get_devices()){
   auto devices = cl::sycl::device::get_devices();
   std::vector<cl::sycl::device>::iterator it =devices.begin();
   while(it!=devices.end()) {
-    /// get_devices returns all the available opencl devices. Either use device_selector or exclude devices that computecpp does not support (AMD OpenCL for CPU )
-    auto s=  (*it).template get_info<cl::sycl::info::device::vendor>();
-    std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-    if((*it).is_cpu() && s.find("amd")!=std::string::npos && s.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
-      it=devices.erase(it);
+    ///FIXME: Currently there is a bug in amd cpu OpenCL
+    auto name = (*it).template get_info<cl::sycl::info::device::name>();
+    std::transform(name.begin(), name.end(), name.begin(), ::tolower);
+    auto vendor = (*it).template get_info<cl::sycl::info::device::vendor>();
+    std::transform(vendor.begin(), vendor.end(), vendor.begin(), ::tolower);
+
+    if((*it).is_cpu() && vendor.find("amd")!=std::string::npos && vendor.find("apu") == std::string::npos){ // remove amd cpu as it is not supported by computecpp allow APUs
+      it = devices.erase(it);
+      //FIXME: currently there is a bug in intel gpu driver regarding memory allignment issue.
+    }else if((*it).is_gpu() && name.find("intel")!=std::string::npos){
+      it = devices.erase(it);
     }
     else{
       ++it;
@@ -69,18 +105,8 @@ EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device
   return devices;
 }
 
-struct QueueInterface {
-  /// class members:
-  bool exception_caught_ = false;
-
-  mutable std::mutex mutex_;
-
-  /// 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.
-  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
-  /// sycl queue
-  mutable cl::sycl::queue m_queue;
+class QueueInterface {
+public:
   /// creating device by using cl::sycl::selector or cl::sycl::device both are the same and can be captured through dev_Selector typename
   /// SyclStreamDevice is not owned. it is the caller's responsibility to destroy it.
   template<typename dev_Selector> explicit QueueInterface(const dev_Selector& s):
@@ -116,11 +142,11 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
   /// use this pointer as a key in our buffer_map and we make sure that we dedicate only one buffer only for this pointer.
   /// The device pointer would be deleted by calling deallocate function.
   EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
-    auto buf = cl::sycl::buffer<uint8_t,1>(cl::sycl::range<1>(num_bytes));
+    std::lock_guard<std::mutex> lock(mutex_);
+    auto buf = cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >(cl::sycl::range<1>(num_bytes));
     auto ptr =buf.get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>().get_pointer();
     buf.set_final_data(nullptr);
-    std::lock_guard<std::mutex> lock(mutex_);
-    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(static_cast<const uint8_t*>(ptr),buf));
+    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >(static_cast<const uint8_t*>(ptr),buf));
     return static_cast<void*>(ptr);
   }
 
@@ -138,62 +164,113 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
     std::lock_guard<std::mutex> lock(mutex_);
     buffer_map.clear();
   }
-
-  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
-    std::lock_guard<std::mutex> lock(mutex_);
-    auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
-    if (it1 != buffer_map.end()){
-      return it1;
-    }
-    else{
-      for(std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
-        auto size = it->second.get_size();
-        if((it->first <  (static_cast<const uint8_t*>(ptr))) && ((static_cast<const uint8_t*>(ptr)) < (it->first + size)) ) return it;
-      }
-    }
-    std::cerr << "No sycl buffer found. Make sure that you have allocated memory for your buffer by calling malloc-ed function."<< std::endl;
-    abort();
+  /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
+  /// pointer created as a key we find the sycl buffer and get the host accessor with write mode
+  /// on it. Then we use the memcpy to copy the data to the host accessor. The first time that
+  /// this buffer is accessed, the data will be copied to the device.
+  /// In this case we can separate the kernel actual execution from data transfer which is required for benchmark
+  /// Also, this is faster as it uses the map_allocator instead of memcpy
+  template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
+    auto it =find_buffer(dst);
+    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(dst))- it->first;
+    offset/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+      auto src_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(static_cast<void*>(const_cast<Index*>(src))), cl::sycl::range<1>(n));
+      m_queue.submit([&](cl::sycl::handler &cgh) {
+        auto dst_acc= it->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+        auto src_acc =src_buf.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+        typedef decltype(src_acc) read_accessor;
+        typedef decltype(dst_acc) write_accessor;
+        cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, offset, 0));
+      });
+      synchronize();
   }
-
-  // This function checks if the runtime recorded an error for the
-  // underlying stream device.
-  EIGEN_STRONG_INLINE bool ok() const {
-    if (!exception_caught_) {
-      m_queue.wait_and_throw();
-    }
-    return !exception_caught_;
+  /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
+  /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
+  /// lifespan of the memcpyDeviceToHost function. We create a kernel to copy the data, from the device- only source buffer to the destination
+  /// buffer with map_allocator on the gpu in parallel. At the end of the function call the destination buffer would be destroyed and the data
+  /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
+  /// to the cpu only once per function call.
+  template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
+    auto it =find_buffer(src);
+    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
+    offset/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+      auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
+      m_queue.submit([&](cl::sycl::handler &cgh) {
+        auto src_acc= it->second.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);
+        typedef decltype(src_acc) read_accessor;
+        typedef decltype(dst_acc) write_accessor;
+        cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
+      });
+      synchronize();
   }
 
-  // destructor
-  ~QueueInterface() { buffer_map.clear(); }
-};
+  /// the memcpy function
+  template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
+    auto it1 = find_buffer(static_cast<const void*>(src));
+    auto it2 = find_buffer(dst);
+    auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
+    auto i= (static_cast<const uint8_t*>(dst)) - it2->first;
+    offset/=sizeof(Index);
+    i/=sizeof(Index);
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
+    m_queue.submit([&](cl::sycl::handler &cgh) {
+      auto src_acc =it1->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
+      auto dst_acc =it2->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
+      typedef decltype(src_acc) read_accessor;
+      typedef decltype(dst_acc) write_accessor;
+      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, i, offset));
+    });
+    synchronize();
+  }
 
-struct SyclDevice {
-  // class member.
-  QueueInterface* m_queue_stream;
-  /// QueueInterface is not owned. it is the caller's responsibility to destroy it.
-  explicit SyclDevice(QueueInterface* queue_stream) : m_queue_stream(queue_stream){}
+  EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
+    size_t rng, GRange, tileSize;
+    parallel_for_setup(n, tileSize, rng, GRange);
+    auto it1 = find_buffer(static_cast<const void*>(data));
+    ptrdiff_t buff_offset= (static_cast<const uint8_t*>(data)) - it1->first;
+    m_queue.submit(memsetCghFunctor(it1->second, buff_offset, rng, GRange, tileSize, c ));
+    synchronize();
+  }
 
   /// 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> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
   get_sycl_accessor(cl::sycl::handler &cgh, const void* ptr) const {
-    return (get_sycl_buffer(ptr).template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
+    return (find_buffer(ptr)->second.template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
   }
 
   /// Accessing the created sycl device buffer for the device pointer
-  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(const void * ptr) const {
-    return m_queue_stream->find_buffer(ptr)->second;
+  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& get_sycl_buffer(const void * ptr) const {
+    return find_buffer(ptr)->second;
+  }
+
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    return (static_cast<const uint8_t*>(ptr))-(find_buffer(ptr)->first);
+  }
+
+  EIGEN_STRONG_INLINE void synchronize() const {
+    m_queue.wait_and_throw(); //pass
+  }
+
+  EIGEN_STRONG_INLINE void asynchronousExec() const {
+    ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
+    //sycl_queue().throw_asynchronous();// FIXME::does not pass. Temporarily disabled
+    m_queue.wait_and_throw(); //pass
   }
 
-  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
-    tileSize =static_cast<Index>(sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
-    auto s=  sycl_queue().get_device().template get_info<cl::sycl::info::device::vendor>();
+    tileSize =static_cast<Index>(m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>());
+    auto s=  m_queue.get_device().template get_info<cl::sycl::info::device::vendor>();
     std::transform(s.begin(), s.end(), s.begin(), ::tolower);
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
       tileSize=std::min(static_cast<Index>(256), static_cast<Index>(tileSize));
     }
     rng = n;
@@ -210,7 +287,7 @@ struct SyclDevice {
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
     Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
       max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
     }
     Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
@@ -234,13 +311,11 @@ struct SyclDevice {
     }
   }
 
-
-
   /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
   template<typename Index>
   EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
     Index max_workgroup_Size = static_cast<Index>(maxSyclThreadsPerBlock());
-    if(sycl_queue().get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
+    if(m_queue.get_device().is_cpu()){ // intel doesnot allow to use max workgroup size
       max_workgroup_Size=std::min(static_cast<Index>(256), static_cast<Index>(max_workgroup_Size));
     }
     Index pow_of_2 = static_cast<Index>(std::log2(max_workgroup_Size));
@@ -273,6 +348,108 @@ struct SyclDevice {
       if (xMode != 0) GRange0 += static_cast<Index>(tileSize0 - xMode);
     }
   }
+
+  EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
+  }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+  }
+
+  /// No need for sycl it should act the same as CPU version
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
+
+  EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
+    // OpenCL doesnot have such concept
+    return 2;
+  }
+
+  EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
+    return m_queue.get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
+  }
+
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue;}
+
+  // This function checks if the runtime recorded an error for the
+  // underlying stream device.
+  EIGEN_STRONG_INLINE bool ok() const {
+    if (!exception_caught_) {
+      m_queue.wait_and_throw();
+    }
+    return !exception_caught_;
+  }
+
+  // destructor
+  ~QueueInterface() { buffer_map.clear(); }
+
+private:
+  /// class members:
+  bool exception_caught_ = false;
+
+  mutable std::mutex mutex_;
+
+  /// 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.
+  mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > > buffer_map;
+  /// sycl queue
+  mutable cl::sycl::queue m_queue;
+
+  EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >::iterator find_buffer(const void* ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
+    if (it1 != buffer_map.end()){
+      return it1;
+    }
+    else{
+      for(std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> > >::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
+        auto size = it->second.get_size();
+        if((it->first <  (static_cast<const uint8_t*>(ptr))) && ((static_cast<const uint8_t*>(ptr)) < (it->first + size)) ) return it;
+      }
+    }
+    std::cerr << "No sycl buffer found. Make sure that you have allocated memory for your buffer by calling malloc-ed function."<< std::endl;
+    abort();
+  }
+};
+
+// Here is a sycl deviuce struct which accept the sycl queue interface
+// as an input
+struct SyclDevice {
+  // class member.
+  QueueInterface* m_queue_stream;
+  /// QueueInterface is not owned. it is the caller's responsibility to destroy it.
+  explicit SyclDevice(QueueInterface* queue_stream) : m_queue_stream(queue_stream){}
+
+  // get sycl accessor
+  template <cl::sycl::access::mode AcMd> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
+  get_sycl_accessor(cl::sycl::handler &cgh, const void* ptr) const {
+    return m_queue_stream->template get_sycl_accessor<AcMd>(cgh, ptr);
+  }
+
+  /// Accessing the created sycl device buffer for the device pointer
+  EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1, SyclAllocator<uint8_t, EIGEN_MAX_ALIGN_BYTES> >& get_sycl_buffer(const void * ptr) const {
+    return m_queue_stream->get_sycl_buffer(ptr);
+  }
+
+  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index n, Index &tileSize, Index &rng, Index &GRange)  const {
+  m_queue_stream->parallel_for_setup(n, tileSize, rng, GRange);
+  }
+
+  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileSize0, Index &tileSize1, Index &rng0, Index &rng1, Index &GRange0, Index &GRange1)  const {
+    m_queue_stream->parallel_for_setup(dim0, dim1, tileSize0, tileSize1, rng0, rng1, GRange0, GRange1);
+  }
+
+  /// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
+  template<typename Index>
+  EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, Index &tileSize0, Index &tileSize1, Index &tileSize2, Index &rng0, Index &rng1, Index &rng2, Index &GRange0, Index &GRange1, Index &GRange2)  const {
+    m_queue_stream->parallel_for_setup(dim0, dim1, dim2, tileSize0, tileSize1, tileSize2, rng0, rng1, rng2, GRange0, GRange1, GRange2);
+
+  }
   /// allocate device memory
   EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
       return m_queue_stream->allocate(num_bytes);
@@ -287,78 +464,27 @@ struct SyclDevice {
 
   /// the memcpy function
   template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
-    auto it1 = m_queue_stream->find_buffer(static_cast<const void*>(src));
-    auto it2 = m_queue_stream->find_buffer(dst);
-    auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
-    auto i= (static_cast<const uint8_t*>(dst)) - it2->first;
-    offset/=sizeof(Index);
-    i/=sizeof(Index);
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
-    sycl_queue().submit([&](cl::sycl::handler &cgh) {
-      auto src_acc =it1->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
-      auto dst_acc =it2->second.template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer>(cgh);
-      typedef decltype(src_acc) read_accessor;
-      typedef decltype(dst_acc) write_accessor;
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, i, offset));
-    });
-    synchronize();
+    m_queue_stream->memcpy(dst,src,n);
   }
 
-  /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
-  /// pointer created as a key we find the sycl buffer and get the host accessor with discard_write mode
-  /// on it. Using a discard_write accessor guarantees that we do not bring back the current value of the
-  /// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
-  /// this buffer is accessed, the data will be copied to the device.
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    return m_queue_stream->get_offset(ptr);
+
+  }
+// memcpyHostToDevice
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
-    auto host_acc= get_sycl_buffer(dst). template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>();
-    ::memcpy(host_acc.get_pointer(), src, n);
+     m_queue_stream->memcpyHostToDevice(dst,src,n);
   }
-  /// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
-  /// buffer with map_allocator for the destination pointer with a discard_write accessor on it. The lifespan of the buffer is bound to the
-  /// lifespan of the memcpyDeviceToHost function. We create a kernel to copy the data, from the device- only source buffer to the destination
-  /// buffer with map_allocator on the gpu in parallel. At the end of the function call the destination buffer would be destroyed and the data
-  /// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
-  /// to the cpu only once per function call.
+/// here is the memcpyDeviceToHost
   template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
-    auto it = m_queue_stream->find_buffer(src);
-    auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
-    offset/=sizeof(Index);
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
-    // Assuming that the dst is the start of the destination pointer
-    auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(n));
-    sycl_queue().submit([&](cl::sycl::handler &cgh) {
-      auto src_acc= it->second.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);
-      typedef decltype(src_acc) read_accessor;
-      typedef decltype(dst_acc) write_accessor;
-      cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<Index, read_accessor, write_accessor>(src_acc, dst_acc, rng, 0, offset));
-    });
-    synchronize();
+    m_queue_stream->memcpyDeviceToHost(dst,src,n);
   }
-  /// returning the sycl queue
-  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->m_queue;}
   /// Here is the implementation of memset function on sycl.
   EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
-    size_t rng, GRange, tileSize;
-    parallel_for_setup(n, tileSize, rng, GRange);
-    sycl_queue().submit(memsetCghFunctor(get_sycl_buffer(static_cast<uint8_t*>(static_cast<void*>(data))),rng, GRange, tileSize, c ));
-    synchronize();
+    m_queue_stream->memset(data,c,n);
   }
-
-  struct memsetCghFunctor{
-    cl::sycl::buffer<uint8_t, 1>& m_buf;
-    const size_t& rng , GRange, tileSize;
-    const int  &c;
-    memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& buff,  const size_t& rng_,  const size_t& GRange_,  const size_t& tileSize_, const int& c_)
-    :m_buf(buff), rng(rng_), GRange(GRange_), tileSize(tileSize_), c(c_){}
-
-    void operator()(cl::sycl::handler &cgh) const {
-      auto buf_acc = m_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
-      cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), memsetkernelFunctor(buf_acc, rng, c));
-    }
-  };
+  /// returning the sycl queue
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->sycl_queue();}
 
   EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
     // FIXME
@@ -367,39 +493,32 @@ struct SyclDevice {
 
   EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
     // We won't try to take advantage of the l2 cache for the time being, and
-    // there is no l3 cache on cuda devices.
+    // there is no l3 cache on sycl devices.
     return firstLevelCacheSize();
   }
   EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
-  //  return stream_->deviceProperties().multiProcessorCount;
+    return m_queue_stream->getNumSyclMultiProcessors();
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
-
-  //  return stream_->deviceProperties().maxThreadsPerBlock;
+    return m_queue_stream->maxSyclThreadsPerBlock();
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerMultiProcessor() const {
     // OpenCL doesnot have such concept
-    return 2;//sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+    return m_queue_stream->maxSyclThreadsPerMultiProcessor();
   //  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
   }
   EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
-  //  return stream_->deviceProperties().sharedMemPerBlock;
+    return m_queue_stream->sharedMemPerBlock();
   }
   /// No need for sycl it should act the same as CPU version
-  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
+  EIGEN_STRONG_INLINE int majorDeviceVersion() const { return m_queue_stream->majorDeviceVersion(); }
 
   EIGEN_STRONG_INLINE void synchronize() const {
-    sycl_queue().wait_and_throw(); //pass
+    m_queue_stream->synchronize(); //pass
   }
 
   EIGEN_STRONG_INLINE void asynchronousExec() const {
-    ///FIXEDME:: currently there is a race condition regarding the asynch scheduler.
-    //sycl_queue().throw_asynchronous();// does not pass. Temporarily disabled
-    sycl_queue().wait_and_throw(); //pass
-
+    m_queue_stream->asynchronousExec();
   }
   // This function checks if the runtime recorded an error for the
   // underlying stream device.
@@ -407,8 +526,10 @@ struct SyclDevice {
     return m_queue_stream->ok();
   }
 };
-
-
+// This is used as a distingushable device inside the kernel as the sycl device class is not Standard layout.
+// This is internal and must not be used by user. This dummy device allow us to specialise the tensor evaluator
+// inside the kenrel. So we can have two types of eval for host and device. This is required for TensorArgMax operation
+struct SyclKernelDevice:DefaultDevice{};
 
 }  // end namespace Eigen