Fixing potential race condition on sycl device.

unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h

@@ -41,6 +41,7 @@ namespace Eigen {
     size_t m_i;
     size_t m_offset;
   };
+
 template<typename AccType>
   struct memsetkernelFunctor{
    AccType m_acc;
@@ -54,6 +55,21 @@ template<typename AccType>
 
   };
 
+struct memsetCghFunctor{
+  cl::sycl::buffer<uint8_t, 1>& m_buf;
+  const ptrdiff_t& buff_offset;
+  const size_t& rng , GRange, tileSize;
+  const int  &c;
+  memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& 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();
@@ -75,18 +91,8 @@ EIGEN_STRONG_INLINE auto get_sycl_supported_devices()->decltype(cl::sycl::device
   return devices;
 }
 
-struct QueueInterface {
+class QueueInterface {
-  /// class members:
+public:
-  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;
   /// 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):
@@ -115,16 +121,53 @@ m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
 }))
 #endif
   {}
+  /// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
+  /// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
+  /// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
+  /// 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 {
+    std::lock_guard<std::mutex> lock(mutex_);
+    auto buf = cl::sycl::buffer<uint8_t,1>(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);
+    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(static_cast<const uint8_t*>(ptr),buf));
+    return static_cast<void*>(ptr);
+  }
+
+  /// This is used to deallocate the device pointer. p is used as a key inside
+  /// the map to find the device buffer and delete it.
+  EIGEN_STRONG_INLINE void deallocate(void *p) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    auto it = buffer_map.find(static_cast<const uint8_t*>(p));
+    if (it != buffer_map.end()) {
+      buffer_map.erase(it);
+    }
+  }
+
+  EIGEN_STRONG_INLINE void deallocate_all() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    buffer_map.clear();
+  }
   //FIXME: currently we have to switch back to write as discard_write doesnot work in forloop
+  /// 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.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
         std::lock_guard<std::mutex> lock(mutex_);
         auto host_acc= find_buffer(dst)->second. template get_access<cl::sycl::access::mode::write, cl::sycl::access::target::host_buffer>();
         ::memcpy(host_acc.get_pointer(), 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.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const Index *src, size_t n) const {
     std::lock_guard<std::mutex> lock(mutex_);
-  // Assuming that the dst is the start of the destination pointer
     auto it =find_buffer(src);
     auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
     offset/=sizeof(Index);
@@ -139,19 +182,67 @@ parallel_for_setup(n/sizeof(Index), tileSize, rng, GRange);
         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();
+  }
 
+  /// the memcpy function
+  template<typename Index> EIGEN_STRONG_INLINE void memcpy(void *dst, const Index *src, size_t n) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    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();
+  }
+
+  EIGEN_STRONG_INLINE void memset(void *data, int c, size_t n) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    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 {
+    std::lock_guard<std::mutex> lock(mutex_);
+    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 {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return find_buffer(ptr)->second;
+  }
+
+  EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return (static_cast<const uint8_t*>(ptr))-(find_buffer(ptr)->first);
   }
 
   EIGEN_STRONG_INLINE void synchronize() const {
-  std::lock_guard<std::mutex> lock(mutex_);
     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();// does not pass. Temporarily disabled
+    //sycl_queue().throw_asynchronous();// FIXME::does not pass. Temporarily disabled
-  std::lock_guard<std::mutex> lock(mutex_);
     m_queue.wait_and_throw(); //pass
-
   }
 
   template<typename Index>
@@ -200,8 +291,6 @@ EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1, Index &tileS
     }
   }
 
-
-
   /// 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 {
@@ -240,61 +329,53 @@ EIGEN_STRONG_INLINE void parallel_for_setup(Index dim0, Index dim1,Index dim2, I
     }
   }
 
-
   EIGEN_STRONG_INLINE unsigned long getNumSyclMultiProcessors() const {
-  std::lock_guard<std::mutex> lock(mutex_);
     return m_queue.get_device(). template get_info<cl::sycl::info::device::max_compute_units>();
-//  return stream_->deviceProperties().multiProcessorCount;
   }
+
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
-  std::lock_guard<std::mutex> lock(mutex_);
     return m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
-
-//  return stream_->deviceProperties().maxThreadsPerBlock;
   }
+
+  /// 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 {
-  std::lock_guard<std::mutex> lock(mutex_);
     // OpenCL doesnot have such concept
-  return 2;//sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+    return 2;
-//  return stream_->deviceProperties().maxThreadsPerMultiProcessor;
   }
+
   EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
-  std::lock_guard<std::mutex> lock(mutex_);
     return m_queue.get_device(). template get_info<cl::sycl::info::device::local_mem_size>();
-//  return stream_->deviceProperties().sharedMemPerBlock;
   }
 
-  /// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue;}
-  /// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
+
-  /// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
+  // This function checks if the runtime recorded an error for the
-  /// use this pointer as a key in our buffer_map and we make sure that we dedicate only one buffer only for this pointer.
+  // underlying stream device.
-  /// The device pointer would be deleted by calling deallocate function.
+  EIGEN_STRONG_INLINE bool ok() const {
-  EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
+    if (!exception_caught_) {
-    std::lock_guard<std::mutex> lock(mutex_);
+      m_queue.wait_and_throw();
-    auto buf = cl::sycl::buffer<uint8_t,1>(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();
+    return !exception_caught_;
-    buf.set_final_data(nullptr);
-    buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(static_cast<const uint8_t*>(ptr),buf));
-    return static_cast<void*>(ptr);
   }
 
-  /// This is used to deallocate the device pointer. p is used as a key inside
+  // destructor
-  /// the map to find the device buffer and delete it.
+  ~QueueInterface() { buffer_map.clear(); }
-  EIGEN_STRONG_INLINE void deallocate(void *p) const {
-    std::lock_guard<std::mutex> lock(mutex_);
-    auto it = buffer_map.find(static_cast<const uint8_t*>(p));
-    if (it != buffer_map.end()) {
-      buffer_map.erase(it);
-    }
-  }
 
-  EIGEN_STRONG_INLINE void deallocate_all() const {
+private:
-    std::lock_guard<std::mutex> lock(mutex_);
+  /// class members:
-    buffer_map.clear();
+  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;
   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;
@@ -308,37 +389,25 @@ EIGEN_STRONG_INLINE size_t sharedMemPerBlock() const {
     std::cerr << "No sycl buffer found. Make sure that you have allocated memory for your buffer by calling malloc-ed function."<< std::endl;
     abort();
   }
-
-  // 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(); }
 };
 
+// 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){}
 
-  /// Creation of sycl accessor for a buffer. This function first tries to find
+  // get sycl accessor
-  /// 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 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>& get_sycl_buffer(const void * ptr) const {
-    return m_queue_stream->find_buffer(ptr)->second;
+    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
@@ -353,8 +422,6 @@ struct SyclDevice {
     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 {
@@ -375,72 +442,27 @@ 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));
+    m_queue_stream->memcpy(dst,src,n);
-    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();
   }
 
   EIGEN_STRONG_INLINE ptrdiff_t get_offset(const void *ptr) const {
-    auto it = m_queue_stream->find_buffer(ptr);
+    return m_queue_stream->get_offset(ptr);
-    return (static_cast<const uint8_t*>(ptr))-it->first;
 
   }
-  /// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
+// memcpyHostToDevice
-  /// 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.
   template<typename Index> EIGEN_STRONG_INLINE void memcpyHostToDevice(Index *dst, const Index *src, size_t n) const {
      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
+/// here is the memcpyDeviceToHost
-  /// 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 {
     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;
+    m_queue_stream->memset(data,c,n);
-    parallel_for_setup(n, tileSize, rng, GRange);
-    auto it1 = m_queue_stream->find_buffer(static_cast<const void*>(data));
-    ptrdiff_t buff_offset= (static_cast<const uint8_t*>(data)) - it1->first;
-    sycl_queue().submit(memsetCghFunctor(it1->second, buff_offset, rng, GRange, tileSize, c ));
-    synchronize();
   }
-
+  /// returning the sycl queue
-  struct memsetCghFunctor{
+  EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->sycl_queue();}
-    cl::sycl::buffer<uint8_t, 1>& m_buf;
-    const ptrdiff_t& buff_offset;
-    const size_t& rng , GRange, tileSize;
-    const int  &c;
-    memsetCghFunctor(cl::sycl::buffer<uint8_t, 1>& 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));
-    }
-  };
 
   EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
     // FIXME
@@ -449,37 +471,31 @@ 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 m_queue_stream->getNumSyclMultiProcessors();
-  //  return stream_->deviceProperties().multiProcessorCount;
   }
   EIGEN_STRONG_INLINE unsigned long maxSyclThreadsPerBlock() const {
-    return sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>();
+    return m_queue_stream->maxSyclThreadsPerBlock();
-
-  //  return stream_->deviceProperties().maxThreadsPerBlock;
   }
   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 m_queue_stream->sharedMemPerBlock();
-  //  return stream_->deviceProperties().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 {
     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
     m_queue_stream->asynchronousExec();
   }
   // This function checks if the runtime recorded an error for the