Adding Memset; optimising MecopyDeviceToHost by removing double copying;

3 files changed, 65 insertions, 32 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
index 7c039890e..e767d8965 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
@@ -72,9 +72,14 @@ struct SyclDevice {
 
   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 is not initialised!!")
+    }
     return ret;
   }
 
@@ -83,36 +88,77 @@ struct SyclDevice {
   }
 
   /// allocating memory on the cpu
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void *allocate(size_t) const {
+  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 memcpy(void *dst, const void *src, size_t n) const {
+  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> 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 {
+
+ 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> 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>  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 {
+  int majorDeviceVersion() const {
   return 1;
   }
 };
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
index 3daecb045..db23bd7b0 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
@@ -188,15 +188,8 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
     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 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
         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 7914b6fad..724eebd83 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
@@ -37,18 +37,12 @@ void run(Expr &expr, Dev &dev) {
     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;