Merge pull request #312 from CNugteren/CLBlast-311-missing-event-in-trsv-trsm

Missing events in TRSV and TRSM

src/clpp11.hpp

@@ -729,9 +729,10 @@ class Buffer {
   }
 
   // Copies the contents of this buffer into another device buffer
-  void CopyToAsync(const Queue &queue, const size_t size, const Buffer<T> &destination) const {
+  void CopyToAsync(const Queue &queue, const size_t size, const Buffer<T> &destination,
+                   EventPointer event = nullptr) const {
     CheckError(clEnqueueCopyBuffer(queue(), *buffer_, destination(), 0, 0, size*sizeof(T), 0,
-                                   nullptr, nullptr));
+                                   nullptr, event));
   }
   void CopyTo(const Queue &queue, const size_t size, const Buffer<T> &destination) const {
     CopyToAsync(queue, size, destination);

src/routines/level2/xtrsv.cpp

@@ -33,7 +33,8 @@ void Xtrsv<T>::Substitution(const Layout layout, const Triangle triangle,
                             const size_t n,
                             const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
                             const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_inc,
-                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) {
+                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
+                            EventPointer event) {
 
   if (n > db_["TRSV_BLOCK_SIZE"]) { throw BLASError(StatusCode::kUnexpectedError); };
 
@@ -69,9 +70,7 @@ void Xtrsv<T>::Substitution(const Layout layout, const Triangle triangle,
   // Launches the kernel
   const auto local = std::vector<size_t>{db_["TRSV_BLOCK_SIZE"]};
   const auto global = std::vector<size_t>{Ceil(n, db_["TRSV_BLOCK_SIZE"])};
-  auto event = Event();
-  RunKernel(kernel, queue_, device_, global, local, event.pointer());
-  event.WaitForCompletion();
+  RunKernel(kernel, queue_, device_, global, local, event);
 }
 
 // =================================================================================================
@@ -146,14 +145,16 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
     }
 
     // Runs the triangular substitution for the block size
+    auto sub_event = Event();
     Substitution(layout, triangle, a_transpose, diagonal, block_size,
                  a_buffer, a_offset + col + col*a_ld, a_ld,
                  b_buffer, b_offset + col*b_inc, b_inc,
-                 x_buffer, x_offset + col*x_inc, x_inc);
+                 x_buffer, x_offset + col*x_inc, x_inc, sub_event.pointer());
+    sub_event.WaitForCompletion();
   }
 
   // Retrieves the results
-  x_buffer.CopyTo(queue_, x_size, b_buffer);
+  x_buffer.CopyToAsync(queue_, x_size, b_buffer, event_);
 }
 
 // =================================================================================================

src/routines/level2/xtrsv.hpp

@@ -32,6 +32,7 @@ class Xtrsv: public Xgemv<T> {
   using Xgemv<T>::device_;
   using Xgemv<T>::db_;
   using Xgemv<T>::program_;
+  using Xgemv<T>::event_;
   using Xgemv<T>::DoGemv;
 
   // Constructor
@@ -50,7 +51,8 @@ class Xtrsv: public Xgemv<T> {
                     const size_t n,
                     const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
                     const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_inc,
-                    const Buffer<T> &x_buffer, const size_t offset_x, const size_t x_inc);
+                    const Buffer<T> &x_buffer, const size_t offset_x, const size_t x_inc,
+                    EventPointer event);
 };
 
 // =================================================================================================

src/routines/level3/xtrsm.cpp

@@ -246,7 +246,7 @@ void Xtrsm<T>::TrsmColMajor(const Side side, const Triangle triangle,
   }
 
   // Retrieves the results
-  x_buffer.CopyTo(queue_, b_size, b_buffer);
+  x_buffer.CopyToAsync(queue_, b_size, b_buffer, event_);
 }
 
 // =================================================================================================

src/routines/level3/xtrsm.hpp

@@ -31,6 +31,7 @@ class Xtrsm: public Xgemm<T> {
   using Xgemm<T>::device_;
   using Xgemm<T>::db_;
   using Xgemm<T>::program_;
+  using Xgemm<T>::event_;
   using Xgemm<T>::DoGemm;
 
   // Constructor