fix move construct performance and conversion MagmaQueue->CudaStream

CompFUSE · Aug 13, 2020 · 1a2c079 · 1a2c079
1 parent c9e8f79
commit 1a2c079
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 15 deletions.
diff --git a/include/dca/linalg/util/magma_queue.hpp b/include/dca/linalg/util/magma_queue.hpp
@@ -32,19 +32,20 @@ class MagmaQueue {
     cusparseCreate(&cusparse_handle_);
     int device;
     cudaGetDevice(&device);
-    magma_queue_create_from_cuda(device, stream_, cublas_handle_,
-                                 cusparse_handle_, &queue_);
+    magma_queue_create_from_cuda(device, stream_, cublas_handle_, cusparse_handle_, &queue_);
   }
 
   MagmaQueue(const MagmaQueue& rhs) = delete;
   MagmaQueue& operator=(const MagmaQueue& rhs) = delete;
 
-  MagmaQueue(MagmaQueue&& rhs) noexcept {
-    swapMembers(rhs);
+  MagmaQueue(MagmaQueue&& rhs) noexcept : queue_(std::move(rhs.queue_)) {
+    std::swap(cublas_handle_, rhs.cublas_handle_);
+    std::swap(cusparse_handle_, rhs.cusparse_handle_);
+    std::swap(queue_, rhs.queue_);
   }
 
   MagmaQueue& operator=(MagmaQueue&& rhs) noexcept {
-    swapMembers(rhs);
+    swap(rhs);
     return *this;
   }
 
@@ -62,21 +63,21 @@ class MagmaQueue {
   // take a MagmaQueue, this makes all this code less intelligible
   // but less verbose.  Consider this carefully.
   operator cudaStream_t() const {
-    return stream_;
+    return static_cast<cudaStream_t>(stream_);
   }
-  
+
   const CudaStream& getStream() const {
     return stream_;
   }
-
-private:
-  void swapMembers(MagmaQueue& rhs) noexcept {
+
+  void swap(MagmaQueue& rhs) noexcept {
     std::swap(stream_, rhs.stream_);
     std::swap(cublas_handle_, rhs.cublas_handle_);
     std::swap(cusparse_handle_, rhs.cusparse_handle_);
     std::swap(queue_, rhs.queue_);
   }
 
+private:
   CudaStream stream_;
   magma_queue_t queue_ = nullptr;
   cublasHandle_t cublas_handle_ = nullptr;

diff --git a/include/dca/phys/dca_step/cluster_solver/ctint/accumulator/ctint_accumulator.hpp b/include/dca/phys/dca_step/cluster_solver/ctint/accumulator/ctint_accumulator.hpp
@@ -100,7 +100,7 @@ class CtintAccumulator {
   MatrixConfiguration configuration_;
   int sign_ = 0;
 
-  std::vector<linalg::util::CudaStream*> streams_;
+  std::vector<const linalg::util::CudaStream*> streams_;
   linalg::util::CudaEvent event_;
 
   util::Accumulator<int> accumulated_sign_;

diff --git a/include/dca/phys/dca_step/cluster_solver/shared_tools/accumulation/tp/tp_accumulator.hpp b/include/dca/phys/dca_step/cluster_solver/shared_tools/accumulation/tp/tp_accumulator.hpp
@@ -123,8 +123,8 @@ class TpAccumulator<Parameters, linalg::CPU, DT> {
     return 0;
   }
 
-  linalg::util::CudaStream* get_stream() const {
-    static dca::linalg::util::CudaStream mock_stream;
+  const linalg::util::CudaStream* get_stream() const {
+    static const dca::linalg::util::CudaStream mock_stream;
     return &mock_stream;
   }
 

diff --git a/include/dca/phys/dca_step/cluster_solver/shared_tools/accumulation/tp/tp_accumulator_gpu.hpp b/include/dca/phys/dca_step/cluster_solver/shared_tools/accumulation/tp/tp_accumulator_gpu.hpp
@@ -87,8 +87,8 @@ class TpAccumulator<Parameters, linalg::GPU> : public TpAccumulator<Parameters,
   // other_acc.
   void sumTo(this_type& other_acc);
 
-  linalg::util::CudaStream* get_stream() {
-    return &queues_[0];
+  const linalg::util::CudaStream* get_stream() {
+    return &queues_[0].getStream();
   }
 
   void synchronizeCopy() {