use reference instead of copy for several command queue calls for OpenCL

Author: SteveBronder

stan/math/opencl/copy.hpp

@@ -97,7 +97,7 @@ inline auto from_matrix_cl(const T& src) {
   } else {
     try {
       cl::Event copy_event;
-      const cl::CommandQueue queue = opencl_context.queue();
+      const cl::CommandQueue& queue = opencl_context.queue();
       std::vector<cl::Event> copy_write_events(src.write_events().begin(),
                                                src.write_events().end());
       queue.enqueueReadBuffer(src.buffer(), opencl_context.in_order(), 0,
@@ -152,7 +152,7 @@ inline T_dst from_matrix_cl(const matrix_cl<T>& src) {
                    "dst.cols()", 1);
   try {
     cl::Event copy_event;
-    const cl::CommandQueue queue = opencl_context.queue();
+    const cl::CommandQueue& queue = opencl_context.queue();
     std::vector<cl::Event> copy_write_events(src.write_events().begin(),
                                              src.write_events().end());
     queue.enqueueReadBuffer(src.buffer(), opencl_context.in_order(), 0,
@@ -186,7 +186,7 @@ inline T_dst from_matrix_cl(const matrix_cl<T>& src) {
   }
   try {
     cl::Event copy_event;
-    const cl::CommandQueue queue = opencl_context.queue();
+    const cl::CommandQueue& queue = opencl_context.queue();
     std::vector<cl::Event> copy_write_events(src.write_events().begin(),
                                              src.write_events().end());
     queue.enqueueReadBuffer(src.buffer(), opencl_context.in_order(), 0,
@@ -257,7 +257,7 @@ inline auto packed_copy(const T& src) {
     return dst;
   }
   try {
-    const cl::CommandQueue queue = opencl_context.queue();
+    const cl::CommandQueue& queue = opencl_context.queue();
     matrix_cl<T_val> packed(packed_size, 1);
     stan::math::opencl_kernels::pack(cl::NDRange(src.rows(), src.rows()),
                                      packed, src, src.rows(), src.rows(),
@@ -310,7 +310,7 @@ inline matrix_cl<Vec_scalar> packed_copy(Vec&& src, int rows) {
   try {
     matrix_cl<Vec_scalar> packed(packed_size, 1);
     cl::Event packed_event;
-    const cl::CommandQueue queue = opencl_context.queue();
+    const cl::CommandQueue& queue = opencl_context.queue();
     queue.enqueueWriteBuffer(
         packed.buffer(),
         opencl_context.in_order() || std::is_rvalue_reference<Vec&&>::value, 0,

stan/math/opencl/opencl_context.hpp

@@ -161,7 +161,7 @@ class opencl_context_base {
                                   // the device
   bool in_order_;                 // Whether to use out of order execution.
   // Holds Default parameter values for each Kernel.
-  using map_base_opts = std::map<std::string, int>;
+  using map_base_opts = std::unordered_map<std::string, int>;
   map_base_opts base_opts_
       = {{"LOWER", static_cast<int>(matrix_cl_view::Lower)},
          {"UPPER", static_cast<int>(matrix_cl_view::Upper)},
@@ -194,7 +194,7 @@ class opencl_context_base {
   } tuning_opts_;
 
  protected:
-  static opencl_context_base& getInstance() {
+  static opencl_context_base& getInstance() noexcept {
     static opencl_context_base instance_;
     return instance_;
   }
@@ -352,21 +352,21 @@ class opencl_context {
    * objects. For stan, there should only be one context, queue, device, and
    * program with multiple kernels.
    */
-  inline cl::Context& context() {
+  inline cl::Context& context() noexcept {
     return opencl_context_base::getInstance().context_;
   }
   /** \ingroup opencl_context_group
    * Returns the reference to the active OpenCL command queue for the device.
    * One command queue will exist per device where
    * kernels are placed on the command queue and by default executed in order.
    */
-  inline cl::CommandQueue& queue() {
+  inline cl::CommandQueue& queue() noexcept {
     return opencl_context_base::getInstance().command_queue_;
   }
   /** \ingroup opencl_context_group
    * Returns a copy of the map of kernel defines
    */
-  inline opencl_context_base::map_base_opts& base_opts() {
+  inline opencl_context_base::map_base_opts& base_opts() noexcept {
     return opencl_context_base::getInstance().base_opts_;
   }
   /** \ingroup opencl_context_group
@@ -376,35 +376,35 @@ class opencl_context {
    * max workgroup of 256 would allow thread blocks of sizes (16,16), (128,2),
    * (8, 32), etc.
    */
-  inline int max_thread_block_size() {
+  inline int max_thread_block_size() noexcept {
     return opencl_context_base::getInstance().max_thread_block_size_;
   }
 
   /** \ingroup opencl_context_group
    * Returns the thread block size for the Cholesky Decompositions L_11.
    */
-  inline opencl_context_base::tuning_struct& tuning_opts() {
+  inline opencl_context_base::tuning_struct& tuning_opts() noexcept {
     return opencl_context_base::getInstance().tuning_opts_;
   }
 
   /** \ingroup opencl_context_group
    * Returns a vector containing the OpenCL device used to create the context
    */
-  inline std::vector<cl::Device>& device() {
+  inline std::vector<cl::Device>& device() noexcept {
     return opencl_context_base::getInstance().device_;
   }
 
   /** \ingroup opencl_context_group
    * Returns a vector containing the OpenCL platform used to create the context
    */
-  inline std::vector<cl::Platform>& platform() {
+  inline std::vector<cl::Platform>& platform() noexcept {
     return opencl_context_base::getInstance().platform_;
   }
   /** \ingroup opencl_context_group
    * Return a bool representing whether the write to the OpenCL device are
    * blocking
    */
-  inline bool& in_order() {
+  inline bool& in_order() noexcept {
     return opencl_context_base::getInstance().in_order_;
   }
 

stan/math/opencl/zeros_strict_tri.hpp

@@ -44,7 +44,7 @@ inline void matrix_cl<T>::zeros_strict_tri() try {
     return;
   }
   this->view_ = both(this->view_, invert(matrix_view));
-  cl::CommandQueue cmdQueue = opencl_context.queue();
+  cl::CommandQueue& cmdQueue = opencl_context.queue();
   opencl_kernels::fill_strict_tri(cl::NDRange(this->rows(), this->cols()),
                                   *this, 0.0, this->rows(), this->cols(),
                                   matrix_view);