Merge pull request #4 from CNugteren/host_code_perf

Host-code performance

CHANGELOG

@@ -1,6 +1,7 @@
 
 Development version (next release)
 - Added support for complex conjugate transpose
+- Some host-code performance improvements
 - Added level-2 routines:
     SGEMV/DGEMV/CGEMV/ZGEMV
 - Added level-3 routines:

include/internal/clpp11.h

@@ -134,8 +134,7 @@ class Platform: public Object {
   }
 
   // Accessors to the private data-member
-  cl_platform_id operator()() const { return platform_; }
+  const cl_platform_id& operator()() const { return platform_; }
-  cl_platform_id& operator()() { return platform_; }
  private:
   cl_platform_id platform_;
 };
@@ -193,8 +192,7 @@ class Device: public Object {
   }
 
   // Accessors to the private data-member
-  cl_device_id operator()() const { return device_; }
+  const cl_device_id& operator()() const { return device_; }
-  cl_device_id& operator()() { return device_; }
  private:
 
   // Helper functions
@@ -259,8 +257,7 @@ class Context: public ObjectWithState {
   }
 
   // Accessors to the private data-member
-  cl_context operator()() const { return context_; }
+  const cl_context& operator()() const { return context_; }
-  cl_context& operator()() { return context_; }
  private:
   cl_context context_;
 };
@@ -296,16 +293,6 @@ class Program: public ObjectWithState {
     swap(*this, other);
     return *this;
   }
-  /*
-  TODO: Implement move construction/assignment?
-  Program(Program &&other) {
-    clRetainProgram(program_);
-    swap(*this, other);
-  }
-  Program& operator=(Program &&other) {
-    swap(*this, other);
-    return *this;
-  }*/
   friend void swap(Program &first, Program &second) {
     std::swap(first.length_, second.length_);
     std::swap(first.source_, second.source_);
@@ -327,8 +314,7 @@ class Program: public ObjectWithState {
   }
 
   // Accessors to the private data-member
-  cl_program operator()() const { return program_; }
+  const cl_program& operator()() const { return program_; }
-  cl_program& operator()() { return program_; }
  private:
   size_t length_;
   std::vector<char> source_;
@@ -382,8 +368,7 @@ class Kernel: public ObjectWithState {
   }
 
   // Accessors to the private data-member
-  cl_kernel operator()() const { return kernel_; }
+  const cl_kernel& operator()() const { return kernel_; }
-  cl_kernel& operator()() { return kernel_; }
  private:
   cl_kernel kernel_;
 };
@@ -445,8 +430,7 @@ class CommandQueue: public ObjectWithState {
   }
 
   // Accessors to the private data-member
-  cl_command_queue operator()() const { return queue_; }
+  const cl_command_queue& operator()() const { return queue_; }
-  cl_command_queue& operator()() { return queue_; }
  private:
   cl_command_queue queue_;
 };
@@ -511,8 +495,7 @@ class Buffer: public ObjectWithState {
   }
 
   // Accessors to the private data-member
-  cl_mem operator()() const { return buffer_; }
+  const cl_mem& operator()() const { return buffer_; }
-  cl_mem& operator()() { return buffer_; }
  private:
   cl_mem buffer_;
 };

include/internal/routine.h

@@ -33,7 +33,7 @@ class Routine {
   const std::string kKhronosHalfPrecision = "cl_khr_fp16";
   const std::string kKhronosDoublePrecision = "cl_khr_fp64";
 
-  // New data-type:tThe cache of compiled OpenCL programs, along with some meta-data
+  // The cache of compiled OpenCL programs, along with some meta-data
   struct ProgramCache {
     Program program;
     std::string device_name;
@@ -101,7 +101,7 @@ class Routine {
   // Queries the cache and retrieve either a matching program or a boolean whether a match exists.
   // The first assumes that the program is available in the cache and will throw an exception
   // otherwise.
-  Program GetProgramFromCache() const;
+  const Program& GetProgramFromCache() const;
   bool ProgramIsInCache() const;
 
   // Non-static variable for the precision. Note that the same variable (but static) might exist in

src/routine.cc

@@ -308,7 +308,7 @@ StatusCode Routine::PadCopyTransposeMatrix(const size_t src_one, const size_t sr
 
 // Queries the cache and retrieves a matching program. Assumes that the match is available, throws
 // otherwise.
-Program Routine::GetProgramFromCache() const {
+const Program& Routine::GetProgramFromCache() const {
   for (auto &cached_program: program_cache_) {
     if (cached_program.MatchInCache(device_name_, precision_, routines_)) {
       return cached_program.program;

src/routines/xaxpy.cc

@@ -60,7 +60,7 @@ StatusCode Xaxpy<T>::DoAxpy(const size_t n, const T alpha,
 
   // Retrieves the Xaxpy kernel from the compiled binary
   try {
-    auto program = GetProgramFromCache();
+    auto& program = GetProgramFromCache();
     auto kernel = Kernel(program, kernel_name);
 
     // Sets the kernel arguments

src/routines/xgemm.cc

@@ -102,7 +102,7 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
     auto temp_c = Buffer(context_, CL_MEM_READ_WRITE, m_ceiled*n_ceiled*sizeof(T));
 
     // Loads the program from the database
-    auto program = GetProgramFromCache();
+    auto& program = GetProgramFromCache();
 
     // Runs the pre-processing kernels. This transposes the matrices, but also pads zeros to fill
     // them up until they reach a certain multiple of size (kernel parameter dependent).

src/routines/xgemv.cc

@@ -100,7 +100,7 @@ StatusCode Xgemv<T>::DoGemv(const Layout layout, const Transpose a_transpose,
 
   // Retrieves the Xgemv kernel from the compiled binary
   try {
-    auto program = GetProgramFromCache();
+    auto& program = GetProgramFromCache();
     auto kernel = Kernel(program, kernel_name);
 
     // Sets the kernel arguments

src/routines/xsymm.cc

@@ -61,7 +61,7 @@ StatusCode Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle
     // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemm
     // routine afterwards
     try {
-      auto program = GetProgramFromCache();
+      auto& program = GetProgramFromCache();
       auto kernel = Kernel(program, kernel_name);
 
       // Sets the arguments for the symmetric-to-squared kernel