Routine, Cache: generalize, reduce amount of copying in fast path

Implement a generalized Cache<K, V>. Two variants are provided: the
first one is based on std::map, using C++14-specific transparent
std::less<> and generalized std::map::find() to allow searching by tuple
of references. The second one is based on std::vector and O(n) lookup,
but remains C++11-compliant.

8 files changed, 12 insertions, 20 deletions

diff --git a/src/routines/level1/xamax.cpp b/src/routines/level1/xamax.cpp
index e9efa1a7..40a66517 100644
--- a/src/routines/level1/xamax.cpp
+++ b/src/routines/level1/xamax.cpp
@@ -43,9 +43,8 @@ void Xamax<T>::DoAmax(const size_t n,
   TestVectorIndex(1, imax_buffer, imax_offset);
 
   // Retrieves the Xamax kernels from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel1 = Kernel(program, "Xamax");
-  auto kernel2 = Kernel(program, "XamaxEpilogue");
+  auto kernel1 = Kernel(program_, "Xamax");
+  auto kernel2 = Kernel(program_, "XamaxEpilogue");
 
   // Creates the buffer for intermediate values
   auto temp_size = 2*db_["WGS2"];
diff --git a/src/routines/level1/xasum.cpp b/src/routines/level1/xasum.cpp
index a242a5fa..b93b271c 100644
--- a/src/routines/level1/xasum.cpp
+++ b/src/routines/level1/xasum.cpp
@@ -43,9 +43,8 @@ void Xasum<T>::DoAsum(const size_t n,
   TestVectorScalar(1, asum_buffer, asum_offset);
 
   // Retrieves the Xasum kernels from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel1 = Kernel(program, "Xasum");
-  auto kernel2 = Kernel(program, "XasumEpilogue");
+  auto kernel1 = Kernel(program_, "Xasum");
+  auto kernel2 = Kernel(program_, "XasumEpilogue");
 
   // Creates the buffer for intermediate values
   auto temp_size = 2*db_["WGS2"];
diff --git a/src/routines/level1/xaxpy.cpp b/src/routines/level1/xaxpy.cpp
index 5436c5b7..39f61ef4 100644
--- a/src/routines/level1/xaxpy.cpp
+++ b/src/routines/level1/xaxpy.cpp
@@ -52,8 +52,7 @@ void Xaxpy<T>::DoAxpy(const size_t n, const T alpha,
   auto kernel_name = (use_fast_kernel) ? "XaxpyFast" : "Xaxpy";
 
   // Retrieves the Xaxpy kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   if (use_fast_kernel) {
diff --git a/src/routines/level1/xcopy.cpp b/src/routines/level1/xcopy.cpp
index d86200c0..62889764 100644
--- a/src/routines/level1/xcopy.cpp
+++ b/src/routines/level1/xcopy.cpp
@@ -52,8 +52,7 @@ void Xcopy<T>::DoCopy(const size_t n,
   auto kernel_name = (use_fast_kernel) ? "XcopyFast" : "Xcopy";
 
   // Retrieves the Xcopy kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   if (use_fast_kernel) {
diff --git a/src/routines/level1/xdot.cpp b/src/routines/level1/xdot.cpp
index 9d718913..9f9c0590 100644
--- a/src/routines/level1/xdot.cpp
+++ b/src/routines/level1/xdot.cpp
@@ -46,9 +46,8 @@ void Xdot<T>::DoDot(const size_t n,
   TestVectorScalar(1, dot_buffer, dot_offset);
 
   // Retrieves the Xdot kernels from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel1 = Kernel(program, "Xdot");
-  auto kernel2 = Kernel(program, "XdotEpilogue");
+  auto kernel1 = Kernel(program_, "Xdot");
+  auto kernel2 = Kernel(program_, "XdotEpilogue");
 
   // Creates the buffer for intermediate values
   auto temp_size = 2*db_["WGS2"];
diff --git a/src/routines/level1/xnrm2.cpp b/src/routines/level1/xnrm2.cpp
index 373820a4..aa341aff 100644
--- a/src/routines/level1/xnrm2.cpp
+++ b/src/routines/level1/xnrm2.cpp
@@ -43,9 +43,8 @@ void Xnrm2<T>::DoNrm2(const size_t n,
   TestVectorScalar(1, nrm2_buffer, nrm2_offset);
 
   // Retrieves the Xnrm2 kernels from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel1 = Kernel(program, "Xnrm2");
-  auto kernel2 = Kernel(program, "Xnrm2Epilogue");
+  auto kernel1 = Kernel(program_, "Xnrm2");
+  auto kernel2 = Kernel(program_, "Xnrm2Epilogue");
 
   // Creates the buffer for intermediate values
   auto temp_size = 2*db_["WGS2"];
diff --git a/src/routines/level1/xscal.cpp b/src/routines/level1/xscal.cpp
index 0521b1e5..9bc096e5 100644
--- a/src/routines/level1/xscal.cpp
+++ b/src/routines/level1/xscal.cpp
@@ -49,8 +49,7 @@ void Xscal<T>::DoScal(const size_t n, const T alpha,
   auto kernel_name = (use_fast_kernel) ? "XscalFast" : "Xscal";
 
   // Retrieves the Xscal kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   if (use_fast_kernel) {
diff --git a/src/routines/level1/xswap.cpp b/src/routines/level1/xswap.cpp
index c9b97dc9..f046575f 100644
--- a/src/routines/level1/xswap.cpp
+++ b/src/routines/level1/xswap.cpp
@@ -52,8 +52,7 @@ void Xswap<T>::DoSwap(const size_t n,
   auto kernel_name = (use_fast_kernel) ? "XswapFast" : "Xswap";
 
   // Retrieves the Xswap kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   if (use_fast_kernel) {