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.

1 files changed, 5 insertions, 8 deletions

diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp
index cb0e0461..fdef43dc 100644
--- a/src/routines/level3/xsyr2k.cpp
+++ b/src/routines/level3/xsyr2k.cpp
@@ -77,9 +77,6 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   // Decides which kernel to run: the upper-triangular or lower-triangular version
   auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    ab_rotated == false;
@@ -103,7 +100,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, false);
     eventWaitList.push_back(eventProcessA);
   }
@@ -112,7 +109,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, false);
     eventWaitList.push_back(eventProcessB);
   }
@@ -123,12 +120,12 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                          n, n, c_ld, c_offset, c_buffer,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          true, c_rotated, false);
   eventWaitList.push_back(eventProcessC);
 
   // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n_ceiled));
@@ -168,7 +165,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
                          n, n, c_ld, c_offset, c_buffer,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          false, c_rotated, false, upper, lower, false);
 }